text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
A Design Methodology for Efficient Implementation
of Deconvolutional Neural Networks on an FPGA
Xinyu Zhang, Srinjoy Das, Ojash Neopane and Ken Kreutz-Delgado
arXiv:1705.02583v1 [cs.LG] 7 May 2017
University of California, San Diego
Email: [email protected], [email protected], [email protected], [email protected]
Abstract—In recent years deep learning algorithms have
shown extremely high performance on machine learning
tasks such as image classification and speech recognition.
In support of such applications, various FPGA accelerator
architectures have been proposed for convolutional neural
networks (CNNs) that enable high performance for classification
tasks at lower power than CPU and GPU processors. However,
to date, there has been little research on the use of
FPGA implementations of deconvolutional neural networks
(DCNNs). DCNNs, also known as generative CNNs, encode
high-dimensional probability distributions and have been widely
used for computer vision applications such as scene completion,
scene segmentation, image creation, image denoising, and
super-resolution imaging. We propose an FPGA architecture
for deconvolutional networks built around an accelerator which
effectively handles the complex memory access patterns needed
to perform strided deconvolutions, and that supports convolution
as well. We also develop a three-step design optimization
method that systematically exploits statistical analysis, design
space exploration and VLSI optimization. To verify our FPGA
deconvolutional accelerator design methodology we train DCNNs
offline on two representative datasets using the generative
adversarial network method (GAN) run on Tensorflow, and
then map these DCNNs to an FPGA DCNN-plus-accelerator
implementation to perform generative inference on a Xilinx
Zynq-7000 FPGA. Our DCNN implementation achieves a peak
performance density of 0.012 GOPs/DSP.
Keywords—FPGA,
Acceleration
Deconvolution,
I.
Generative
Model,
I NTRODUCTION
Deep learning algorithms have shown extremely high
performance on machine learning tasks. In particular,
convolutional neural networks (CNNs) have become the
state-of-the-art for applications like computer vision and audio
recognition [1] [2] [3]. To address the increasing demand for
applications that require running neural network algorithms
in real time on embedded devices, various high performance
hardware platforms for discriminative CNN implementations
have been proposed, including the use of distributed GPUs
or customized accelerators like FPGAs and ASICs [4] [5].
In particular, FPGA-based accelerators have been proposed
because they have lower latency and consume less power than
GPUs while being more flexible and configurable than ASICs
[6] [7].
However, current FPGA accelerators focus on enhancing
the performance of convolutional neural networks (CNNs),
not deconvolutional neural networks (DCNNs). Unlike
discriminative CNNs that effectively “downsample” the input
to produce classification [1], DCNNs are generative models
capable of generating data by “upsampling” the input using
deconvolution layers [8]. There are many applications of
DCNNs, including multi-modal data modeling [9], super
resolution [10] and image-to-image translation [11] [12]
(see Fig. 1). Such applications motivate us to design
an FPGA-based accelerator with the ability to execute
deconvolution operations with high throughput and low cost.
Fig. 1. DCNNs work for pattern completion/generation (Images from [9]
[10] [12]).
There are several issues that must be addressed to design
an FPGA-based deconvolution accelerator. First, a direct
translation of CPU-optimized deconvolution algorithms to
an FPGA will generally lead to inefficient implementations.
A suitable adaptation of the deconvolution operation to a
hardware substrate such as FPGA is therefore necessary in
order to achieve high performance with low implementation
complexity. In addition, although recent research shows that
discriminative CNNs are robust to low bitwidth quantization
[13] [14], it is important to be able to systematically study
the effects of such bitwidth reductions on the quality of
inference from a generative model such as DCNN implemented
with finite precision on FPGA. Thus it is necessary to use
metrics which quantify the effects of such approximations in
DCNNs in order to achieve an efficient design optimized for
performance and power.
To address the issues described above, we make the
following contributions in this paper. 1) We create a
deconvolution accelerator with reverse looping and stride hole
skipping to efficiently implement deconvolution on an FPGA,
where our proposed solution, in a nontrivial way, reuses the
same computational architecture proposed for implementing
a convolution accelerator in [6]. 2) We propose a three-step
procedure to design the deconvolution accelerator as follows.
A) At the highest design level, we train DCNNs using
the generative adversarial network method (GAN) [15] and
use statistical tests to quantitatively analyze the generative
quality under different bitwidth precisions to select the most
cost-efficient bitwidth. B) We use the roofline model proposed
in [6] to explore the design space in order to find the set of
high-level constraints that achieves the best tradeoff between
memory bandwidth and accelerator throughput. C) We use loop
unrolling and pipelining, memory partitioning, and register
insertion to further optimize performance. 3) We validate our
procedure via two implementations on a Xilinx Zynq-7000
FPGA.
The rest of this paper is organized as follows: Section II
provides background on the DCNN and the deconvolution
layers. Section III presents our methodology for efficiently
implementing an FPGA-based deconvolution accelerator.
Section IV explains our three-step design methodology.
Section V shows our experimental results. Section VI
concludes the paper.
II.
D ECONVOLUTIONAL N EURAL N ETWORK
A deconvolutional neural network (DCNN) converts
latent space representations to high-dimensional data similar
to the training set by applying successive deconvolution
operations in multiple layers [16]. The latent space contains
low-dimensional latent variables that provide a succinct
(“conceptual”) representations of the possible outputs (e.g. an
image). Thus a latent variable may correspond to “chair” with
the associated output being the image of a chair “generated”
by the DCNN (see Fig. 1). Fig. 2 shows a 5-layer DCNN
developed in [17] that consists of 4 deconvolutional layers.
The first layer is fully-connected and transforms an input size
of 1x100 to an output size of 1024x4x4; layers 2 to 5 are
deconvolution layers that project low-dimensional feature maps
into corresponding high-dimensional ones through successive
layers.
Fig. 3. Visualization of a Single Deconvolution Layer. The four steps required
to implement the deconvolutional layer are: (1) multiply a single input pixel
ih , iw by a K × K kernel; (2) add the result of step 1 to a local area in the
output feature map that starts at ih × S, iw × S; (3) repeat 1 and 2 for all
input pixels; (4) remove elements from output feature maps in the border by
zero padding of size P .
Fig. 4.
Visualization of Algorithm 1 with loop variables.
The relation of the input size IH × IW to output size
OH × OW after applying stride and padding are given in the
following equations [19]:
OH = S × (IH − 1) + K − 2P
OW = S × (IW − 1) + K − 2P
III.
(1)
D ECONVOLUTION H ARDWARE D ESIGN
An FPGA accelerator usually consists of processing
elements (PEs), registers, and local memory elements referred
to as block RAMs (BRAMs). Processing elements operate on
data provided by the local memory, which communicates with
external dual data rate (DDR) memory using direct memory
access (DMA). Fig. 5 shows a traditional implementation of
deconvolution, where TIH , TIW , TIC , TOH , TOW , and TOC
are the dimensions of the input and output block. Replacing
IH , OH with TIH , TOH in Eq. 1, we have:
TOH = S × (TIH − 1) + K − 2P
(2)
Here the zero padding P = 0 because blocks are inside input
feature maps. However, Eq. 3 shows that deconvolution results
of input blocks overlap with each other:
IH
× TOH > OH
(3)
TIH
Fig. 2. A DCNN that generates realistic 64x64 indoor scenes based on the
use of four deconvolution layers that was trained on the Large-scale Scene
Understanding (LSUN) Dataset [17] [18] (Image is taken and adapted from
reference [17].)
Fig. 3 shows how a typical deconvolution layer works,
where S and P denote the chosen values of stride and
padding respectively for a given layer. The pseudo code of
a deconvolution layer as implemented in CPU is shown in
Algorithm. 1 which uses the loop variables defined in Fig. 4.
By convention we use capital letters e.g. OH to denote specific parameters
of the DCNN whereas small letters e.g. oh to denote its corresponding loop
variable.
Deconvolution arithmetic requires overlapping regions
between output blocks to be summed together [19] which
can be realized in processor-based implementations. However
handling such operations in FPGAs requires either the design
Algorithm 1 Deconvolution in CPU
1: procedure D ECONVOLUTION
2:
for ic = 0 to IC − 1 do
3:
for ih = 0 to IH − 1 do
4:
for iw = 0 to IW − 1 do
5:
for oc = 0 to OC − 1 do
6:
for kh = 0 to K − 1 do
7:
for kw = 0 to K − 1 do
8:
oh ← S × ih + kh − P
9:
ow ← S × i w + k w − P
10:
out[oc ][oh ][ow ] ← (in[ic ][ih ][iw ]
× kernel[oc ][ic ][kh ][kw ])
of additional hardware blocks which creates overhead or
communicating with a host processor which can increase
system latencies thereby precluding real-time applications.
aforementioned problem. First note that a sufficient condition
for ih to be an integer in Eq. 5 is:
(oh + P − kh ) mod S = 0
(6)
OH
S
Assuming
is an integer (OH is defined in Eq. 1), we can
recast oh as follows:
oh = S × o0h + fh ,
fh ∈ {0, 1, ..., S − 1}
(7)
OH
o0h ∈ {0, 1, ...,
− 1}
S
Using the definition of oh in Eq. 6, we can recast the sufficient
condition Eq. 6 in terms of fh as below:
Fig. 5. Traditional implementation of deconvolution. The input feature map
is first divided into separate blocks and PEs read each block from DDR and
process the deconvolution operations on this block. Finally the results are
stored back to the DDR. This procedure is inefficient and can be circumvented
as described in the text.
(fh + P − kh ) mod S = 0
(8)
Eq. 7 implies that we can rewrite fh as:
fh = S − ((P − kh ) mod S)
(9)
This can be verified by plugging in Eq. 9 into Eq. 8 which
yields the following identity:
A. Reverse Looping
To avoid the overlapping sum problem, we propose a
technique called reverse looping, where instead of directly
deconvolving the input space, we use the output space
to determine which input blocks to deconvolve and thus
eliminating the need for the additional summation operations
described above. This procedure is indicated in Fig. 6.
(P − kh − (P − kh ) mod S) mod S = 0
(10)
To prevent fh from taking a value equal to S, we enforce
the additional condition:
fh = (S − ((P − kh ) mod S)) mod S
(11)
By using Eq. 11 to choose values for fh , we can ensure
that oh computed from Eq. 7 meets the condition in Eq. 6.
Therefore we can avoid the previously mentioned issue of
discarding fractional values of ih that we would otherwise
encounter from a direct application of Eq. 5. The pseudo code
for deconvolution on FPGA is shown in Algorithm 2.
Fig. 6. An efficient way to deconvolve. We first take a block in the output
space and determine which inputs are needed to calculate the values in the
block. Then, for each block, the input is deconvolved and the appropriate
output is extracted. This is done sequentially until values have been computed
for the entire output space.
The loop iterations over ih and iw in the CPU
implementation shown in Algorithm 1 need to be recast over
oh and ow . Referring to Algorithm 1 and Fig. 4, we have:
oh = ih × S + kh − P
(4)
Rearranging terms, we get:
oh + P − k h
(5)
S
Unfortunately Eq. 5 generally results in a non-integer value for
the loop variable ih , which is invalid [19]. One way to address
this problem would be to monitor ih so that fractional values
can be discarded. However this would consume additional
hardware resources and create unnecessary latencies in the
system.
ih =
B. Stride Hole Skipping
In this section, we propose a technique called stride hole
skipping to ensure ih of Eq. 5 is an integer. Toward this
end, we recast oh in terms of two new variables, o0h and fh
and show that this leads to an effective way of solving the
Algorithm 2 Our FPGA Implementation of Deconvolution
1: procedure R EVERSE D ECONVOLUTION
2:
for kh = 0 to K − 1 do
3:
for kw = 0 to K − 1 do
T
4:
for o0h = 0 to OSH − 1 do
T
5:
for o0w = 0 to OSW − 1 do . loop TOW
6:
for oc = 0 to TOC − 1 do . loop TOC
7:
for ic = 0 to TIC − 1 do. loop TIC
8:
9:
10:
11:
12:
13:
14:
15:
16:
COMPUTE(kh , kw , o0h , o0w , oc , ic )
procedure C OMPUTE(kh , kw , o0h , o0w , oc , ic )
fh ← (S − ((P − kh ) mod S)) mod S
fw ← (S − ((P − kw ) mod S)) mod S
oh = o0h × S + P + fh
ow = o0w × S + P + fw
ih ← (oh − kh )/S
iw ← (ow − kw )/S
out[oc ][oh ][ow ] ← in[ic ][ih ][iw ] × kernel[oc ][ic ][kh ][kw ]
IV.
T HREE -S TEP D ESIGN M ETHODOLOGY
A. Statistical Analysis
It is important to study the effect of bitwidth reduction
on the quality of inference from the generative model. To
find out the most cost-efficient bitwidth for DCNNs, we
fix TOH , TOW , TOC , TIC , and study the trade-off between
generative quality and implementation complexity over a
range of bitwidths using statistical analysis. Quantifying
generative models using traditional techniques such as
Kullback-Leibler divergence and log-likelihood are not feasible
in high-dimensional settings such as the typical setting
deconvolutional neural networks are used in. To overcome
this drawback, we apply nonparametric goodness of fit
testing. Specifically, we apply the Relative Maximum Mean
Discrepancy (RMMD) Test proposed by [20] to measure and
compare the performance of our system at different bitwidths.
The RMMD is an extension of the Maximum Mean
Discrepancy (MMD) two sample test proposed by [21]. Given
n
samples {Xi }m
i=1 and {Yi }i=1 from distributions Px and Py
the MMD test statistic is given by:
m
MMD2 (X, Y ) =
m
XX
1
k(xi , xj )
m(m − 1) i=1
j6=i
n
+
n
m
n
XX
1
2 XX
k(yi , yj ) −
k(xi , yj )
n(n − 1) i=1
mn i=1 j=1
j6=i
the null hypothesis H0 : Px = Py is tested versus alternative
H1 : Px 6= Py . In the above equation, k is the Radial Basis
Function given by
k(x, y) = exp ||x − y||
The RMMD test builds upon the standard MMD framework
by computing the MMD test statistic between two pairs of
n
r
distributions. Given samples {Xi }m
i=1 , {Yi }i=1 , and {Zi }i=1
respectively from the training data, low-bitwidth DCNN,
and full-precision DCNN, RMMD tests the null hypothesis
H0 : M M D2 (X, Y ) < M M D2 (X, Z) against the alternative
H1 : M M D2 (X, Z) < M M D2 (X, Y ). [20] shows that the
p-values for testing H0 against H1 are given by:
Fig. 7.
1) Computation to Communication Ratio: Let αin , αw ,
αout and Bin , Bw , Bout denote the trip counts and buffer sizes
of memory accesses to input/output feature maps, weights,
respectively. The CTC is given by:
CTC =
MMD2u (Xm , Yn ) − MMD2u (Xm , Zr )
p
)
p ≤ Φ(−
2
2
σXY
+ σXZ
− 2σXY XZ
where Φ is the Normal Cumulative Distribution Function. The
p-value in the above equation indicates the probability that,
based on the observed samples, the distribution based on the
low bitwidth DCNN is closer to the training data than the
distribution based on the full precision DCNN is to the training
data. Using this interpretation:
Roofline Model, adopted from [6]
In this drawing, A, B and C correspond to designs
of accelerator with different values of TOH , TOW , TOC , TIC .
Design A transfers too much data, so computation speed is low,
and therefore falls well beneath the computation roof. Design
B lies well beneath the bandwidth roof, which means the
system performance is dominated by memory transfers. Design
C is more efficient than A and B with its balance between
computation speed and memory bandwidth. This technique
is described in [6] and is used for the design of convolution
accelerator. We apply roofline analysis to design deconvolution
accelerator and estimate the computation to communication
ratio (CTC) and computational roof (CR) for a given layer.
total number of operations
total amount of external memory access
2 × IC × OC × IH × IW × K 2
=
αin Bin + αw Bw + αout Bout
OC OH
IC
, αin = αw =
αout
TOC TOH
TIC
TOW + K
TOH + K
= TIC
S
S
(12)
αout =
(13)
Bin
(14)
Bout = TOC TOH TOW ,
Bweight = TOC TIC K 2
(15)
•
a p-value > 0.5 indicates the low bitwidth DCNN is
more similar to the training data
0 ≤ Bin + Bw + Bout ≤ BRAMcapacity
•
a p-value < 0.5 indicates the full precision DCNN is
more similar to the training data
2) Computation Roof: Let PD denotes the pipeline depth
and II is the number of cycles between the start of each loop
iteration TOW , the CR is given by:
total number of operations
number of execution cycles
2 × IC × OC × IH × IW × K 2
=
αin K 2 TOH (PD + II(TOW − 1))
B. Roofline Analysis
The generative quality is determined by choosing the
optimal bitwidth using the previously described procedure.
Following this we turn to further increasing the throughput by
optimizing with respect to TOH , TOW , TOC , and TIC , which
are the height, width, channel size of output block, and channel
size of input block respectively (see Fig. 5). This is done
using roofline analysis [6]. Fig. 7 shows an example roofline
plot where the X axis denotes the number of operations per
memory access and Y axis denotes the number of operations
per cycle.
(16)
CR =
where
0 ≤ TOC TIC ≤ (# of DSPs)
0 < TIC ≤ IC
0 < TOC ≤ OC
0 < TOH ≤ OH
0 < TOW ≤ OW
(17)
Note that 0 ≤ TOC TIC ≤ (# of DSPs) will not hold true
when the bitwidth is greater than 18, because the maximum
bitwidth of the multipliers used in our implementation is 18-bit
[22]. Since we use a bitwidth of 12 in all our experiments this
constraint is therefore valid.
function of bitwidths. The two curves are shown in Fig. 10.
Both curves peak at bitwidth 12, which we take to be a good
choice because it represents a high p-value (generative quality)
with a low power consumption and high minimum slack.
C. VLSI Level Optimization
1) Loop Unrolling and Pipelining: Loop unrolling is a key
technique of high level synthesis [23]. It works by generating
parallel hardware to accelerate FPGA program execution. The
innermost loop TOC and TIC in Algorithm 2 are unrolled and
can be executed in a constant amount of cycles P , which forms
the processing engine as shown in Fig. 8. We also pipeline the
loop TOW with carried dependency of 2.
(a) p-value/power vs bitwidth
(b) p-value×slack vs bitwidth
Fig. 10. Approximate concave curves based on trade-off between generative
quality and implementation complexity.
B. Hardware System
Fig. 8.
Processing Engine
2) Register Insertion: The critical path length and pipeline
interval are constrained by the on-chip local memory
bandwidth, especially when the size of the processing engine
is large. To further improve performance, we insert registers
to economize local memory bandwidth, which is illustrated in
Fig. 9.
Fig. 9.
We implemented the deconvolution accelerator IP with
Vivado HLS (v2016.2). We use ap fixed.h from Vivado
Math Library to implement fixed point arithmetic operations
with arbitrary bitwidth precision, and use hls stream.h
& ap axi sdata.h to model streaming data structure. The
hardware system is built on a Zynq-7000 FPGA XZ7020 with
Vivado Design Suite and Xilinx SDK. The FPGA 7Z020 is
programed with our accelerator IP and the ARM processor is
used to initialize the accelerator, set parameters, and transfer
data for each layer. An overview of the implementation block
diagram is in Fig. 11.
Insert register to reduce local memory (BRAM) writes
V.
E VALUATION
A. Statistical Analysis
Previous work such as that described in [24] has shown the
effectiveness of using high-dimensional nonparametric tests
to determine optimal parameters for generative inference in
hardware. For designing the deconvolution accelerator we
follow a similar approach and use the RMMD test framework
outlined in Section IV A to choose the optimal bitwidth for our
system. For this purpose, we trained two DCNNs through the
method described in [17] on the MNIST and CelebA Human
Face datasets [25]. To study the trade-off between generative
quality and system complexity over a range of bitwidths, we
determine p-value × minimum slack and p-value/power as a
Fig. 11.
Overview of Implementation Block Diagram.
C. Experimental Results
Fig. 12 shows some generated faces and digits from our
trained DCNNs. Fig. 13 shows the output of DCNNs under
different bitwidths for the same input. Visually evaluating
degradation of image quality is only feasible in the cases
of extremely low bitwidth such as 8 bits. Our proposed
methodology provides an analytical framework for quantifying
the trade-off between image quality and implementation
complexity over a range of bitwidths.
Fig. 14 shows all constraint-admissible design solutions for
the first layer of our CelebA DCNN, where the best design is
TABLE II.
[7]
[6]
Ours
[2]
[3]
[4]
Fig. 12. Sample MNIST and CelebA images generated by the full precision
DCNN.
[5]
[6]
Fig. 13.
Images generated by different bitwidth DCNNs.
shown as located at the left corner of the roof. Table I shows
the utilization rate after place and route, and we compare our
DCNN performance with some existing CNN accelerators for
reference in table II. The performance can be further improved
by implementing a ping-pong buffer in our system.
[7]
[8]
[9]
[10]
[11]
[12]
[13]
Fig. 14. Design space Exploration for a layer with input 10x2x2 and output
64x4x4.
TABLE I.
FPGA R ESOURCE U TILIZATION
DSP
95%
LUT
48%
VI.
FF
29%
BRAM
48%
[15]
C ONCLUSION
In this work, we develop an FPGA-based deconvolution
accelerator for deconvolutional neural networks and propose
a three-step design methodology which first uses statistical
analysis to find out the most cost-efficient bitwidth, then
explore the design space with roofline model [6] and use VLSI
optimization methods to produce the final design. Finally, we
implement our method on a Zynq-7000 FPGA and realize a
performance density of 0.012 GOPs/DSP.
R EFERENCES
[1]
[14]
J. Schmidhuber, “Deep learning in neural networks: An overview,”
Neural networks, vol. 61, pp. 85–117, 2015.
[16]
[17]
[18]
[19]
[20]
Chip
VLX240T
VX485T
7Z020
C OMPARISON TO PREVIOUS IMPLEMENTATIONS
Precision
Fixed
Float
12Fixed
#DSP
768
2800
220
Freq
150M
100M
100M
GOPS
17
61.62
2.6
GOPS/DSP
0.022
0.022
0.012
Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
Press, 2016, http://www.deeplearningbook.org.
S. Chakradhar, M. Sankaradas, V. Jakkula, and S. Cadambi,
“A dynamically configurable coprocessor for convolutional neural
networks,” in ACM SIGARCH Computer Architecture News, vol. 38,
no. 3. ACM, 2010, pp. 247–257.
Y. Chen, T. Luo, S. Liu, S. Zhang, L. He, J. Wang, L. Li, T. Chen,
Z. Xu, N. Sun et al., “Dadiannao: A machine-learning supercomputer,”
in Proceedings of the 47th Annual IEEE/ACM International Symposium
on Microarchitecture. IEEE Computer Society, 2014, pp. 609–622.
C. Zhang, P. Li, G. Sun, Y. Guan, B. Xiao, and J. Cong, “Optimizing
fpga-based accelerator design for deep convolutional neural networks,”
in Proceedings of the 2015 ACM/SIGDA International Symposium on
Field-Programmable Gate Arrays. ACM, 2015, pp. 161–170.
M. Peemen, A. A. Setio, B. Mesman, and H. Corporaal,
“Memory-centric accelerator design for convolutional neural networks,”
in Computer Design (ICCD), 2013 IEEE 31st International Conference
on. IEEE, 2013, pp. 13–19.
M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus,
“Deconvolutional networks,” in Computer Vision and Pattern
Recognition (CVPR), 2010 IEEE Conference on. IEEE, 2010, pp.
2528–2535.
J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum,
“Learning a probabilistic latent space of object shapes via 3d
generative-adversarial modeling,” in Advances in Neural Information
Processing Systems, 2016, pp. 82–90.
W. Shi, J. Caballero, F. Huszár, J. Totz, A. P. Aitken, R. Bishop,
D. Rueckert, and Z. Wang, “Real-time single image and video
super-resolution using an efficient sub-pixel convolutional neural
network,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 1874–1883.
P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image
translation with conditional adversarial networks,” arXiv preprint
arXiv:1611.07004, 2016.
V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep
convolutional encoder-decoder architecture for image segmentation,”
arXiv preprint arXiv:1511.00561, 2015.
J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng, “Quantized
convolutional neural networks for mobile devices,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2016,
pp. 4820–4828.
G. Dundar and K. Rose, “The effects of quantization on multilayer
neural networks,” IEEE Transactions on Neural Networks, vol. 6, no. 6,
pp. 1446–1451, 1995.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,”
in Advances in neural information processing systems, 2014, pp.
2672–2680.
H. Noh, S. Hong, and B. Han, “Learning deconvolution network
for semantic segmentation,” in Proceedings of the IEEE International
Conference on Computer Vision, 2015, pp. 1520–1528.
A. Radford, L. Metz, and S. Chintala, “Unsupervised representation
learning with deep convolutional generative adversarial networks,”
arXiv preprint arXiv:1511.06434, 2015.
F. Yu, A. Seff, Y. Zhang, S. Song, T. Funkhouser, and J. Xiao, “Lsun:
Construction of a large-scale image dataset using deep learning with
humans in the loop,” arXiv preprint arXiv:1506.03365, 2015.
V. Dumoulin and F. Visin, “A guide to convolution arithmetic for deep
learning,” arXiv preprint arXiv:1603.07285, 2016.
W. Bounliphone, E. Belilovsky, M. B. Blaschko, I. Antonoglou, and
A. Gretton, “A test of relative similarity for model selection in
generative models,” arXiv preprint arXiv:1511.04581, 2015.
[21]
[22]
[23]
[24]
[25]
A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola,
“A kernel two-sample test,” Journal of Machine Learning Research,
vol. 13, no. Mar, pp. 723–773, 2012.
7 Series DSP48E1 Slice, XILINX INC, 9 2016, rev. 1.9.
P. Coussy and A. Morawiec, High-Level Synthesis: From Algorithm to
Digital Circuit, 1st ed. Springer Publishing Company, Incorporated,
2008.
O. Neopane, S. Das, E. Arias-Castro, and K. Kreutz-Delgado, “A
nonparametric framework for quantifying generative inference on
neuromorphic systems,” in Circuits and Systems (ISCAS), 2016 IEEE
International Symposium on. IEEE, 2016, pp. 1346–1349.
Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes
in the wild,” in Proceedings of International Conference on Computer
Vision (ICCV), 2015.
| 9cs.NE
|
Differentiable Scheduled Sampling for Credit Assignment
Kartik Goyal
Carnegie Mellon University
Pittsbrugh, PA, USA
[email protected]
Chris Dyer
DeepMind
London, UK
[email protected]
arXiv:1704.06970v1 [cs.CL] 23 Apr 2017
Abstract
We demonstrate that a continuous relaxation of the argmax operation can be
used to create a differentiable approximation to greedy decoding for sequence-tosequence (seq2seq) models. By incorporating this approximation into the scheduled sampling training procedure (Bengio
et al., 2015)–a well-known technique for
correcting exposure bias–we introduce a
new training objective that is continuous
and differentiable everywhere and that can
provide informative gradients near points
where previous decoding decisions change
their value. In addition, by using a related
approximation, we demonstrate a similar
approach to sampled-based training. Finally, we show that our approach outperforms cross-entropy training and scheduled sampling procedures in two sequence
prediction tasks: named entity recognition
and machine translation.
1
Introduction
Sequence-to-Sequence (seq2seq) models have
demonstrated excellent performance in several
tasks including machine translation (Sutskever
et al., 2014), summarization (Rush et al., 2015),
dialogue generation (Serban et al., 2015), and
image captioning (Xu et al., 2015). However,
the standard cross-entropy training procedure for
these models suffers from the well-known problem of exposure bias: because cross-entropy training always uses gold contexts, the states and contexts encountered during training do not match
those encountered at test time. This issue has been
addressed using several approaches that try to incorporate awareness of decoding choices into the
training optimization. These include reinforcement learning (Ranzato et al., 2016; Bahdanau
Taylor Berg-Kirkpatrick
Carnegie Mellon University
Pittsburgh, PA, USA
[email protected]
et al., 2017), imitation learning (Daumé et al.,
2009; Ross et al., 2011; Bengio et al., 2015),
and beam-based approaches (Wiseman and Rush,
2016; Andor et al., 2016; Daumé III and Marcu,
2005). In this paper, we focus on one the simplest
to implement and least computationally expensive
approaches, scheduled sampling (Bengio et al.,
2015), which stochastically incorporates contexts
from previous decoding decisions into training.
While scheduled sampling has been empirically successful, its training objective has a drawback: because the procedure directly incorporates
greedy decisions at each time step, the objective is
discontinuous at parameter settings where previous decisions change their value. As a result, gradients near these points are non-informative and
scheduled sampling has difficulty assigning credit
for errors. In particular, the gradient does not provide information useful in distinguishing between
local errors without future consequences and cascading errors which are more serious.
Here, we propose a novel approach based on
scheduled sampling that uses a differentiable approximation of previous greedy decoding decisions inside the training objective by incorporating a continuous relaxation of argmax. As a result, our end-to-end relaxed greedy training objective is differentiable everywhere and fully continuous. By making the objective continuous at
points where previous decisions change value, our
approach provides gradients that can respond to
cascading errors. In addition, we demonstrate a
related approximation and reparametrization for
sample-based training (another training scenario
considered by scheduled sampling (Bengio et al.,
2015)) that can yield stochastic gradients with
lower variance than in standard scheduled sampling. In our experiments on two different tasks,
machine translation (MT) and named entity recognition (NER), we show that our approach outperforms both cross-entropy training and standard
ŷi
1 (✓)
= ‘kitten’
ŷi
1 (✓)
= ‘dog’
ŷi
1 (✓)
ŷi
= ‘cat’
1
= dog
argmax
objective(✓)
✓
Figure 1: Discontinuous scheduled sampling objective (red)
and continuous relaxations (blue and purple).
scheduled sampling procedures with greedy and
sampled-based training.
2
Discontinuity in Scheduled Sampling
While scheduled sampling (Bengio et al., 2015) is
an effective way to rectify exposure bias, it cannot differentiate between cascading errors, which
can lead to a sequence of bad decisions, and local
errors, which have more benign effects. Specifically, scheduled sampling focuses on learning optimal behavior in the current step given the fixed
decoding decision of the previous step. If a previous bad decision is largely responsible for the current error, the training procedure has difficulty adjusting the parameters accordingly. The following
machine translation example highlights this credit
assignment issue:
Ref: The cat purrs . Pred: The dog barks .
At step 3, the model prefers the word ‘barks’ after incorrectly predicting ‘dog’ at step 2. To correct this error, the scheduled sampling procedure
would increase the score of ‘purrs’ at step 3, conditioned on the fact that the model predicted (incorrectly) ‘dog’ at step 2, which is not the ideal
learning behaviour. Ideally, the model should be
able to backpropagate the error from step 3 to the
source of the problem which occurred at step 2,
where ‘dog’ was predicted instead of ‘cat’.
The lack of credit assignment during training
is a result of discontinuity in the objective function used by scheduled sampling, as illustrated in
Figure 1. We denote the ground truth target symbol at step i by yi∗ , the embedding representation of word y by e(y), and the hidden state of
a seq2seq decoder at step i as hi . Standard crossentropy training defines the loss at each step to be
∗ ), h
log p(yi∗ |hi (e(yi−1
i−1 )), while scheduled sampling uses loss log p(yi∗ |hi (e(ŷi−1 ), hi−1 )), where
hi
{
e(dog)
e(kitten)
e(cat)
si
si
si
1 (dog)
1 (kitten)
1 (cat)
peaked softmax
{
↵ = 10
↵=1
X
y
e(y) ·
exp [↵ · si
Z
1 (y)]
↵-soft argmax
ēi
1
1
hi
Figure 2: Relaxed greedy decoder that uses a continuous
approximation of argmax as input to the decoder state at next
time step.
ŷi−1 refers the model’s prediction at the previous step.1 Here, the model prediction ŷi−1 is
obtained by argmaxing over the output softmax
layer. Hence, in addition to the intermediate hidden states and final softmax scores, the previous model prediction, ŷi−1 , itself depends on the
model parameters, θ, and ideally, should be backpropagated through, unlike the gold target symbol
∗
yi−1
which is independent of model parameters.
However, the argmax operation is discontinuous,
and thus the training objective (depicted in Figure 1 as the red line) exhibits discontinuities at parameter settings where the previous decoding decisions change value (depicted as changes from ‘kitten’ to ‘dog’ to ‘cat’). Because these change points
represent discontinuities, their gradients are undefined and the effect of correcting an earlier mistake
(for example ‘dog’ to ‘cat’) as the training procedure approaches such a point is essentially hidden.
In our approach, described in detail in the next
section, we attempt to fix this problem by incorporating a continuous relaxation of the argmax operation into the scheduled sampling procedure in
order to form an approximate but fully continuous
objective. Our relaxed approximate objective is
depicted in Figure 1 as blue and purple lines, depending on temperature parameter α which tradesoff smoothness and quality of approximation.
3
Credit Assignment via Relaxation
In this section we explain in detail the continuous relaxation of greedy decoding that we will use
to build a fully continuous training objective. We
also introduce a related approach for sample-based
training.
1
For the sake of simplicity, the ‘always sample’ variant of
scheduled sampling is described (Bengio et al., 2015).
3.1
Soft Argmax
In scheduled sampling, the embedding for the best
scoring word at the previous step is passed as an
input to the current step. This operation2 can be
expressed as
êi−1 =
X
y
e(y)1[∀y 0 6= y si−1 (y) > si−1 (y 0 )]
where y is a word in the vocabulary, si−1 (y) is the
output score of that word at the previous step, and
êi−1 is the embedding passed to the next step. This
operation can be relaxed by replacing the indicator
function with a peaked softmax function with hyperparameter α to define a soft argmax procedure:
ēi−1 =
X
y
exp (α si−1 (y))
0
y 0 exp (α si−1 (y ))
e(y) · P
As α → ∞, the equation above approaches the
true argmax embedding. Hence, with a finite and
large α, we get a linear combination of all the
words (and therefore a continuous function of the
parameters) that is dominated heavily by the word
with maximum score.
3.2
Soft Reparametrized Sampling
Another variant of scheduled sampling is to pass
a sampled embedding from the softmax distribution at the previous step to the current step instead
of the argmax. This is expected to enable better
exploration of the search space during optimization due to the added randomness and hence result in a more robust model. In this section, we
discuss and review an approximation to the Gumbel reparametrization trick that we use as a module in our sample-based decoder. This approximation was proposed by Maddison et al. (2017) and
Jang et al. (2016), who showed that the same soft
argmax operation introduced above can be used
for reducing variance of stochastic gradients when
sampling from softmax distributions. Unlike soft
argmax, this approach is not a fully continuous approximation to the sampling operation, but it does
result in much more informative gradients compared to naive scheduled sampling procedure.
The Gumbel reparametrization trick shows that
sampling from a categorical distribution can be
refactored into sampling from a simple distribution followed by a deterministic transformation
2
Assuming there are no ties for the sake of simplicity.
as follows: (i) sampling an independent Gumbel noise G for each element in the categorical distribution, typically done by transforming
a sample from the uniform distribution: U ∼
U nif orm(0, 1) as G = −log(−log U ), then
(ii) adding it componentwise to the unnormalized
score of each element, and finally (iii) taking an
argmax over the vector. Using the same argmax
softening procedure as above, they arrive at an approximation to the reparametrization trick which
mitigates some of the gradient’s variance introduced by sampling. The approximation is3 :
ẽi−1 =
X
y
exp (α (si−1 (y) + Gy ))
0
y 0 exp (α (si−1 (y ) + Gy 0 ))
e(y) · P
We will use this ‘concrete’ approximation of softmax sampling in our relaxation of scheduled sampling with a sample-based decoder. We discuss
details in the next section. Note that our original motivation based on removing discontinuity
does not strictly apply to this sampling procedure,
which still yields a stochastic gradient due to sampling from the Gumbel distribution. However, this
approach is conceptually related to greedy relaxations since, here, the soft argmax reparametrization reduces gradient variance which may yield a
more informative training signal. Intuitively, this
approach results in the gradient of the loss to be
more aware of the sampling procedure compared
to naive scheduled sampling and hence carries forward information about decisions made at previous steps. The empirical results, discussed later,
show similar gains to the greedy scenario.
3.3
Differentiable Relaxed Decoders
With the argmax relaxation introduced above, we
have a recipe for a fully differentiable greedy decoder designed to produce informative gradients
near change points. Our final training network for
scheduled sampling with relaxed greedy decoding is shown in Figure 2. Instead of conditioning
the current hidden state, hi , on the argmax embedding from the previous step, êi−1 , we use the
α-soft argmax embedding, ēi−1 , defined in Section 3.1. This removes the discontinuity in the
original greedy scheduled sampling objective by
passing a linear combination of embeddings, dominated by the argmax, to the next step. Figure 1
3
This is different from using the expected softmax embedding because our approach approximates the actual sampling
process instead of linearly weighting the embeddings by their
softmax probabilities
illustrates the effect of varying α. As α increases,
we more closely approximate the greedy decoder.
As in standard scheduled sampling, here we
minimize the cross-entropy based loss at each time
step. Hence the computational complexity of our
approach is comparable to standard seq2seq training. As we discuss in Section 5, mixing model
predictions randomly with ground truth symbols
during training (Bengio et al., 2015; Daumé et al.,
2009; Ross et al., 2011), while annealing the probability of using the ground truth with each epoch,
results in better models and more stable training. As a result, training is reliant on the annealing schedule of two important hyperparameters: i)
ground truth mixing probability and ii) the α parameter used for approximating the argmax function. For output prediction, at each time step, we
can still output the hard argmax, depicted in Figure 2.
For the case of scheduled sampling with
sample-based training–where decisions are sampled rather than chosen greedily (Bengio et al.,
2015)–we conduct experiments using a related
training procedure. Instead of using soft argmax,
we use the soft sample embedding, ẽi−1 , defined
in Section 3.2. Apart from this difference, training
is carried out using the same procedure.
4
Related Work
Gormley et al. (2015)’s approximation-aware
training is conceptually related, but focuses on
variational decoding procedures. Hoang et al.
(2017) also propose continuous relaxations of decoders, but are focused on developing better inference procedures. Grefenstette et al. (2015) successfully use a soft approximation to argmax in
neural stack mechanisms. Finally, Ranzato et al.
(2016) experiment with a similarly motivated objective that was not fully continuous, but found it
performed worse than the standard training.
5
Experimental Setup
We perform experiments with machine translation
(MT) and named entity recognition (NER).
Data: For MT, we use the same dataset (the
German-English portion of the IWSLT 2014 machine translation evaluation campaign (Cettolo
et al., 2014)), preprocessing and data splits as Ranzato et al. (2016). For named entity recognition,
we use the CONLL 2003 shared task data (Tjong
Kim Sang and De Meulder, 2003) for German language and use the provided data splits. We perform no preprocessing on the data.The output vocabulary length for MT is 32000 and 10 for NER.
Implementation details: For MT, we use a
seq2seq model with a simple attention mechanism
(Bahdanau et al., 2015), a bidirectional LSTM encoder (1 layer, 256 units), and an LSTM decoder
(1 layer, 256 units). For NER, we use a seq2seq
model with an LSTM encoder (1 layer, 64 units)
and an LSTM decoder (1 layer, 64 units) with a
fixed attention mechanism that deterministically
attends to the ith input token when decoding the
ith output, and hence does not involve learning of
attention parameters. 4
Hyperparameter tuning: We start by training
with actual ground truth sequences for the first
epoch and decay the probability of selecting the
ground truth token as an inverse sigmoid (Bengio
et al., 2015) of epochs with a decay strength parameter k. We also tuned for different values of α
and explore the effect of varying α exponentially
(annealing) with the epochs. In table 1, we report
results for the best performing configuration of decay parameter and the α parameter on the validation set. To account for variance across randomly
started runs, we ran multiple random restarts (RR)
for all the systems evaluated and always used the
RR with the best validation set score to calculate
test performance.
Comparison We report validation and test metrics for NER and MT tasks in Table 1, F1 and
BLEU respectively. ‘Greedy’ in the table refers
to scheduled sampling with soft argmax decisions
(either soft or hard) and ‘Sample’ refers the corresponding reparametrized sample-based decoding scenario. We compare our approach with two
baselines: standard cross-entropy loss minimization for seq2seq models (‘Baseline CE’) and the
standard scheduled sampling procedure (Bengio
et al. (2015)). We report results for two variants
of our approach: one with a fixed α parameter
throughout the training procedure (α-soft fixed),
and the other in which we vary α exponentially
with the number of epochs (α-soft annealed).
4
Fixed attention refers to the scenario when we use the
bidirectional LSTM encoder representation of the source sequence token at time step t while decoding at time step t instead of using a linear combination of all the input sequences
weighted according to the attention parameters in the standard attention mechanism based models.
Table 1: Result on NER and MT. We compare our approach (α-soft argmax with fixed and annealed temperature) with standard
cross entropy training (Baseline CE) and discontinuous scheduled sampling (Bengio et al. (2015)). ‘Greedy’ and ‘Sample’ refer
to Section 3.1 and Section 3.2.
Training procedure
Baseline CE
Bengio et al. (2015)
α-soft fixed
α-soft annealed
6
NER (F1)
Dev
49.43
Test
53.32
Dev
20.35
Test
19.11
Greedy
Dev
Test
Sample
Dev
Test
Greedy
Dev
Test
Sample
Dev
Test
49.75
51.65
51.43
50.90
51.13
50.99
20.52
21.32
21.28
20.40
20.48
21.36
54.83
55.88
56.33
Results
All three approaches improve over the standard
cross-entropy based seq2seq training. Moreover,
both approaches using continuous relaxations
(greedy and sample-based) outperform standard
scheduled sampling (Bengio et al., 2015). The
best results for NER were obtained with the relaxed greedy decoder with annealed α which
yielded an F1 gain of +3.1 over the standard
seq2seq baseline and a gain of +1.5 F1 over standard scheduled sampling. For MT, we obtain
the best results with the relaxed sample-based decoder, which yielded a gain of +1.5 BLEU over
standard seq2seq and a gain of +0.75 BLEU over
standard scheduled sampling.
We observe that the reparametrized samplebased method, although not fully continuous endto-end unlike the soft greedy approach, results
in good performance on both the tasks, particularly MT. This might be an effect of stochastic exploration of the search space over the output sequences during training and hence we expect MT to benefit from sampling due to a much
larger search space associated with it. We also observe that annealing α results in good performance
which suggests that a smoother approximation to
the loss function in the initial stages of training is
helpful in guiding the learning in the right direction. However, in our experiments we noticed that
k
NER (F1)
100
56.33
10
55.88
MT (BLEU)
1
55.30
Always
54.83
Table 2: Effect of different schedules for scheduled sampling
on NER. k is the decay strength parameter. Higher k corresponds to gentler decay schedules. Always refers to the
case when predictions at the previous predictions are always
passed on as inputs to the next step.
54.60
56.25
54.20
19.85
20.28
20.18
19.69
19.69
20.60
the performance while annealing α was sensitive
to the hyperparameter associated with the annealing schedule of the mixing probability in scheduled sampling during training.
The computational complexity of our approach
is comparable to that of standard seq2seq training. However, instead of a vocabulary-sized max
and lookup, our approach requires a matrix multiplication. Practically, we observed that on GPU
hardware, all the models for both the tasks had
similar speeds which suggests that our approach
leads to accuracy gains without compromising
run-time. Moreover, as shown in Table 2, we observe that a gradual decay of mixing probability
consistently compared favorably to more aggressive decay schedules. We also observed that the
‘always sample’ case of relaxed greedy decoding,
in which we never mix in ground truth inputs (see
Bengio et al. (2015)), worked well for NER but
resulted in unstable training for MT. We reckon
that this is an effect of large difference between
the search space associated with NER and MT.
7
Conclusion
Our positive results indicate that mechanisms for
credit assignment can be useful when added to
the models that aim to ameliorate exposure bias.
Further, our results suggest that continuous relaxations of the argmax operation can be used as effective approximations to hard decoding during
training.
Acknowledgements
We thank Graham Neubig for helpful discussions.
We also thank the three anonymous reviewers for
their valuable feedback.
References
Daniel Andor, Chris Alberti, David Weiss, Aliaksei
Severyn, Alessandro Presta, Kuzman Ganchev, Slav
Petrov, and Michael Collins. 2016. Globally normalized transition-based neural networks. In Association for Computational Linguistics.
Dzmitry Bahdanau, Philemon Brakel, Kelvin Xu,
Anirudh Goyal, Ryan Lowe, Joelle Pineau, Aaron
Courville, and Yoshua Bengio. 2017. An actor-critic
algorithm for sequence prediction. In International
Conference on Learning Representations.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly
learning to align and translate. In International Conference on Learning Representations.
Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and
Noam Shazeer. 2015. Scheduled sampling for sequence prediction with recurrent neural networks.
In Advances in Neural Information Processing Systems. pages 1171–1179.
Mauro Cettolo, Jan Niehues, Sebastian Stüker, Luisa
Bentivogli, and Marcello Federico. 2014. Report on
the 11th iwslt evaluation campaign, iwslt 2014. In
Proceedings of the International Workshop on Spoken Language Translation, Hanoi, Vietnam.
Hal Daumé, John Langford, and Daniel Marcu. 2009.
Search-based structured prediction. Machine learning 75(3):297–325.
Hal Daumé III and Daniel Marcu. 2005. Learning
as search optimization: Approximate large margin
methods for structured prediction. In Proceedings
of the 22nd international conference on Machine
learning. ACM, pages 169–176.
Matthew R. Gormley, Mark Dredze, and Jason Eisner.
2015. Approximation-aware dependency parsing by
belief propagation. Transactions of the Association
for Computational Linguistics (TACL) .
Edward Grefenstette, Karl Moritz Hermann, Mustafa
Suleyman, and Phil Blunsom. 2015. Learning to
transduce with unbounded memory. In Advances
in Neural Information Processing Systems. pages
1828–1836.
Cong Duy Vu Hoang, Gholamreza Haffari, and Trevor
Cohn. 2017. Decoding as continuous optimization in neural machine translation. arXiv preprint
arXiv:1701.02854 .
Eric Jang, Shixiang Gu, and Ben Poole. 2016. Categorical reparameterization with gumbel-softmax. In
International Conference on Learning Representations.
Chris J Maddison, Andriy Mnih, and Yee Whye Teh.
2017. The concrete distribution: A continuous relaxation of discrete random variables. In International Conference on Learning Representations.
Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli,
and Wojciech Zaremba. 2016. Sequence level training with recurrent neural networks. In International
Conference on Learning Representations.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell.
2011. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS. volume 1, page 6.
Alexander M Rush, Sumit Chopra, and Jason Weston.
2015. A neural attention model for abstractive sentence summarization. In Empirical Methods in Natural Language Processing.
Iulian V Serban, Alessandro Sordoni, Yoshua Bengio,
Aaron Courville, and Joelle Pineau. 2015. Building
end-to-end dialogue systems using generative hierarchical neural network models. In AAAI’16 Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural networks. In Advances in neural information processing systems. pages 3104–3112.
Erik F Tjong Kim Sang and Fien De Meulder.
2003. Introduction to the conll-2003 shared task:
Language-independent named entity recognition. In
Proceedings of the seventh conference on Natural
language learning at HLT-NAACL 2003-Volume 4.
Association for Computational Linguistics, pages
142–147.
Sam Wiseman and Alexander M Rush. 2016.
Sequence-to-sequence learning as beam-search optimization. In Empirical Methods in Natural Language Processing.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho,
Aaron C Courville, Ruslan Salakhutdinov, Richard S
Zemel, and Yoshua Bengio. 2015. Show, attend and
tell: Neural image caption generation with visual attention. In ICML. volume 14, pages 77–81.
| 9cs.NE
|
1
Coordinated Linear Precoding in Downlink
Multicell MU-MISO OFDMA Networks
Mirza Golam Kibria, Hidekazu Murata and Susumu Yoshida
arXiv:1708.04750v1 [cs.IT] 16 Aug 2017
Graduate School of Informatics, Kyoto University, Kyoto, Japan
Email: [email protected]
Abstract
This paper considers coordinated linear precoding in downlink multicell multiuser orthogonal frequencydivision multiple access (OFDMA) network. A less-complex, fast and provably convergent algorithm that
maximizes the weighted sum-rate with per base station (BS) transmit power constraint is formulated.
We approximate the nonconvex weighted sum-rate maximization (WSRM) problem with a solvable
convex form by means of sequential parametric convex approximation (SPCA) approach. The second
order cone program (SOCP) formulations of the objective function and constraints of the optimization
problem are derived through proper change of variables, first order linear approximation and hyperbolic
constraints transformation, etc. The algorithm converges to the suboptimal solution taking fewer number
of iterations in comparison to other known iterative WSRM algorithms. Finally, numerical results are
presented to justify the effectiveness and superiority of the proposed algorithm.
Index Terms
Weighted sum-rate maximization, Coordinated linear precoding, Convex approximation.
I. Introduction
The weighted sum-rate maximization (WSRM) problem is known to be nonconvex and NPhard [1], [2], even for single antenna users. Though the beamforming design methods presented
in [3], [4] achieve optimum capacity, these methods may be practically inapplicable since the
complexity evolves exponentially with the optimization problem size. Therefore, computationally
inexpensive suboptimal beamforming design is very appealing. Beamforming design based on
achieving necessary conditions of optimality has been studied thoroughly in [1], [2]. Importantly,
2
the authors of [3] numerically prove that the performances of the suboptimal beamforming
techniques that achieve the necessary optimality conditions are indeed very close to optimal
beamforming design.
In [1], the authors proposed iterative coordinated beamforming design based on Karush-KuhnTucker (KKT) optimality conditions, which is not provably convergent. Alternating maximization
(AM) algorithm for WSRM optimization problem is proposed in [2], which is based on alternating updation between a closed-form posterior conditional probability and the beamforming
vectors. In [5]–[7], the authors establish a relationship between weighted sum-rate and weighted
minimum mean-square error (WMMSE), and solve the WSRM optimization problem based
on alternating optimization. Discrete power control based WSRM has been proposed in [8].
However, all these iterative WSRM optimization designs exhibit relatively slower convergence
rate in comparison to our proposed design.
In this paper, we formulate and propose a WSRM optimization solution with faster convergence
for multicell orthogonal frequency division multiple access (OFDMA) system. This iterative
design manipulates the sequential parametric convex approximation (SPCA) technique explored
in [9]. The SPCA based WSRM algorithm approaches the local optimal solution within a few
iterations, iteratively approximating the nonconvex problem with solvable convex structure. At
each step of this iterative process, the nonconvex problem is approximated with a solvable convex
form and updating the acting variables until convergence. With appropriate change of variables,
introducing additional optimization variable, making use of first-order linear approximation
and hyperbolic constraints transformations, we iteratively approximate the WSRM optimization
problem as a second order cone program (SOCP) [12].
The reminder of the paper is organized as follows. The multicell multiuser OFDMA network
model and WSRM optimization framework are presented in Section II. Section III explains the
process of sequential convex approximation of the nonconvex optimization problem. In Section
IV, we discuss the simulation parameters and numerical results found in this work. Section V
concludes the paper.
Notations: (·)H /(·)T stand for Hermitian-transpose/transpose operation. Gaussian distributions of
real/complex random variables with mean µ and variance σ2 is defined as RN(µ, σ2 )/CN(µ, σ2 ).
Boldface lower-case/upper-case letter defines a vector/ matrix. Operator vec(·) stacks all the
elements of the argument into a column vector and diag(·) puts the diagonal elements of a
3
matrix in a column vector. R and C define real and complex spaces, respectively. | · | and || · ||2
refer to absolute value and l2 norm of the arguments, respectively.
II. Problem Formulation
A. System Model
We consider an interference-limited cellular system of M cells with K users per cell. OFDMA
multiplexing scheme with N subcarriers over a fixed bandwidth is employed, while the subcarrier
assignments among users within each cell are non-overlapping. Therefore, there is no intracell interference, only inter-cell interference is experienced by the users. The coordinated base
stations (BSs) are equipped with Nt antennas and they are interconnected via high-capacity
backhaul links. The non-cooperative users have single antenna each. Coordinated linear multiuser
downlink precoding is employed at each BS. The assignment function f (m, n) determines the
downlink user scheduling. The assignment of user k from mth BS on the nth subcarrier is defined
as k = f (m, n). The received data of user k from cell m on the nth subcarrier is given by
X
ykmn = hkmn gkmn dkmn +
hkm0 n gk0 m0 n dk0 m0 n + zkmn
(1)
m0 ∈S\m
k0 = f (m0 ,n)
where k = f (m, n) and S is the set of all BSs. ykmn ∈ C denotes the received symbol for user
k. hkmn ∈ C1×Nt is the complex channel vector between BS m and user k. The beamformer
formed by BS m to transmit data on subcarrier n is denoted by gkmn ∈ CNt ×1 . zkmn ∼ CN(0, 1) is
the additive white Gaussian noise (AWGN) at user k. dkmn ∼ CN(0, 1) denotes the transmitted
symbol from BS m to user k on subcarrier n.
B. Transmit Precoding Problem
This paper emphasizes the linear beamformer design for sum-rate optimization in multicell
multiuser OFDMA network. The design objective is to maximize the weighted sum-rate under
per BS power constraints. The signal-to-interference-plus-noise ratio (SINR) of the kth user from
cell m scheduled on subcarrier n is given by
γkmn =
1+
H
hkmn gkmn gkmn
hHkmn
.
P
hkm0 n gk0 m0 n gkH0 m0 n hHkm0 n
m0 ∈S\m
k0 = f (m0 ,n)
(2)
4
The instantaneous downlink rate achieved by the kth user from cell m on subcarrier n is ckmn =
P
log2 (1 + γkmn ), and the instantaneous rate for user k over all the subcarriers is Rkm = n∈Skm ckmn ,
where the summation is over all the subcarriers assigned to user k from cell m, i.e., n ∈ Skm ,
where Skm = {n|k = f (m, n)}. Let wkm be the weight of user k in cell m. The weight corresponding
to a particular user may reflect the quality of the service it requests or its priority in the system.
Then the WSRM problem is defined as
maximize
G
subject to
N
M X
X
wkm ckmn
m=1 n=1
N
X
(3)
||gkmn ||22 ≤ Pm,max , m = 1, ..., M
n=1
where G := {gkmn ; m ∈ M, n ∈ N} is the set of all beam forming vectors and Pm,max is the
transmit power constraint of cell m. Let Gm be the set of beamformers for cell m. Since the
optimization problem in (7) is nonconvex, finding the global optimal solution is difficult and
complex enough. Therefore, we focus on local optimal solution in this paper.
C. Review of Second Order Cone Programming
Recently, substantial progress and development have been achieved for solving a large class of
optimization problems. In order to apply these algorithms, one needs to reformulate the problem
into the standard form that the algorithms are capable of dealing with. Conic programs, i.e.,
linear programs [12] with generalized inequalities, are subjected to special attention. One such
standard conic program is SOCP, which is of the form
minimize Real(aH x)
x
H
SOCP :
c
x
+
d
i
i
0, i = 1, ..., U
subject to
M
H
D x + bi
(4)
i
where x is the vector of optimization variables and Di , bi , ci , ai are parameters with appropriate
sizes. The notation M defines the generalized inequalities:
h
i
(v s)T M 0 ⇔ ||s||2 ≤ v.
(5)
Hyperbolic constraints play an important role in the SOCP formulation of WSRM objective
function and constraints. The hyperbolic constraints w2 ≤ xy, x ≥ 0, y ≥ 0 with w ∈ R1×e ,
5
x, y ∈ R and the equivalent SOCP is given by [12]
wT w ≤ xy, x ≥ 0, y ≥ 0 ⇔
2w
x−y
≤ x + y.
(6)
2
III. Sequential Parametric Convex Approximation for WSRM Problem
As a step toward transforming the nonconvex WSRM optimization problem to SOCP1 form,
we reformulate the problem (3) into a standard form that SOCP programming is capable of
dealing with. We rewrite (3) as
M
N
X
X
maximize
G
m=1
subject to
N
X
δkmn log2 (1 + γkmn )
n=1
k= f (m,n)
(7)
||gkmn ||22 ≤ Pm,max , m = 1, ..., M
n=1
where δkmn = wkm , ∀n. Let L := {kmn, ∀m, n | k = f (m, n)} and T = MN. Therefore, the objective
function becomes a function of T variables and can be expressed as
max
G
T
X
δLt log2 (1 + γLt ) = max
G
t=1
T
Y
(1 + γLt )δLt
(8)
t=1
δ Lt
where Lt is the tth set in L. Setting rLt = (1 + γLt ) , we get
maximize
G,rLt
T
Y
rLt
t=1
subject to C1 :
N
X
||gkmn ||22 ≤ Pm,max , m = 1, ..., M
(9)
n=1
q
C2 : rLLt t ≤ γLt + 1, ∀Lt ∈ L, t = 1, ..., T
where qLt = 1/δLt and the constraints in C2 are active at the optimum. Per BS transmit power
p
constraint C1 of (9) can be reformulated using vec(·) as ||vec (Gm ) ||2 ≤ Pm,max , for which the
equivalent SOC according to (5) is expressed as
p
Pm,max
vec (G ) M 0.
m
1
(10)
SOCP constraints are convex and can be solved using convex optimization tools such as SeDuMi [13]. Also note that the
non-overlapping subcarrier allocation in each cell does not restrict the applicability of the proposed algorithm for multiple active
users in one subcarrier in one cell.
6
Further, introducing slack variables ζLt ≥ 0, we can reformulate (9) using (2) as given below
maximize
G,rLt ,ζLt
T
Y
rLt
t=1
p
Pm,max
0, m = 1, ..., M
subject to C1 :
M
vec (Gm )
q
C2 : ζLt (rLLt t − 1)1/2 ≤ hLt gLt
(11)
C32 : I{hLt gLt } = 0, I {x} = Imaginary part of x
v
X
u
t
hkm0 n gk0 m0 n gkH0 m0 n hHkm0 n ≤ ζLt
C4 : 1 +
m0 ∈S\m
k0 = f (m0 ,n)
Let Hint ∈ C(M−1)×Nt and Gint ∈ CNt ×(M−1) be the collected channel and beamforming matrices, respectively, containing the channels from all interfering BSs and beamforming vectors
corresponding to the constraint C4 of (11). Therefore, we can write the constraint C4 as
T
≤ ζLt , which is equivalent to the SOCP constraint
1 diag (Hint Gint )
2
ζ
L
t
T
M 0.
1 diag (Hint Gint )
(12)
Constraints C1, C3-C4 of (11) are convex, hence require no approximation. However, C2 is still
nonconvex. To make use of the SPCA technique to approximate C2 as a convex constraint, we
break C2 of (11) and reformulate as
v1/2
Lt ζLt ≤ hLt gLt , ∀Lt ∈ L
q
rLLt t ≤ vLt + 1.
(13)
(14)
Though both (13) and (14) are still nonconvex, yet this formulation facilitates us to use the
established convex approximation methods. First, we consider the convex approximation of (13).
Defining Q(ζLt , vLt ) = v1/2
Lt ζLt with vLt , ζLt ≥ 0, we approximate Q(ζLt , vLt ) with its convex upper
estimate function [9] G(ζLt , vLt , θLt ) as
!
1 vLt
2
+ θLt ζLt .
G(ζLt , vLt , θLt ) ,
2 θLt
2
(15)
For any φ, we have |hLt gLt |2 = |hLt gLt e jφ |2 . Therefore, by choosing φ such that I{hLt gLt } = 0 does not affect the optimality
of (11).
7
Hence, Q(ζLt , vLt ) ≤ G(ζLt , vLt , θLt ), ∀θLt ≥ 0. At the optimum, Q(ζLt , vLt ) = G(ζLt , vLt , θLt ) when
√
θLt = vLt /ζLt . This point can be reached in an iterative way by intuitively updating the variables
until we obtain the KKT points of (11). Convex overestimation of Q(ζLt , vLt ) allows us to express
equation (13) as hyperbolic constraints, and the SOCP representation for the corresponding
hyperbolic constraints is given by
r
T
v Lt
θ Lt
ζLt
(hLt gLt −
− 1)]
2
2θLt
≤ (hLt gLt −
2
v Lt
+ 1)
2θLt
(16)
which can be equivalently expressed as SOCP constraint as
vLt
+
1
h
g
−
L
L
t
t
2θLt
" q
#
T
M 0.
θ Lt
v
ζ
(h g − Lt − 1)
Lt
Lt Lt
2
(17)
2θLt
q
Now, let us turn our focus on (14) and we notice that the term rLLt t is a differentiable function. To
q
arrive at SOCP, we scale all qLt such that qLt < 1 so as to make the function rLLt t concave. For a
differentiable function V with (∀x, y ∈ domain (V)), the first order condition for concavity says
that a function V is concave if and only if the gradient line is the global over-estimator of the
function [12]. The function V(x) + ∇ x V(x)T (y − x) is defined as the first order approximation
to the function at x, where (∇ x V(x))i =
∂V(x)
.
∂xi
Correspondingly, we approximate rkqk with its
concave over-estimator as follows
q
q
q −1
rLLt t − rLLt ,it ≤ qLt rLLt ,it (rLt − rLt ,i )
q −1
(18)
q
i.e., vLt ≥ qLt rLLt ,it (rLt − rLt ,i ) + rLLt ,it − 1 (using (14))
q
and iteratively solve until convergence in parallel with (16). In fact, it is the linearization of rLLt t
around the point rLt ,i , where rLt ,i is the value of rLt at the ith iteration. Both (16) and (18) are
increasing function; however, they are upper bounded by the per BS power constraints. Now,
we turn our attention to the objective function. There are two possible ways to convexify the
objective function of (11) and the methods are as follows
Method 1: The geometric mean (GM) of the optimization variables χ = (rL1 rL2 ...rLT )1/T is
concave when rLt 0, ∀Lt . Maximizing the GM of the optimization variables will serve the
same weighted sum-rate as maximizing the product of the optimization variables as long as the
variables are nonnegative affine [12], hence we can rewrite the objective function as
T
T
Y
Y
maximize
rLt :⇔ maximize
(rLt )1/T .
G,rLt ,ζLt
t=1
G,rLt ,ζLt
t=1
(19)
8
Using the CVX [14] solver with SeduMi, a disciplined convex programming, we can directly use
the GM of the optimization variables as objective function. We refer this method as SPCA-GM
and it is not in SOCP form.
Method 2: The second approach is based on transforming the product of the optimization
variables into hyperbolic constraints, which also admit SOCP representation. Thus, we require
to reformulate the problem by introducing new variables and by incorporating hyperbolic constraints. Let us define the set of new variables as ψ. During the transformation process the
variables are assigned values at log2 T stages. For simplified analysis, let T = 2 p , where p is a
real positive quantity. The transformation procedure is provided below.
Procedure 1: for hyperbolic constraints transformation
Initialize: ψtp = rLt , t = 1, ..., T and p = log2 (T )
for j = p, p − 1, ..., 1
j−1 2
j
ψi
≤ ψ2i−1
ψ2ij , i = 1, ..., 2 j−1
end
9
The overall SPCA-WSRM algorithm is summarized here:
SPCA-WSRM algorithm:
1. Initialize: Imax , (θiLt , rLi t , ζLi t ), i = 0
2. repeat
3. solve the following:
maximize χ (if GM approach (Method 1) is used) or
G,rLt ,ζLt
maximize ψ0 (if SOCP approach (Method 2) is used)
G,rLt ,ζLt ,vLt ,ψLt
subject to
C1 : Procedure 1with (6) (ignore if Method 1 is used).
p
Pm,max
M 0, m = 1, ..., M
C2 :
vec (Gm )
v Lt
+
1
h
g
−
Lt Lt
2θiL
t
#T M 0
C3 : " q θi
v Lt
ζ
Lt
(h
−
1)
g
−
Lt
L
L
i
t
t
2
2θ
Lt
C4 : I{hLt gLt } = 0
q −1
q
C5 : vLt ≥ qLt rLLt ,it (rLt − rLt ,i ) + rLLt ,it − 1
ζ
Lt
T M 0
C6 :
1 diag (Hint Gint )
C7 : ψLt ≥ 0, rLt ≥ 0 implicit constraints
4. denote (rLi+1
, ζLi+1
, vi+1
Lt ) = optimal values at step 3.
t
t
q
i+1
vi+1
5. θi+1
Lt =
Lt /ζLt , i = i + 1
6. until convergence or i = Imax
The objective function emerges to be a one variable function defined as ψ01 = ψ0 , which is
obtained at the final stage of hyperbolic constraint formulation described in Procedure 1. Finally,
applying (6) yields the SOCP formulations for 2 p − 1 hyperbolic equations of Method 2. It is
worth noting that this algorithm is inspired by [9]–[11] and is similar to [10], which proposes the
10
SPCA based algorithm for multicell MU-MISO networks. However, we formulate and propose
the SPCA based algorithm with GM approach for multicell OFDMA networks and resolve two
practical limiting factors related to the algorithm implementation, which are not addressed in
[10] to make the algorithm more general, especially when the problem size is comparatively
larger.
The initial θLt s are very crucial to the feasibility and convergence of the SPCA-WSRM
algorithm. It could be possible that for some cases, the randomly generated θLt s can lead to
infeasible solution at the first iteration. To make sure that the algorithm is feasible at the first
step, we follow the steps in Procedure 2 to find good initial θLt s.
The other numerical issue that is not addressed in [10] is the situation when one or some of
the vLt s become zero, i.e., no power on that or those particular subcarriers of the corresponding
cell. It is usual that some of the subcarriers may not get any power due to limited BS power
if we recall the mechanism of water-filling algorithm. However, when such situation arises, we
have noticed numeral instability. We encounter the problem of dividing by zero since we need to
calculate 1/θLt . In order to avoid this situation we slightly modify the imposed constraints on vLt
such as vLt ≥ ε (e.g., ε=0.0001) so that we bypass the numerical problem. By this constraint, the
algorithm yields a solution that is close to the original one without encountering the numerical
instability.
Procedure 2: Proposal for generating initial values of θLt
Step 1: Generate channel-matched beamforming vectors so
that per BS power constraint is satisfied for all cells, i.e.,
p
gkmn = Pm,max /N(hkmn /||hkmn ||2 ), ∀m, n and k = f (m, n)
Step 2: Use C4 of (11) to find ζL0t by replacing inequality
with equality.
Step 3: Calculate rL0 t from C2 of (11) putting the absolute
value of hLt gLt .
Step 4: Find vLt using (13). Finally the initial value of θLt
is obtained as θ0Lt =
√
vLt /ζL0t .
11
Fig. 1. Convergence rate comparison for different WSRM algorithms.
IV. Numerical Results
The performance of the proposed algorithm is analyzed on a cellular network with 3 coordinated BSs and 2 users per cell, with 1-cell frequency reuse factor, via Monte-Carlo simulations.
The distance between adjacent BSs is 1000 m. The users are uniformly distributed around its own
BS within a circular annulus of external and internal radii of 1000 m and 500 m, respectively.
Like the paper [1], frequency-selective channel coefficients over 64 subcarriers are modeled as
!3.5
1
hkmn = 200
Φkmn Λkmn
(20)
lkm
where lkm is the distance between BS m and user k. 10log10 (Φkmn ) is distributed as RN(0, 8),
accounting for log-normal shadowing and Λkmn ∼ CN(0, 1) accounts for Rayleigh fading. All
the BSs are subjected to the equal maximum power constraint, i.e., Pm,max = Pmax , ∀m. We also
consider that perfect channel state information (CSI) is available both at the BSs and users. The
initial-user assignment is performed randomly. We consider Nt = 2 and use CVX [14] package
for specifying and solving convex programs.
In Fig. 1, we compare the WSR achieved by all schemes as a function of the number of
iterations required to acquire steady output for a random channel realization. The maximum
power limit for all the BS is set to 20 dBW, i.e., Pmax = 20 dBW. It is easily noticed that SPCAWSRM algorithm converges within few iterations, while the AM and WMMSE are still far
away from convergence level of SPCA-WSRM. This phenomenon may be attributed to the fact
12
Fig. 2. (a). Performance comparison between Method 1 and Method 2, (b). Average sum-rate performances for different
WSRM algorithms.
that AM-WSRM optimization requires alternation between a closed-form posterior conditional
probability update and updating the beamforming vectors, while the WMMSE algorithm relies
on the relationship between mutual information and minimum mean-square error (MMSE), and
alternates between updating of transmit and receive beamformers. As a result, comparatively
slower convergences are observed. However, good initial values for the variables involved in
WMMSE accelerate the convergence rate. Though SIN algorithm, which is also based on convex
approximation of the precoder covariance matrices, has similar convergence performance to
SPCA-WSRM. However, the per iteration running time is much higher.
Fig. 2a compares the WSR performances for the two different methods described in the
previous section. For both methods, we generate the initial values of θLt s using Procedure 2
and modify the constraints on vLt s as we discussed. Although both methods exhibit same WSR
performance for higher values of ε, the per iteration running time for Method 1 is much longer
than Method 2. This is attributed to the fact that the solver internally transforms the GM to
hyperbolic constraints in each iteration. We have observed that the algorithm provides feasible
solution to the optimization problem all the times. It is obvious that the larger the value of ε,
the bigger performance gap between Method 2 and Method 2 evolves.
Finally, in Fig. 2b, we compare the average sum-rate (wkm =1) performances for various
precoding strategies as a function of per BS transmit power. The suboptimal solutions achieved
13
by SPCA-WSRM algorithm and other techniques such as AM and WMMSE are indeed very
close to the optimal precoding performance obtained from [3]. However, AM and WMMSE
require a large number of iterations to reach their respective suboptimal levels.
V. Conclusions
In this paper, we study the WSRM optimization problem for a multicell OFDMA multiplexing
system. We formulate and propose an SPCA based convex approximation of the optimization
problem, which is known to be nonconvex and NP-hard. This iterative SOCP optimization
is provably convergent to the local optimal solution. Some numerical issues related to the
algorithm implementation are also discussed. Particularly, in terms of convergence rate, this
algorithm exhibits excellent performance and outperforms some previously analyzed solutions
to the WSRM optimization problem.
References
[1] L. Venturino, N. Prasad, and X. Wang, “Coordinated linear beamforming in downlink multi-cell wireless networks,” IEEE
Trans. on Wireless Commun., vol. 9, no. 4, pp. 1451-1461, Apr. 2010.
[2] K. Wang, X. Wang, W. Xu, and X. Zhang, “Coordinated Linear Pecoding in Downlink Multicell MIMO-OFDMA Networks,”
IEEE Trans. on Signal Process., vol. 60, no. 8, pp. 4264-4277, Aug. 2012.
[3] S. Joshi, P. Weeraddana, M. Codreanu, and M. Latva-aho, “Weighted sum-rate maximization for MISO downlink cellular
networks via branch and bound,” IEEE Trans. on Signal Process., vol. 60, no. 4, pp. 2090-2095, Apr. 2012.
[4] L. Liu, R. Zhang, and K.-C. Chua, “Achieving global optimality for weighted sum-rate maximization in the K-user Gaussian
interference channel with multiple antennas” IEEE Trans. on Wireless Commun., vol. 11, no. 5, pp. 1933-1945, May. 2012.
[5] S. S. Christensen, R. Agarwal, E. Carvaldho, and J. Cioffi, “Weighted sum-rate maximization using weighted MMSE for
MIMO-BC beamforming design,” IEEE Trans. on Wireless Commun., vol. 7, no. 12, pp. 4792-4799, Dec. 2008.
[6] F. Sun and D. Carvalho, “Weighted MMSE Beamforming Design for Weighted Sum-rate Maximization in Coordinated
Multi-Cell MIMO Systems,” in Proc. IEEE Vehicular Technology Conference., Quebec, Canada, Sep. 2012, pp. 1-5.
[7] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility
maximization for a MIMO interfering broadcast channel,” IEEE Trans. on Signal Process., vol. 59, no. 9, pp. 4331-4340,
Sep. 2011.
[8] H. Zhang, Venturino, N. Prasad, P. Li, S. Rangarajan, and X. Wang, “Weighted sum-rate maximization in multi-cell networks
via coordinated scheduling and discrete power control,” IEEE J. Sel. Areas Commun., vol. 29, no. 6, pp. 1214-1224,
Jun. 2011.
[9] A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric convex approximation method with applications to
nonconvex truss topology design problems,” Journal of Global Optimization, vol. 47, no. 1, pp. 29-51, 2010.
[10] L. Tran, M. F. Hanif, A. Tolli, and M. Juntti, “Fast Converging Algorithm for Weighted Sum Rate Maximization in
Multicell MISO Downlink,h IEEE Signal Process. Letters, vol. 19, no. 12, pp. 872-875, Dec. 2012.
14
[11] Chris T. K. Ng and H. Huang, “Linear precoding in cooperative MIMO cellular networks with limited coordination
clusters,h IEEE J. Sel. Areas Commun., vol. 28, no. 9, pp. 1446-1454, Dec. 2010.
[12] M. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, “Applications of second-order cone programming,” Linear Algebra
and its Applications, vol. 248, pp. 193-228, Nov. 1998.
[13] J. F. Strum, “Using Sedumi 1.02, a Matlab toolbox for optimization over symmetric cones,” Optimization Methods and
Software, vol. 11, no. 12, pp. 625-693, 1999.
[14] CVX Research, Inc, “CVX: Matlab software for disciplined convex programming, version 2.0 beta.” http://cvxr.com/cvx,
Sep. 2012.
| 7cs.IT
|
arXiv:1708.04739v2 [math.CO] 27 Sep 2017
THE SLACK REALIZATION SPACE OF A POLYTOPE
JOÃO GOUVEIA, ANTONIO MACCHIA, REKHA R. THOMAS, AND AMY WIEBE
Abstract. In this paper we introduce a natural model for the realization
space of a polytope up to projective equivalence which we call the slack realization space of the polytope. The model arises from the positive part of
an algebraic variety determined by the slack ideal of the polytope. This is a
saturated determinantal ideal that encodes the combinatorics of the polytope.
We also derive a new model of the realization space of a polytope from the
positive part of the variety of a related ideal. The slack ideal offers an effective
computational framework for several classical questions about polytopes such
as rational realizability, projective uniqueness, non-prescribability of faces, and
realizability of combinatorial polytopes.
The simplest slack ideals are toric. We identify any toric ideal that arises
from a projectively unique polytope as the toric ideal of the bipartite graph
given by the vertex-facet non-incidences of the polytope. Several new and
classical examples illuminate the relationships between projective uniqueness
and toric slack ideals.
1. Introduction
An important focus in the study of polytopes is the investigation of their realization spaces. Given a d-polytope P ⊂ Rd , its face lattice determines its combinatorial type. A realization space of P is, roughly speaking, the set of all geometric
realizations of the combinatorial type of P . This set, usually defined by fixing an
affinely independent set of vertices in every realization of P , is a primary basic
semialgebraic set, meaning that it is defined by a finite set of polynomial equations
and strict inequalities.
Foundational questions about polytopes such as whether there is a polytope with
rational vertices in the combinatorial class of P , whether a combinatorial type has
any realization at all as a convex polytope, or whether faces of a polytope can be
freely prescribed, are all questions about realization spaces. In general, many of
these questions are hard to settle and there is no straightforward way to answer
them by working directly with realization spaces. Each instance of such a question
often requires a clever new strategy; indeed, the polytope literature contains many
ingenious methods to find the desired answers.
Date: September 29, 2017.
Key words and phrases. polytopes; slack matrix; slack ideal; realization spaces; binomial ideal;
toric ideal; projective uniqueness.
Gouveia was partially supported by the Centre for Mathematics of the University of Coimbra
– UID/MAT/00324/2013, funded by the Portuguese Government through FCT/MEC and cofunded by the European Regional Development Fund through the Partnership Agreement PT2020,
Macchia was supported by INdAM, Thomas by the U.S. National Science Foundation grant DMS1418728, and Wiebe by NSERC.
1
2
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Further obscuring the matter is the fact that realization spaces of polytopes
can be arbitrarily complicated. The celebrated Universality Theorem of RichterGebert [RG96a] states that every primary basic semialgebraic set defined over Z is
stably equivalent to the realization space of a 4-dimensional polytope. An earlier
Universality Theorem of Mnëv [Mnë88] makes the same claim for realization spaces
of d-dimensional polytopes with d + 4 vertices.
In this paper, we introduce a model for the realization space of a polytope in
a given combinatorial class modulo projective transformations. This space arises
from the positive part of an algebraic variety called the slack variety of the polytope.
An explicit model for the realization space of the projective equivalence classes of
a polytope does not exist in the literature, although several authors have implicitly
worked modulo projective transformations [AP17, APT15, RG96b]. Using a related
idea, we also construct a model for the realization space for a polytope that is
rationally equivalent to the classical model for the realization space of the polytope.
The ideal giving rise to the slack variety is called the slack ideal of the polytope
and was introduced in [GPRT17]. The slack ideal in turn was inspired by the slack
matrix of a polytope. This is a nonnegative real matrix with rows (and columns)
indexed by the vertices (and facets) of the polytope and with (i, j)-entry equal to the
slack of the ith vertex in the jth facet inequality. Each vertex/facet representation
of a d-polytope P gives rise to a slack matrix SP of rank d + 1. Slack matrices
have found remarkable use in the theory of extended formulations of polytopes (see
for example, [Yan91], [FMP+ 12], [Rot14], [GPT13], [LRS15]). Slack matrices were
characterized in [GGK+ 13].
1.1. Our contribution. By passing to a symbolic version of the slack matrix
SP , wherein we replace every positive entry by a distinct variable in the vector
of variables x, one gets a symbolic matrix SP (x). The slack ideal IP is the ideal
obtained by saturating the ideal of (d + 2)-minors of SP (x) with respect to all
variables. The complex variety of IP , V(IP ), is the slack variety of P . We prove
that modulo a group action, the positive part of V(IP ) is a realization space for
the projective equivalences classes of polytopes that are combinatorially equivalent
to P . This is the slack realization space of P and it provides a new model for
the realizations of a polytope modulo projective transformations. Working with a
slightly modified ideal called the affine slack ideal of P , we also obtain a realization
space for P that is rationally equivalent to the classical realization space of P . We
call this the affine slack realization space of P . By the positive part of a complex
variety we mean the intersection of the variety with the positive real orthant of the
ambient space.
The slack realization space has several nice features. The inequalities in its
description are simply nonnegativities of variables in place of the determinantal
inequalities in the classical model. By forgetting these inequalities one can study
the entire slack variety, which is a natural algebraic relaxation of the realization
space. The slack realization space naturally mods out affine equivalence among
polytopes and, unlike in the classical construction, does not depend on a choice
of affine basis. It also offers a natural way to study polytopes up to projective
equivalence.
Additionally, the slack ideal provides a computational engine for establishing
several types of results one can ask about the combinatorial class of a polytope. We
exhibit three concrete applications of this machinery to determine non-rationality,
THE SLACK REALIZATION SPACE OF A POLYTOPE
3
non-prescribability of faces, and non-realizability of polytopes. We expect that
several other applications will be amenable to this algebraic framework.
A well-studied question about a combinatorial class is whether there is only one
realization up to projective equivalence. We say that P is projectively unique if that
is the case. The only known characterizations of projectively unique polytopes are
in dimensions two and three [Grü03, Exercise 4.8.30, p. 68]. We show that when
the slack ideal is toric, it is easy to determine projective uniqueness, thus creating a
methodology for checking this property in a specific class of polytopes in arbitrary
dimension.
Since slack ideals do not contain monomials, the simplest slack ideals we can
hope for are binomial ideals. An especially nice class of binomial ideals is the
class of toric ideals, which come with a rich combinatorial structure. The vertexfacet incidence structure of a polytope P gives rise to a bipartite graph whose
toric ideal, TP , plays a special role in this context. We call TP the toric ideal of
the non-incidence graph of P , and say that IP is graphic if it coincides with TP .
In Theorem 6.3 we prove that IP is graphic if and only if IP is toric and P is
projectively unique. This does not characterize all projectively unique polytopes,
as there are infinitely many combinatorial types in high enough dimension that
are projectively unique but do not have toric slack ideals. On the other hand,
non-projectively unique polytopes can have toric slack ideals. We illustrate these
features with several concrete examples.
The toric ideal TP has other interesting geometric connections. We prove that
whenever IP is contained in TP , the polytope P is morally 2-level, which is a polarinvariant property of a polytope that generalizes the notion of 2-level polytopes
[Sta80], [BFF+ 17], [FFM16], [GS17]. Theorem 5.13 characterizes morally 2-level
polytopes in terms of the slack variety.
Many open questions and further avenues of research remain and we hope that
the framework we introduce here will create an algebraic geometry based approach
to the important and difficult subject of realization spaces of polytopes, and related
objects. We describe some of these questions at the end of the paper.
1.2. Organization of the paper. In Section 2, we summarize the results on slack
matrices needed in this paper. We also define the slack ideal and affine slack ideal of
a polytope. In Section 3, we construct the slack and affine slack realization spaces of
a polytope. We show that the affine slack realization space is rationally equivalent
to the classical realization space of the polytope. In Section 4 we illustrate how the
slack ideal provides a computational framework for many classical questions about
polytopes such as convex realizability of combinatorial polytopes, rationality, and
prescribability of faces. In Section 5 we introduce the toric ideal of the non-incidence
graph of a polytope and show its relationship to pure difference binomial slack ideals
and morally 2-level polytopes. We prove in Section 6 that slack ideals are graphic
if and only if they are toric and the underlying polytope is projectively unique. We
also illustrate toric slack ideals that do not come from projectively unique polytopes
and projectively unique polytopes that do not have toric slack ideals. A range of
examples are used to show these results. We conclude the paper in Section 7 with
a list of open problems related to this work.
1.3. Acknowledgements. We thank Arnau Padrol and Günter Ziegler for helpful
pointers to the literature and valuable comments on the first draft of this paper. We
4
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
also thank Marco Macchia for providing us with a list of known 2-level polytopes
(available at http://homepages.ulb.ac.be/~mmacchia/data.html) that helped
us find interesting examples and counterexamples.
The SageMath and Macaulay2 software systems were invaluable in the development of the results below. All computations described in this paper were done with
one of these two systems [Dev17], [GS].
2. Background: Slack Matrices and Ideals of Polytopes
In this section we first present several known results about slack matrices of
polytopes needed in this paper. Many of these results come from [GGK+ 13]. We
then recall the slack ideal of a polytope from [GPRT17] which will be our main
computational engine. While much of this section is background, we also present
new objects and results that play an important role in later sections.
Suppose we are given a polytope P ⊂ Rd with v labelled vertices and f labelled
facet inequalities. Assume that P is a d-polytope, meaning that dim(P ) = d. Recall
that P has two usual representations: a V-representation P = conv{p1 , . . . , pv } as
the convex hull of vertices, and an H-representation P = {x ∈ Rd : W x ≤ w} as the
common intersection of the half spaces defined by the facet inequalities Wj x ≤ wj ,
j = 1, . . . , f , where Wj denotes the jth row of W ∈ Rf ×d . Let V ∈ Rv×d be the
matrix with rows p1 ⊤ , . . . , pv ⊤ , and let 1 denote a vector (of appropriate size) with
all entries equal to 1. Then the combined data of the two representations yields a
slack matrix of P , defined as
(1)
SP := 1 V
w⊤
−W ⊤
∈ Rv×f .
The name comes from the fact that the (i, j)-entry of SP is w j −Wj pi which is the
slack of the ith vertex pi of P with respect to the jth facet inequality Wj⊤ x ≤ wj of
P . Since P is a d-polytope, rank( 1 V ) = d + 1, and hence, rank (SP ) = d + 1.
Also, 1 is in the column span of SP . While the V-representation of P is unique,
the H-representation is not, as each facet inequality Wj x ≤ wj is equivalent to
the scaled inequality λWj x ≤ λw j for λ > 0, and hence P has infinitely many
slack matrices obtained by positive scalings of the columns of SP . Let Dt denote
a diagonal matrix of size t × t with all positive diagonal entries. Then all slack
matrices of P are of the form SP Df for some Df .
A polytope Q is affinely equivalent to P if there exists an invertible affine transformation ψ such that Q = ψ(P ). If Q is affinely equivalent to P , then SP is a slack
matrix of Q and thus P and Q have the same slack matrices (see Example 2.6 for
this calculation). In fact, a slack matrix of P offers a representation of the affine
equivalence class of P by the following result.
Lemma 2.1 ([GGK+ 13, Theorem 14]). If S is any slack matrix of P , then the
polytope Q = conv(rows(S)), is affinely equivalent to P .
By the above discussion, we may translate P so that 0 ∈ int(P ) without changing
its slack matrices. Subsequently, we may scale facet inequalities to set w = 1. Then
the affine equivalence class of P can be associated to the slack matrix
1
1
(2)
SP = [1 V ]
−W ⊤
THE SLACK REALIZATION SPACE OF A POLYTOPE
5
which has the special feature that the all-ones vector of the appropriate size is
present in both its row space and column space. Again this matrix is not unique
as it depends on the position of 0 ∈ int(P ).
Recall that the polar of P is P ◦ = {y ∈ (Rd )∗ : hx, yi ≤ 1 ∀ x ∈ P }. Under
the assumption that 0 ∈ int(P ) and that w = 1, P ◦ is again a polytope with 0 in
its interior and representations [Zie95, Theorem 2.11]:
P ◦ = conv{W1⊤ , . . . , Wf⊤ } = {y ∈ (Rd )∗ : V y ≤ 1}.
This implies that (SP1 )⊤ is a slack matrix of P ◦ and all slack matrices of P ◦ are of
the form (Dv SP1 )⊤ .
We now pass from the fixed polytope P to its combinatorial class. Note that
the zero-pattern in a slack matrix of P , or equivalently, the support of SP , encodes
the vertex-facet incidence structure of P , and hence the entire combinatorics (face
lattice) of P [JKPZ01]. A labelled polytope Q is combinatorially equivalent to P if
P and Q have the same face lattice under the identification of vertex pi in P with
vertex q i in Q and the identification of facet inequality fj in P with facet inequality
gj in Q. The combinatorial class of P is the set of all labelled polytopes that are
combinatorially equivalent to P . A realization of P is a polytope Q, embedded in
some Rk , that is combinatorially equivalent to P . By our labelling assumptions, all
realizations of P have slack matrices with the same support as SP . Further, since
each realization Q of P is again a d-polytope, all its slack matrices have rank d + 1
and contain 1 in their column span. Interestingly, the converse is also true and is
a consequence of [GGK+ 13, Theorem 22].
Theorem 2.2. A nonnegative matrix S is a slack matrix of some realization of the
labelled d-polytope P if and only if all of the following hold:
(1) supp(S) = supp(SP )
(2) rank (S) = rank (SP ) = d + 1
(3) 1 lies in the column span of S.
This theorem will play a central role in this paper. It allows us to identify the
combinatorial class of P with the set of nonnegative matrices having the three listed
properties.
A polytope Q is projectively equivalent to P if there exists a projective transformation φ such that Q = φ(P ). Recall that a projective transformation is a
map
Bx + b
φ : Rd → Rd , x 7→ ⊤
c x+γ
for some B ∈ Rd×d , b, c ∈ Rd , γ ∈ R such that
B b
(3)
det
6= 0.
c⊤ γ
The polytopes P and Q = φ(P ) are combinatorially equivalent. Projective equivalence within a combinatorial class can be characterized in terms of slack matrices.
Lemma 2.3 ([GPRT17, Corollary 1.5]). Two polytopes P and Q are projectively
equivalent if and only if Dv SP Df is a slack matrix of Q for some positive diagonal
matrices Dv , Df .
6
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Notice that Lemma 2.3 does not say that every positive scaling of rows and
columns of SP is a slack matrix of a polytope projectively equivalent to P , but
rather that there is some scaling of rows and columns of SP that produces a slack
matrix of Q. In particular, condition (3) of Theorem 2.2 requires 1 to be in the
column span of the scaled matrix. Not all row scalings will preserve 1 in the
column span. Regardless, we will be interested in all row and column scalings of
slack matrices.
Definition 2.4. A generalized slack matrix of P is any matrix of the form Dv SQ Df ,
where Q is a polytope that is combinatorially equivalent to P and Dv , Df are
diagonal matrices with positive entries on the diagonal. Let SP denote the set of
all generalized slack matrices of P .
Theorem 2.5. The set SP of generalized slack matrices of P consists precisely of
the nonnegative matrices that satisfy conditions (1) and (2) of Theorem 2.2.
Proof. By construction, every matrix in SP satisfies conditions (1) and (2) of Theorem 2.2. To see the converse, we need to argue that if S is a nonnegative matrix
that satisfies conditions (1) and (2) of Theorem 2.2, then there exists some Dv , Df
such that S = Dv SQ Df for some polytope Q that is combinatorially equivalent to
P , or equivalently, that there is some row scaling of S that turns it into a slack
matrix of a polytope combinatorially equivalent to P . By Theorem 2.2, this is
equivalent to showing that 1 lies in the column span of Dv−1 S. Choose the diagonal
matrix Dv−1 so that Dv−1 S divides each row of S by the sum of the entries in that
row. Note that this operation is well-defined, as a row of all zeros would correspond
to a vertex which is part of every facet. Then the sum of the columns of Dv−1 S
is 1 making Dv−1 S satisfy all three conditions of Theorem 2.2. Therefore, by the
theorem, Dv−1 S = SQ for some polytope Q in the combinatorial class of P .
We illustrate the above results on a simple example.
Example 2.6. Consider two realizations of a quadrilateral in R2 ,
P1 = conv{(0, 0), (1, 0), (1, 1), (0, 1)}, and
P2 = conv{(1, −2), (1, 2), (−1, 2), (−1, −2)},
h
i
h i
0 −2
1
where P2 = ψ(P1 ) for the affine transformation ψ(x) = 4 0 x + −2 . The
most obvious choice of facet representation for P1 yields the slack matrix
SP1
1
1
=
1
1
0
1
1
0
0
0 1
0
0 −1
1
1 0
1
0
1 0
0
0 1 =
1
−1 0
1
1
0
0
1
1
1
0
0
0
1
.
1
0
THE SLACK REALIZATION SPACE OF A POLYTOPE
Since P2 = ψ(P1 ) = {y ∈ Rd : W ψ −1 (y) ≤ w},
1 0 0
1
1 1 0 1 1 −2 1 2
0 0
4
SP1 =
0 0
1 1 1
0 −2 0
0 41
1 0 1
|
{z
}
|
1
1
=
1
1
1
2
1
2
1
2
0
− 41
0
1
2
1
1 0
−1 0 1
1 0 −1 0
{z
}
1
0
2
− 12 0
0
[w −W ψ −1 ]⊤
[1 ψ(V )]
1 −2 1
2
1
2
− 1
2
−1 2
0
−1 −2
7
0 = SP2
1
4
Notice that the first equality follows from the fact that the second and third matrices
in the first product of four matrices are inverses of each other.
Since P2 also contains the origin in its interior, we can scale its H-representation
from above to obtain
1 1 −2
0 2 2 0
1
1
1
1
1 1
2
0 0 2 2
SP1 2 =
1 −1 2 −1 01 1 01 = 2 0 0 2 .
0 −2 0 2
1 −1 −2
2 2 0 0
Finally, consider the following nonnegative
0 1
0 0
S=
1 0
1 2
matrix
1 0
1 1
.
0 2
0 0
Since S satisfies all three conditions of Theorem 2.2, it must be the slack matrix
of some realization of a quadrilateral. Notice that S can be obtained from SP1 by
scaling its columns and rows,
1
0 0 0
1 0 0 0
2
0 1 0 0
2
SP1 0 2 0 0
S=
0 0 2 0
0 0 1 0
0 0 0 2
0 0 0 1
so that by Lemma 2.3 it is SP3 , where P3 = φ(P1 ) for some projective transformation
φ. One can check this holds for
2 0
x
0 1
.
φ(x) =
(0, −1)x + 2
We now recall the symbolic slack matrix and slack ideal of P which were defined
in [GPRT17]. Given a d-polytope P , its symbolic slack matrix SP (x) is the sparse
generic matrix obtained by replacing each nonzero entry of SP by a distinct variable.
Suppose there are t variables in SP (x). The slack ideal of P is the saturation of
the ideal generated by the (d + 2)-minors of SP (x), namely
!∞
t
Y
(4)
IP := h(d + 2)-minors of SP (x)i :
xi
⊂ C[x] := C[x1 , . . . , xt ].
i=1
8
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
The slack variety of P is the complex variety V(IP ) ⊂ Ct . The saturation of IP by
the product of all variables guarantees that there are no components in V(IP ) that
live entirely in coordinate hyperplanes. If s ∈ Ct is a zero of IP , then we identify
it with the matrix SP (s).
Lemma 2.7. The set SP of generalized slack matrices is contained in the real part
of the slack variety V(IP ).
Proof. By Theorem 2.5, all matrices in SP have real entries, support equal to
supp(SP ), and rank d + 1. Therefore, SP is contained in the real part of V(IP ).
To focus on “true slack matrices” of polytopes in the combinatorial class of P ,
meaning matrices that satisfy all conditions of Theorem 2.2, we define the affine
slack ideal
!∞
t
Y
e
⊂ C[x],
xi
(5)
IP = h(d + 2)-minors of [SP (x) 1]i :
i=1
where [SP (x) 1] is the symbolic slack matrix with a column of ones appended. By
construction, V(IeP ) is a subvariety of V(IP ).
e P denote the set of true slack matrices of polytopes in the
Definition 2.8. Let S
combinatorial class of P , or equivalently, the set of all nonnegative matrices that
satisfy the three conditions of Theorem 2.2.
e P of true slack matrices is contained in the real part of
Lemma 2.9. The set S
e
V(IP ).
e P have real entries and supp(S) = supp(SP ).
Proof. By definition, all elements S ∈ S
It remains to show that rank ([S 1]) ≤ d + 1. This follows immediately from the
fact that S satisfies properties (2) and (3) of Theorem 2.2.
Example 2.10. For our quadrilateral P1 from Example 2.6 and in fact any quadrilateral P labeled in the same way as P1 , we have
0 x1 x2 0
0
0 x3 x4
.
SP (x) =
x5 0
0 x6
x7 x8 0
0
Its slack ideal is
IP = h4-minors of SP (x)i :
The affine slack ideal of P is
8
Y
i=1
IeP = h4-minors of [SP (x) 1]i :
xi
!∞
8
Y
i=1
xi
= hx2 x4 x5 x8 − x1 x3 x6 x7 i ⊂ C[x1 , . . . , x8 ].
!∞
= hx1 x3 x6 − x2 x4 x8 + x2 x6 x8 − x3 x6 x8 ,
x2 x4 x5 − x2 x4 x7 + x2 x6 x7 − x3 x6 x7 ,
x1 x4 x5 − x1 x4 x7 + x1 x6 x7 − x4 x5 x8 ,
x1 x3 x5 − x1 x3 x7 + x2 x5 x8 − x3 x5 x8 i.
Notice, for example, that the generalized slack matrix which corresponds to
s = (2, 2, 2, 1, 8, 2, 2, 1) is a zero of IP but not of IeP and indeed 1 is not in the
column span of SP (s).
THE SLACK REALIZATION SPACE OF A POLYTOPE
9
3. Realization spaces from Slack Varieties
Recall that a realization of a d-polytope P ⊂ Rd is a polytope Q that is combinatorially equivalent to P . A realization space of P is, essentially, the set of all
polytopes Q which are realizations of P , or equivalently, the set of all “geometrically
distinct” polytopes which are combinatorially equivalent to P . We say “essentially”
since it is typical to mod out by affine equivalence within the combinatorial class.
The standard construction of a realization space of P = conv{p1 , . . . , pv } is
as follows (see [RG96a]). Fix an affine basis of P , that is, d + 1 vertex labels
B = {b0 , . . . , bd } such that the vertices {pb }b∈B are necessarily affinely independent
in every realization of P . Then the realization space of P with respect to B is
R(P, B) = {realizations Q = conv{q 1 , . . . , q v } of P with q i = pi for all i ∈ B}.
Fixing an affine basis ensures that just one Q from each affine equivalence class in
the combinatorial class of P occurs in R(P, B).
Realization spaces of polytopes are primary basic semialgebraic sets, that is, they
are defined by finitely many polynomial equations and strict inequalities. Recording
each realization Q by its vertices, we can think of R(P, B) as lying in Rd·v .
Let X ⊆ Rm and Y ⊆ Rm+n be primary basic semialgebraic sets with π(Y ) = X
for the canonical coordinate projection π : Rm+n → Rm . Call X a stable projection
of Y if Y has the form
Y = (v, v ′ ) ∈ Rm+n : v ∈ X and φvi (v ′ ) > 0, ψjv (v ′ ) = 0 for all i ∈ I, j ∈ J ,
where φvi , ψjv are affine functionals whose parameters depend polynomially on X
and that are indexed by finite sets I, J, respectively. In other words, the fibers
π −1 (v) ⊂ Y are relative interiors of polyhedra. Call two primary basic semialgebraic sets X, Y rationally equivalent if there exists a homeomorphism f : X → Y
such that both f and f −1 are rational functions. Finally, two primary basic semialgebraic sets X, Y are stably equivalent, denoted X ≈ Y , if they lie in the same
equivalence class under stable projections and rational equivalence. The important
result for us is that if B1 , B2 are two affine bases of a polytope P , then R(P, B1 )
and R(P, B2 ) are rationally equivalent [RG96a, Lemma 2.5.4]. Thus one can call
R(P, B) ⊂ Rd·v , the realization space of P .
The main goal of this section is to construct models of realization spaces for P
from the slack variety V(IP ) ⊂ Ct and affine slack variety V(IeP ) defined in Section 2.
Recall that we identify an element s in either variety with the matrix SP (s). Then
by Lemma 2.7, SP , the set of all generalized slack matrices of all polytopes in
eP,
the combinatorial class of P , is contained in V(IP ). Similarly, by Lemma 2.9, S
the set of all true slack matrices of polytopes in the combinatorial class of P , is
contained in V(IeP ). In fact, SP is contained in the positive part of V(IP ), defined
as
(6)
V+ (IP ) := V(IP ) ∩ Rt>0
e P is contained in the positive part of V(IeP ) defined as
and S
(7)
V+ (IeP ) := V(IeP ) ∩ Rt>0 .
These positive spaces are going to lead to realization spaces of P . In order to
get there, we first describe these sets more explicitly. We start with a well-known
lemma, whose proof we include for later reference.
10
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Lemma 3.1. Let S be a matrix with the same support as SP . Then rank (S) ≥ d+1.
Proof. Consider a flag of P , i.e., a maximal chain of faces in the face lattice of P .
Choose a sequence of facets F0 , F1 , . . . , Fd so that the flag is
∅ = F0 ∩ · · · ∩ Fd ⊂ F1 ∩ · · · ∩ Fd ⊂ · · · ⊂ Fd−1 ∩ Fd ⊂ Fd ⊂ P.
Next choose a sequence of vertices so that v0 = F1 ∩ · · · ∩ Fd is the 0-face in the
flag, making v0 6∈ F0 . Then choose v1 ∈ F2 ∩ · · · ∩ Fd but v1 6∈ F1 , v2 ∈ F3 ∩ · · · ∩ Fd
but v2 6∈ F2 and so on, until vd−1 ∈ Fd but not in Fd−1 . Finally, choose vd so that
vd 6∈ Fd . Then the (d + 1) × (d + 1) submatrix of SP indexed by the chosen vertices
and facets is lower triangular with a nonzero diagonal, hence has rank d + 1.
Now if S is a matrix with supp(S) = supp(SP ), S will also have this lower
triangular submatrix in it, thus rank (S) ≥ d + 1.
We remark that the vertices chosen from the flag in the above proof form a
suitable affine basis to fix in the construction of R(P, B).
Theorem 3.2. The positive part of the slack variety, V+ (IP ), coincides with SP ,
eP,
the set of generalized slack matrices of P . Similarly, V+ (IeP ) coincides with S
the set of true slack matrices of P .
Proof. We saw that SP ⊆ V+ (IP ) and by Theorem 2.5, SP is precisely the set
of nonnegative matrices with the same support as SP and rank d + 1. On the
other hand, if s ∈ V+ (IP ), then SP (s) is nonnegative and supp(SP (s)) = supp(SP ).
Therefore, by Lemma 3.1, rank (SP (s)) = d + 1. Thus, V+ (IP ) = SP .
e P ⊆ V+ (IeP ). Also recall that V+ (IeP ) is contained in V+ (IP ).
We saw that S
Therefore, by the first statement of the theorem, if s ∈ V+ (IeP ), then SP (s) is nonnegative, supp(SP (s)) = supp(SP ) and rank (SP (s)) = d+ 1. From the definition of
IeP , we have rank ([SP (s) 1]) ≤ d + 1, so it follows that rank ([SP (s) 1]) = d + 1, or
equivalently, 1 lies in the column span of SP (s). Therefore, the matrices in V+ (IeP )
eP.
satisfy all three conditions of Theorem 2.2, hence V+ (IeP ) = S
Since positive row and column scalings of a generalized slack matrix of P give
another generalized slack matrix of P , we immediately get that V+ (IP ) is closed
under row and column scalings. Similarly, V+ (IeP ) is closed under column scalings.
Corollary 3.3.
(1) If s ∈ V+ (IP ), then Dv sDf ∈ V+ (IP ), for all positive diagonal matrices
Dv , Df .
(2) Similarly, if s ∈ V+ (IeP ), then sDf ∈ V+ (IeP ), for all positive diagonal
matrices Df .
Corollary 3.3 tells us that the groups Rv>0 × Rf>0 and Rf>0 act on V+ (IP ) and
V+ (IeP ), respectively, via multiplication by positive diagonal matrices. Modding out
these actions is the same as setting some choice of variables in the symbolic slack
matrix to 1, which means that we may choose a representative of each equivalence
class (affine or projective) with ones in some prescribed positions.
Corollary 3.4.
(1) Given a polytope P , there is a bijection between the elements of
V+ (IP )/(Rv>0 × Rf>0 ) and the classes of projectively equivalent polytopes
THE SLACK REALIZATION SPACE OF A POLYTOPE
11
of the same combinatorial type as P . In particular, each class contains a
true slack matrix.
(2) Given a polytope P , there is a bijection between the elements of V+ (IeP )/Rf>0
and the classes of affinely equivalent polytopes of the same combinatorial
type as P .
The last statement in Corollary 3.4 (1) follows from the fact that every generalized slack matrix admits a row scaling that makes it satisfy all three conditions of
Theorem 2.2, thereby making it a true slack matrix. An explicit example of such a
scaling can be seen in the proof of Theorem 2.5.
By the above results we have that V+ (IP )/(Rv>0 × Rf>0 ) and V+ (IeP )/Rf>0 are
parameter spaces for the projective (respectively, affine) equivalence classes of polytopes in the combinatorial class of P . Thus they can be thought of as realization
spaces of P .
Definition 3.5. Call V+ (IP )/(Rv>0 × Rf>0 ) the slack realization space of the polytope P , and V+ (IeP )/Rf>0 the affine slack realization space of the polytope P .
We will see below that the affine slack realization space V+ (IeP )/Rf>0 is rationally
equivalent to the classical model of realization space R(P, B) of the polytope P . On
the other hand, our main object, the slack realization space V+ (IP )/(Rv>0 × Rf>0 ),
does not have an analog in the polytope literature. This is partly because in every
realization of P , fixing a projective basis does not guarantee that the remaining
vertices in the realization are not at infinity. The slack realization space is a natural
model for the realization space of projective equivalence classes of polytopes. We
note that in [GPS17] the authors investigate the projective realization space of
combinatorial hypersimplices and find an upper bound for its dimension. However
they do not present an explicit model for it.
Theorem 3.6. The affine slack realization space V+ (IeP )/Rf>0 is rationally equivalent to the classical realization space R(P, B) of the polytope P .
Proof. We will show that V+ (IeP )/Rf>0 is rationally equivalent to R(P, B) for a
particular choice of B. By [RG96a, Lemma 2.5.4], this is sufficient to show rational
equivalence for any choice of basis.
We have already shown that realizations of P modulo affine transformations are
in bijective correspondence with the elements of both V+ (IeP )/Rf>0 and R(P, B).
So we just have to prove that this bijection induces a rational equivalence between
these spaces, i.e., both the map and its inverse are rational.
We will start by showing the map sending a polytope in R(P, B) to its slack
matrix is rational. Fix a flag in P , as in the proof of Lemma 3.1. Suppose the
sequence of vertices and facets chosen from the flag in the proof are indexed by the
sets I and J respectively. The vertices {pi }i∈I are affinely independent, so that
B = I is an affine basis of P . Moreover, by applying an affine transformation to
P , we may assume that 0 is in the convex hull of {pi }i∈I , hence is in the interior
of every element of R(P, B). Consider the map
g : R(P, B) → V+ (IeP ), Q 7→ S 1 .
Q
The polytope Q is recorded in R(P, B) by its list of vertices,
which in turn are the
1
1
. To prove that g is
rows of the matrix V . Also, recall that SQ
= [1 V ]
−W ⊤
12
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
a rational map, we need to show that the matrix of facet normals W is a rational
function of V . Since we know the combinatorial type of P , we know the set of
vertices that lie on each facet. For facet j, let V (j) be the submatrix of V whose
rows are the vertices on this facet. Then the normal of facet j, or equivalently Wj ,
is obtained by solving the linear system V (j) · x = 1 which proves that Wj is a
rational function of V . Then e
g = π ◦ g is the desired rational map from R(P, B)
to V+ (IeP )/Rf>0 , where π is the standard quotient map π : V+ (IeP ) → V+ (IeP )/Rf>0 .
It sends the representative in R(P, B) of an affine equivalence class of polytopes in
the combinatorial class of P to the representative of that class in V+ (IeP )/Rf>0 .
For the reverse map, we have to send a slack matrix SQ of a realization Q
of P to the representative of its affine equivalence class in R(P, B). We saw in
Lemma 2.1 that the rows of SQ are the vertices of a realization Q′ of P that is
affinely equivalent to Q. So we just have to show that Q′ can be rationally mapped
to the representative of Q in R(P, B). To do that, denote by SbQ the (d+ 1)× (d+ 1)
lower triangular submatrix of SQ from our flag, with rows indexed by I and columns
indexed by J. Then (SbQ )−1 consists of rational functions in the entries of SbQ . Let
B be the (d + 1) × d matrix whose rows are the vertices of P indexed by B. Recall
that these vertices are common to all elements of R(P, B), and in particular, they
form an affine basis for the representative of Q in R(P, B). Then the linear map
b−1
ψSQ : Rf → Rd , x 7→ x⊤
J SQ B,
where xJ is the restriction of x ∈ Rf to the coordinates indexed by J, is defined
rationally in terms of the entries of SQ , and maps row i of SQ to the affine basis
vertex pi , for all i ∈ I. Now since ψSQ is a linear map, ψSQ (Q′ ) is affinely equivalent
to Q′ which is itself affinely equivalent to Q. Furthermore, ψSQ sends an affine
basis of Q′ to the corresponding affine basis in Q, so in fact it must be a bijection
between the two polytopes. Hence, ψSQ (rows of SQ ) equals the representative of Q
in R(P, B), completing our proof.
Example 3.7. Let us return to the realization space of the unit square P1 from
Example 2.6. Suppose we fix the affine basis B = {1, 2, 4}, where we had p1 =
(0, 0), p2 = (1, 0) and p4 = (0, 1). Then the classical realization space R(P1 , B)
consists of all quadrilaterals Q = conv{p1 , p2 , (a, b), p4 }, where a, b ∈ R must satisfy
a, b > 0 and a + b > 1 in order for Q to be convex.
In the slack realization spaces, modding out by row and column scalings is equivalent to fixing some variables in SP (x) to 1. So for example, we could start with
the following scaled symbolic slack and affine slack matrices
0 1 1 0
0
1
1
0 1
0 0 1 1
0
0 x3 x4 1
.
SP (x) =
1 0 0 1 , [SP (x) 1] = 1
0
0
1 1
1 x8 0 0
x7 x8 0
0 1
Computing the 4-minors of these scaled symbolic slack matrices and saturating
with all variables produces the scaled slack ideals
IPscaled = hx8 − 1i, and
IePscaled = hx3 x8 + x4 x8 − x3 − x8 , x4 x7 + x4 x8 − x4 − x7 , x3 x7 − x4 x8 i.
THE SLACK REALIZATION SPACE OF A POLYTOPE
13
Therefore the slack realization space, V+ (IP )/(R4>0 × R4>0 ), has the unique element (1, 1, 1, 1, 1, 1, 1, 1), and indeed, all convex quadrilaterals are projectively
equivalent to P1 .
From the generators of IePscaled one sees that the affine slack realization space,
V(IeP )/R4>0 , is two-dimensional and parametrized by x3 , x4 with
x7 =
x4
x3
and x8 =
.
x3 + x4 − 1
x3 + x4 − 1
Since all the four variables have to take on positive values in a (scaled) slack matrix
of a quadrilateral, we get that the realization space V+ (IeP )/R4>0 is cut out by the
inequalities x3 > 0, x4 > 0, x3 + x4 > 1. This description coincides exactly with
that of R(P1 , B) that we saw earlier.
Example 3.8. Consider the 5-polytope P with vertices p1 , . . . , p8 given by
e1 , e2 , e3 , e4 , −e1 − 2e2 − e3 , −2e1 − e2 − e4 , −2e1 − 2e2 + e5 , −2e1 − 2e2 − e5
where e1 , . . . , e5 are the standard basis vectors in R5 . It can be obtained by splitting
the distinguished vertex v of the vertex sum of two squares, (, v) ⊕ (, v) in the
notation of [McM76]. This polytope has 8 vertices and 12 facets and its symbolic
slack matrix has the zero-pattern below
0
0
0
∗
∗
0
∗
0
∗
0
0
0
∗
0
0
0
0
0
0
0
∗
∗
∗
0
0
∗
0
∗
0
0
0
0
0
∗
0
0
0
∗
0
0
0
0
∗
∗
0
0
∗
0
∗
0
∗
0
0
0
0
0
0
0
∗
0
0
∗
∗
0
0
0
0
∗
∗
0
0
∗
0
0
0
0
∗
∗
0
∗
0
0
∗
0
.
0
∗
0
∗
0
0
∗
∗
0
0
0
∗
By [McM76, Theorem 5.3] P is not projectively unique, meaning its slack realization
space V+ (IP )/(Rv>0 ×Rf>0 ) will consist of more than a single point. Indeed, by fixing
ones in the maximum number of positions, marked in bold face below, we find
that V+ (IP )/(Rv>0 × Rf>0 ) is a one-dimensional space of projectively inequivalent
realizations parametrized by slack matrices of the following form
0
0
0
1
SP (a) =
1
0
1
0
1
0
0
0
1
0
0
0
0 0
0 1
0 0
0 1
1 0
1 0
1 0
0 0
0
1
0
0
0
1
0
0
0
0
1
1
0
0
1
0
1
0
1
0
0
0
0
0
0 0
0 0
1 0
0 a
0 1
1 0
1 0
0 1
0
0
0
0
1
a
0
1
0
0
1
a
0
0
0
1
0
0
1
0
.
0
a
0
1
If we wish to look at a representative of each equivalence class which is a true slack
matrix, then we can scale the above to guarantee that 1 is in the column space to
14
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
get the following
0
0
0
1
SP (a) =
1
0
2
0
where SP (a) · 0 0
1
0
0
0
1
0
0
0
1
2
0
0
0
0
1
2
a+1
2
0
1
0
1
0
1
0
0
0
0
0 0
0
1
0
0
0
2
a+1
0
0
1 0
0
0
1
1
0
0
2
0
1
0
1
0
0
0
0
0
0
0
0
1
0
0
2
a+1
2
0
1
2
0
0
0
0
a
1
0
0
2
⊤
0
0
0
0
0
1
2a
a+1
0
2
= 1.
0
0
1
a
0
0
0
2
0
0
1
0
0
,
2a
a+1
0
2
We end this section with an illustration of the simplicity of the slack realization
space by considering P ◦ , the polar of P . It is not immediately obvious from the
standard model of a realization space, how R(P, B1 ) and R(P ◦ , B2 ) are related.
Indeed the proof of Theorem 2.6.3 in [RG96a], that the realization spaces of P and
P ◦ are stably equivalent, is non-trivial. Now consider the slack model. Recall we
know that one slack matrix of P ◦ is (SP1 )⊤ , so that SP ◦ (x) = SP (x)⊤ . In particular,
this means that IP ◦ = IP , so that the slack varieties and realization spaces of P
and P ◦ are actually the same when considered as subsets of Rt . We simply need
to interpret s ∈ V+ (IP ) = V+ (IP ◦ ) as a realization of P or P ◦ by assigning its
coordinates to SP (x) along rows or columns.
4. Three applications
In this section we illustrate the computational power of the slack ideal in answering three types of questions that one can ask about realizations of polytopes.
We anticipate further applications.
4.1. Abstract polytope with no realizations. Checking if an abstract polytopal complex is the boundary of an actual polytope is the classical Steinitz problem, and an important ingredient in cataloging polytopes with few vertices. In
[AS85], Altshuler and Steinberg enumerated all 4-polytopes and 3-spheres with
8 vertices. The first non-polytopal 3-sphere in [AS85, Table 2] has simplices and
square pyramids as facets, and these facets have the following vertex sets
12345, 12346, 12578, 12678, 14568, 34578, 2357, 2367, 3467, 4678.
If there was a polytope P
0
0
0
0
0
0
0
0
SP (x) =
0 x18
x23
0
x27 x28
x30 x31
with these facets, its symbolic slack matrix would be
0
0
0
x1 x2 x3 x4 x5
0
0
x6 x7
0
0
x8 x9
x10 x11 x12
0
0
0
0 x13
x14 x15
0
0 x16 x17
0
0
.
0 x19
0
0
0 x20 x21 x22
x24
0
0 x25 x26
0
0
0
0
0 x29
0
0
0
0
0
0
0
0
0 x32 x33 x34
0
One can compute that the would-be slack ideal IP in this case is trivial, meaning
that there is no rank five matrix with the support of SP (x). In particular, there is
no polytope with the given facial structure. In fact, there is not even a hyperplanepoint arrangement in R4 or C4 with the given incidence structure.
THE SLACK REALIZATION SPACE OF A POLYTOPE
15
In some other cases, one can obtain non-empty slack varieties that have no
positive part. A simple example of that behaviour can be seen in the tetrahemihexahedron, a polyhedralization of the real projective plane with 6 vertices, and facets
with vertex sets 235, 346, 145, 126, 2456, 1356, 1234. Its slack matrix is therefore
x1 x2
0
0 x3
0
0
0
x4 x5
0
0 x6
0
0
0
x
x
x
0
0
7
8
9
.
SP (x) =
x
0
0
x
0
x
0
11
12
10
0 x13
0 x14 0
0 x15
x16
0 x17
0
0
0 x18
Computing the slack ideal from the 5-minors of SP (x) we find that IP is generated
by the binomials
x8 x15 x17 + x7 x14 x18 x4 x15 x17 + x5 x13 x18 x11 x15 x16 + x10 x14 x18
x2 x15 x16 + x1 x13 x18 x5 x12 x16 + x6 x10 x17 x7 x11 x16 − x8 x10 x17
x3 x7 x16 + x1 x9 x17
x2 x5 x16 − x1 x4 x17
x6 x11 x13 + x4 x12 x14
x1 x11 x13 − x2 x10 x14 x5 x8 x13 − x4 x7 x14
x3 x8 x13 + x2 x9 x14
x6 x7 x11 + x5 x8 x12
x3 x8 x10 + x1 x9 x11
x2 x6 x10 + x1 x4 x12
x6 x11 x15 x17 − x5 x12 x14 x18 x2 x9 x15 x17 − x3 x7 x13 x18
x4 x12 x15 x16 − x6 x10 x13 x18 x3 x8 x15 x16 − x1 x9 x14 x18
x2 x7 x14 x16 − x1 x8 x13 x17 x5 x11 x13 x16 − x4 x10 x14 x17
x2 x6 x9 x11 − x3 x4 x8 x12
x2 x5 x8 x10 − x1 x4 x7 x11
x3 x6 x7 x10 − x1 x5 x9 x12
x3 x4 x7 + x2 x5 x9
Since the slack ideal contains binomials whose coefficients are both positive, it
has no positive zeros. In fact, by fixing some coordinates to one, it has a unique
zero up to row and column scalings, where all entries are either 1 or −1.
4.2. Non-prescribable faces of polytopes. Another classical question about
polytopes is whether a face can be freely prescribed in a realization of a polytope
with given combinatorics.
We begin by observing that there is a natural relationship between the slack
matrix/ideal of a polytope and those of each of its faces. For instance, if F is
a facet of a d-polytope P , a symbolic slack matrix SF (x) of F is the submatrix
of SP (x) indexed by the vertices of F and the facets of P that intersect F in its
(d − 2)-dimensional faces. Let xF denote the vector of variables in that submatrix.
All (d + 1)-minors of SF (x) belong to the slack ideal IP . To see this, consider a
(d + 2)-submatrix of SP (x) obtained by enlarging the given (d + 1)-submatrix of
SF (x) by a row indexed by a vertex p 6∈ F and the column indexed by F . The
column of F in this bigger submatrix has all zero entries except in position (p, F ).
The minor of this (d + 2)-submatrix in SP (x) after saturating out the variable in
position (p, F ), is the (d + 1)-minor of SF (x) that we started with. Therefore,
IF ⊆ IP ∩ C[xF ].
By induction on the dimension, this containment is true for all faces F of P .
A face F of a polytope P is prescribable if, given any realization of F , we can
complete it to a realization of P . In our language, a face F is prescribable in P if
and only if
V+ (IF ) = V+ (IP ∩ C[xF ]).
16
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Consider the four-dimensional prism over a square pyramid, for which it was
shown in [Bar87] that its only cube facet F is non-prescribable. This polytope P
has 10 vertices and 7 facets and its symbolic slack matrix is
x1
0
0
0
x2
x3
0
x4
0
0
0
0
x5
x6
x7
0
x8
0
0
x9
0
x10
0 x11 x12
0
0
0
x13
0
x14
0
0
0
0
.
SP (x) =
0 x15
0
0 x16 x17
0
0 x18
0
0
0 x19 x20
0 x21
0 x22
0
0 x23
0 x24
0 x25 x26
0
0
0
x27 x28
0
0
0
0
In bold we mark SF (x) sitting inside SP (x). Computing IP and intersecting with
C[xF ], we obtain an ideal of dimension 15. On the other hand, the slack ideal of a
cube has dimension 16, suggesting an extra degree of freedom for the realizations of
a cube, and the possibility that the cubical facet F cannot be arbitrarily prescribed
in a realization of P . However, we need more of an argument to conclude this, since
IF 6= IP ∩ C[xF ] does not immediately mean that V+ (IF ) 6= V+ (IP ∩ C[xF ]). We
need to compute further to get Barnette’s result.
We first note that one can scale the rows and columns of SF (x) to set 13 of its
24 variables to one, say x1 , x2 , x3 , x4 , x6 , x7 , x8 , x10 , x15 , x16 , x18 , x21 , x24 . Guided
by the resulting slack ideal we further set x20 = 1, x11 = 21 , x17 = 2 and x25 = 1.
Now solving for the remaining variables from the equations of the slack ideal, we
get the following true slack matrix of a cube:
1 0
0 1 1 0
1 0
0 0 2 1
1 0
1 0 0 1
1 0 1/2 1 0 0
.
0 1
0 1 2 0
0 1
0 0 3 1
0 1 3/2 0 0 1
0 1
1 1 0 0
However, making the above-mentioned substitutions for
x1 , x2 , x3 , x4 , x6 , x7 , x8 , x10 , x11 , x15 , x16 , x17 , x18 , x20 , x21 , x24 , x25
in SP (x) and eliminating x13 , x14 , x27 and x28 from the slack ideal results in the
trivial ideal showing that the cube on its own admits further realizations than are
possible as a face of P .
4.3. Non-rational polytopes. A combinatorial polytope is said to be rational if
it has a realization in which all vertices have rational entries. This has a very simple
interpretation in terms of slack varieties.
Lemma 4.1. A polytope P is rational if and only if V+ (IP ) has a rational point.
The proof is trivial since any rational realization gives rise to a rational slack
matrix and any rational slack matrix is itself a rational realization of the polytope
P . Recall that any point in V+ (IP ) can be row scaled to be a true slack matrix
THE SLACK REALIZATION SPACE OF A POLYTOPE
17
Figure 1. Non-rational line-point configuration
by dividing each row by the sum of its entries, so a rational point in V+ (IP ) will
provide a true rational slack matrix of P .
Unfortunately, the usual examples of non-rational polytopes tend to be too large
for direct computations, so we illustrate our point on the non-rational point-line
arrangement in the plane shown in [Grü03, Figure 5.5.1]. A true non-rational
polytope can be obtained from this point-line arrangement by Lawrence lifting.
We will show the non-rationality of this configuration by computing its slack ideal
as if it were a 2-polytope. Its symbolic slack matrix is the 9 × 9 matrix
x1
0
x2
0
x3 x4 x5 x6
0
x7 x8 x9
0 x10
0
0 x11 x12
x13 x14
0 x15 x16 x17 x18
0
0
x19 x20
0 x21
0
0 x22 x23 x24
0 x26 x27
0 x28
0
0 x29
S(x) =
x25
.
0
0
x
x
x
0
x
x
x
30
31
32
33
34
35
0 x36
0 x37 x38 x39
0 x40 x41
0 x42 x43
0 x44 x45 x46
0 x47
0 x48 x49 x50
0 x51 x52 x53
0
One can simplify the computations by scaling rows and columns to fix xi = 1
for i = 1, 2, 8, 14, 20, 26, 30, 36, 42, 44, 47, 48, 50, 51, 52, 53, as this does not affect
rationality. Then
one sees that the polynomial x246 + x46 − 1 is in the slack ideal,
√
−1± 5
, and there are no rational realizations of this configuration.
so x46 =
2
Remark 4.2. We note that as illustrated by the above example, the slack matrix
and slack ideal constructions are not limited to the setting of polytopes, but in fact,
are applicable to the more general setting of any point/hyperplane configuration.
5. The toric ideal of the non-incidence graph of a polytope
We constructed the slack realization space of a polytope P through its slack
ideal IP . Looking back, a key algebraic question about this realization space is
whether IP is indeed the vanishing ideal of V+ (IP ). In general, we do not know
the answer to this question. If V+ (IP ) is Zariski dense in the complex variety
of IP then indeed, IP would be the vanishing ideal of V+ (IP ) up to radicality.
Work arounds to make the algebra correct would be to replace IP by its radical
or real radical, or even better, to take as our main algebraic object the vanishing
ideal of the Zariski closure of V+ (IP ). However, all of these ideals would grind all
computations to a halt, and computational ability with the ideal is an important
aspect of our realization space model, as we saw in the previous section.
18
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
An important situation in which everything works well is when the slack ideal is
toric since the positive part of a toric variety is Zariski dense in its complex variety.
Since slack ideals do not contain monomials, the simplest non-trivial slack ideals we
can hope for would be generated by binomials. Prime binomial ideals are precisely
the toric ideals. In this section we define the toric ideal TP of the non-incidence
graph of a polytope P and relate it to the slack ideal IP . In the next section we
examine when the slack ideal equals the toric ideal TP .
First we recall the definition of a toric ideal. Let A = {a1 , . . . , an } be a point
configuration in Zd . Sometimes we will identify A with the d × n matrix whose
columns are the vectors ai . Consider the C-algebra homomorphism
±1
aj
π : C[x1 , . . . , xn ] → C[t±1
1 , . . . , td ], such that xj 7→ t .
The kernel of π, denoted by IA , is called the toric ideal of A. The ideal IA is
binomial and prime (see [Stu96, Chapter 4]). More precisely, IA is generated by
homogeneous binomials:
(8)
+
−
IA = hxu − xu ∈ C[x1 , . . . , xn ] : u ∈ kerZ (A)i,
where kerZ (A) = {u ∈ Zn : Au = 0}, u = u+ −u− , with u+ , u− ∈ Zn≥0 the positive
and the negative part of u.
Let IA be a toric ideal and VA = VC (IA ) be its complex affine toric variety which
is the Zariski closure of the set of points {(ta1 , . . . , tan ) : t ∈ (C∗ )d }. Define
φA : (C∗ )d → Cn ,
t 7→ (ta1 , . . . , tan ),
so that VA = φA ((C∗ )d ). We are interested in the positive part of VA , namely,
VA ∩ Rn>0 . Note that this set contains φA (Rd>0 ).
The following result follows from the Zariski density of the positive part of a
toric variety in its complex variety. However, we write an independent proof.
Lemma 5.1. Let IA be a toric ideal in C[x1 , . . . , xn ]. If u, v ∈ Nn and xu − xv
vanishes on the set of points φA (Rd>0 ), then xu − xv ∈ IA .
Proof. Notice that xu − xv evaluated at any point (ta1 , . . . , tan ) ∈ φA ((C∗ )d ) is
just tAu − tAv . Then, since xu − xv vanishes on φA (Rd>0 ), we have that tAu = tAv
for all t ∈ Rd>0 . Thus, if we fix i ∈ {1, . . . , d} and specialize to tj = 1 for all j 6= i,
(Av)
(Au)
we get ti i = ti i for all ti ∈ R>0 , which means we must have (Au)i = (Av)i .
Since this holds for all i, it follows that Au = Av, hence xu − xv ∈ IA by (8).
We now define a special toric ideal associated to a polytope P .
Definition 5.2. Let P be a d-polytope in Rd .
(1) Define the non-incidence graph of P , denoted as GP , to be the undirected
bipartite graph on the vertices and facets of P with an edge connecting
vertex i to facet j if and only if i does not lie on j.
(2) Let TP be the toric ideal of AP , the vertex-edge incidence matrix of GP .
We call TP the toric ideal of the non-incidence graph of P .
Note that GP records the support of a slack matrix of P , and so we can think
of its edges as being labelled by the corresponding entry of SP (x). Toric ideals
of bipartite graphs have been studied in the literature and the following result is
well-known.
THE SLACK REALIZATION SPACE OF A POLYTOPE
F1
F2
F3
F4
F5
F6
F7
p1
p2
p3
p4
p5
p6
p7
19
Figure 2. Non-incidence graph GP
Lemma 5.3 ([OH99, Lemma 1.1], [Vil15, Theorem 10.1.5]). The ideal TP is gen+
−
erated by all binomials of the form xC − xC , where C is an (even) chordless cycle
in GP , and C + , C − ∈ Z|E| are the incidence vectors of the two sets of edges that
partition C into alternate edges (that is, if we orient edges from vertices to facets in
GP , then C + consists of the forward edges in a traversal of C and C − the backward
edges). Thus, for every even closed walk W in GP , and indeed any union of such,
+
−
xW − xW ∈ TP .
Example 5.4. Consider the 4-polytope P = conv(0, 2e1 , 2e2 , 2e3 , e1 +e2 −e3 , e4 , e3 +
e4 ) [GPRT17, Table 1. #3] where ei is the standard unit vector in R4 . This polytope is a projectively unique polytope with f -vector (7,17,17,7). It has symbolic
slack matrix
0 x1
0
0
0
x2
0
x3 0
0
0
0
x4
0
x5 0 x6
0
0
0
x7
0
0
0 x10
SP (x) =
.
0 x8 x9
0
0
0
0
x
0
x
11
12
0
0
0 x13 x14 x15
0
0
0 x16 x17
0
0
0
Its non-incidence graph GP is given in Figure 2. Notice each edge of GP can be
naturally labeled with the corresponding xi from SP (x). Under this labeling, the
chordless cycle marked with dashed lines in Figure 2 corresponds to the binomial
x2 x3 x6 x8 − x1 x4 x5 x9 ∈ TP . One can check that the remaining generators of TP ,
corresponding to chordless cycles of GP , are
x7 x9 − x6 x10 ,
x7 x11 x13 x16 − x6 x12 x14 x17 ,
x4 x5 x13 x16 − x3 x6 x15 x17 ,
x4 x5 x12 x14 − x3 x7 x11 x15 ,
x10 x11 x13 x16 − x9 x12 x14 x17 ,
x2 x8 x13 x16 − x1 x9 x15 x17 ,
x2 x8 x12 x14 − x1 x10 x11 x15 ,
x2 x3 x7 x8 − x1 x4 x5 x10 .
The introduction of the graph GP allows us to make a precise statement on a
technique we have been using throughout the paper. We have been scaling entries
in slack matrices to ones in a mostly ad hoc way. We can now characterize precisely
which sets of variables can be fixed in this way.
Lemma 5.5. Given a polytope P , we may scale the rows and columns of its slack
matrix so that it has ones in the entries indexed by the edges in a maximal spanning
forest F of the graph GP .
20
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Proof. For every tree T in the forest, pick a vertex to be its root, and orient the
edges away from it. Now for each tree, pick the edges leaving the root and set to
one the corresponding entry of SP by scaling the row or column corresponding to
the destination vertex of the edge. Continue the process with the edges leaving the
vertices just used and so on, until the trees are exhausted. Notice that once we
fix an entry, the only way for us to change it again is by scaling either its row or
column, which would mean in the graph that we would revisit one of the nodes of
its corresponding edge. But this would imply the existence of a cycle in F , so by
the time this process ends we have precisely the intended variables set to one.
It turns out that the toric ideal TP can arise as a slack ideal.
Proposition 5.6. Let P be a polytope in Rd with d + 2 vertices or facets. Then its
slack ideal IP equals the toric ideal TP .
Proof. Up to polarity we may consider P to be a polytope with d + 2 vertices. In
this case P is combinatorially equivalent to a repeated pyramid over a free sum of
two simplices, pyrr (∆k ⊕ ∆ℓ ), with k, ℓ ≥ 1, r ≥ 0 and r + k + ℓ = d [Grü03, Section
6.1]. Since taking pyramids preserves the slack ideal, it is enough to study the slack
ideals of free sums of simplices (respectively, product of simplices). By [GPRT17,
Lemma 5.7], if P = ∆k ⊕ ∆ℓ , then SP (x) has the zero pattern of the vertex-edge
incidence matrix of the complete bipartite graph Kk+1,ℓ+1 .
From [GPRT17, Proposition 5.9], it follows that IP is generated by the binomials
det(MC ) =
x1
x3
..
.
0
x4
..
.
0
0
..
.
···
···
..
.
0
0
..
.
x2
0
..
.
0
0
0
0
0
0
···
···
x2c−2
x2c−1
0
x2c
,
where MC is a c × c symbolic matrix whose support is the vertex-edge incidence
matrix of the simple cycle C (of size c) in Kk+1,ℓ+1 .
+
−
On the other hand, TP is generated by the binomials xD − xD corresponding
to chordless cycles D of the non-incidence graph GP by Lemma 5.3. Thus, it
suffices to show that there exists a bijection between simple cycles C in Kk+1,l+1
+
−
and chordless cycles D in GP such that det(MC ) = xD − xD .
Let v1 , . . . , vk+ℓ+2 be the vertices of P and F1 , . . . , F(k+1)(ℓ+1) be its facets.
Since SP (x) has the support of the vertex-edge incidence matrix of Kk+1,ℓ+1 , we
can consider Kk+1,ℓ+1 to be a bipartite graph on the vertices v1 , . . . , vk+ℓ+2 where
each edge {vi1 , vi2 } corresponds exactly to the facet Fj of P containing neither vi1
nor vi2 . Notice that the non-incidence graph GP can be obtained by subdividing
each edge {vi1 , vi2 } of Kk+1,ℓ+1 into two edges {vi1 , Fj } and {Fj , vi2 }.
Now, let C be a simple cycle of size c in Kk+1,ℓ+1 with vertices vi1 , vi2 , . . . , vic and
assume that Fj1 , Fj2 , . . . , Fjc are the facets corresponding to the edges of C. Then
in GP there is a cycle D of size 2c on vertices vi1 , Fj1 , vi2 , Fj2 , . . . , Fjc−1 , vic , Fjc . In
fact, one can see that the subgraph induced by these vertices is exactly a chordless
cycle in GP . This is because from the support of SP we know each facet in P
corresponds to a vertex of degree 2 in GP ; furthermore, every edge in GP must be
between a vertex and a facet, but since every facet already has degree 2 in the cycle
D, this subgraph must consist only of this cycle. Hence from a simple cycle C in
THE SLACK REALIZATION SPACE OF A POLYTOPE
21
Kk+1,ℓ+1 , we get a chordless cycle D in GP , as desired. The reverse correspondence
is analogous.
However, the class of polytopes for which IP = TP is larger than those with d + 2
vertices or facets.
Example 5.7. For the polytope given in Example 5.4, which was 4-dimensional
but with 7 vertices and 7 facets, one can check that IP is the toric ideal TP .
We postpone the exact characterization of slack ideals that coincide with TP to
the next section where we will see that the projective uniqueness of P plays an
important role. In the rest of this section, we will focus on TP and establish several
results that connect the slack ideal IP and the toric ideal TP .
An ideal is said to be a pure difference binomial ideal if it is generated by binomials of the form xa − xb . It follows from (8) that toric ideals are pure difference
binomial ideals. We now prove that if IP is toric, or more generally, a pure difference
binomial ideal, then IP is always contained in TP .
Lemma 5.8. If a binomial xa − xb belongs to IP , then it also belongs to TP .
Proof. Let p = xa − xb . Each component ai of a and bj of b appears as the
exponent of a variable in the symbolic slack matrix SP (x) and is hence indexed
by an edge of GP . Recall that all matrices obtained by scaling rows and columns
of SP by positive scalars also lie in the real variety of IP , and hence must vanish
on p. This implies that the sum of the components of a appearing as exponents
of variables in a row (column) of SP (x) equals the sum of the components of b
appearing as exponents of variables in the same row (column).
Now think of the edges of GP in the support of a as oriented from vertices of P
to facets of P and edges in the support of b as oriented in the opposite way. Then
the previous statement is equivalent to saying that p is supported on an oriented
subgraph of GP (possibly with repeated edges) with the property that the in-degree
and out-degree of every node in the subgraph are equal. Therefore, this subgraph
is the vertex-disjoint union of closed walks in GP , which by Lemma 5.3 implies that
p is in TP .
Corollary 5.9. If IP is a pure difference binomial ideal, then IP ⊆ TP .
At first glance, the converse might seem true, but this is not the case in general.
Example 5.10. For the 3-cube, IP ( TP . The toric ideal TP is minimally generated by 80 binomials, each corresponding to a chordless cycle in GP , while IP is
minimally generated by 222 polynomials many of which are not binomials.
Even if IP is a pure difference binomial ideal, hence contained in TP , it can still
be the case that IP 6= TP .
Example 5.11. Consider the 5-polytope P from Example 3.8. One can check
using Macaulay2 [GS] that IP is toric and IP ( TP . In fact, dim C[x]/IP = 20,
while dim C[x]/TP = 19.
It turns out that one can attach a geometric meaning to polytopes for which
IP ⊆ TP . A polytope P is said to be 2-level if it has a slack matrix in which every
positive entry is one, i.e., SP (1) is a slack matrix of P .
22
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Definition 5.12. We call a polytope P morally 2-level if SP (1) lies in the slack
variety of P .
All regular d-cubes are 2-level and hence any polytope that is combinatorially
a d-cube is morally 2-level. However, in general, being morally 2-level does not
require that there is a polytope in the combinatorial class of P that is a 2-level
polytope. For example, a bisimplex in R3 is morally 2-level, but no polytope in its
combinatorial class is 2-level. This is since SP (1) can lie in the slack variety of P
even though it may not have the all-ones vector in its column space.
Notice that, unlike the set of 2-level polytopes, the set of morally 2-level polytopes is closed under polarity.
Theorem 5.13. A polytope P is morally 2-level if and only if IP ⊆ TP .
Proof. Notice that the ideal JP = h(d+2)-minors of SP (x)i is contained in the slack
ideal IP . Suppose that SP (1) ∈ V(IP ). Then any (d + 2)-minor p of SP (x) must
have the same number of monomials with coefficient +1 as those with coefficient
−1 since p must vanish on SP (1), which sets each monomial to one. This implies
that we can write p as a sum of pure difference binomials. Since p is a minor, each
of these pure difference binomials corresponds to a pair of permutations that induce
two perfect matchings on the same set of vertices. The union of these two matchings
is a subgraph of GP , which we can view as a directed graph by orienting the two
matchings in opposite directions. Then each vertex will have equal in-degree and
out-degree, which shows that these edges form a union of closed walks in GP , and
thus the corresponding binomial is in TP by Lemma 5.3. Therefore p ∈ TP , so that
JP ⊆ TP . Since toric ideals are saturated with respect to all variables, the result
follows.
Conversely, suppose IP ⊆ TP . Since TP is generated by pure difference binomials,
which vanish when evaluated at SP (1), we have SP (1) ∈ V(TP ). But IP ⊆ TP
implies that V(IP ) ⊇ V(TP ) ∋ SP (1), which is the desired result.
A useful corollary of the above result is the following.
Corollary 5.14. Let P be a polytope in Rd with no rational realization. Then IP
cannot be a pure difference binomial ideal and, in particular, cannot be toric.
Proof. If P has no rational realization, then SP (1) does not lie in the slack variety
of P by Lemma 4.1. Therefore, by Theorem 5.13, IP is not contained in TP . Now
applying Corollary 5.9, we can conclude that IP is not a pure difference binomial
ideal and, in particular, is not toric.
We have talked about pure difference binomial slack ideals as a superset of toric
slack ideals. A slack ideal is binomial if it is generated by binomials of the form
xa − γxb , where γ is a non-zero scalar. Therefore, one might extend the study of
toric slack ideals to the following hierarchy of binomial slack ideals:
toric ⊆ pure difference binomial ⊆ binomial.
So far, we have not encountered a pure difference binomial slack ideal that is not
toric, nor a binomial slack ideal which is not pure difference, but it might be possible
that all containments are strict. It follows from Corollaries 2.2 and 2.5 in [ES96]
that, if the slack ideal IP is binomial, then it is a radical lattice ideal. This implies
that the slack variety is a union of scaled toric varieties.
THE SLACK REALIZATION SPACE OF A POLYTOPE
23
6. Projective uniqueness and toric slack ideals
Recall that a polytope P is said to be projectively unique if any polytope Q that
is combinatorially equivalent to P is also projectively equivalent to P , i.e., there
is a projective transformation that sends Q to P . This corresponds to saying that
the slack realization space of P is a single positive point.
Example 6.1.
(1) Every d-polytope with d+ 2 vertices or facets is projectively unique [Grü03,
Exercise 4.8.30 (i), p. 68]. In particular, all products of simplices are
projectively unique. We saw in Proposition 5.6 that a d-polytope with
d + 2 vertices or facets has a toric slack ideal.
(2) In R2 the only projectively unique polytopes are triangles and squares. In
R3 there are four combinatorial classes of projectively unique polytopes
— tetrahedra, square pyramids, triangular prisms and bisimplices. The
number of projectively unique 4-polytopes is currently unknown. There
are 11 known combinatorial classes, attributed to Shephard by McMullen
[McM76], and listed in full in [AZ15]. Beyond the 4-polytopes with 4+2 = 6
vertices or facets, this list has three additional combinatorial classes. One
of them is the polytope seen in Example 5.7. It was shown in [GPRT17]
that all of the 11 known projectively unique polytopes in R4 have toric
slack ideals.
The above examples suggest that there might be a connection between projective
uniqueness of a polytope and its slack ideal being toric. In this section we establish
the precise result. The toric ideal TP of the non-incidence graph GP from the
previous section will play an important role.
Definition 6.2. We say that the slack ideal IP of a polytope P is graphic if it is
equal to the toric ideal TP .
Theorem 6.3. The slack ideal IP of a polytope P is graphic if and only if P is
projectively unique and IP is toric.
Proof. Suppose that IP is graphic. Then, IP is toric, so we only need to show that
P is projectively unique. Pick a maximal spanning forest F of the bipartite graph
GP . By Lemma 5.5 we may scale the rows and columns of SP so that it has ones
in the entries indexed by F . Take an edge of GP outside of F and consider the
binomial corresponding to the unique cycle this edge forms together with F . Since
IP = TP , this binomial is in IP , therefore it must vanish on the above scaled slack
matrix of P . This implies that the entry in the slack matrix indexed by the chosen
edge must also be 1. Repeating this argument we see that the entire slack matrix
has 1 in every non-zero entry which implies that there is only one possible slack
matrix for P up to scalings, hence only one polytope in the combinatorial class of
P up to projective equivalence.
Conversely, suppose that P is projectively unique and IP is toric, say IP = IA
for some point configuration A. Let xu − xv be a generator of TP . Notice this
+
−
generator vanishes when each xi = 1, and by Lemma 5.3, xu − xv = xC − xC
for some chordless cycle C of GP . Now, since IP is toric, by Corollary 5.9 we have
that IP ⊆ TP , and then by Theorem 5.13, SP (1) ∈ V(IP ). Since P is projectively
unique, every element of V+ (IP ) is obtained by positive row and column scalings of
SP (1). Therefore, φA (Rd>0 ) ⊆ V+ (IP ) consists of row and column scalings of SP (1).
24
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
0 0 0 ∗ ∗ ∗ 0 0 0 0 0 0 0 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 0 0 0 0 0 0
0
0
0
0
∗
0
0
∗
0
0
0 0 ∗ 0 0 ∗ ∗ ∗ 0 0 0 0 ∗ ∗ ∗ ∗ 0 0 0 0 ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 0 0
0 0 0 0 0 ∗ 0 0 ∗ ∗ 0 0 ∗ ∗ 0 0 ∗ ∗ 0 0 ∗ 0 0 0 ∗ ∗ 0 0 0 0 0 0 0
0 0 0 ∗ 0 0 0 0 0 0 ∗ ∗ 0 0 ∗ ∗ 0 0 ∗ 0 0 ∗ ∗ 0 0 0 ∗ ∗ ∗ 0 0 0 0
0 0 0 0 0 0 ∗ 0 ∗ 0 ∗ 0 0 0 0 0 0 0 0 0 ∗ ∗ 0 0 ∗ 0 ∗ ∗ 0 ∗ ∗ ∗ 0
0 0 0 0 0 0 0 0 0 ∗ 0 ∗ 0 0 0 0 ∗ 0 ∗ 0 0 0 0 0 0 ∗ 0 0 ∗ 0 0 0 ∗
∗ 0 0 0 0 0 0 ∗ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ∗ ∗ 0 0 0 0 ∗ ∗ 0 0 ∗
0 ∗ 0 0 ∗ 0 0 0 0 0 0 0 0 0 0 0 0 ∗ 0 ∗ 0 0 0 0 ∗ 0 0 0 0 ∗ ∗ ∗ ∗
0 0 ∗ 0 0 0 ∗ 0 0 0 0 0 0 0 0 0 0 0 0 ∗ 0 0 0 ∗ 0 0 ∗ 0 0 0 ∗ ∗ 0
∗ 0 0 ∗ 0 0 0 0 ∗ 0 0 0 ∗ 0 0 0 ∗ ∗ 0 ∗ 0 0 0 0 0 ∗ ∗ ∗ 0 0 ∗ 0 0
0 ∗ 0 0 0 ∗ 0 0 0 0 0 ∗ 0 0 ∗ 0 0 0 0 0 0 ∗ ∗ ∗ 0 ∗ 0 ∗ 0 0 0 0 0
0 0 0 0 0 ∗ 0 0 ∗ 0 ∗ ∗ 0 0 ∗ 0 ∗ 0 0 ∗ 0 ∗ 0 0 0 ∗ 0 0 0 ∗ ∗ 0 ∗ ∗
Figure 3. Slack support of the Perles polytope.
+
−
Since a binomial of the form xC − xC , where C is a chordless cycle, contains in
each of its monomials exactly one variable from each row and column of SP (x) on
which it is supported, it must also vanish on all row and column scalings of SP (1).
It follows that the generator xu − xv vanishes on φA (Rd>0 ). By Lemma 5.1, this
means that xu − xv ∈ IP , thus all generators of TP are contained in IP , which
completes the proof.
Theorem 6.3 naturally leads to the question whether P can have a toric slack ideal
even if it is not projectively unique and whether all projectively unique polytopes
have toric slack ideals. In the rest of this section, we discuss these two questions.
All d-polytopes with toric slack ideals were found for d ≤ 4 in [GPRT17]. We
saw in Example 6.1 that these polytopes all happen to be projectively unique, and
hence have graphic slack ideals. Therefore the first possible non-graphic toric slack
ideal has to come from a polytope of dimension at least 5. Indeed, we saw that
the polytope in Examples 3.8 and 5.11 has toric slack ideal but is not projectively
unique, thus is not graphic.
Going back to Theorem 6.3 we can now ask if there are projectively unique
polytopes that do not have a toric slack ideal.
Example 6.4 (The Perles polytope). A famous example of a projectively unique
polytope with no rational realization is due to Perles [Grü03, p. 94]. This is an
8-polytope with 12 vertices and 34 facets with the additional feature that it has a
face that is not projectively unique. The symbolic slack matrix of this polytope has
support as shown in Figure 3.
By Corollary 5.14, the slack ideal of the Perles polytope is not toric. Indeed, one
can check independently that rank (SP (1)) = 12 6= 9 = d + 1; thus by Corollary 5.9
and Theorem 5.13, IP cannot be toric.
Furthermore, this is not an isolated instance as there are infinitely many such
examples in high enough dimension.
Proposition 6.5. For d ≥ 69 there exist infinitely many projectively unique dpolytopes that do not have a toric (even pure difference binomial) slack ideal.
Proof. In [AZ15], Adiprasito and Ziegler have shown that for d ≥ 69 there are
infinitely many projectively unique d-polytopes. On the other hand, it follows from
THE SLACK REALIZATION SPACE OF A POLYTOPE
25
results in [GPRT17] concerning semidefinite lifts of polytopes that in any dimension,
there can only be finitely many combinatorial classes of polytopes whose slack ideal
is a pure difference binomial ideal.
7. Conclusion
In this paper we introduced the slack realization space of a polytope P which
is inherently algebraic, and coordinate free. The slack ideal IP which defines this
space can be used to answer many questions about realizations of P via the tools
of computational algebraic geometry, thus creating a computational framework for
attacking such questions. We exhibited examples of this utility.
Although V+ (IP ) works out as a model for the realization space of P , it could
be that IP is not the biggest ideal that vanishes on its Zariski closure. In other
words, we do not know if V+ (IP ) is Zariski dense in the slack variety V(IP ). So as
a first open problem we raise the following.
Problem 7.1. Is IP the vanishing ideal of the Zariski closure of V+ (IP )? If not,
what is the vanishing ideal of V+ (IP )?
Finding the vanishing ideal of V+ (IP ) allows one to transfer invariants from the
variety of this ideal to the realization space. For instance, computing the dimension
of a realization space is largely open and having the right ideal would provide an
algebraic tool for answering this question. On the other hand, we suspect that this
vanishing ideal, if it is different from IP , is also likely to be more complicated than
IP , which could make computations less feasible. What is the ideal sweet spot?
Related to the above question is a more basic one regarding the commutative
algebra of slack ideals.
Problem 7.2. Is IP a prime or radical ideal? If not, what are the simplest counterexamples?
In the case that IP is toric, we know that IP is indeed the vanishing ideal of
V+ (IP ), providing a perfect correspondence between algebra and geometry. Many
further questions remain. In particular, what sort of polytope combinatorics lead
to simple algebraic structure in slack ideals?
Problem 7.3. What conditions on P make its slack ideal toric, or pure difference
binomial, or binomial?
In fact, so far, we have not been able to find any slack ideal that is in one of
these classes but not in the others.
Problem 7.4. Is any of the inclusions
toric ⊆ pure difference binomial ⊆ binomial
of classes of slack ideals strict?
We also characterized the toric slack ideals that come from projectively unique
polytopes as being TP , the toric ideal of GP , the graph of vertex-facet non-incidences
of P . Such slack ideals were called graphic and in Theorem 5.13 we saw that if
IP ⊆ TP , then P is morally 2-level. The fact that testing and certifying projective
uniqueness is easy for toric slack ideals is very interesting, as that is in general a
hard problem. This raises the question of finding other classes of polytopes for
which one can certify projective uniqueness easily.
26
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
Problem 7.5. Is there another class of polytopes, beyond those with graphic slack
ideals, for which one can characterize projective uniqueness?
We saw that the slack ideal of the Perles polytope is not toric (Example 6.4). In
this example the slack matrix is too large for the slack ideal to be computed easily
and we do not know its ideal currently. Since the Perles polytope does not have a
rational realization, we know that SP (1) is not in its slack variety. Therefore, its
slack ideal is not a pure difference binomial ideal.
Problem 7.6. Is the slack ideal of the Perles polytope a non-pure difference binomial
ideal? Is it the vanishing ideal of V+ (IP )?
Despite not knowing the slack ideal, the fact that the Perles polytope is projectively unique gives us total knowledge of V+ (IP ). More precisely, once we have
one concrete slack matrix of the Perles polytope, every other point in V+ (IP ) can
be obtained by row and column rescalings of it. In particular, any polynomial
that vanishes on a concrete slack matrix and is invariant under scalings, i.e., it is
homogeneous with respect to every row and column, will vanish on all of V+ (IP ).
Notice that in the slack matrix of Example 6.4 we have, for instance, the 2 × 2
submatrix of rows 3 and 10 and columns 18 and 19 being a square matrix with all
x3,18 x10,19
entries non-zero. The quotient x3,19
x10,18 is invariant under scalings so it should
be constant on all elements of V+ (IP ). By computing a concrete slack√matrix of
the Perles polytope and evaluating the quotient we get that it is α = ( 5 − 1)/2.
This means that the binomial x3,18 x10,19 − αx3,19 x10,18 vanishes on V+ (IP ) so it
is either in the slack ideal or the slack ideal is not the vanishing ideal of V+ (IP ).
Since projective uniqueness is such a strong restriction, knowing more about how
it affects the slack ideal would be enlightening.
Apart from these concrete questions, many others could be formulated. There
are, in particular, two general directions of study that can potentially be very
fruitful: strengthening the correspondence between algebraic invariants and combinatorial properties, and revisiting the literature on realization spaces in our new
language to see if further insights can be gained or open questions can be answered.
References
[AP17]
K.A. Adiprasito and A. Padrol. The universality theorem for neighborly polytopes.
Combinatorica, 37(2):129–136, 2017.
[APT15] K.A. Adiprasito, A. Padrol, and L. Theran. Universality theorems for inscribed polytopes and Delaunay triangulations. Discrete Comput. Geom., 54(2):412–431, 2015.
[AS85]
A. Altshuler and L. Steinberg. The complete enumeration of the 4-polytopes and 3spheres with eight vertices. Pacific J. Math., 117(1):1–16, 1985.
[AZ15]
K. A. Adiprasito and G. M. Ziegler. Many projectively unique polytopes. Invent.
Math., 199(3):581–652, 2015.
[Bar87]
D. W. Barnette. Two “simple” 3-spheres. Discrete Math., 67(1):97–99, 1987.
[BFF+ 17] A. Bohn, Y. Faenza, S. Fiorini, V. Fisikopoulos, M. Macchia, and K. Pashkovich.
Enumeration of 2-level polytopes. Preprint, https://arxiv.org/abs/1703.01943, 2017.
[Dev17]
The Sage Developers. SageMath, the Sage Mathematics Software System, 2017.
http://www.sagemath.org.
[ES96]
D. Eisenbud and B. Sturmfels. Binomial ideals. Duke Math. J., 84(1):1–45, 1996.
[FFM16] S. Fiorini, V. Fisikopoulos, and M. Macchia. Two-level polytopes with a prescribed
facet. In Combinatorial optimization, volume 9849 of Lecture Notes in Comput. Sci.,
pages 285–296. Springer, [Cham], 2016.
[FMP+ 12] S. Fiorini, S. Massar, S. Pokutta, H. R. Tiwary, and R. de Wolf. Linear vs. semidefinite extended formulations: exponential separation and strong lower bounds. In
THE SLACK REALIZATION SPACE OF A POLYTOPE
27
STOC’12—Proceedings of the 2012 ACM Symposium on Theory of Computing, pages
95–106. ACM, New York, 2012.
[GGK+ 13] J. Gouveia, R. Grappe, V. Kaibel, K. Pashkovich, R. Z. Robinson, and R. R. Thomas.
Which nonnegative matrices are slack matrices? Linear Algebra Appl., 439(10):2921–
2933, 2013.
[GPRT17] J. Gouveia, K. Pashkovich, R. Z. Robinson, and R. R. Thomas. Four-dimensional
polytopes of minimum positive semidefinite rank. J. Combin. Theory Ser. A, 145:184–
226, 2017.
[GPS17]
F. Grande, A. Padrol, and R. Sanyal. Extension complexity and realization spaces of
hypersimplices. Discrete Comput. Geom., pages 1–22, 2017.
[GPT13] J. Gouveia, P. A. Parrilo, and R. R. Thomas. Lifts of convex sets and cone factorizations. Math. Oper. Res., 38(2):248–264, 2013.
[Grü03]
B. Grünbaum. Convex Polytopes. Springer GTM, New York, 2003. Second Edition.
[GS]
D. R. Grayson and M. E. Stillman. Macaulay 2, a software system for research in
algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[GS17]
F. Grande and R. Sanyal. Theta rank, levelness, and matroid minors. J. Combin.
Theory Ser. B, 123:1–31, 2017.
[JKPZ01] M. Joswig, V. Kaibel, M. E. Pfetsch, and G. M. Ziegler. Vertex-facet incidences of
unbounded polyhedra. Adv. Geom., 1(1):23–36, 2001.
[LRS15]
J. R. Lee, P. Raghavendra, and D. Steurer. Lower bounds on the size of semidefinite
programming relaxations. In STOC’15—Proceedings of the 2015 ACM Symposium on
Theory of Computing, pages 567–576. ACM, New York, 2015.
[McM76] P. McMullen. Constructions for projectively unique polytopes. Discrete Math.,
14(4):347–358, 1976.
[Mnë88]
N. E. Mnëv. The universality theorems on the classification problem of configuration
varieties and convex polytopes varieties. In Topology and geometry—Rohlin Seminar,
volume 1346 of Lecture Notes in Math., pages 527–543. Springer, Berlin, 1988.
[OH99]
H. Ohsugi and T. Hibi. Toric ideals generated by quadratic binomials. J. Algebra,
218(2):509–527, 1999.
[RG96a]
J. Richter-Gebert. Realization Spaces of Polytopes, volume 1643 of Lecture Notes in
Mathematics. Springer-Verlag, Berlin, 1996.
[RG96b]
J. Richter-Gebert. Two interesting oriented matroids. Doc. Math., 1:No. 07, 137–148,
1996.
[Rot14]
T. Rothvoss. The matching polytope has exponential extension complexity. In
STOC’14—Proceedings of the 2014 ACM Symposium on Theory of Computing, pages
263–272. ACM, New York, 2014.
[Sta80]
R. P. Stanley. Decompositions of rational convex polytopes. Ann. Discrete Math.,
6:333–342, 1980. Combinatorial mathematics, optimal designs and their applications
(Proc. Sympos. Combin. Math. and Optimal Design, Colorado State Univ., Fort
Collins, Colo., 1978).
[Stu96]
B. Sturmfels. Gröbner Bases and Convex Polytopes, volume 8 of University Lecture
Series. American Mathematical Society, Providence, RI, 1996.
[Vil15]
R. H. Villarreal. Monomial Algebras. Monographs and Research Notes in Mathematics.
CRC Press, Boca Raton, FL, second edition, 2015.
[Yan91]
M. Yannakakis. Expressing combinatorial optimization problems by linear programs.
J. Comput. System Sci., 43(3):441–466, 1991.
[Zie95]
G. M. Ziegler. Lectures on Polytopes, volume 152 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1995.
28
GOUVEIA, MACCHIA, THOMAS, AND WIEBE
CMUC, Department of Mathematics, University of Coimbra, 3001-454 Coimbra, Portugal
E-mail address: [email protected]
Dipartimento di Matematica, Università degli Studi di Bari “Aldo Moro”, Via Orabona
4, 70125 Bari, Italy
E-mail address: [email protected]
Department of Mathematics, University of Washington, Box 354350, Seattle, WA
98195, USA
E-mail address: [email protected]
Department of Mathematics, University of Washington, Box 354350, Seattle, WA
98195, USA
E-mail address: [email protected]
| 0math.AC
|
Traffic Lights with Auction-Based Controllers:
Algorithms and Real-World Data
Shumeet Baluja, Michele Covell, Rahul Sukthankar
Google, Inc.
1600 Amphitheatre Parkway
Mountain View, CA 94043
arXiv:1702.01205v1 [cs.AI] 3 Feb 2017
Abstract
Real-time optimization of traffic flow addresses important practical problems: reducing a driver’s wasted time, improving citywide efficiency, reducing gas emissions and improving air quality. Much of the current research in traffic-light optimization relies
on extending the capabilities of traffic lights to either communicate with each other or communicate with vehicles. However,
before such capabilities become ubiquitous, opportunities exist to improve traffic lights by being more responsive to current traffic
situations within the current, already deployed, infrastructure. In this paper, we introduce a traffic light controller that employs
bidding within micro-auctions to efficiently incorporate traffic sensor information; no other outside sources of information are
assumed. We train and test traffic light controllers on large-scale data collected from opted-in Android cell-phone users over
a period of several months in Mountain View, California and the River North neighborhood of Chicago, Illinois. The learned
auction-based controllers surpass (in both the relevant metrics of road-capacity and mean travel time) the currently deployed lights,
optimized static-program lights, and longer-term planning approaches, in both cities, measured using real user driving data.
Keywords: Traffic Lights, Traffic Flow, Signal Timing, Traffic Estimation, Adaptive Traffic Management, Light Schedule
Estimation, Machine Learning, Stochastic Search
1. Introduction
gorithms are methods that use reinforcement learning to discover optimal policies for the lights and/or cars [5, 6, 7, 8].
The methods referenced above have not only been applied
to fixed-policy lights, but also to lights that use vehicle sensors
to provide real-time traffic state information to the light controllers. Additionally, many researchers have looked into future possibilities where car-to-car, light-to-car, and car-to-light
communication exists [9, 10, 11]. Such communication allows
for better light control through the possibility of explicitly and
directly communicating light schedules to cars and car arrival
times to lights. Further, if we consider communication from
lights to other lights, such that each light could communicate
its schedule as well as the flows it is observing, more detailed
planning and scheduling schemes can be created [12, 13, 14].
A good overview can be found in [15].
It should be mentioned that approaches relying on a centralized controller or hierarchies of controllers have also been
explored in the research literature. However, the more coordination that is assumed, the greater the difficulties encountered
in scalability. Further, the problem of system “nervousness”
grows when central coordination is assumed — small changes
in the overall state of the system may require large changes at
the lower levels [16]. For scalability, we concentrate only on
local-decision making in this paper.
In the bidding-based traffic light controllers presented here,
external information is provided solely through local sensors
(the most common being physically nearby in-roadway induc-
Traffic congestion is a practical problem resulting in substantial delays, extra fuel costs, and unnecessary harmful gas
emissions. In urban areas, traffic is largely controlled by traffic
lights. Improving their control and responsiveness to existing
travel flows holds immense potential for alleviating congestion
and its associated problems.
Inefficient configuration of traffic lights remains a common
problem in many urban areas – one of the largest complaints of
commuters in the Mountain View, California area is the amount
of traffic they face during the morning and evening rush hours.
One of the problem areas, controlled by seven main lights, is
shown in Figure 1 (Top). The goal of our project is to evaluate
the timing of the traffic lights on these intersections and also
improve them through either better control algorithms or improved sensors. For example, many traffic lights are based on
fixed cycles, which means that they are set to green, yellow and
red for fixed amounts of time. Rarely is this an optimal solution, as real-time traffic situations are not considered, and can
leave cars waiting in long queues to satisfy shorter queues or
even empty queues. Nonetheless, even assuming fixed-length,
non-traffic-responsive lights, it is possible to optimize the light
timings to utilize historic knowledge of average (or worst case)
flows that have been observed. Such approaches are often tackled through the use of genetic algorithms to optimize the light
timings (phases and offsets) [1, 2, 3, 4]. An alternate approach
to the learning/optimization approaches of stochastic search al1
tion loop sensors1 or cameras). Rather than creating ad hoc
rules for controlling light-changes, the sensor information is
placed within the framework of a micro-auction. When a light
(phase) change is permitted, the light controller collects bids
from all the phases and conducts a micro-auction to determine
the next phase. Each phase’s bids are set by current readings
from local induction-loop sensors. The bidding process and the
weights of the bids are learned by our system through the use of
large-scale, real, historic data collected from opted-in Android
cell phone users.
This paper builds upon two branches of work: our initial
studies with micro-auction based controllers [18] and the datadriven estimation of deployed traffic lights [19]. To avoid potential confusion, it should be remarked that the auctions in this
paper substantially differs from auction-related approaches in
which drivers and/or automated cars bid for the right of way [20,
21, 22] to gain favorable light timings. In this work, the microauctions serve as a unifying internal mechanism to handle the
complexities of prioritizing the different phases (colors/settings)
of the lights.
There are two primary contributions of this work. The first,
as described above, is the introduction of an auction-based controller that relies primarily on currently deployed sensors and
does not assume the existence of future communication between sensors. The proposed controller is tested on months of
real data in two large cities. The second contribution of this
paper is to provide a novel method for establishing a realistic
baseline of performance based on the currently deployed traffic controllers. This addresses an often overlooked pragmatic
issue — in many real environments, access to the programs of
the existing light deployed throughout cities is not available.
Without this information, however, realistically assessing improvements over baselines is impossible. We address this problem through a novel method of creating models of the behavior
of the currently deployed lights based on observed traffic data.
This yields a realistic baseline of performance with which to
compare proposed improvements — including those of our own
traffic controllers.
To provide the necessary background for understanding traffic light control, in the next section, we briefly describe how
to simulate unconstrained transitioning between light phases,
which is fundamental for fully responsive controllers. We also
describe a powerful recent state of the art planning-based traffic controller [12] which was has been successfully deployed
in Pittsburgh, Pennsylvania. This provides a strong competitor to our approach. Section 3 details our micro-auction controller. All of the methods presented in this paper require learning many parameters. A simple learning/optimization procedure, termed Next Ascent Stochastic Hillclimbing (NASH), is
presented in Section 4. Experiments with alternative optimization procedures, such as Genetic Algorithms and Probabilistic
Model Based Optimization, are also described.
Figure 1: Top: Area to be optimized in Mountain View, California. A rush hour traffic flow shown (using Google Maps with
Traffic Overlay.) Bottom: Area to be optimized in Chicago,
Illinois.
1 Induction
loop sensors are placed inside the roadway’s pavement and, at
a high-level, work by creating an electromagnetic field around the loop area.
As vehicles enter and exit the field, fluctuations in the field are recorded as an
indication that a car has passed over [17]. They are widely deployed, and have
been in use since the 1960s.
2
traffic, we allow lights to make out-of-sequence transitions between green phases. This allows more complete use of the intersection capacity. In addition, when the traffic-light logic includes advance planning, there is a second advantage to out-ofsequence transitions by shortening the planning horizon. With
planning-based logic, forcing a round-robin sequencing necessitates long planning horizons to avoid making short-sighted
decisions — the light must plan ahead for the combined duration of all the phases in the cycle. In contrast, with unconstrained sequences, the light only needs to plan ahead for the
duration of a single yellow phase.
Which of the signal lights needs to be yellow depends on
the specific combination of the previous and subsequent green
phases. Since we allow out-of-sequence transitions between
green phases, we create a grid of the correct yellow phases to
place between each transition. During the simulation setup [23],
for each pair of green phases, if any green state needs to change
to a red state, the intermediate yellow phase is simply constructed by setting those soon-to-be-red states as yellow. The
second step is setting the duration of the yellow. We can look
at the states that are marked as yellow and trace backwards
from them to the maximum of the speed limits for the lanes
that those signal states control. The yellow for that pair is set to
the maximum speed divided by the Department Of Transportation (DOT)-recommended safe deceleration rate of 3 m/s2 [24]
plus a reaction time of 1 second.
As an implementation note, we remark here that all experiments were conducted within the SUMO [23] simulator.
SUMO (Simulation of Urban MObility) is a traffic micro simulation package [25], that uses discrete time steps (1 msec each)
in its simulations but keeps a continuous representation of location, distance, and speed. 2 This continuous spatial representation improves simulations that include congested surface
streets, compared to many discrete-space alternatives [26].
Figure 2: An example four-phase traffic light
As mentioned earlier, a real baseline of performance, one
which considers the programs/schedules of the currently deployed lights, is difficult to obtain. Before we present our final
results, in Section 5, we describe a practical method to estimate
the behavior of the currently deployed traffic light schedules
based on collected traffic data. Although this section is not part
of the auction-based algorithm itself, the procedures presented
in this section are vital to accurately estimate the real expected
differences in performance over current traffic-light programs.
With the baselines obtained through the methods described
in Section 5, we can provide confident measurements of the
effects of bidding-based traffic controllers. Large scale, extensive, empirical results are presented in Sections 6 & 7, for
Mountain View California and Chicago, Illinois. The auctionbased controllers yield improvements in the both the important
measurements of road-capacity and mean travel time within
both urban environments. The paper concludes with directions
for future work in Section 8.
2.1. Prior Work: Long-Term Planning Based Approaches
The micro-auction based traffic lights that are the focus of
this paper are reactive by design; they do not have a planning
mechanism for future events. Complementary approaches, such
as planning based traffic lights that attempt to anticipate the
traffic that will arrive in the future, provide additional avenues
to explore. We have implemented planning based lights to compare their effectiveness with micro-auction lights. One implementation of planning based lights is described here.
Inspired by the exemplary results of the Carnegie Mellon
University Pittsburgh traffic-light deployment [12], we implemented a traffic light controller that uses remote sensors and
planning for controlling when traffic phase switches occur. With
the planning-based approach, we solve the phase scheduling
2. Background Context: Traffic Light Control with Unconstrained Transitions
At a high level, whenever there is a transition from one
traffic-light phase to another, where some of the lights that were
green become red, there is the need for an intermediate phase, a
yellow phase, to provide warning to the affected drivers to stop
or to proceed, based on their speed and distance to the intersection. In static traffic lights, the sequence of green phases is
simply round-robin (e.g., in Figure 2, repeatedly going through
the four indicated green phases in counting order). In the case
of round-robin transitions, since the sequence of greens is fixed
and pre-defined, it follows that there is only one yellow phase
that will follow any given green phase. However, when we employ traffic sensor inputs, we are given the flexibility for outof-sequence transitions (e.g., phase 1, then phase 3, and so on),
based on the readings of the current intersection’s local induction loops. This may yield improved traffic flow, but it also
introduces the need for additional yellow-phase logic.
In our experiments, both with planning-based lights and
micro-auction based lights, in order to be responsive to sensed-
2 Two important considerations in selecting SUMO were: (1) SUMO is written in a combination of C++ and Python that can be easily modified; this was
important for being able to handle non-standard light logic. (2) The free license allows us to parallelize it on multiple machines; in our tests, thousands of
tests have been carried out simultaneously — allowing deeper exploration than
would be possible with more restrictive licenses.
3
problem by collecting the traffic data from a sequence of induction loops that are tied to a given phase of the traffic light,
with the loops spaced at about 3 seconds of expected travel time
apart from each other. The sensors are placed in the roadway,
spaced at distances dictated by the speed limit and this target
separation time. In the planning-based approach, we use these
sensors for both occupancy and speed data [27]. We assume
that they have been placed on all lanes that could lead traffic
to the controlled intersection in 15 seconds (or less) of travel.
When intersections are closely spaced, these non-local sensors
may be on the other side of other intersections, leading to a
large web of sensors and distributed communication. While
this likely includes more sensors (and communication) than is
minimally required, it enables us to create a detailed speed and
occupancy profile for each section of road, which is useful for
planning.
Internal to our planning-based traffic lights, we maintain
time lines of when we expect cars (observed by local and remote induction loops) to arrive at the controlled intersection.
The car counts (from induction loops) are scaled by a learned
weighting factor, based on the historically observed turning ratios between the sensor and the traffic light, as well as the historically observed rates for which the phase will be needed by
the incoming traffic (e.g., straight or left turn at the controlled
intersection itself). The distance-to-time mapping used to put
the car onto the time line is based on the observed speed profile
for the road (also from the known induction loops). Finally, we
use dynamic programming to solve for the best phase sequence
and transition times. Since this is a crucial part of the planning
process, it is described in more detail.
The dynamic-programming search is initialized with a single potential schedule that has the start time of the current phase
as its (already past) starting point. If there are cars waiting at the
traffic light on other phases, it increases the solution space by
considering a phase change to each of the other phases that are
in demand, taking into account the yellow-duration lead time
needed for scheduling. For example, considering Figure 2, if
phase 3 is the current phase and there is traffic waiting for phase
1 and phase 2, the space of possible solutions will expand to
three possibilities: (a) remaining on phase 3; (b) changing to
phase 1; and (c) changing to phase 2. This expansion of the
possible solution space continues with new possible branches
being introduced each time a new car arrives.
The timing of the phase changes is restricted so that the
minimum time given for each phase of the light (plus the yellow duration) is respected: e.g. if phase 4 requires a minimum
duration of 3 seconds, then the next possible phase change will
be postponed until 3 seconds plus the yellow duration after the
start of phase 4. Also, if the current phase is expected to be
“empty” before the arrival of the new traffic (that is, no cars
waiting as measured by the local induction loop sensors and
no cars arriving as measured by the remote induction loop sensors), the proposed phase change for the newly arriving traffic
is moved up to be as early as possible, such that the yellow will
start just after the current phase is empty. This early change has
the advantage of reducing the amount of time that the intersection remains under utilized.
Given the complete description of how phases can change,
the scheduling solution is then selected based on minimizing
a combination of three penalties (the automatic setting of the
weights of these penalties is described in Section 4).
1. Speed-loss penalty: This is the penalty for forcing a car
to stop. The penalty is scaled by the speed limit of the
road that the car will be traveling onto. That speed was
chosen since it best reflects the acceleration that the car
will require, that it otherwise would not have needed had
it not been forced to stop.
2. Waiting-time penalty: This is a penalty that is linear in
the amount of time that each car must wait at a red light.
3. Phase-change penalty: This is a penalty that is added for
each proposed phase change. It increases the penalty on
the schedules that unnecessarily cycle through the different phases when there is no waiting or incoming traffic.
The solution space is repeatedly expanded and then pruned,
using the approach described in [14]. It is expanded by moving forward through all of the arrival times of incoming traffic.
After each expansion, the solution space is checked for schedules that can be pruned. Pruning of candidate solutions happens with solutions that end on the same phase and have the
same number of waiting cars as other solutions, but have higher
associated partial costs (longer times). Experiments with the
planning-based controllers are presented in Section 6.
In contrast to the approaches described in this section, the
alternatives that we propose, auction-based controllers, do not
create plans and are purely reactive. They are described next.
3. Auction-Based Controllers
In this section, a detailed description of the new auctionbased controllers is provided. In contrast with the planningbased traffic-light control [14], the auction-based approach does
not require the use of remote sensors since no planning is required. Instead, only the typical induction loops that are placed
at the entrance lanes to the controlled intersection are used. Because we rely only on existing roadway infrastructure, these
controllers should be easier to deploy than those that require
remote communication.
At a high level, each phase of the auction-based traffic light
logic has three time-separated behaviors (see Figure 3). The
timing and inputs used for these behaviors are optimized to the
general traffic patterns that are expected a given time of day
(e.g., morning commute hours). Based on our tests, it is the
combination of these three behaviors along with the parameter optimization that are responsible for the improved efficiency
witnessed over static timed and planning-based approaches. The
behaviors are described below and the parameter optimization
is explained in the next section.
In auction-based logic, each traffic-light phase definition includes a weighted list of sensors that is to be used by that phase
to determine its bid for the cycle at any given time. The weights
can be positive or negative and are used to effectively scale
the number of cars observed on that sensor. For example, a
4
Figure 3: Phase change logic in auction-based traffic lights.
5
phase for a lower-priority road could include in its sensor list a
negative-weighted induction loop from in front of the higherpriority road, so that the lower-priority road would be more
likely to release the phase when cars arrive from the higherpriority direction. Similarly, a more heavily-used phase could
use a larger positive weight for some of its own sensors, to allow it to out-bid the lighter-traffic direction. Instead of having
these weighted lists of sensors be given as a fixed input that is
manually specified, the learning procedure selects both which
of the local sensors to use and their associated weight for each
phase. Formally, a bid is the weighted sum of the current, local induction-loop
measurements (sj ): the bid bi for phase i is
P
bi =
w
s
.
The
weights (wij ) are selected in the learnj ij j
ing process (see Section 4). When the learning process sets
wij = 0, we refer to that as having removed sensor j from
phase i.
As shown in Figure 3, the way in which the traffic light decides whether to change phase depends on how long the current
phase has been active. We use the terms minimum duration,
priority duration, and release duration to separate the time intervals for these different behaviors. As suggested by its name,
minimum duration is the minimum amount of time that a phase
must be green before possibly changing to yellow. The minimum duration is given as an input to the simulation and can be
set to be different for each phase of the light. For simplicity,
we start with all of the minimum durations as 3 seconds but allow the learning procedure to adjust that to be larger, if needed
to avoid too-fast switching between phases in light traffic. No
phase changes can occur before minimum duration.
For each second between minimum duration and priority
duration, the current phase has priority on the traffic light. If its
bid for the cycle is non-negative (indicating that it would like
to have the cycle), then it will keep the cycle, no matter what
the bids of the other phases are. While this greedy approach
may seem to be suboptimal, it has the advantage of increasing
the average duration of the cycles and reducing the amount of
time spent switching between green phases (and thereby reducing amount of time wasted on yellow lights). Again, we allow
the learning procedure to adjust the priority duration to the expected traffic demands.
For each second between priority duration and release duration, an auction is held between the different phases. Each
phase bids according to the weighted sum of the sensors that
have been selected for that phase. If the highest bid is negative,
then the current phase is the default winner of the cycle until
one or more of the bids change. Otherwise, the phase with the
highest bid will get the cycle (after the appropriate yellow). If
multiple phases have the same winning bid, the winner is selected by simple round-robin selection.
For each second after the release duration, it is important to
make phase switching less restrictive. The same type of auction
is held with the added constraint that the current phase cannot
bid an amount above zero. This non-positive bid by the current
phase likely releases the cycle. Any other phase that would like
to have the cycle will win the auction away from the current
phase. If more than one of the other phases have positive bids,
the auction process will pick the strongest bidder. Only if all the
bids are strongly negative, is it possible for the current phase to
stay green.
The progression restarts after the start of each green phase,
progressing from non-negotiable (below minimum duration) to
greedy (below priority duration) to auctioned (below release duration) to a handicapped auction (above release duration).
Because the learning procedure is allowed to remove any or
all of the sensor inputs to any given phases bid, we designed our
control logic to handle these cases. When a phase has no sensor
inputs, it will continually place a zero bid for the light. If the
learning procedure removes all sensor inputs from all phases
of the traffic light, the auction logic results in the light behaving as a static light, using round-robin cycling and using each
phase’s priority duration. Interestingly, in practice, the nosensor configuration is automatically instantiated through our
learning procedure for multiple lights; this will be discussed in
the results sections.
4. Learning/Optimization Algorithms and Data Gathered
No matter which approach is used for traffic light control,
fixed-schedule, long term planning, or micro-auctions, each has
numerous internal parameters that must be set. For example,
even with a simple fixed-schedule controller, the length of the
phase and offsets of each light have a large impact on the performance of the overall system. Table 1 summarizes the parameters that characterize the behavior for a given traffic light in
each of the three approaches explored here. Genetic Algorithms
(GA) [28] have been most commonly used to set the numeric
and enumerable values associated with traffic lights [1, 2, 3, 4].
Following the published research, genetic algorithms were
first attempted. A thorough overview of GAs can be found in
[28]. Since genetic algorithms are themselves characterized
by a number of control parameters, over 40 different combinations were tried. In the GA variants explored, the control
parameters that were varied included: the mutation rate was
varied between 0.5% - 10%, crossover type employed: uniform and single/multi point, the use of elitist selection (preserving best solution from one generation to the next), population
sizes were set between 10-1000, and numerous crossover rates
were attempted. Two settings consistently had the largest impact. The largest effect on the performance of the algorithm
was observed in the setting of the mutation rate. Regardless
of the other parameters, the entire population converged to far
from-optimal candidate solutions that were then substantially
improved through random mutations. Even when crossover was
turned off (probability = 0.0%), the results were not statistically
different than when crossover was used. The other control parameter, whether elitist selection was turned on, was important.
In the runs where it was used, the final solutions found were
consistently better than when it was not used.
In addition to GAs, probabilistic optimization approaches
such as Population Based Incremental Learning (PBIL) [29],
and variants that include explicit modeling of inter-parameter
dependencies were also attempted [30, 31]. These probabilistic models provide a method to explicitly maintain the statistics
that a genetic algorithm’s population implicitly maintains. A
6
Table 1: Parameter Categories Optimized for Each Approach
Fixed-Schedule
Planning
Micro-Auctions
Phase Lengths
Phase Offsets
Speed Loss Penalty Weight
Waiting Time Penalty Weight
Phase-Change Penalty Weight
Detector Weights
Detectors to Use
Durations (Minimum, Priority, Release)
parameters in S, it can generate a setting S 0 that violates constraints over parameter subsets. For instance,
increasing a phase length in a given fixed-schedule traffic light controller will break synchronization unless another phase length for the same light is reduced to keep
the overall cycle constant. To address this, we employ
an approach-specific repair operation over S 0 to adjust
parameter settings (e.g., normalizing parameters for each
light).
3. Evaluate objective: With the parameter modifications,
the new S 0 is evaluated with the desired objective function. The standard objective functions that have been
used in the literature include: minimizing the overall or
average wait time, maximum wait time, emissions, stop
time, or maximizing throughput, speed, etc. For our studies, we set the objective function to minimize total travel
time of all the cars that enter the region of interest during
the observation interval.
Our implementation assigns a pre-generated route and
entry time for each car c ∈ C in the simulation. These
are drawn directly from the (anonymized) real gathered
data of Android users, as will be described next, in Section 4.1. The exit time for the car depends on the traffic
flow in the region of interest; in general, poor settings
of S result in the car exiting the region of interest later.
Thus, the objective of total travel time is given by
X
O=
cexit − centry
good overview of probabilistic optimization techniques is provided in [32]. For our experiments, the inter-parameter dependencies that were modeled were pair-wise dependencies using
tree-shaped networks. As shown in [33], a simple algorithm
can be employed to select the optimal tree-shaped network for
a maximum-likelihood model of the data (in our case, the data is
the high performing parameter settings found through search).
This tree-model is then sampled to generate the next candidate
solution in an analogous step to the crossover operation in a GA.
Like the GAs attempted, a large driver of improvement was the
mutation operator. 3
Based on the consistently large effects of the mutation operator on the quality of the solution obtained, a simpler search
method Next-Ascent Stochastic Hillclimbing (NASH), was tried.
It is described below.
Formally, we define S to be the set of parameters that fully
specify a given scenario. For instance, a micro-auction simulation containing 10 traffic lights would require setting |S| = 50
parameters (5 parameters per light). Since different lights will
(in general) employ different parameter settings, their behavior
will not be identical. The goal of our optimization is to identify
parameter set Ŝ that minimizes the total travel time for the cars
in the simulation (detailed more formally below). NASH operates in a manner similar to other stochastic hillclimbing methods:
1. Perturb S: Since S is an aggregation of parameters, we
randomly select a parameter s ∈ S (drawn with uniform
probability) to perturb as:
c∈C
0
s ← s + U (−0.05s, 0.05s) for continuous-valued s;
s0 ← U (options(s) \ s)
for discrete-valued s.
where centry P
and cexit denote the entry and exit times for
car c. Since c∈C centry is constant across iterations (deterministic arrival times and routes), and cexit is determined by the parameter settings S, our optimization problem can be concisely stated as
X
minimize
cexit
where U (., .) denotes a draw from the uniform distribution and options(s) denotes the set of legal values for
a discrete parameter. In other words, we perturb continuous parameters by ±5% and assign a random legal
value to enumerable parameters. We do not simultaneously perturb every parameter in s; the number of parameters changed in one operation is given by sampling
U (1, 0.05|S|) (i.e., perturb no more than 5% of the total parameters in the simulation in each iteration). The
perturbed parameter set is denoted as S 0 .
2. Project S 0 onto valid subspace: Since the perturbation
operation described above operates independently over
S
c∈C
4. Accept S’: If the perturbed parameters S 0 result in an
improvement over S on the objective, we accept the perturbed settings, S ← S 0 . Otherwise, we discard S 0 .
This process is iterated until either a satisfactory solution
is found or time expires. Though extremely simple, NASH
worked as effectively as the other optimization methods tried,
consistently outperforming GAs. Further, it was simpler to implement and faster in practice than GAs and building probabilistic models. This somewhat counter-intuitive result has also
been observed by other researchers in exploring the trade-offs
3 Brief attempts with other heuristic optimization procedures were also conducted. Although they sometimes outperformed the GAs used, they did not
provide any statistical benefit over simple hillclimbing. For simplicity, we use
only hillclimbing throughout the remainder of this paper.
7
between genetic algorithms and stochastic hillclimbing techniques [34, 35, 36, 37]. This finding is especially pronounced
in problems in which mutation (as opposed to the Genetic Algorithm’s primary search operator, crossover) is the main driver
for improvement in the solution — as we have empirically found
to be the case in this problem.
We employed a straightforward objective function that meets
the goal of minimizing overall travel times. However, it is
worth noting that many other objective functions could have
been used. Two common ones include minimizing driver idle or
wait-times and minimizing emissions. Although both of these
can be used here with no change to the algorithms or optimization procedures (other than the objective function); their effects
can be large. For example, in minimizing emissions, reducing
stop times at lights and consistency of speed may take higher
precedence than minimizing overall travel time. Although this
may also reduce overall travel time, that reduction will be a
side-effect of reducing the emissions. Additionally, other interesting, more specialized objective functions can be used as well,
such as minimizing travel times on specified roads or routes, or
even minimizing congestion at a specific light.
traffic light control and examine the pragmatic issue of how to
obtain the existing traffic light schedules and behaviors from
which to create a realistic baseline of performance.
Unfortunately, for deployed traffic lights, the running light
schedules/programs are often not known. Rarely are they kept
in a central database, and even when they are, they are often
not easily obtainable. The most straight-forward solution is
to manually watch and time each traffic light in the city to be
optimized. For lights on fixed schedules this may be theoretically feasible, but will likely be prohibitively expensive and
time-consuming to do at scale. Nonetheless, without the existing schedule information, it is difficult to ascertain any realimprovements that new traffic light algorithms and approaches
will have in reality.
We take a behaviorist approach to discover lights’ currently
deployed schedules. We attempt to discover the light schedules
for all the lights in the system we wish to model. This is done
by creating models that match known car travel paths and timings that were observed in a large collection of gathered traveltracks. If we can determine the lights’ programs accurately,
when simulating the known travel-tracks through the city, the
timings will match observed timings. With this derived model,
we can simulate their performance light’s programs with varying traffic conditions to test how they will scale in comparison
to any of the alternative light controllers that we propose.
Unlike the rest of this paper, in which which we use NASH
to optimize controllers to improve traffic lights’ schedules, in
this section, the goal is to obtain a light schedule that matches,
as closely as possible, the behavior of the current pre-existing
schedule. Recall that NASH relies on a repeated stochastic generation and evaluation methodology to guide search. The evaluation of the candidate traffic light schedules is controlled by
setting an objective function (in the description of the NASH algorithm presented in Section 4, see step 3). Instead of using the
commonly employed traffic-optimization objective functions to
improve some aspect of flow throughput, we specify the following objective function based on similarity (formulated as a
minimization task):
4.1. Anonymous Real-World Traffic Data
For the real-world experiments presented in this paper, two
sets of data are needed: the roadway information (layout, speed
limits, etc.), and travel-track information. To gather the road
information, we combined the data available from Google maps
and OpenStreetMap [38]. The results provided roadways as
well as traffic light locations, as shown in Figure 4.
In addition to accurate road information, demand for each
road section must be modeled. We created a demand profile
through anonymized location data collected from opted-in Android cell phone users [39][40]. The data was collected over
several months. The raw data, which itself does not include personally identifiable information, was additionally scrubbed by
segmenting the travel-tracks to prevent association of trip origins and destinations to even further reduce identifiability risks.
See Figure 4(Enlarged Section) for a sample of the data collected.
From this data, we further sub-select the data with traveltracks that intersect with the map area shown. We also filter
by time, limiting to looking at a given start and end date — in
particular around rush hour periods. We then filter all the given
times down to the weekday and time of day (e.g., Tuesday 7am
local time). This provides a close-to-realistic profile of the road
demand for Android users. By folding/overlapping the gathered
data over weeks, we compensated for the fact that all travelers
are not Android users who provide their data.
f (h) =
X
|JourneyTimeh (c) − JourneyTimea (c)|,
(1)
c∈C
where h and a are the hypothesized and actual light settings,
respectively, and C denotes the set of cars in the simulation.
Minimizing this objective allows us to determine light settings that generate a simulated traffic flow that closely mirrors
the actual flow. Note that although the constraint is not explicitly specified in Eqn. 1, the cars are all introduced into the
system in the same order and at the same times as they were
observed in the actual data; this is crucial to ensure that traffic
jams and roadway usage are emulated correctly. As a reminder,
note that this similarity measurement is suitable only for determining the currently deployed light schedules. Similarity is not
the correct measurement to use when optimizing the light controllers to minimizing wait time, emissions, total travel time,
etc. For those objectives, we will simply minimize the specific
objective instead.
5. Establishing a Baseline: Modeling the Behavior of Currently Deployed Lights
Before conducting the experiments, we need a method of
determining whether the new controllers will improve real traffic flow beyond the currently deployed traffic light controllers.
In this section, we step away from discussing algorithms for
8
Figure 4: Roadway data imported into SUMO simulator [23]. Left: Overview of the area we are considering in Mountain View,
CA. The seven lights considered are numbered (and appear in red). Right: Enlarged section with traffic at Shoreline Blvd. and
101/85 exit ramps at 9am on a summer Tuesday. Each yellow triangle represents one vehicle.
5.1. Validating the Estimates on Controlled Data
Because estimating deployed traffic light parameters is a
novel part of this study, we validate the approach with a series
of experiments using synthetic data before turning to real-world
data. Using a smaller, synthetic, data set has the advantages of
being noise-free and completely within our control to modify
in order to examine different aspects of matching. For the synthetic data, 3200 cars were instantiated over a period of 4000
seconds to travel along the simple grid shown in Figure 5. The
speed limit for each segment was chosen independently and
randomly, and seven types of car were instantiated, with differing profiles in terms of acceleration/deceleration, following
distance, length, etc.
Each simulated car’s path was chosen to start and end at randomly selected edge nodes. The path was constrained such that
no intersection was visited twice. Vehicle launch times were
uniformly randomly distributed over the 4000 seconds.
The 3200 cars were simulated in SUMO with a pseudorandomized 4 light setting, this is referred to as the Target-LightSetting. For the synthetic data experiments, the Target-LightSetting corresponds to the settings for traffic lights that we seek
to estimate using the procedures in this section. For simplicity
of exposition, here we assume that the actual traffic lights can
be modeled with fixed schedules. As discussed later, more com-
Figure 5: The synthetic roadway. The number of lanes varies
between 1 and 3 in each direction. Cars can be introduced and
exit at any of the twelve outer edges. Lights are at located at
each of the nine intersections. Enlarged region shows typical
traffic.
4 This light setting was significantly perturbed from SUMO’s default light
setting to ensure that it could not “accidentally” be found by NASH by initializing SUMO and making tiny perturbations to their default light settings.
9
Figure 6: Match Error over successive trials. 2000 trials shown.
Y-axis the average difference in journey times for cars under
the hypothesized light setting and the actual light setting (in
seconds). X-axis: trial number.
Figure 7: Distribution of travel times for 3200 cars. (Top)
Lights calibrated with NASH. (Bottom) uncalibrated lights.
These are lights with random phases and offsets — presented
for comparison.
plex controllers that incorporate induction loop sensors also fit
naturally within this procedure.
NASH was then applied to the uncalibrated, randomly initialized, lights to adjust them such that the cars had approximately the same travel times as in the original distribution. The
objective function was to match, for each travel path, the arrival
time of the car as closely as possible to what SUMO yielded
with the Target-Light-Setting. Recall that because we want to
emulate not having any a priori information about the actual
light settings, we start NASH with random settings for all the
lights in the system. The hope is that through learning, the
light settings that are found will behave the same as those in
Target-Light-Setting. It is interesting to note that there may be
many possible light settings that generate similar aggregate traffic flow behaviors. As will be shown in later experiments, when
multiple NASH runs are conducted, different, but equally well
matching, light settings are found.
The progress of the matching algorithm is shown in Figure 6. Two lines are shown, the bottom line (in red) shows
the best light setting that was found in the search to that point.
The top line (in blue) shows the evaluation of each candidate
light setting as search progresses. From the red line, observe
that in the beginning of learning, the average difference (in seconds) between when a car reaches its destination with the original Target-Light-Setting and the approximated light-setting was
over 80 seconds. By the end of learning, this is reduced to 31.8
seconds. Also notice that the blue line is riddled with spikes.
This means that the majority of perturbations to the best-light
setting found to that point yielded matches that performed significantly worse; small mutations in the light settings caused
drastic changes in the overall traffic flow. Only a few trials,
those where the red line took a step downwards, revealed an
improved performance. These are the steps in which the NASH
algorithm accepted a new baseline from which the learning then
proceeded.
To compare the distributions of travel times to the actual
travel times, see Figure 7 (Top). The actual travel time dis-
tribution is shown with dark bars, the travel time distribution
obtained with the calibrated light settings is shown with lighter
(green) bars. As can be seen, the distribution of travel time appear similar. For comparison, if we plot the travel times for random light settings (uncalibrated), the distributions appear quite
different Figure 7 (Bottom).
We can also measure the correlation of journey times under the hypothesized light system and the target light system.
Here, we correlate each car’s travel times under both scenarios.
Results are shown in Figure 8. Instead of just correlating a single trial, Figure 8 also shows the results of 24 additional tests.
We reran the entire NASH-matching algorithm from scratch 25
times. Because NASH is stochastic and is initialized with random seeds, we fully expect different schedules to be found in
each run; the hope is that at the end of each run, the aggregate behavior approximates the target light system — even if
the light settings themselves are not the same. The correlations
to the actual travel times of all 25 calibrated systems (found
through NASH) remains high for all trials.
For comparison, also shown in Figure 8 are the correlation
of 25 uncalibrated (random) light systems to actual travel times.
It might seem surprising that for random systems there is any
correlation; however, since we are measuring travel times, even
with random light settings, as cars travel through the entire system, longer paths are likely to have longer travel times, regardless of the light settings. Nonetheless, as can be seen, by matching the lights through NASH, the correlation of travel times increases dramatically.
Next we posit the question: how robust are these matches?
What happens when we severely alter the underlying traffic profile? If we change all the routes and their distributions, how
do the travel times of the new cars compare in the NASH10
Figure 8: Correlation of calibrated and uncalibrated lights with
actual travel times. Average calibrated correlation = 0.8, average uncalibrated correlation = 0.5.
Figure 10: Correlations of 25×calibrated and 25×uncalibrated
lights to each other and across sets. The lighter quadrant (upperleft corner) is the 25 calibrated lights compared to each other.
They exhibit a large degree of correlation. The darker quadrant
(lower-right corner) is the uncalibrated lights, exhibiting far less
correlation. The white diagonal line is each setting’s correlation
with itself (1.0).
shows the 25x25 correlations with the calibrated lights. As can
be seen, they exhibit high correlations with each other. When
they are compared to the non-calibrated lights, the correlations
drop precipitously. The bottom-right corner shows the correlations between the uncalibrated lights; far less correlation is
present.
In summary, the results are what we hoped: despite not
being optimized to be exactly the same light-programs (since
the true light program may not be known), the behavior of
the calibrated lights are highly correlated with each other and
also highly correlated to the actual observed timings. This is
precisely what is required to obtain a realistic baseline performance.
Next, we return our attention to optimizing the traffic lights
to improve capacity and throughput. To measure our performance, we use the methods described in this section to establish
credible baseline performances in both cities examined.
Figure 9: Correlations of calibrated and uncalibrated Lights to
actual timings when all of the routes are replaced with never
previously seen routes.
calibrated light systems to the original Target-Light-Setting?
If the NASH calibrated light systems were truly close to the
Target-Light-Setting, we would expect a high correlation to remain under any traffic load, not just the load on which it was
trained.5 To test this, all of the routes are replaced with randomly created new routes. We measure the correlation of timings of the cars under the target-light-schedule and the NASHderived-light schedules, and the uncalibrated light schedules
(random). The results are shown in Figure 9.
The results are positive: under both seen (Figure 8) and unseen traffic loads (Figure 9), after NASH is used to estimate the
settings of the lights, we observe that the correlation of travel
times to the original Target-Light-Setting remains high.
As a final test, we visualize how correlated the light settings
are with each other. If they all capture the same conditions
well, they should exhibit high correlation to each other. We
also compare how correlated the 25 uncalibrated light settings
are to both the calibrated lights and uncalibrated lights. The results are shown in Figure 10. The upper-left quadrant (brighter)
6. Results: Mountain View, California
In this section, we examine the performance of static lights,
long-term planning based lights, and auction based lights, on
real traffic from Mountain View, California.
6.1. Calibrating to Real Lights
The results from Section 5 revealed that matching the behavior of lights was possible, at least with simulated data. In
this section, we apply the same approach to the seven traffic
lights in Mountain View, California. This is a significantly
larger and more complex simulation than the synthetic data
5 This
additional test is only possible for the data in Section 5, since we are
using purely synthesized traffic, and we therefore have the actual traffic light
settings that we are trying to discover with NASH. In real usage (e.g. the next
two sections) this test would not be possible without having ground-truth light
settings.
11
not have unnecessary delays in start/stop times due to exogenous factors such as distraction, change of plans, etc. Second,
in the synthetic data, drivers behaved uniformly at yellow and
red lights; this was not the case for real drivers. Even if the exact correct light settings were found, these factors would lead to
lower correlations since they would not be modeled. Third, the
real data is significantly more noisy; a problem we did not have
with synthetic data. Recall how the data was acquired and aggregated over the period of many months (done to account for
the fact that not all drivers are opted-in Android phone users).
This aggregation process leads to noise in the real data which
we are trying to mimic. Fortunately, as usage of cell-phone
GPS/maps increases and more travel-track data becomes available, these estimates will improve. Despite the above difficulties, we were able to significantly reduce the average discrepancy between actual and predicted times from approximately 7
minutes to 51 seconds.
As a final test to ensure that NASH actually learned something about the lights and did not overfit the exact traffic on
which it was trained, we repeated the experiment with a more
difficult setup. For NASH, we only looked at the data from
the first 1.5 months of collection. Then, for testing the timings we tested on a non-overlapping, subsequent, 1.5 months
of data. This also represents another common use and test scenario where a period of time is devoted to training and then the
learned model is used in the future. What we found indicated
that NASH captured the behavior of the lights and did not overfit the data. The results did not change from using the entire
3 months of data. Both the mean error and overall correlation
remained the same as when the entire 3 month period was used
for modeling the lights.
Figure 11: Progress on matching the real traffic using NASH.
Y-axis the average difference in journey times for cars under
the hypothesized light setting and the real-data (in seconds). Xaxis: trial number. First 500 trials shown - the remaining 1500
trials (not shown) showed continued, but smaller magnitude,
improvements.
use earlier. In the previous section, the synthetic simulations
used 3200 travel-tracks. For these experiments, we use approximately 67,000 tracks. Additionally, unlike in the synthetic experiments, the distribution of paths is far from uniform. Backups happen non-uniformly — only on certain streets and in certain directions. Small perturbations in a single critical traffic
light’s schedule can lead to drastic changes in throughput while
large perturbations to the schedule of a less busy traffic light
may generate little observable impact.
We begin the calibration procedure similarly to the synthetic
case — we use NASH to match the timings. The progress is
shown in Figure 11. Note that the mean error in times drops
quickly. In the beginning of the learning process, the error was
over over 400 seconds (in this case, the default SUMO traffic light settings were used for initialization). By completion,
NASH lowered the error between actual and matched times to
approximately 51 seconds.
Next, similar to the analysis conducted with synthetic data,
we examine the distribution of travel times for the 67,000 tracks.
Figure 12 shows the distributions of the tracks for the real data
and the times obtained through a simulation in SUMO using the
NASH-calibrated lights.
As can be seen the distributions are similar, but as expected,
not as close a match as with the simulated data. The correlation
between the predicted and real times is 0.5. For reference, when
a random light setting is used (the default SUMO settings), the
correlation with the real data is only 0.18. Recall that with the
synthetic data, the correlation of NASH-Calibrated lights to the
actual timings was 0.8 and the random light settings to the actual timings was 0.5.
There are several distinguishing attributes of this data that
explain why it was possible to correlate the synthetic data better than the real data. First, in the synthetic data, drivers did
6.2. Testing the New Traffic Light Controllers
Thus far, we have used NASH to match the timings of the
lights in simulation to those actually deployed on the roadways.
Now, we use NASH in a manner that is more familiar — to optimize the internal parameters of the light controllers to reduce
the mean travel times. NASH is used to optimize the parameters of the three types of controllers. For fixed-schedule controllers, the phase lengths and offsets are optimized, for longterm planning controllers the penalty weights are optimized,
and for micro-auction based controllers the detector weights
and their durations (minimum, priority and release) are optimized. The objective function (step 3 in the NASH algorithm)
is set to minimizing the mean travel time of the cars. Each
learning run is given a total of 2000 simulations – that is, 2000
perturbations of the traffic controller’s parameters were tested
in simulation.
The traffic-data used for the learning process were based
on the data collected over months in Mountain View, California. Rather than using this exact data for the learning, multiple
new, reality-based, data sets were created based on this data.
Each new data set was created by perturbing the original data
in the number of cars released on each route and the car-release
schedules. Small changes in the order of car release or in the
number of cars can have very large effects on the overall traffic
and timings of the cars in the simulation. Therefore, instead of
12
Figure 12: Distributions of Real (Dark) and NASH-calibrated light travel times (Light/Green).
comparison of optimized static control to the NASH-calibrated
control. The optimized static control uses the phase durations
and relative offsets that were found to be best for an averagemorning commute period (weekday 7am–11am), using NASH.
There is improved throughput throughout the commute cycle,
with the largest improvements seen when it is most needed,
during the 8:00-9:30am peak period. For example, we could increase the peak period load by as much as 16% without increasing the travel times, using this learned static controller instead
of the NASH-calibrated controller. This improvement demonstrates the power of the learning procedure. It is interesting to
see this level of improvement using static light timing, even after the light synchronization work was done on Shoreline Blvd.
in 2012–13 [41]. It may be that the increase in traffic in those
two years is enough to change the best choices for the traffic
light durations.
A potential concern with the witnessed improvements is
whether the improvements come at the expense of infrequently
traveled routes experiencing severe delays that they previously
did not? These trade-offs were examined in detail. The vast
majority of cars reduced their travel times. Of the ones that
did not, most increased their travel times by under 4 seconds.
Only a few cars experienced longer delays under the optimized
controllers; however, even with the extra delays, their resulting
times were in the lowest 10% of the times experienced without
learning.
For planning-based lights (Table 3), we did not see a consistent change in capacity across the commute periods: the implied changes in capacity, compared to the NASH-calibrated
control, were all small and likely not to be significant. This is
most probably due to the situation that was studied. The majority of the traffic that was involved in the simulation were cars
exiting from a freeway onto a congested arterial road. Most of
the traffic was not in “clusters”, as observed in the Pittsburgh
study [12]. Another potential contributing factor is that there
were only three parameters (the penalty weights) that were optimized. We made this choice because of the obvious physical
optimizing the signal timings on a single data set, simulations
were run for all the data sets (10 dataset perturbations were created and used). The lights were optimized to work well on all
the data sets by incorporating them into the objective function.
During the learning process, to determine whether a candidate
light setting was better than its predecessor (Steps 4 & 5 in the
NASH algorithm), not only did the average travel times across
data sets have to be lower than its predecessor, but of the N data
sets that were used, at least N/2 had to have a lower mean travel
times. This ensured that if a candidate light setting worked exceedingly well on a single data set at the expense of working
well on the rest, it was discarded, despite potentially having a
lower overall average across all data sets.
The benefits of using the multiple reality-based data sets
were two-fold. First, multiple data sets helps account for naturally occurring variability in traffic patterns on different days.
Since small changes cause large distortions in traffic flow, this
is vital. Second, by not using the actual data for training the signals, this left the pristine, real, data for the final round of tests.
The clean data was used only as a final test to measure effects
of improved controllers on capacity and mean time of travel.
The results, comparing mean travel times, and expected capacity, are shown in Tables 2 and 3, respectively. We compare different approaches to traffic light control during different commuter-traffic profiles. We perform the comparison by
considering as our baseline the travel times under the NASHcalibrated light controls.
When comparing an alternative approach (e.g., optimized
static-phase controls) to the baseline, we run the simulation using the alternative control approach, always using the same distribution of routes as was used for the baseline. For Table 2,
we keep the same number of cars as well and simply compare
the mean travel-time (MTT) changes. For Table 3, we scale
the number of cars up or down (with the same distribution of
routes), until we match its average travel time to the baseline
travel time.
The first lines of Tables 2 and 3 report the results of our
13
Table 2: Mean-travel-time (MTT) changes under matched demand - Mountain View, California
Optimized static lights
Planning-based lights
Learned Auction lights
Peak rush (8:00–9:30 am)
21861 observed cars
668.68 sec observed MTT
Average rush (9:00–10:30 am)
20139 observed cars
173.65 sec observed MTT
Low rush (9:30–11:00 am)
18154 observed cars
110.37 sec observed MTT
44% faster
21% faster
79% faster
30% faster
8% slower
38% faster
6% faster
6% slower
10% faster
(376.56 sec MTT)
(525.82 sec MTT)
(140.87 sec MTT)
(121.37 sec MTT)
(187.18 sec MTT)
(108.01 sec MTT)
(103.58 sec MTT)
(116.54 sec MTT)
(99.57 sec MTT)
Table 3: Capacity changes under matched MTTs - Mountain View, California
Optimized static lights
Planning-based lights
Learned Auction lights
Peak rush (8:00–9:30 am)
21861 observed cars
668.68 sec observed MTT
Average rush (9:00–10:30 am)
20139 observed cars
173.65 sec observed MTT
Low rush (9:30–11:00 am)
18154 observed cars
110.37 sec observed MTT
+16%
+3%
+47%
+9%
-4%
+46%
+8%
-2%
+11%
(25306 cars)
(22473 cars)
(32082 cars)
(22037 cars)
(19253 cars)
(29499 cars)
(19661 cars)
(17736 cars)
(20204 cars)
the seven lights, the learning procedure removed all sensors
from all the phases, resulting in the lights acting as static lights:
lights 1, 2, 5, and 7 (see Figure 4 for numbering). The phase
durations seen on all of these lights except light 5 were nearly
identical to the durations found when we optimized static-control
lights: the other three were within 2 sec (out of 86 sec total cycle length) of the optimized static-control light durations. At
least this small amount of variation (if not more) is expected
from a stochastic search process, such as NASH.
The learning procedure also found several different ways
to, in effect, remove dedicated-left phases, especially at the
most congested intersections (for example, lights 3 and 6 in
Figure 4).6 This behavior seems especially interesting, given
the reduced capacity that is generally seen at intersections with
many dedicated lefts [42].
One approach that the system automatically discovered to
remove dedicated lefts from the less-used roads was to match
the bids of the dedicated left to the direct-through phase that
was just before it, in the round-robin ordering. Specifically, on
light 6 (Figure 4), phases 1 and 2 share a matched set of sensor
weights and phases 3 and 4 share another matched set of sensor
weights. This means that the direct-through/dedicated-left pair
will always tie on any auctions. Since the direct-through phases
are always before the dedicated left phases in the round-robin
cycle when starting from an opposing direction (e.g., phase 1
will be before phase 2, when starting from either phase 3 or
phase 4), this has the interesting property that cycling between
directions (cycling from north-south to east-west or vice-versa)
will always go to the through phase.
In visual examination of this intersection during simulation,
it often happened that phases 1 and 2 started giving high bids
for the light, while phase 3 (east-west direct-through) was active but still greedy. As soon as phase 3 is slightly longer than
its priority duration, the light changes to using phase 1 (northsouth direct-through), without going through the phase 4 (the
correlates for all of the other parameters (e.g. observed speed
profiles for sensor-to-light delays and turning ratios for sensorto-phase weights).
Finally, our auction-based traffic lights provided the largest
gains in capacity over the matched-to-current controls. The
biggest gains were during the peak rush period, when it is most
needed: we were able to allow 47% more traffic into the road
network (using the same distribution of routes) during these
times, without increasing travel times, when we used auctionbased traffic controls compared to the matched-to-current lights,
a significant accomplishment. In contrast, if we increased the
traffic load during peak hours by 47%, the currently deployed
lights (as modeled by the procedures described in Section 5),
we expect to see the travel times increase by over 200% (to
2203.6 sec). Similarly, impressive reduction in travel times
were achieved across the board, especially during peak rush
hour.
6.3. Detailed Discussion
Because of the stochastic nature of NASH, as well as the
complex interactions possible between a system of traffic lights
and real traffic, it is difficult to a priori predict what behaviors will emerge after learning. From the previous tables, it
was clear that throughput and mean travel time improvements
were found. In this section, we attempt to give some intuition into (the sometimes non-obvious) ways in which they were
achieved.
Somewhat surprisingly, the planning-based approach was
the least successful of the systems that we tested. Due to the
strong physical interpretation of each of the parameters (physical distances, etc), it was the one that least lent itself to optimization. A future research direction is to test this approach
using fully optimized parameters that may not correspond to
physical interpretations, but may reveal improved performance
within the context of the full system.
Examining the auction-based lights in more detail, the decisions learned are both interesting and revealing. In four of
6 The largest amount of congestion is at light 4. However, due to the complexity of that intersection, there was no dedicated-left phase to remove.
14
dedicated left). Similar switching is only occasionally seen for
the opposite pairs. Since Shoreline has much heavier traffic
than Terra Bella (the east-west road), this happened much less
often. Both sets of dedicated lefts happen less often but are still
possible.
The learned controllers also used the time differences between release durations to favor traffic on the phases that were
more prone to long lines. For example, at light 4 (Figure 4), the
priority duration given to the 101-Freeway exit ramp was 10×
longer than the priority duration given to La Avenida (the small
road that enters the intersection from the east). With this 10×
weighting, we see the traffic delays suffered by the La Avenida
traffic approximately equal (on a per-car basis) to the delays
seen by the 101-exit-ramp traffic.
Finally, the learned lights used the number of incoming lanes
allowed to pass through a phase as another way to bias the auction towards the dominant direction of traffic. For example, at
light 3, the combined weight given to the sensors of Shoreline
traffic is five times that given to the Pear-Ave traffic (the eastwest road at that intersection). This was done by giving all of
the sensors the maximum-allowed positive or negative weights
and simply relying on the number of sensors (and hence the
number of lanes) to provide priority to larger streets.
7. Results: Chicago, Illinois
In Chicago, we considered twenty lights in the River North
area shown in Figure 1 (Bottom). The area, as imported into
SUMO, is shown in Figure 13.
The steps in the optimization of the twenty traffic lights in
Chicago proceeded in the same manner as the optimization of
the lights in the Mountain View area. To avoid redundancy, we
will provide only a summary of the procedure and results and
refer the reader to Section 6 for details.
7.1. Calibrating to Real Lights
In order to be able to establish a realistic baseline, as we
did with Mountain View, we once again used NASH to calibrate the lights in our simulations (see Section 4). Similarly
to the previous experiment, anonymized, fully privacy preserving, travel-tracks of Android users were collected over several
months. For our tests, approximately 18,000 cars were used in
the simulation. By using NASH to adjust the lights schedules,
we were able to reduce the average discrepancy between the actual and the predicted travel times to approximately 43 seconds.
For comparison, in the analogous procedure using the Mountain
View, California data, the error was reduced to 51 seconds (see
Figure 11).
Figure 13: Chicago roadways as imported in to SUMO. The
twenty lights to be optimized are marked.
7.2. Testing the New Traffic Light Controllers
Once the settings for the deployed lights are estimated, the
next step is to attempt to change the light settings to improve
the traffic flow. In the first experiment, NASH is employed to
modify the phase durations and offsets of the 20 lights (e.g. optimize static lights). In the second experiment, NASH is used to
optimize the parameters of the micro-auction based controllers.
15
tained. With the auction-based light controllers, the learning
process discovered settings that reduced the mean travel time
by 25%.
As with the study conducted with the Mountain View, California, the settings for the light controllers are found by training
with multiple reality-based data sets. This helps to account for
variability in traffic flows as well as leaves the actual data untouched for testing. For the final results, reported below, we use
the actual data.
Tables 4 and 5 compare the mean-travel times and the capacity changes to NASH-calibrated lights. As witnessed in
Mountain View, both the optimized static and the auction-based
traffic lights provide improvements in the mean-travel times and
in the capacity of the streets: up to 25% faster mean-travel time
for the auction-based traffic lights and, during the main commute hours, up to 150% more capacity during low rush hours.
The traffic problems addressed in this region of Chicago are
different than those in Mountain View, California. In Mountain
View, most of the traffic delay was associated with long, persistent, queues at intersection 4 in Figure 4 (Shoreline and the
northbound exit from 101 and 85) and with the addition of traffic onto Shoreline from that exit. In the Chicago study, there is a
busy freeway exit at intersection 12 in Figure 13, but the traffic
clears from the ramp on most light cycles. Instead, the delays
in traversing this section are largely due to repeated stops at
the traffic lights along W. Ohio (intersections 12, 13, and 14 in
Figure 13) and along W. Ontario (intersections 9, 10, and 11).
The optimized static lights are able to increase capacity and
decrease travel times by changing the phase duration allocation
to better match the dominant directions of congestion at the different intersections, most significantly increasing phase durations for east-west traffic as it crosses N. Wells and N. Franklin
and increasing the phase duration for north-south traffic at N.
Orleans and W. Ontario (intersection 9) by 14%. The phase offset between the lights was unchanged from the matched lights
to the optimized static traffic lights. This helps explain why
the travel times had a smaller improvement than the capacity
measurements: many cars still slowed or stopped for the same
sets of lights, but the length of the queues were smaller (even
with increased traffic load), due to the optimized match to the
historically-observed congestion.
The explanation for the improved capacity for the actuated
lights is similar but more indirect. We cannot simply examine
the phase durations, since the duration is demand driven and
changes throughout the simulation. Furthermore, the average
change in phase duration from matched to actuated lights (that
is, the difference in the percent time that a given phase was
green over the full simulation run) was not consistent across
the different time windows. The most noticeable and consistent
change between the static and auction-based light simulations
was that the size of the queues for the major intersections were
shorter in all directions using the actuated traffic lights. The
first car to the intersection would typically still need to slow or
stop but, the following cars in the same wave would typically
make it through the same cycle. Examining the way that the
sensors were used by the actuated lights, there seems to be a
bias towards using all of the sensor readings to establish the bid
Figure 14: Progress of average time of travel as learning of controller parameters occurs for the 20 lights in Chicago. Lower is
better. 1.0 on the Y-Axis is the base performance. Variations
from 1.0 are due to perturbations in the controller’s settings.
TOP: Optimization of fixed light schedules. BOTTOM: Optimization of micro-auction based traffic lights. Each run is allowed 2000 evaluations (controller settings) to try. Note that
many controller settings result in performance worse than the
initial setting; these are discarded. Also note that the learning
of micro-auction controllers allows the average time of travel
to drop substantially more than optimizing the fixed-time controllers.
Unlike the previous experiments with Mountain View, CA. we
do not revisit the planning based approach for this scenario.
Learning, as it progresses, is shown in Figure 14. From the
two graphs, we see several important trends. First, optimizing
the fixed-light controllers (Top) does not dramatically improve
the performance of the fixed-light controllers in terms of the average time to destination. Compared to the default settings, the
best setting found revealed an improvement of approximately
7%.
In contrast, there are much larger variations in the optimization of the auction based controllers. In Figure 14 (Bottom),
we see that in the 2000 trials, many times the settings are significantly worse than the original. As explained earlier, these
settings are tried and then discarded. Most importantly, however, in the cases in which better controller settings are found,
a substantial improvement over the fixed-light controllers is ob16
Table 4: Chicago: Mean-travel-time (MTT) changes under matched demand
Optimized static lights
Learned Auction lights
Peak rush (8:00–9:30 am)
7516 observed cars
87.18 sec observed MTT
Average rush (9:00–10:30 am)
6834 observed cars
86.42 sec observed MTT
Low rush (9:30–11:00 am)
6390 observed cars
85.71 sec observed MTT
2% faster
24% faster
3% faster
23% faster
3% faster
25% faster
(85.63 sec MTT)
(66.40 sec MTT)
(84.02 sec MTT)
(66.57 sec MTT)
(83.17 sec MTT)
(64.41 sec MTT)
Table 5: Chicago: Capacity changes under matched MTTs
Optimized static lights
Learned Auction lights
Peak rush (8:00–9:30 am)
7516 observed cars
87.18 sec observed MTT
Average rush (9:00–10:30 am)
6834 observed cars
86.42 sec observed MTT
Low rush (9:30–11:00 am)
6390 observed cars
85.71 sec observed MTT
+9%
+26%
+26%
+95%
+31%
+150%
(8195 cars)
(9505 cars)
of the phase for the dominant direction of travel (intersections
1, 2, 3, 9, 11, 12, 14, 15), with the weighting on the sensors
leading the bid to pay more attention to traffic in the dominant
direction but to yield the light if that traffic is less than seen on
the non-dominant direction sensors.
Interestingly, as with many of the optimized Mountain View
lights, most of the protected-left phases were effectively removed from the light sequence. In most cases, this was done by
having the through-traffic phases with sensor weights that are
“matched opposites” (that is, all the sensor weights for phase 3
are the negative of the weights for phase 1, in Figure 2). Here,
the protected-left can be effectively removed by omitting sensors from that phase (lights 1, 3, 4, 7, 9, 16, 19) or by matching
the sensor weights for the protected left to the previous throughtraffic phase (lights 2, 3, 7, 9, 15, 16, 19).7
(8601 cars)
(13360 cars)
(8362 cars)
(15965 cars)
demonstrated that with auction-based controllers, the improvements in capacity and mean travel times in both cities were extremely promising. Perhaps equally as interesting as the quantitative results are the common ways in which the learning procedures worked with the auction-based controllers in finding
similar strategies for improving traffic in both cities: eliminating protected lefts, weighting bids higher from dominant roads,
and even negatively weighting bids from less traveled roads.
These strategies were automatically determined; none were preprogrammed.
The second contribution of this work is a method through
which the program schedules of currently installed traffic light
can be approximated. This addresses an important practical issue in measuring the benefits of new traffic light controllers. For
the traffic researcher, two classes of crucial data have become
increasingly available. First, detailed maps of the streets and
the precise locations of the traffic lights is publicly available
through a number of sources. Second, through the increased
usage of personal cell-phone based GPS systems, an enormous
amount of travel-tracks have been amassed. What is often lacking, however, is detailed knowledge of the existing traffic light
schedules and traffic light response behaviors. This paper has
presented a simulation-based approach to approximate the behavior of installed lights. This provides a method to create a
solid baseline from which to quantitatively measure improvements.
There are several areas of future research that are natural
next steps. First, with respect to our procedures used to determine the schedules of the currently deployed lights, in this
study, we only considered matching the timings of static-lights.
Nonetheless, it should readily possible to adapt the same methods presented here to learning the parameters for lights that incorporate induction loop sensors. We suspect that this will provide improvements even in the scenarios considered here.
Second, in this study, the planning-based lights did not perform as well as the auction-based system. In the future, if
planning-based lights are reexamined, further parameterization
of the sensor placement and readings should be considered, despite the potential for a priori setting weights based on physi-
8. Conclusions and Future Work
There are two primary contributions of this work. First, we
have introduced the micro-auction based traffic light controllers
and a simple learning/optimization procedure to make them effective. The controllers do not necessitate communication between lights or between cars and lights; only local induction
loops sensors are used. Rather than coming up with ad hoc rules
for traffic priorities and scheduling, each sensor’s information is
placed within the framework of a micro-auction system. When
a phase change is permitted, the light controller collects bids
from all the phases and conducts a real-time micro-auction.
We tested the controllers on real traffic loads observed over
months of traffic collection. In Mountain View, California, the
load was dominated by the confluence of a freeway exit ramp
with a congested arterial road. Next, we studied real traffic
in Chicago, Illinois, to determine the controller’s performance
in within-city grids. For both cities, we employed the exact
same method to estimate the performance of the deployed traffic lights with no changes to the algorithms. We successfully
7 Lights 3, 7, 9, 16, and 19 appear on both of these lists since their two
protected-left phases used both of the strategies.
17
cal constraints. This may require new, more complex learning
algorithms that explicitly model the interactions between parameters; many have been explored in genetic algorithm literature [30, 32, 31], among other places.
Third, a large amount of research has been conducted towards lights that communicate between each other and with
cars. If this form of communication is available, it can be easily
incorporated into auction based models. In the simplest manner, the induction loop sensors can be replaced with counters
revealing the presence of cars through communication channels. However, more interesting interactions are also possible.
If there is higher priority traffic (for example emergency vehicles or lanes that need right-of-way for event traffic management), in an auction based system their weights can be appropriately accounted for in a phase’s bidding process to reflect the
urgency of travel.
The focus of this paper was not in exposing the auction mechanics externally, but rather using the auction as a guiding
principle for internal light controls. In the future, if lights do
communicate with cars, it is conceptually easy to incorporate
tolls and payment priorities into our auction process. If each
car has the ability to virtually “bid” on how much it is willing
to pay to make it through the lights, the auction-based mechanisms support this by altering the weights of the bids that are
received into the micro-auction. Of course, setting the weights
and the algorithms to be fair will require careful consideration;
nonetheless, the mechanisms to support monetary bidding are
provided through these learned auctions-based controllers.
[9] V. Gradinescu, C. Gorgorin, R. Diaconescu, V. Cristea, L. Iftode, Adaptive traffic lights using car-to-car communication, in: Vehicular Technology Conference, 2007. VTC2007-Spring. IEEE 65th, IEEE, 2007, pp.
21–25.
[10] T. Tielert, M. Killat, H. Hartenstein, R. Luz, S. Hausberger, T. Benz, The
impact of traffic-light-to-vehicle communication on fuel consumption and
emissions, in: Internet of Things (IOT), 2010, IEEE, 2010, pp. 1–8.
[11] B. Asadi, A. Vahidi, Predictive cruise control: Utilizing upcoming traffic
signal information for improving fuel economy and reducing trip time,
Control Systems Technology, IEEE Transactions on 19 (3) (2011) 707–
714.
[12] S. F. Smith, G. J. Barlow, X.-F. Xie, Z. B. Rubinstein, Surtrac: Scalable
urban traffic control, in: Transportation Research Board Annual Meeting,
2013.
[13] X.-F. Xie, S. F. Smith, G. J. Barlow, Schedule-driven coordination for
real-time traffic network control, in: International Conference on Automated Planning and Scheduling, 2012.
[14] X.-F. Xie, S. F. Smith, L. Lu, G. J. Barlow, Schedule-driven intersection
control, Transportation Research Part C 24.
[15] K. Dresner, P. Stone, Traffic intersections of the future, in: AAAI-2006,
2006, pp. 1593–1596.
[16] G. Heisig, Nervousness in material requirements planning systems, in:
Planning Stability in Material Requirements Planning Systems, Springer,
2002, pp. 21–64.
[17] D. Christensen, D. Newcomb, Evaluation of induction loop installation
procedures in flexible pavements, in: WA-RD 76.1, Washington State
Dept. of Tranportation, 1985.
[18] M. Covell, S. Baluja, R. Sukthankar, Micro-auction based traffic light
control: Responsive, local decision making, in: IEEE Intelligent Transportation Systems Conference-2015, 2015.
[19] S. Baluja, M. Covell, R. Sukthankar, Approximating the effects of installed traffic lights:a behaviorist approach based on travel tracks, in:
IEEE Intelligent Transportation Systems Conference-2015, 2015.
[20] D. Carlino, S. D. Boyles, P. Stone, Auction-based autonomous intersection management, in: 16th Intelligent Transportation Systems-(ITSC),
2013, IEEE, 2013, pp. 529–534.
[21] H. Schepperle, K. Böhm, Agent-based traffic control using auctions, in:
International Workshop on Cooperative Information Agents, Springer,
2007, pp. 119–133.
[22] H. Schepperle, K. Böhm, S. Forster, Traffic management based on negotiations between vehicles–a feasibility demonstration using agents, in:
Agent-Mediated Electronic Commerce and Trading Agent Design and
Analysis, Springer, 2008, pp. 90–104.
[23] D. Krajzewicz, J. Erdmann, M. Behrisch, L. Bieker, Recent development
and applications of sumo - simulation of urban mobility, International
Journal On Advances in Systems and Measurements 5.
[24] U.S. Department of Transportation Federal Highway Administration,
Traffic signal timing manual: Chapter 5 basic signal timing procedure and controller parameters, http://ops.fhwa.dot.gov/publications/fhwahop08024/chapter5.htm (2013).
[25] DLR Institute of Transportation, Sumo - simulation of urban mobility, http://www.dlr.de/ts/en/desktopdefault.aspx/tabid-9883/16931 read41000 (2015).
[26] O. K. Tonguz, W. Viriyasitavat, F. Bai, Modeling urban traffic: A cellular
automata approach, IEEE Communications Magazine.
[27] P. V. Thankuriah, D. G. Geers, Transportation and Information: Trends in
Technology and Policy, Springer Science & Business Media, 2013.
[28] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley Longman Publishing Co., Inc., 1989.
[29] S. Baluja, Genetic algorithms and explicit search statistics, Advances in
Neural Information Processing Systems (1997) 319–325.
[30] S. Baluja, S. Davies, Fast probabilistic modeling for combinatorial optimization, in: AAAI/IAAI, 1998, pp. 469–476.
[31] G. Harik, Linkage learning via probabilistic modeling in the ecga, Urbana
51 (61) (1999) 801.
[32] M. Pelikan, D. E. Goldberg, F. G. Lobo, A survey of optimization by
building and using probabilistic models, Computational optimization and
applications 21 (1) (2002) 5–20.
[33] C. Chow, C. Liu, Approximating discrete probability distributions with
dependence trees, IEEE transactions on Information Theory 14 (3) (1968)
462–467.
Acknowledgments
The authors would like to gratefully acknowledge Kevin
Wang for his helpful comments on this paper and many aspects
of this project over the last year.
References
[1] J. J. Sanchez, M. Galan, E. Rubio, Genetic algorithms and cellular
automata: a new architecture for traffic light cycles optimization, in:
Congress on Evolutionary Computation, Vol. 2, IEEE, 2004, pp. 1668–
1674.
[2] B. Park, C. J. Messer, T. Urbanik II, Enhanced genetic algorithm for
signal-timing optimization of oversaturated intersections, Transportation
Research Record: Journal of the Transportation Research Board 1727 (1)
(2000) 32–41.
[3] A. M. Turky, M. S. Ahmad, M. Z. M. Yusoff, The use of genetic algorithm
for traffic light and pedestrian crossing control, Int. J. Comput. Sci. Netw.
Security 9 (2) (2009) 88–96.
[4] T. Kalganova, G. Russell, A. Cumming, Multiple traffic signal control
using a genetic algorithm, in: Artificial Neural Nets and Genetic Algorithms, Springer, 1999, pp. 220–228.
[5] M. Wiering, Multi-agent reinforcement learning for traffic light control,
in: ICML, 2000, pp. 1151–1158.
[6] I. Arel, C. Liu, T. Urbanik, A. Kohls, Reinforcement learning-based
multi-agent system for network traffic signal control, Intelligent Transport Systems, IET 4 (2) (2010) 128–135.
[7] D. E. Moriarty, P. Langley, Learning cooperative lane selection strategies
for highways, in: AAAI/IAAI,1998, 1998, pp. 684–691.
[8] B. Abdulhai, R. Pringle, G. J. Karakoulas, Reinforcement learning for
true adaptive traffic signal control, Journal of Transportation Engineering
129 (3) (2003) 278–285.
18
[34] A. Juels, M. Wattenberg, Stochastic hillclimbing as a baseline method
for evaluating genetic algorithms, in: Advances in Neural Information
Processing Systems (NIPS), 1995.
[35] M. Mitchell, J. H. Holland, S. Forrest, When will a genetic algorithm
outperform hill climbing, in: J. D. Cowan, G. Tesauro, J. Alspector
(Eds.), Advances in Neural Information Processing Systems 6, MorganKaufmann, 1994, pp. 51–58.
[36] M. Harman, P. McMinn, A theoretical & empirical analysis of evolutionary testing and hill climbing for structural test data generation, in: Proceedings of the 2007 international symposium on Software testing and
analysis, ACM, 2007, pp. 73–83.
[37] A. Rosete-Suárez, A. Ochoa-Rodrıguez, M. Sebag, Automatic graph
drawing and stochastic hill climbing, in: Proceedings of the Genetic and
Evolutionary Computation Conference, Vol. 2, 1999, pp. 1699–1706.
[38] M. Haklay, P. Weber, Openstreetmap: User-generated street maps, IEEE
Pervasive Computing 7.
[39] D. Barth, The bright side of sitting in traffic: Crowdsourcing road congestion data, http://googleblog.blogspot.com/2009/08/bright-side-of-sittingin-traffic.html (2009).
[40] A. Dobie, Understanding Google’s Android location tracking,
http://www.androidcentral.com/understanding-googles-android-locationtracking (2014).
[41] City of Mountain View and Caltrans’ Metropolitan Transportation Commission, Program for arterial system synchronization (pass) fy12/13
shoreline blvd traffic signal timing project, http://files.mtc.ca.gov/pdf/pass/2012-13 Factsheets/City of Mountain View Caltrans.pdf (2013).
[42] T. Vanderbilt, Traffic: Why We Drive the Way We Do (and what it Says
about Us), Vintage Books, 2009.
19
| 3cs.SY
|
Map-aided Dead-reckoning — A Study on
Locational Privacy in Insurance Telematics
arXiv:1611.07910v1 [cs.CR] 14 Nov 2016
Johan Wahlström, Isaac Skog, Member, IEEE, João G. P. Rodrigues, Student Member, IEEE, Peter Händel, Senior
Member, IEEE, and Ana Aguiar, Member, IEEE
Abstract—We present a particle-based framework for estimating the position of a vehicle using map information and measurements of speed. Two measurement functions are considered.
The first is based on the assumption that the lateral force on
the vehicle does not exceed critical limits derived from physical
constraints. The second is based on the assumption that the
driver approaches a target speed derived from the speed limits
along the upcoming trajectory. Performance evaluations of the
proposed method indicate that end destinations often can be
estimated with an accuracy in the order of 100 [m]. These results
expose the sensitivity and commercial value of data collected in
many of today’s insurance telematics programs, and thereby have
privacy implications for millions of policyholders. We end by
discussing the strengths and weaknesses of different methods for
anonymization and privacy preservation in telematics programs.
Index Terms—Vehicle positioning, dead-reckoning, insurance
telematics, map-matching, usage-based-insurance, privacy.
I. I NTRODUCTION
The emergence of connected vehicles has opened up possibilities for services which combine the power of cloud
computing with the expanding capabilities of modern vehicle
technology. Connected vehicles not only enable remote analysis of vehicle condition and driving behavior, but also allow
drivers to benefit from a multitude of convenience-related
applications such as theft tracking and accident detection.
However, as the amounts of generated data and the means
of connectivity increases, so do the dangers related to privacy
and security [1]. Recently, connected vehicles and associated
aftermarket devices have been shown to be vulnerable to
wireless hacking via a broad range of attack vectors [2]–[4]. In
the worst case, the hacker will be able to remotely manipulate
any of the vehicle’s electronic control units (ECUs). Moreover,
privacy concerns have been raised about the large databases
of sensitive data collected from vehicle-installed telematics
units [5]–[8]. As illustrated in Fig. 1, a telematics insurer uses
data on driving behavior to set the premium offered to each
policyholder [9]–[11]. Typically, the risk profile of each driver
will be based on e.g., the number of speeding or harsh braking
events, as detected using measurements from vehicle-installed
black boxes or from the vehicle’s on-board-diagnostics (OBD)
system [12], [13]. This study examines the locational privacy
of such insurance telematics programs.
J. Wahlström, I. Skog, and P. Händel are with the ACCESS Linnaeus Center,
Dept. of Signal Processing, KTH Royal Institute of Technology, Stockholm,
Sweden. (e-mail: {jwahlst, skog, ph}@kth.se).
J. G. P. Rodrigues and A. Aguiar are with the Instituto de Telecomunicações,
Departamento de Engenharia Eletrotécnica e de Computadores, Faculdade
de Engenharia da Universidade do Porto, 4200-465 Porto, Portugal. (e-mail:
{joao.g.p.rodrigues, anaa}@fe.up.pt)
sk
Insurer
Individual
premium
sk
sk
Fig. 1. Several insurance telematics providers are today collecting measurements of speed, here denoted by sk , from vehicles’ wheel speed sensors to
adjust premiums and provide feedback to drivers.
A. Locational Privacy
The digital revolution has made violations of locational
privacy both cheaper and harder to detect [14]. Temporally
isolated digital position inferences were first made possible
by the widespread usage of credit cards and access cards.
Today, the abundance of e.g., global navigation satellite system (GNSS) receivers in smartphones and vehicles enables
continuous streams of position measurements to be sent to
central servers [15], [16]. In addition, it is often possible
to perform smartphone-based WiFi-localization by exploiting
system permissions required by many popular apps [17].
Although some might argue that consumers are willingly
giving up personal information in exchange for services and
products, the actual privacy implications of sharing specific
data are often obscured. This is especially true when the
data itself does not include position measurements, but rather,
information which in combination with additional databases
can be used for positional inference.
Initially, privacy and security were not major concerns in the
insurance telematics industry, and the amount of data collected
was primarily determined by restrictions on transmission and
storage. However, these topics have received more attention
as the industry has matured, and privacy concerns have even
been cited as a contributing reason to suspend telematics trials
[18]–[21]. Today, several telematics providers only collect
measurements of speed. When responding to privacy concerns,
these companies will refer to the fact that no position data is
recorded. However, as shown in [7], location-based information can still be extracted from many trips.
B. Contributions and Outline
In this article, we use map-aided dead-reckoning (DR) to
geographically locate a vehicle purely based on its speed
profile and digital map information. Since the study utilizes
the same type of information that is available for many of the
current insurance telematics providers, the results have privacy
implications for millions of policyholders. The estimation
is carried out using a particle filter, and thereby parallels
previous filtering-based approaches where measurements of
the vehicle’s yaw rate have been available [22]–[25]. Though
the idea of estimating a vehicle’s position using only speed
measurements and map information is not new [7], [8], this
is the first time that it has been formulated as a filtering
problem. In addition to making it possible to motivate the
implementation using optimality arguments, our hope is that
this also will make the method more accessible, and facilitate
extensions based on established methods for particle filtering.
Previous algorithms have been very reliant on the matching
of vehicle stops in the speed data with intersections or traffic
signals in the map. For example, in [8], data where the vehicle
made a large number of stops due to traffic congestion or
similar was removed altogether since this type of driving
was not accounted for in the algorithm. In this study, we
avoid these issues by processing the speed measurements oneby-one, rather than in segments delimited by vehicle stops.
Moreover, the performance evaluations are made on data sets
with significantly longer trip lengths than in previous studies.
Section II reviews the state-of-the-art in map-aided deadreckoning and discusses how the problem is altered depending
on what sensor information is available. The employed system
model and the filter implementation are detailed in Section
III. Section IV illustrates the performance characteristics of
the proposed method by studying the accuracy of estimated
end positions and the dependence on a priori knowledge
of the initial position. The results indicate that GNSS or
OBD measurements of speed often are sufficient to extract a
substantial amount personal information on daily activities and
location-based behavior. Section V discusses the implications
of the presented results and possible ways forward for privacyaware telematics providers. Finally, the article is concluded in
Section VI.
II. P RELIMINARIES ON M AP - AIDED D EAD - RECKONING
DR refers to the process of recursively estimating the
position of an object by numerically integrating measurements or estimates of velocity. The general concept has been
widely used to navigate e.g., aircrafts, ships, submarines,
robots, automobiles, and pedestrians [26]. Obviously, perfect
integration requires perfect knowledge of the velocity at all
time instances. In practice, this means that stand-alone deadreckoning systems always will be subject to position errors
that accumulate with time. To prevent the position error from
growing without bound, additional information is required.
This type of information can be provided by, e.g., satellitebased positioning systems [12], WiFi fingerprinting systems
[27], magnetic fingerprinting systems [28], terrain measurements [29], or road maps.
A. Map-aided DR with Sensor-based Yaw Information
DR aided only by information from road maps (and some
a priori knowledge of the initial position) have been the
topic of several studies within land-vehicle navigation [30].
Historically, the primary motivation of these studies have been
to increase the accuracy and integrity of existing low-cost
navigation systems during failures or performance deteriorations of the employed positioning systems (e.g., during GNSS
outages). Assuming that measurements of both left and right
wheel speeds are available from the controller area network
(CAN) bus, estimates of the vehicle’s longitudinal velocity and
yaw rate can be obtained by means of differential geometry
[22]. These estimates are then used for DR. If only a scalar
speed measurement is available at each sampling instance,
measurements of the yaw angle or yaw rate can instead be
obtained from the steering wheel angle sensor, from a vehiclemounted gyroscope, or from a magnetometer [31]. The error
growth of the navigation solution is typically mitigated by
comparing the current position and yaw estimates with the
corresponding quantities at the nearest map segment [23],
[32]. To avoid the logistics of accessing measurements of
wheel speed, DR can instead be performed by means of
inertial navigation. However, this will increase the number
of integration steps, and thereby also the rate of the position
error growth. As a result, the position estimates will be very
sensitive to e.g., mounting misalignments and sensor biases. In
general, the performance of inertial measurement unit (IMU)based methods can be expected to be inferior to that of
odometric methods [24].
B. Map-aided DR without Sensor-based Yaw Information
Since the standard parameters available from the vehicle’s
OBD system do not include any information on the vehicle’s
yaw rate or yaw angle, it is interesting to see what level
of positioning accuracy that can be obtained by using only
OBD measurements of speed. (As previously mentioned, the
industry of insurance telematics today includes several actors, e.g., Progressive, Geico, and Allstate, that collect speed
measurements from their policyholders [33], [34].) Related
studies were conducted in [7] and [8], which estimated the
position of a vehicle by matching intersections in the map
with detected vehicle stops in the speed measurements. A
list of possible map traces was continuously updated by
adding new traces after each vehicle stop, and deleting traces
deemed infeasible after 1) comparing their lengths with the
corresponding distance implied by the speed measurements;
2) comparing the measured speeds with speed limits; and 3)
comparing the measured speeds with the maximum speeds
as implied by the curvature of the road. The estimation was
not statistically motivated and was not formulated within any
established estimation framework. In this study, we aim to fill
this gap.
The DR problem arising when only speed measurements
are available is fundamentally different from the DR problem
when also sensor-based yaw information is available. In the
latter case, the position estimate can be propagated in two dimensions using only sensor measurements. Spatial information
from the road map is then conveniently incorporated into the
estimation as soft constraints. Usually, this information will be
sufficient to avoid filter divergence, and additional information,
Possible end positions
n(2)
0 [s]
30 [s]
60 [s]
90 [s]
n(N )
rk+1
Latitude
n(1)
rk+1
rk+1
rk
Longitude
Fig. 2. Possible positions 0 [s], 30 [s], 60 [s], and 90 [s] after initializing
DR (using only measurements of speed and map information) from a known
starting point. The example is taken from a trip in Porto, Portugal, and the
figure shows an area of about 3.5 [(km)2 ].
such as speed limits, will be of secondary importance [22],
[23]. When only measurements of speed are available, the
position estimate cannot be propagated solely based on sensor
measurements. Instead, the vehicle’s yaw angle has to be
estimated from the road map. In this case, sensor measurements and spatial information from the road map will never be
sufficient as the posterior distribution will continue to ”spread
out” at intersections (see Fig. 2). Additional models which
relate the expected vehicle speed to the vehicle’s position on
the map are required, and speed limits will become of utmost
importance.
III. M ODEL AND ESTIMATION FRAMEWORK
We present a filtering-based approach to the problem of estimating a vehicle’s position using map information and measurements of speed. This section describes both the employed
system model and the particle-based filter implementation.
A. Overview of System Model
The map data from OpenStreetMap (OSM) can be described
as a mathematical graph where each vertex (node) defines a
latitude-longitude pair. The connectivity of the graph is defined
by a set of ways, i.e., ordered lists of nodes indicating that
there is an edge (link) between the first and the second node,
between the second and the third node, etc. Each way can be
characterized by attributes specifying road type, speed limits,
restrictions on travel direction, etc. The problem at hand is
to use information from a digital map such as OSM, together
with measurements of speed, to estimate a vehicle’s position.
The considered estimation framework is based on the model
xk+1 ∼ p(xk+1 |xk , sk ),
y(S) ∼ p(y(S)|X).
(1a)
(1b)
Here, x is a state vector comprising the vehicle’s twodimensional position r, and the vehicle’s direction of travel.
Assuming that the vehicle always travels along links in the
graph, the state vector can be given as an ordered list of two
connected nodes and a constant indicating the ratio of the
vehicle’s distance to the first node and the distance between the
Fig. 3. Illustration of the process model when the vehicle passes a road
junction. If the probability of the initial position rk is p, the probability of
each of the updated positions rk+1 are p/N where N denotes the number
of links along which rk+1 can be expected. The distance between rk and
(any) rk+1 , measured along the links, is equal to ∆tk sk .
two nodes. The vehicle’s speed is denoted by s (no notational
distinction is made between the true and measured speed). As
can be seen, the speed measurements act both as input in the
state equation (1a) and as measurements in the measurement
equation (1b). Moreover, p(·|·) is used to denote a conditional
probability density function (pdf). All pdfs are implicitly
conditioned on the available map information. Throughout the
paper, we will follow the convention of using a subindex k to
denote quantities at sampling instance k. The complete set of
∆
state vectors and speed measurements are then given by X =
∆
{..., xk−1 , xk , xk+1 , ...} and S = {..., sk−1 , sk , sk+1 , ...},
respectively. The employed measurements are collected in a
generic function y, and will be discussed more thoroughly
in Section III-C. For clarity, we will in the text differentiate
between s and y by referring to speed measurements and
measurements, respectively.
B. Process Model
The process model (1a) describes the time development of
the vehicle’s position, using speed as input. The update is
performed by distributing the travelled distance ∆tk sk along
the upcoming links. We have here used ∆tk to denote the
sampling interval between sampling instances k and k + 1.
When the vehicle passes a road junction, it is modeled to
continue along any of the connected links (excluding the link
it has just traversed) with equal probability. This is illustrated
in Fig. 3. As may be realized, less probable events such as unexpected u-turns have been disregarded. If needed, the model
can easily be modified to account for errors in the estimated
travelled distance. Discrepancies between the estimated traveled distance ∆tk sk and the true travelled distance can arise
due to map errors and off-road driving [25]. In addition, GNSS
speed measurements are subject to multipath propagation [35],
whereas OBD measurements are affected by wheel slips,
quantization errors in the measured speed [36], and scale factor
errors dependent on tire pressure [22]. Previous attempts to
include the scale factor in the state vector did not result in any
improvement in navigation performance, but rather, increased
the risk of filter divergence [23], [25]. We further note that the
speed measured by the OBD system is the absolute value of
the vehicle’s three-dimensional velocity, whereas the studied
model only considers two spatial dimensions (most nodes in
OSM do not include any altitude information). As a result, the
travelled distance in the two-dimensional plane of the map will
tend to be overestimated when travelling along steep inclines
[23]. For example, if the vehicle is traveling with a speed of
25 [m/s] along a road where the angle of inclination is 5 [ ◦ ],
the speed in the two-dimensional plane of the map will be
25 · cos(5 [ ◦ ]) [m/s] ≈ 24.62 [m/s]. This error contribution
can be compensated for by using a commercial digital map
with altitude information and extending the dimension of the
state vector accordingly.
Obviously, the estimated travelled distance ∆tk sk is subject
to several modeling and measurement errors. Despite this, the
resulting accumulated position error will in many cases be
negligible in comparison to the position error resulting from
the uncertainty of the vehicle path, i.e., the list of nodes or
links that the vehicle has traversed.
of y (1) is heuristically modeled as
p(y (1) (sk , sk+1 )|x1 , ..., xk )
1,
y (1) < g (1)
(1)
y − g (2)
(1)
≤ y (1) < g (2) .
∝
(1)
(2) , g
g
−
g
0,
g (2) ≤ y (2)
The design parameters g (1) and g (2) are chosen so that the
lateral forces fall below g (1) during the larger part of the
driving while allowing for outliers giving estimated lateral
forces up to g (2) .
To derive the second type of measurements, we note that
drivers tend to maintain a speed close to the speed limit for
the larger part of the route. Hence, we propose the use of the
measurements
∆
y (2) (sk , sk+1 ) =
ak + csk
C. Measurement Model
When propagating the state distribution as described in the
preceding subsection, the number of feasible vehicle paths
grows exponentially as a function of the traveled distance
[7]. To bound the number of paths, measurement updates
must be employed. We will give two examples of possible
measurements. First, we the study the lateral forces exerted
on the vehicle. When cornering, the lateral g-forces on the
vehicle at sampling instance k can be approximated by [37]
q
∆
a2k + s2k ωk2
(2)
y (1) (sk , sk+1 ) =
where a denotes the vehicle’s longitudinal acceleration and ω
denotes the vehicle’s yaw rate. The acceleration is approximated by the difference quotient
∆ sk+1 − sk
ak =
.
(3)
∆tk
The yaw rate ω must be estimated from map data. We choose
to estimate the yaw rate in a separate Kalman filter based on
the state-space model
(5)
(6)
distributed according to
p(y (2) (sk , sk+1 )|x1 , ..., xk+K ) = pC (y (2) , c s̄k+K , σ)
(7)
where K ∈ N and pC denotes the pdf of the Cauchy
distribution with location and scale parameters given by the
second and third arguments, respectively. The probabilistic
assumption in (7) is derived from the model
ak = c(s̄k+K − sk ) + sk
(8)
which says that the vehicle’s acceleration will be approximately proportional to the difference between the current
speed and the target speed s̄k+K . As should be obvious, we
have assumed that sk is Cauchy distributed with location 0
and scale σ. The design parameters c and σ describe the
characteristics of each individual driver and car. Modeling the
error as distributed according to the the heavy-tailed Cauchy
distribution will make the estimation robust to measurement
errors caused by unexpected driving behavior.
The target speed is, just as the yaw rate, estimated in a
separate Kalman filter. In this case, the state-space model is
θk+1 = θk + ∆tk ωk ,
(4a)
ωk+1 = ωk + ηkω ,
(4b)
s̄k+1 = s̄k + ηks .
(9a)
(4c)
sk .
(9b)
θ̃k = θk +
ηkω
θk
θk .
Here, the noises
and
are assumed to be white with
variances ∆tk σω2 and σθ2 , respectively. Further, θ is the vehicle’s yaw angle, and θ̃ is the direction of the current link.
Due to the periodicity of the yaw angle, the measurement
residuals must be chosen within the interval [−180, 180] [ ◦ ].
An alternative to using a Kalman filter is to estimate the yaw
rate directly from the difference quotients (θ̃k+1 − θ̃k )/∆tk .
However, due to the limited resolution of the nodes in OSM
(many sharp corners will have links meeting at a 90 [ ◦ ] angle)
these difference quotients will be fairly unstable.
Typically, passenger cars are not designed for g-forces
higher than 0.8 g, and will seldom exceed 0.6 g during normal
driving. (A sufficiently high lateral force will cause the vehicle
to rollover [37].) Here, g denotes the gravitational acceleration
at the surface of the earth. Motivated by this, the distribution
s̃k = s̄k +
2
Here, the noises ηks and sk are white with variances ∆tk σs1
2
and σs2 , respectively, while s̃ is the speed limit of the current
link. By letting the target speed be dependent on the speed
limits on the road ahead, we are able to capture the forwardlooking behavior of typical driving. At each iteration, the
estimated target speed is projected to the closest speed in the
space
q
(10)
{sk : a2k + s2k ωk2 < g (1) }.
In other words, the estimated target speed is adjusted so as
to never imply an excessively large lateral force. This means
that particles which pass sharp corners while the vehicle slows
down often will have their weights increased, and hence, we
can detect the vehicle turning even though all nearby links
have the same speed limit.
Algorithm 1 : Particle-filter algorithm.
1: for k = 0, 1, . . . do
2:
Time update: Update each particle state xk and its
associated particle weight wk by distributing the travelled
distance ∆tk sk along the upcoming links.
3:
Parameter update: Perform a time update in the
Kalman filters estimating the yaw angle, the yaw rate,
and the target speed for each particle. Use the travel
direction and the speed limit at each particle to perform
the associated measurement updates.
4:
Measurement update: Update the particle weights according to equation (11). The estimated yaw rate is used
in the updates based on both y (1) and y (2) , whereas the
estimated target speed only is needed in the update based
on y (2) . Set the weights of all particles which violate
constraints on travel directions to zero.
5:
Particle weight redistribution: Redistribute the prior
weight of particles which have had their weight set to
zero in the measurement update to particle ”siblings” (if
such siblings exists).
6:
Particle elimination: Eliminate all particles with zero
weight. If the total number of particles exceeds a given
threshold, eliminate all particles with insufficient weight.
7:
Particle resampling: If k = n · N0 for some n ∈ N and
some integer parameter N0 , and if the number of particles
is sufficiently low, resample particles using systematic resampling. Displace each sampled particle along its current
link by sampling from a uniformly distributed distance.
Re-initialize the particles so that all particle weights are
equal.
8:
Particle merging: Merge particles that are sufficiently
similar.
9: end for
The algorithm is detailed in Sections III-B, III-C, and III-D.
As an alternative to using the model (8), it is possible
to derive a measurement function based on the assumption
that the current speed is close to the target speed. This
model is obtained as a special case of (8) in the limit of
s̄k+K = sk . However, this model tends to be too restrictive on
discrepancies between the vehicle speed and the target speed,
and often causes the measurement errors to have a significant
temporal correlation (e.g., if the vehicle’s speed exceeds the
target speed at one sampling instance it is very likely to do so
also at the next sampling instance). By using (8), we allow for
temporally correlated discrepancies between the vehicle speed
and the target speed, but still expect the speed to be at least
approaching the target speed.
Both types of measurements y (1) and y (2) are based on
ideas similar to those in [7]. Additional measurements can be
derived from the fact that vehicle stops tend to coincide with
traffic lights or stop signs in the map. This information was
not utilized here as the number of unpredictable stops (e.g.,
stops due to congestion) was very large, and information on
traffic lights and stop signs was lacking over the larger part
of the map area.
Since all speed measurements are used both in the process
model (1a) and in the measurements (2) and (6) one might
argue that the system model includes several error correlations
that must be considered in the estimation. However, the
larger part of the variance in y (1) (sk , sk+1 )|x1 , ..., xk and
y (2) (sk , sk+1 )|x1 , ..., xk+K stem from variations in driving
behavior, rather than from errors in the measured speed (for
example, the absolute OBD measurement error is typically
smaller than 1 [km/h], while the speed at a given position
often can differ by 20 [km/h] depending on the traffic conditions). As a result, the correlations caused by errors in the
measured speed are generally negligible.
D. Particle Filter Implementation
Due to the multimodal characteristics of the conditional
pdfs in the system model (1), assuming unimodal posterior
distributions will typically lead to irrecoverable localization
errors in a very short period of time. To avoid this, we
propose a solution based on the particle filter framework [38].
Each particle must, in addition to the current state vector
xk , also store its estimates of the yaw angle, the yaw rate,
and the target speed. However, as the state-space models for
these parameters are time-invariant (up to variations in the
sampling interval), the covariance matrices only need to be
stored once. The particle propagation is performed according
to the model described in Section III-B. Instead of sampling a
single continuing path when a particle passes a road junction
(i.e., ignoring all but one of the possible links), the particle is
split into several new particles with equal weights (the sum of
which is equal to the weight of the original particle). In this
way, we increase the particle diversity and decrease the risk
of filter divergence due to unlucky sampling. Obviously, this
also means that the number of particles in many cases will
increase during the propagation step.
Aside from the resulting increase in computational cost, the
choice of ”splitting” particles as they pass a road junction
has an additional downside. This becomes clear when the true
path of the vehicle passes a large number of road junctions.
In this case, a particle following this path will have its weight
diminished with each passing road junction. If the weight of
the particle is not to fall short of resampling thresholds, it
must then increase substantially during measurement updates.
In practice, this means that particles often will be eliminated
mainly because they passed many road junctions. To mitigate
this effect, we once again utilize the fact that the sum of the
weights of particles emerging from a particle passing a road
junction should be equal to the weight of the original particle.
Hence, if a particle which has emerged from a ”particle
split” at a road junction is immediately eliminated due to an
estimated lateral force exceeding g (2) , its weight, prior to the
elimination, is distributed to its ”siblings”, i.e., other particles
which originated from the same road junction and particle. The
same rules are applied to particles which violate constraints
on travel directions, i.e., which travel in the wrong direction
on a one-way link. Implementing this modification will tend
to reduce the risk of filter divergence when the vehicle travels
along a main road with a large number of intersections.
As motivated in the last paragraph in Section III-B, the particle propagation is performed without any simulated process
TABLE I
FILTER PARAMETERS .
Parameter
σω
σs1
Measurement noise σθ
σs2
(1)
y update
g (1)
g (2)
(2)
y update
K
c
σ
Particle threshold (resampling trigger)†
Weight threshold (elimination)
Distance threshold (particle merge)
Resampling displacement interval
Process noise
†
Value
√
5 [ ◦ /s2 /√Hz ]
0.5 [m/s2 / Hz ]
15 [ ◦ ]
10 [m/s]
0.55 g
0.65 g
12
0.05 [1/s]
1.5 [m/s2 ]
100
1/200
20 [m]
(−100, 100) [m]
Checked every hundredth second.
noise in the travelled distance. Similar modeling choices have
previously been discussed in [39] and references therein.
In the measurement update, each particle weight wk is
updated according to
wk+1 ∝ p(yk+1 (S)|X)wk
(11)
where the total set of measurements are
(1)
|
∆
yk+1 (S) =
y (sk , sk+1 ) y (2) (sk−K , sk+1−K ) . (12)
Note that the measurements y (2) use earlier speed measurements than y (1) as the conditional distribution of
y (2) (sk−K , sk+1−K ) is modeled as dependent on the target
speed at sampling instance k.
The filter is implemented with systematic resampling [40].
New particles are resampled at equidistant points in time
given that the number of particles is sufficiently low. Each
new particle is displaced with a uniformly distributed distance
along its trajectory. This will make it possible to detect any
eventual accumulated errors in the total travelled distance.
Particles with insufficient weight are eliminated whenever the
total number of particles exceed a given threshold. However,
particles which are given a zero weight are eliminated immediately. In addition, whenever two particles are sufficiently
similar (i.e., are on the same link, have the same parameter
estimates, and have sufficiently similar position estimates) they
are merged into a single particle which will have its weight
equal to the sum of the original weights and its position equal
to the weighted average of the original positions. This will
ease the computational burden somewhat as well as mitigate
problems with particle clustering [22]. The particle filter is
summarized in Algorithm 1.
IV. F IELD STUDY
GNSS and OBD data were collected from five drivers
performing a total of eighteen different trips in both rural
and urban areas in and around Porto, Portugal. No specific
trip was conducted for the purpose of this study, and hence,
the data represents a random sample of typical routes in the
area (however, for practical reasons, the trips were somewhat
skewed towards more lengthy trips). All data was collected
using the SenseMyCity app, which was developed as part
of the Future Cities project [41]. The median and mean trip
lengths were 28.3 [km] and 49.5 [km], respectively. The update
date rate of the GNSS data was 1 [Hz], while the update rate
of the OBD data varied between 1 [Hz] and 3 [Hz] depending
on the car (the majority of the trips had an update rate close to
1 [Hz]). Using a dedicated setup collecting OBD speeds with
a higher update rate may improve the estimation performance.
The GNSS position measurements were used as ground truth.
Assuming knowledge of the vehicle’s initial position (this will
often be the driver’s home address, the driver´s work address,
etc.), the filter was initialized by sampling particles on all
links in a circle of radius 50 [m] centered at the true initial
position. Due to the stochastic nature of the filter, all displayed
performance measures have been averaged over 10 runs. The
parameters used in the field study are shown in Table I.
Occasionally, data was lost from either of the two sensors.
In these cases, the missing data was replaced by data from
the other sensor (GNSS replacing OBD or OBD replacing
GNSS). The replacement data was scaled to take the OBD
scale factor into account. The scale factor can be estimated
using data from sampling instances where measurements from
both sensors are available. While some GNSS outages always
can be expected, we stress that the larger part of the data
collection disruptions occurred in the OBD data and was
related to the chosen communications setup. In other words,
data losses of this kind should not be expected in e.g., a
commercial OBD-based insurance telematics program. In two
(five) of the trips, data losses caused the measurements of the
trip to start (end) at nonzero speeds. If the data starts or ends
close to the motorway, this will usually simplify the estimation
since filter divergence is more prone to occur along roads with
lower speed limits.
A. Results
The performance of the filter when using GNSS and OBD
measurements of speed are shown in Fig. 4 and Fig. 5, respectively. The figures display the empirical distribution functions
(edfs) of the horizontal kr̂−rk and relative horizontal position
errors kr̂ − rk/L of the estimated end position. Here, we have
used L to denote the driving length of the trip. If all particles
were eliminated while running the filter, the horizontal position
error was considered to be equal to the driving length of the
trip. The performance of two estimators are shown. These are
the maximum a posteriori (MAP) estimator, i.e., the position
of the particle with the largest weight, and the minimum mean
square error (MMSE) estimator, i.e., the weighted average
of all particle’s end points. In addition, we also display the
accuracy of the particle with the position estimate that is
closest to the true end position. This is called the best-particle
(BP). Obviously, this is not a realizable estimator as we can
only find the BP by assuming knowledge of the true end
position. However, we still believe the accuracy of the BP is
relevant as it gives an indication of the performance that can be
(a) Horizontal position error
0
10 0
10 1
edf
1
10 3
10 4
10 5
10
10
-3
10
-2
10
-1
10
0
10
1
(a) Horizontal position error
MAP
MMSE
BP
Average trip length
edf
1
10 1
10 2
10 3
10 4
10 5
kr̂ − rk [m]
edf
0.5
0
10 -5
(b) Relative horizontal position error
MAP
MMSE
BP
10 -4
10 3
10 2
10 1
200
400
600
800
1000
1200
1400
1600
1800
2000
Fig. 6. The empirical 0.1 and 0.25 quantiles of the horizontal position error
of the end position of the BP, as dependent on the radius of the circle in
which the initial particles were sampled.
-4
1
edf−1 (0.1) GNSS
edf−1 (0.1) OBD
(b) Relative horizontal position error
Fig. 4. The empirical distribution function of the horizontal and relative horizontal position errors of the end position. The filter used GNSS measurements
of speed.
0
10 0
10 4
Uncertainty radius [m]
kr̂ − rk/L
0.5
edf−1 (0.25) GNSS
edf−1 (0.25) OBD
kr̂ − rk [m]
MAP
MMSE
BP
0.5
0
10 -5
10 2
edf−1 [m]
0.5
Dependence on initial uncertainty
10 5
MAP
MMSE
BP
Average trip length
edf
1
10 -3
10 -2
10 -1
10 0
that even though we are not be able to accurately estimate
the end position of a trip, we might still be able to estimate
the overall direction of travel (e.g., whether the driver drove
towards or away from the city center).
The dependence on the initial uncertainty is illustrated in
Fig. 6 which shows the quantiles edf−1 (0.1) and edf−1 (0.25)
of the BP estimate as dependent on the uncertainty radius,
i.e., the radius of the circle in which the initial particles were
sampled. All other filter parameters were kept constant. Even
with an uncertainty radius of 1 [km], a fourth of the trips have
a horizontal position error below 600 [m]. This demonstrates
the capability to extract positional information from speed
measurements even when there is significant uncertainty in
the inital position.
The estimated OBD scale factor had a sample mean and
sample standard deviation (over the studied trips) of 1.012 and
0.024, respectively. Despite this, the choice of using GNSS or
OBD measurements of speed seem to have a limited effect on
the filter performance.
10 1
kr̂ − rk/L
Fig. 5. The empirical distribution function of the horizontal and relative horizontal position errors of the end position. The filter used OBD measurements
of speed.
expected when additional information on likely end positions
are available (whereas the MAP and MMSE estimators give
an indication of the performance that can be expected when
processing trips independently without access to any additional
personal information). Many drivers tend to start and end most
of their trips at a very limited number of locations. Hence,
when a vehicle is detected to have stopped for an extended
period of time, the current particles can be evaluated based on
personal information, previously extracted driving habits, and
map information (for example, a driver will seldom stop in
the middle of the motorway). Under ideal circumstances, only
one particle will be considered to be a realistic end point.
As can be seen, the BP lies less than 200 [m] from the end
position in about a fourth of the trips, and lies less than 2 [km]
from the end position in about half of the trips. If the true
end position is e.g., the driver’s home, this level of accuracy
will often be sufficient to determine that the driver went home
during the trip. Similarly, the MAP estimator is accurate to
within approximately 200 [m] in a tenth of the trips. Moreover,
the errors of the MAP and MMSE estimators are about a fifth
of the total trip length in about half of the trips. This means
B. Filter Divergence
When no particle is closer than approximately 200 [m] to the
true position, the filter should be considered to have diverged.
In other words, the filter ultimately diverged in a majority
of the trips studied here. However, in many cases the the
filter did not diverge until late into the trip, and hence, the
error of the end position could still be small as compared
to the total driving length. The filter will usually diverge in
urban environments where the road density is high and most
roads have the same speed limit. In these cases, the number of
possible paths grows very fast and the measurement updates
provide little information. While the rate of divergence may
seem rather high, the results (e.g., an MMSE estimator with a
median horizontal error of about a fifth of the driving length)
are comparable to those obtained in [7] where the median trip
length was 7.5 [km] (about four times shorter than in the data
used here), the maximum trip length was 16.0 [km] (shorter
than both the median and mean trip length in the data used
here), and no uncertainty in the initial position was mentioned.
It can also be noted that filter divergence occasionally can
occur even if measurements of yaw rate are available [22]–
[24].
V. D ISCUSSION
In the preceding section, digital map information and measurements of speed were used to geographically locate a
TABLE II
CAPABILITY OF PROPOSED METHOD FOR NAVIGATION .
Inference of end destination
Exact location (∼ 200 [m])
Approximate region (∼ 2 [km])
Share of trips
1/4
1/2
car during everyday trips. The estimates were shown to be
accurate enough to enable violations of the driver’s locational
privacy in a substantial amount of the trips. (A summary of
the indicated capability is provided in Table II.) As a result,
this also means that the data collected in many of the current
insurance telematics programs is far more sensitive than stated
by the insurance companies themselves. While many of these
companies will have privacy policies that regulate their data
practices, the data can easily be misused if it falls into the
wrong hands. Moreover, since the privacy risks will not always
be obvious for someone without experience in navigation, the
data might not be treated as sensitive, thereby increasing the
risk of data theft. Normally, the data is saved indefinitely
without consideration of any ”right to be forgotten”.
Insurance companies offering telematics services will often
have years’ worth of data attributed to individual policyholders, and can therefore make use of previously extracted
personal information (e.g., commonly visited locations) to
improve the position estimation over time. One way to do this
could be to apply more sophisticated weight distributions in
the propagation step presented in Section III-B. Put differently,
the weights of different particles emerging from the crossing of
a road junction could be adjusted based on previous inferences
on driving habits.
The prospect of being able to extract positional information
from speed measurements does not only pose privacy risks, but
also means that insurance companies can use a larger number
of features to distinguish and characterize policyholders. For
example, while speed measurements alone only enable studies of absolute speeding, i.e., whether the driver exceeds a
given speed threshold, accurate knowledge of both speed and
position makes it possible for the insurer to directly identify
eventual speed limit violations. In addition, the insurer can
penalize driving in areas where accidents are prone to happen.
There are several ways to implement a privacy-preserving
insurance telematics program. For example, the data assembled
at the central server does not need to reveal from which vehicle
a specific set of data was gathered. This approach is taken
in [42], which associates each collected data tuple with a
time stamp and a random time-varying vehicle identifier. Any
information extraction (e.g., the computation of a driver score)
requiring data from a single vehicle must then be performed as
a secure multi-party computation involving both the server and
the vehicle-fixed client application. In this case, ideal privacy
preservation means that the server data by itself cannot be used
to extract more information about an individual vehicle’s data
than can be done from the same set of data but without any
vehicle identifiers. However, if position data is collected, it
may still be possible to identify data from policyholders who
live in secluded areas (since they might be the only people
driving in this area) [43], [44]. The corresponding location
traces may then be found by using that location updates close
in time and space are likely to originate from the same vehicle
[45]. Further, location traces from different drivers can be
clustered by studying individual driving characteristics, so
called driver fingerprinting [46]. Obviously, it is possible to
reduce the risk of inference attacks by simply reducing the data
quality, assuming that the primary privacy-preserving information extraction can still be performed with tolerable errors.
This can be done by e.g., omitting samples or deliberately
perturbing measurements [47].
Many insurance telematics providers are not interested in
speed measurements or location updates per se, but rather, in
the driver score or risk profile that can be extracted from them.
Hence, it will often be sufficient to compute the measures
of interest directly on the telematics unit, send these to a
central server, and discard the original measurements [48].
This will both reduce the required communications and the
risks of privacy intrusion. On the downside, this may increase
the computational demands on the telematics units.
VI. C ONCLUSIONS
We have considered the problem of estimating the position
of a vehicle using only map information and measurements of
speed. As opposed to previous studies utilizing the same set
of sensors, the problem was here formulated within the wellknown particle filter framework. Particle-based estimation
makes it straightforward to 1) study the estimation accuracy
using estimators suited for multimodal posteriors; 2) draw use
of experience from previous filter-based approaches where
sensor-based yaw information have been available; and 3)
extend the presented framework with additional measurement
functions. Performance evaluations of the proposed method
indicated that location-based information such as end destinations often can be estimated with an accuracy in the order
of 100 [m]. This level of accuracy will in many cases be
sufficient to extract information on e.g., visited locations or
areas. As a result, the collection of speed measurements from
policyholders in current insurance telematics programs both
poses privacy risks and makes it possible for the insurer to
discriminate among drivers based on visited areas. Telematics
providers who wishes to decrease the risk of privacy infringement for their policyholders can for example make use of timevarying vehicle identifiers, or perform the computations of
relevant driver measures directly in the vehicle-fixed telematics
units.
R EFERENCES
[1] P. Lawson, “The connected car: Who’s in the driver’s seat?” Mar. 2015,
FIPA.
[2] L. Pike, J. Sharp, M. Tullsen, P. C. Hickey, and J. Bielman, “Securing
the automobile: A comprehensive approach,” in Embedded Security in
Cars, Detroit, MI, May 2015.
[3] S. Checkoway, D. McCoy, B. Kantor, D. Anderson, H. Shacham, S. Savage, K. Koscher, A. Czeskis, F. Roesner, and T. Kohno, “Comprehensive
experimental analyses of automotive attack surfaces,” in Proc. 20th
USENIX Conf. Security, Berkeley, CA, Aug. 2011.
[4] I. Foster, A. Prudhomme, K. Koscher, and S. Savage, “Fast and vulnerable: A story of telematic failures,” in Proc. 9th USENIX Workshop on
Offensive Technol., Washington, D.C., Aug. 2015.
[5] S. Duri, J. Elliott, M. Gruteser, X. Liu, P. Moskowitz, R. Perez, M. Singh,
and J.-M. Tang, “Data protection and data sharing in telematics,” Mob.
Netw. Appl., vol. 9, no. 6, pp. 693–701, Dec. 2004.
[6] N. Rizzo, E. Sprissler, Y. Hong, and S. Goel, “Privacy-preserving driving
style recognition,” in Proc. IEEE Int. Conf. Connected Veh. Expo,
Shenzhen, China, Oct. 2015.
[7] X. Gao, B. Firner, S. Sugrim, V. Kaiser-Pendergrast, Y. Yang, and
J. Lindqvist, “Elastic pathing: Your speed is enough to track you,” in
Proc. ACM Int. Joint Conf. Pervasive and Ubiquitous Comput., Seattle,
WA, 2014.
[8] R. Dewri, P. Annadata, W. Eltarjaman, and R. Thurimella, “Inferring trip
destinations from driving habits data,” in Proc. 12th ACM Workshop on
Privacy in the Electron. Soc., Berlin, Germany, Nov. 2013, pp. 267–272.
[9] P. Händel, I. Skog, J. Wahlström, F. Bonawiede, R. Welch, J. Ohlsson,
and M. Ohlsson, “Insurance telematics: Opportunities and challenges
with the smartphone solution,” IEEE Intell. Transport. Syst. Mag., vol. 6,
no. 4, pp. 57–70, Oct. 2014.
[10] J. Wahlström, I. Skog, and P. Händel, “Driving behavior analysis for
smartphone-based insurance telematics,” in Proc. 2nd Workshop on
Physical Analytics, Florence, Italy, May 2015, pp. 19–24.
[11] J. Engelbrecht, M. J. Booysen, G.-J. van Rooyen, and F. J. Bruwer,
“Survey of smartphone-based sensing in vehicles for intelligent transportation system applications,” IET Intell. Transport. Syst., vol. 9, no. 10,
pp. 924–935, Dec. 2015.
[12] J. Wahlström, I. Skog, and P. Händel, “IMU alignment for smartphonebased automotive navigation,” in Proc. 18th IEEE Int. Conf. Inf. Fusion,
Washington, DC, Jul. 2015, pp. 1437–1443.
[13] J. Engelbrecht, M. J. Booysen, and G.-J. van Rooyen, “Recognition of
driving manoeuvres using smartphone-based inertial and GPS measurement,” in Proc.1st Int. Conf. Use of Mobile ICT, Stellenbosch, South
Africa, Dec. 2014.
[14] A. J. Blumberg and P. Eckersley, “On locational privacy, and how to
avoid losing it forever,” Aug. 2009, Electron. Frontier Foundation.
[15] M. Iqbal and S. Lim, “Privacy implications of automated GPS tracking
and profiling,” IEEE Technol. Soc. Mag., vol. 29, no. 2, pp. 39–46, Jun.
2010.
[16] K. Michael, A. McNamee, M. G. Michael, and H. Tootell, “Locationbased intelligence - Modeling behavior in humans using GPS,” in Proc.
IEEE Int. Symp. Technol. Soc., Queens, NY, Jun. 2006, pp. 1–8.
[17] L. Nguyen, Y. Tian, S. Cho, W. Kwak, S. Parab, Y. Kim, P. Tague, and
J. Zhang, “UnLocIn: Unauthorized location inference on smartphones
without being caught,” in Proc. Int. Conf. Privacy and Security in Mobile
Syst., Atlantic City, NJ, Jun. 2013, pp. 1–8.
[18] C. Troncoso, G. Danezis, E. Kosta, J. Balasch, and B. Preneel,
“PriPAYD: Privacy-friendly pay-as-you-drive insurance,” IEEE Trans.
Dependable and Secure Comput., vol. 8, no. 5, pp. 742–755, Sep. 2011.
[19] J. Paefgen, T. Staake, and F. Thiesse, “Evaluation and aggregation
of pay-as-you-drive insurance rate factors: A classification analysis
approach,” Decision Support Syst., vol. 56, pp. 192 – 201, Dec. 2013.
[20] M. Courtney, “Premium binds,” Eng. Technol., vol. 8, no. 6, pp. 68–73,
Jul. 2013.
[21] J. Ohlsson, P. Händel, S. Han, and R. Welch, BPM - Driving Innovation
in a Digital World.
Springer, 2015, ch. Process Innovation with
Disruptive Technology in Auto Insurance: Lessons Learned from a
Smartphone-Based Insurance Telematics Initiative, pp. 85–101.
[22] P. Hall, “A Bayesian approach to map-aided vehicle positioning,”
Master’s thesis, Linköping University, Jan. 2001.
[23] N. Svenzén, “Real time implementation of map aided positioning using
a Bayesian approach,” Master’s thesis, Linköping University, Dec. 2002.
[24] G. Hedlund, “Map aided positioning using an inertial measurement unit,”
Master’s thesis, Linköping University, Nov. 2008.
[25] J. Kronander, “Robust automotive positioning: Integration of GPS and
relative motion sensors,” Master’s thesis, Linköping University, Dec.
2004.
[26] P. D. Groves, Principles of GNSS, inertial, and multisensor integrated
navigation systems, 1st ed. Artech House, 2008.
[27] Y. Chen and H. Kobayashi, “Signal strength based indoor geolocation,”
in Proc. IEEE Int. Conf. Commun., vol. 1, Apr., New York, NY 2002,
pp. 436–439.
[28] J. Haverinen and A. Kemppainen, “Global indoor self-localization based
on the ambient magnetic field,” Robot. Auton. Syst., vol. 57, no. 10, pp.
1028 – 1035, Oct. 2009.
[29] N. Bergman, “A Bayesian approach to terrain-aided navigation,” in IEEE
Proc. 11th Symp. Syst. Identification, Fukuoka, Japan, Sep. 1997, pp.
1531–1536.
[30] F. Gustafsson, U. Orguner, T. Schön, P. Skoglar, and R. Karlsson,
Handbook of Intelligent Vehicles. Springer, 2012, ch. Navigation and
tracking of road-bound vehicles, pp. 397–434.
[31] P. Davidson, J. Collin, J. Raquet, and J. Takala, “Application of particle
filters for vehicle positioning using road maps,” in Proc. 23rd Int. Tech.
Meeting of the Satellite Division of the Institute of Navigation, Portland,
OR, Sep. 2010, pp. 1653–1661.
[32] I. Skog and P. Händel, “In-car positioning and navigation technologies A survey,” IEEE Trans. Intell. Transport. Syst., vol. 10, no. 1, pp. 4–21,
Mar. 2009.
[33] “The state of usage based insurance today,” Apr. 2014, Ptolemus Consulting Group.
[34] S. Derikx, M. D. Reuver, M. Kroesen, and H. Bouwman, “Buying-off
privacy concerns for mobility services in the internet-of-things era: A
discrete choice experiment on the case of mobile insurance,” in 28th
Bled eConference, Bled, Slovenia, Jul. 2015.
[35] S. N. Sadrieh, A. Broumandan, and G. Lachapelle, “Doppler characterization of a mobile GNSS receiver in multipath fading channels,” The
J. Navigation, vol. 65, no. 7, pp. 477–494, Jul. 2012.
[36] F. Gustafsson, “Rotational speed sensors: limitations, pre-processing and
automotive applications,” IEEE Mag. Instrum. Meas., vol. 13, no. 2, pp.
16–23, Apr. 2010.
[37] J. Wahlström, I. Skog, and P. Händel, “Detection of dangerous cornering
in GNSS-data-driven insurance telematics,” IEEE Trans. Intell. Transport. Syst., vol. 16, no. 6, pp. 3073–3083, Dec. 2015.
[38] N. Gordon, D. Salmond, and A. Smith, “Novel approach to
nonlinear/non-Gaussian Bayesian state estimation,” IEEE Proc. Radar
and Signal Process., vol. 140, no. 2, pp. 107–113, Apr. 1993.
[39] J.-O. Nilsson and P. Händel, “Recursive Bayesian initialization of
localization based on ranging and dead reckoning,” in IEEE, Intell.
Robots and Syst. Int. Conf., Tokyo, Japan, Nov. 2013, pp. 1399–1404.
[40] J. D. Hol, “Resampling in particle filters,” Linköping University, Tech.
Rep., May 2004.
[41] J. G. Rodrigues, A. Aguiar, and J. Barros, “SenseMyCity: Crowdsourcing an urban sensor,” CoRR, vol. arXiv:1412.2070, Dec. 2014.
[42] R. A. Popa, H. Balakrishnan, and A. J. Blumberg, “VPriv: Protecting
privacy in location-based vehicular services,” in Proc. 18th Int. Conf.
USENIX Security Symp., Montreal, QC, Aug. 2009, pp. 335–350.
[43] B. Hoh, M. Gruteser, H. Xiong, and A. Alrabady, “Enhancing security
and privacy in traffic-monitoring systems,” IEEE Pervasive Comput.,
vol. 5, no. 4, pp. 38–46, Oct. 2006.
[44] J. Krumm, “Inference attacks on location tracks,” in Proc. 5th Int. Conf.
Pervasive Comput., Berlin, Germany, May 2007, pp. 127–143.
[45] M. Gruteser and B. Hoh, “On the anonymity of periodic location
samples,” in Proc. 2nd Int. Conf. Security in Pervasive Comput., Berlin,
Germany, Apr. 2005, pp. 179–192.
[46] K. K. Miro Enev, Alex Takakuwa and T. Kohno, “Automobile driver
fingerprinting,” in Proc. Privacy Enhancing Technol., vol. 1, Jan. 2016,
pp. 34–50.
[47] B. Hoh, M. Gruteser, H. Xiong, and A. Alrabady, “Preserving privacy
in GPS traces via uncertainty-aware path cloaking,” in Proc. 14th ACM
Int. Conf. Comput. Commun. Security, Alexandria, VA, Oct. 2007, pp.
161–171.
[48] P. Händel, J. Ohlsson, M. Ohlsson, I. Skog, and E. Nygren,
“Smartphone-based measurement systems for road vehicle traffic monitoring and usage-based insurance,” IEEE Syst. J., vol. 8, no. 4, pp. 1238
– 1248, Dec. 2014.
Johan Wahlström received his MSc degree in Engineering Physics from the KTH Royal Institute of
Technology, Stockholm, Sweden, in 2014. He subsequently joined the Signal Processing Department at
KTH, working towards his PhD. His main research
topic is insurance telematics. In 2015, he received
a scholarship from the Sweden-America foundation
and spent six months at Washington University, St.
Louis, USA.
Isaac Skog (S’09-M’10) received the BSc and MSc
degrees in Electrical Engineering from the KTH
Royal Institute of Technology, Stockholm, Sweden,
in 2003 and 2005, respectively. In 2010, he received
the Ph.D. degree in Signal Processing with a thesis
on low-cost navigation systems. In 2009, he spent 5
months at the Mobile Multi-Sensor System research
team, University of Calgary, Canada, as a visiting
scholar and in 2011 he spent 4 months at the Indian
Institute of Science (IISc), Bangalore, India, as a
visiting scholar. He is currently a Researcher at KTH
coordinating the KTH Insurance Telematics Lab. He was a recipient of a Best
Survey Paper Award by the IEEE Intelligent Transportation Systems Society
in 2013.
João G. P. Rodrigues (S’11) received the MSc
degree in electrical and computer engineering from
the University of Porto, Porto, Portugal, in 2009. He
is currently working toward the PhD degree with
the University of Porto. He develops his work at
the Institute for Telecommunications, and the main
topics of his thesis are data gathering and mining
in intelligent transportation systems. His main research interests include sensor networks and intelligent transportation systems. He received a Doctoral
Scholarship from the Portuguese Foundation for
Science and Technology in 2009.
Peter Händel (S’88-M’94-SM’98) received a PhD
degree from Uppsala University, Uppsala, Sweden,
in 1993. From 1987 to 1993, he was with Uppsala
University. From 1993 to 1997, he was with Ericsson
AB, Kista, Sweden. From 1996 to 1997, he was
a Visiting Scholar with the Tampere University of
Technology, Tampere, Finland. Since 1997, he has
been with the KTH Royal Institute of Technology,
Stockholm, Sweden, where he is currently a Professor of Signal Processing and the Head of the
Department of Signal Processing. From 2000 to
2006, he held an adjunct position at the Swedish Defence Research Agency.
He has been a Guest Professor at the Indian Institute of Science (IISc),
Bangalore, India, and at the University of Gävle, Sweden. He is a co-founder
of Movelo AB. Dr. Händel has served as an associate editor for the IEEE
TRANSACTIONS ON SIGNAL PROCESSING. He was a recipient of a Best
Survey Paper Award by the IEEE Intelligent Transportation Systems Society
in 2013.
Ana Aguiar (S’94-M’98-S’02-M’09) received the
Electrical and Computer Engineering degree from
the University of Porto, Porto, Portugal, in 1998,
and the PhD in telecommunication networks from
the Technical University of Berlin, Berlin, Germany,
in 2008. Since 2009, she has been an Assistant Professor with the Faculty of Engineering, University
of Porto. She began her career as an RF Engineer working for cellular operators, and she worked
at Fraunhofer Portugal AICOS on service-oriented
architectures and wireless technologies applied to
ambient assisted living. She is the author of several papers published and
presented in IEEE and ACM journals and conferences, respectively. She
contributes to several interdisciplinary projects in the fields of intelligent
transportation systems and well-being (stress). Her research interests include
wireless networking and mobile sensing systems, specifically vehicular networks, crowd sensing, and machine-to-machine communications. She is a
Reviewer for several IEEE and ACM conferences and journals.
| 3cs.SY
|
1
Delay Performance of Wireless
Communications with Imperfect CSI and Finite
Length Coding
arXiv:1608.08445v2 [cs.IT] 2 Jun 2017
Sebastian Schiessl, Student Member, IEEE,
Hussein Al-Zubaidy, Senior Member, IEEE,
Mikael Skoglund, Senior Member, IEEE, and James Gross, Senior Member, IEEE
Abstract
With the rise of critical machine-to-machine applications, next generation wireless communication
systems must be designed with strict constraints on the latency and reliability. A key question in this
context relates to channel state estimation, which allows the transmitter to adapt the code rate to the
channel. In this work, we characterize the trade-off between the estimation sequence length and data
codeword length: shorter channel estimation leaves more time for the actual payload transmission but
reduces the estimation accuracy and causes more decoding errors. Using lower coding rates can mitigate
this effect, but may result in a higher backlog of data at the transmitter. We analyze this trade-off using
queueing analysis on top of accurate models of the physical layer, which also account for the finite
blocklength of the channel code. Based on a novel closed-form approximation for the error probability
given the rate, we show that finding the optimal rate adaptation strategy becomes a convex problem.
The optimal rate adaptation strategy and the optimal training sequence length, which both depend on
the latency and reliability constraints of the application, can improve the delay performance by an order
of magnitude, compared to suboptimal strategies that do not consider those constraints.
Index Terms
Finite blocklength regime, imperfect CSI, rate adaptation, quasi-static fading, queueing analysis
The authors are with the School of Electrical Engineering, KTH Royal Institute of Technology, Stockholm, Sweden (e-mail:
[email protected], [email protected], [email protected], [email protected]).
2
I. I NTRODUCTION
While wireless networks have traditionally been optimized for the typical requirements of
human-related services (which includes voice communication as well as Internet applications), a
new class of machine-to-machine applications has been arising over the last decade. This class can
be separated into two groups. The first group consists of so called massive machine-to-machine
applications, where for example sensor readings need to be conveyed to a central data collection
point. For such applications, the major challenges are the energy efficiency and the scalability
of the communication service to potentially thousands of terminals. However, the requirements
on the latency and reliability of the communication are only moderate in this case. The second
group contains critical machine-to-machine applications, which are foremost encountered in the
context of industrial automation and are traditionally realized by specialized wired networks.
Due to increasing flexibility demands, and in order to enable entirely new designs of automation
systems, wireless transmissions become more and more attractive. Critical machine-to-machine
applications typically generate small payloads periodically or at event-triggered time points, and
require transmission with very low latency and ultra-high reliability. For instance, automation
applications from manufacturing easily require communication latencies between a sensor and
a control unit below 1 ms, as well as packet delivery ratios (with respect to that deadline)
of 1 − 10−6 and above. This area of critical machine-to-machine communications still poses
significant challenges with respect to wireless network design.
Traditionally, physical layer analysis has been based on the assumption that error-free transmissions can be achieved through channel coding at Shannon’s channel capacity, which is a fairly
accurate model when the blocklength of the channel code is very large. However, with target
latencies below 1 ms, systems will only be able to spend a small number of symbols onto a single
transmission, thus the blocklength becomes small, resulting in a significant performance loss due
to channel coding at finite blocklength. In the finite blocklength regime, transmission errors occur
due to “above average” noise occurrences. Although the transmitter can reduce the probability
of transmission errors by selecting a rate lower than the channel capacity, transmission errors
are inevitable [1]. Understanding the implications from finite blocklength coding in combination
with the fading effects of wireless communication channels is fundamental to the efficient design
of ultra-reliable low latency wireless networks. An important question in this context relates to
the most efficient transmission strategy, in particular, if and how the transmitter should adapt
3
the rate of the channel code to the instantaneous state of the wireless channel. This is still an
open question in the context of low-latency communications, as rate adaptation requires time to
estimate the current channel state, which reduces the already short time for data transmission.
While spending more time on channel estimation improves the accuracy of the estimate and
allows the transmitter to send the payload more reliably, this decreases the amount of time left
for payload transmission even further. The shortened payload transmission duration is penalized
also through a higher sensitivity to the finite blocklength effects. This trade-off between the time
spent on channel estimation and the time spent on payload transmission cannot be characterized
by a purely information-theoretic analysis, as an adaptation to the channel state leads to a
time-varying transmission rate that affects the backlog, and hence the latency, at the transmitter.
Therefore, in addition to the impact of finite blocklength and imperfect channel state information
(CSI) on the physical layer performance, queueing effects on the link layer must be considered to
address this problem. To our best knowledge, a performance analysis that takes all these effects
into account does not exist in the literature.
As this work relates to both information theory and communication networking, we build our
work on literature in both fields. Concerning research in information theory, the comprehensive
analysis of the theoretical limits of finite blocklength channel coding by Polyanskiy et al. [1] was
extended by Yang et al. to block-fading channels [2]–[4]. Surprisingly, the authors found that
for many types of fading channels, the maximum achievable data rate shows little dependency
on the blocklength and is in fact well approximated by the outage capacity. These works even
accounted for the fact that the channel state may not be known a priori, i.e. at the start of the
transmission. However, the data rate was assumed to be fixed and rate adaptation with imperfect
CSI at the transmitter was not considered. Analysis of imperfect CSI often focuses on the receiver
side: imperfect CSI at the receiver (CSIR) will cause an error in the amplitude and phase of the
signal during demodulation and decoding, which can cause errors. Médard [5] investigated this
by computing the mutual information of a system with imperfect CSIR. Hassibi and Hochwald
[6] investigated the impact of imperfect CSIR on the ergodic capacity in multi-antenna systems.
However, those information-theoretic results are based on statistical averages of the estimation
error and therefore only apply when decoding is performed over infinitely many fading blocks,
which would cause infinite delay. Moreover, rate adaptation at the transmitter cannot be analyzed
using such models. The authors in [7], [8] studied rate adaptation at finite blocklength with
restricted input alphabets, as well as performance bounds for binary-input channels, but did not
4
consider imperfect CSI. Yang et al. [9] also studied the finite-blocklength performance when
the transmitter adapts the power (but not the coding rate) to the perfectly known channel state.
Finally, Lim and Lau [10] and Lau et al. [11] studied rate adaptation where the transmitter has
imperfect or outdated CSI, but considered neither finite blocklength effects nor the impact of
transmission errors on the delay.
In the field of communication networks, queueing theory has been used extensively to analyze
the delay performance of wireless networks. While wireless network analysis poses a significant
challenge to traditional queuing theory, several techniques have been developed in the last decade
to address this challenge. Wu and Negi [12] developed the framework of effective capacity that
provides approximations on the delay performance, which are however asymptotic, i.e. only
valid for long delays. Al-Zubaidy et al. [13] used stochastic network calculus in a transform
domain, which not only provides non-asymptotic bounds on the delay performance, but can also
be extended for the analysis of multi-hop wireless links [13], [14]. Finite blocklength effects
and imperfect CSI have been separately studied with respect to their impact on the queueing
performance. Wu and Jindal [15] analyzed a system with automatic repeat requests in fading
channels at finite blocklength. Gursoy [16] computed the effective capacity for block-fading
channels at finite blocklength and showed that there is a unique optimum for the error probability.
In our own previous work [17], we extended this analysis using stochastic network calculus and
provided analytical solutions for finite blocklength coding in Rayleigh block-fading channels.
Nevertheless, none of these works considered imperfect CSI. The queueing performance under
rate adaptation with imperfect/outdated CSI but without finite blocklength effects was analyzed
by Gross [18].
Thus, in this work we address the delay performance of a wireless communication system in
the finite blocklength regime with rate adaptation based on imperfect channel estimation. In this
context we provide three main contributions:
•
Based on stochastic network calculus [13], we characterize the trade-off between the rate
and the error probability with respect to the delay performance as a convex optimization
problem. Thus, the transmitter can efficiently determine the optimal transmission rate, taking
the overall latency and reliability constraints of the application flow into account.
•
This optimization is based on a novel closed-form approximation for the error probability
due to the combined effects of finite blocklength coding and imperfect CSI at the transmitter
in Rayleigh fading channels. Specifically, we derive an approximation for an information-
5
theoretic result from Yang et al. [4]. A key challenge that we overcame in this analysis is
that finite blocklength effects are modeled as variations in the rate, whereas imperfect CSI
corresponds to variations in the SNR. Our approximation is invertible, providing a direct
mapping from the error probability to the rate, and has potential uses beyond the scope of
this paper.
•
We show through numerical analysis that both finite blocklength coding and imperfect CSI
have a significant impact on the performance under strict latency and reliability constraints.
Our results show that rate adaption, despite needing a fraction of the resources for channel
training, significantly outperforms fixed rate transmissions when low latency is required.
Moreover, we find that through an optimal rate adaptation strategy and through an optimal
choice of training duration versus payload transmission time, the system can improve the
overall reliability by one order of magnitude.
This paper is organized as follows: The system model is given in Sec. II. Our main contributions are presented in Sec. III. Numerical results are then presented in Sec. IV, followed by
our conclusions in Sec. V.
Throughout the paper, we utilize the following notation: Uppercase italic letters X generally
refer to random variables, whereas the corresponding lowercase letters x refer to a realization
of that random variable. We write fX|y (x) for the probability density function and FX|y (x) for
the cumulative distribution function of the variable X, conditioned on the value Y = y of a
different random variable Y . For complex values, < {x} describes the real part, ∠(x) describes
the phase, and x∗ is the complex conjugate.
II. S YSTEM M ODEL
We consider data transmission over a point-to-point wireless link. More precisely, a data
flow arrives at a transmitter and needs to be transmitted to a receiver. The incoming data must
be transmitted with probabilistic constraints on the quality-of-service: a target deadline can be
violated with a probability not exceeding a given maximum delay violation probability. In general,
we are interested in data flows as arising in industrial contexts, i.e., with a low constant data
rate, periodic arrivals, short deadlines and very low target violation probabilities. A time-slotted
system with equal slots containing nslot symbols is considered. Each time slot is furthermore
assumed to be split into two phases: The training/estimation phase, where the transmitter sends
a known training sequence of m symbols to the receiver; and the actual data transmission phase
Estim. Cap.
log2 (1 + γ̂)
Training
Rate / Estimated Capacity
6
Selected Rate
r(γ̂)
log2 (1 + γ̂)
log2 (1 + γ̂)
r(γ̂)
m
n
r(γ̂)
nslot
i=0
i=1
i=2
Time
Fig. 1. Wireless transmission for the first three time slots. After the training phase of m symbols, the transmitter receives an
estimate γ̂ of the channel’s SNR as feedback and transmits a codeword at a rate r(γ̂). Rate adaptation is challenging when the
transmitter has only an imperfect estimate log2 (1 + γ̂) of the channel capacity.
of n = nslot − m symbols. At the end of the training phase, the receiver sends the estimated CSI
as feedback to the transmitter. Then, based on the feedback and the amount of data backlogged,
the transmitter attempts to transmit a certain amount of data during the transmission phase.
We assume that the feedback is instantaneous and error-free. Furthermore, the feedback also
includes an acknowledgment of the previous data transmission. Fig. 1 illustrates the basic system
operation: since the transmitter knows only a noisy estimate of the channel capacity, it often
selects a rate below the estimated capacity to decrease the chance of transmission errors, which
can occur due to both imperfect channel knowledge and finite blocklength effects.
Transmitter and receiver are both assumed to be static. A frequency-flat Rayleigh block-fading
channel model is assumed, where the channel remains constant for the duration of one time
slot and changes independently between time slots. This model applies for example to systems
that apply frequency hopping after each time slot. We assume that there is one transmit and
one receive antenna, so the channel can be described by the scalar complex fading coefficient
H. For the Rayleigh block-fading channel, H has circularly symmetric Gaussian distribution
CN (0, 1). In each time slot, the instantaneous signal-to-noise ratio (SNR) at the receiver is
given as Γ = γ̄|H|2 and has exponential distribution with mean γ̄. The average SNR γ̄ is
assumed to be constant and known at the transmitter and the receiver. In the following, we first
provide more details on the two different system operation phases, before we relate these phases
to the queuing performance and the problem statement.
7
A. Training Phase
The receiver estimates the fading coefficient H through a known training sequence of m
symbols. The minimum mean square error (MMSE) estimate for H is given as [6], [19]:
√
γ̄m
γ̄m
H+
N,
Ĥ =
1 + γ̄m
1 + γ̄m
(1)
where N ∼ CN (0, 1) is independent of H. Therefore, the channel estimate Ĥ is distributed
∆
as Ĥ ∼ CN (0, ρ2 ) with ρ2 = γ̄m/(1 + γ̄m), and the estimated SNR Γ̂ = γ̄|Ĥ|2 follows an
exponential distribution with mean ρ2 γ̄. Due to H and Ĥ being jointly Gaussian, H can be
expressed in terms of the estimate Ĥ as [19]
H = Ĥ + Z ,
(2)
2
) is independent of Ĥ with
where the estimation noise Z ∼ CN (0, σN
2
σN
=
1
.
1 + γ̄m
(3)
B. Data Transmission Phase
After the training phase, a noise-free feedback provides the transmitter with the estimated
coefficient Ĥ. In the rest of this section we will consider a single time slot, in which the
transmitter has a specific estimate ĥ of the fading coefficient. In this scenario, the phase of the
channel coefficient is not used at the transmitter, so it is not relevant whether the transmitter
knows ĥ or the estimated SNR γ̂ = γ̄|ĥ|2 . We assume that the transmitter constantly operates
at the maximum allowed transmit power, i.e., it does not perform power allocation but uses
the channel estimate only to select a code rate r and encode n · r data bits into a codeword
of n symbols. However, the actual SNR Γ is unknown and can be lower than the estimated
SNR γ̂, and thus the transmitter may select a code rate that is higher than the channel capacity
C = log2 (1 + Γ). In this case, the channel is said to be in outage. When the blocklength of
the channel code becomes very large (infinite), the outage probability becomes equal to the
probability that the received signal cannot be correctly decoded. Conditioned on the estimate γ̂,
the outage probability is given as
n
o
εout = P log2 (1 + Γ) < r Γ̂ = γ̂ .
(4)
To compute the outage probability, first note that the distribution of H conditioned on a known ĥ
is H = ĥ + Z, according to (2). Thus, when conditioned on the estimate, the channel coefficient
8
H has the same distribution as the fading coefficient of a Rician channel with line-of-sight (LoS)
component ĥ (which is known at transmitter and receiver) and unknown non-LoS component Z.
The SNR Γ = γ̄|H|2 conditioned on the estimate γ̂ then follows a non-central χ2 -distribution
with two degrees of freedom and PDF:
− x+γ̂
1
γ̄·σ 2
N · I
·
e
fΓ|γ̂ (x) =
0
2
γ̄ · σN
√
2 xγ̂
.
2
γ̄ · σN
(5)
The conditional outage probability can be found through the cumulative distribution, which is
given by the Marcum Q-function Q1 (a, b) [18], [20]:
s
s
!
2γ̂
2x
FΓ|γ̂ (x) = 1 − Q1
,
.
2
2
γ̄σN
γ̄σN
(6)
In contrast, when the blocklength of the channel code is rather small (finite), the interpretation
of the conditional channel distribution as that of a Rician fading channel with known LoS
component ĥ enables the use of finite blocklength results for quasi-static fading channels from
Yang et al. [3], [4]. According to the normal approximation [4, (59-61)], the approximation ε
for the error probability for a code with rate r is given by:
#
!
"
log2 (1 + Γ) − r
p
Γ̂ = γ̂ ,
ε≈E Q
V(Γ)/n
where Q(·) is the Gaussian Q-function and the channel dispersion is given as:
1
2
V(γ) = log2 (e) 1 −
.
(1 + γ)2
(7)
(8)
The expectation is taken over the conditional distribution (5) of the SNR Γ given the estimated
SNR γ̂. We finally state two remarks concerning (7):
(I) The authors in [4] assumed perfect CSI at the receiver (CSIR) when proposing (7). In our
scenario, the receiver does not have perfect knowledge of H, as it only knows the estimate ĥ but
not the estimation noise Z. We can verify the accuracy of (7) by numerically computing a lower
bound [3, Cor. 3] on the achievable rate for Rician fading channels that does not depend on
perfect CSIR1 . These results indicate that, in our scenario with a significant amount of training
and rate adaptation at the transmitter, (7) provides a reasonable approximation for the achievable
performance. A second and independent argument to support the accuracy of (7) is that after
1
The achievable rate in [3, Cor. 3] is for fading channels with no instantaneous CSI, which means the realization Z is unknown
at both transmitter and receiver, but the statistics (i.e., the mean ĥ) are known. This corresponds exactly to our model, and can
be computed with the SPECTRE toolbox [21].
9
the training and feedback phase, many wireless communication systems send pilot symbols in
addition to the codeword symbols in the data transmission phase, which does not change the
CSI at the transmitter but significantly improves the CSI at the receiver, such that imperfect CSI
at the receiver will not affect the decoding performance.
(II) The finite blocklength result is an approximation and not accurate for all parameters. For
the non-fading AWGN channel [1], the approximation is considered to be accurate when the
blocklength n is above 200. Furthermore, it may become less accurate at extremely low error
probabilities, e.g., ε < 10−4 . In this work, we will avoid very short blocklengths and very low
error probabilities, as ultra high reliability with respect to a certain target deadline is achieved
through retransmissions. Thus, we will assume that the approximation of ε in (7) is exactly equal
to the error probability.
C. Queueing Model
In the wireless channel model, transmissions fail with probability ε due to imperfections in
the channel state estimation and due to finite blocklength effects. Furthermore, as the data rate r
is adapted to the time-varying channel, it varies from time slot to time slot and may become very
small. Thus, not all the data arriving in a certain time slot can be transmitted successfully. To
avoid data loss, data must be stored in a buffer or queue, in which it will remain for a random time
until the receiver sends an acknowledgment indicating that the data was successfully decoded.
As the arrival rate of the considered data flow is rather small, the buffer is assumed to be large
enough to hold all incoming data, such that all data will eventually be transmitted.
One performance metric for such a transmission scheme is the expected goodput, which is
the expected value of the amount of data that can be successfully transmitted. In each time slot,
r · n bits can be transmitted, but transmissions fail with probability ε. We define the normalized
expected goodput (normalized with respect to the total number of symbols nslot = n + m) as:
h
i
n
∆
r=
E r(Γ̂) · (1 − ε) ,
(9)
nslot
where the expectation is taken over the distribution of the estimated SNR Γ̂, which follows an
exponential distribution according to (1). We specifically denote the rate as r(Γ̂) as it is chosen
according to the estimated SNR Γ̂.
However, for the considered data flow (originating from an industrial application), the expected
goodput is not a suitable performance metric as it cannot characterize the tail of the random
10
delay distribution of the system. Thus, in order to analyze the random delay of data in the queue,
the system is described in terms of its arrival, service and departure processes. In time slot i,
the arrival process Ai is given by the data that is generated at the transmitter side, e.g., from
sensor readings, and enters the queue. The service process Si describes the number of bits that
can potentially be transmitted over the wireless channel in time slot i. In case of transmission
errors, i.e., with probability εi , the service Si is zero:
n · r(Γ̂ )
with prob. (1 − εi )
i
Si =
,
0
with prob. εi
(10)
where the estimated SNR Γ̂i , as well as the corresponding r(Γ̂i ) and εi vary from one time slot
to another. The departure process Di is given as the number of bits leaving the queue, which
is equal to the number of bits that reach the receiver successfully. The departure process is
upper-bounded both by the service process and by the amount of data waiting in the queue. For
the analysis of queueing systems, we also define the cumulative arrival, service, and departure
processes
∆
A(τ, t) =
t−1
X
Ai ,
∆
S(τ, t) =
i=τ
t−1
X
Si ,
∆
D(τ, t) =
i=τ
t−1
X
Di .
(11)
i=τ
The random delay W (t) at time t is then defined as the time it takes before all data that
arrived before time t has actually departed from the queue, i.e. reached the receiver:
∆
W (t) = inf {u ≥ 0 :
A(0, t) ≤ D(0, t + u)} .
(12)
For the considered application, we are interested in conveying the data with respect to a certain
deadline. This translates into the delay violation probability as a performance metric, i.e. the
maximum probability at any time t that the random delay W (t) exceeds a specified target delay
w (i.e., deadline):
∆
pv (w) = sup {P {W (t) > w}} .
(13)
t≥0
D. Problem Statement
The main objective of this work is a characterization of the trade-off between the time spent
on channel training and time spent on actual data transmission. When using a long training sequence of m symbols, the channel estimates become more accurate, allowing transmissions with
higher reliability and hence fewer retransmission attempts. However, a long training sequence
reduces the number of remaining symbols for data transmission n = nslot − m. We thus face
11
a typical trade-off between increasing the transmission reliability and the duration for payload
transmissions. This trade-off becomes in particular interesting when shorter transmission slots are
considered, which is the case in low latency wireless networks for critical industrial applications:
Due to finite blocklength effects, the shortening of the payload transmission deteriorates the
communication performance even more rapidly. What is then the optimal trade-off between m
and n? How is this optimal trade-off related to the required delay violation probability pv (w) of
the system, i.e. the criticality of the conveyed application data? Does it even make sense to use
training in all scenarios, or is it sometimes better to skip training (m = 0) and use the entire
time slot for data transmissions at fixed rate?
The same kind of trade-off can be observed between the code rate r and the corresponding
error probability ε: Lower data rates lead to lower error probability ε but also to reduced data
throughput. Therefore, our secondary objective is to determine the optimal trade-off with respect
to the delay performance between code rate r and error probability ε.
III. A NALYSIS
In this section, we present our main contribution: based on stochastic network calculus [13],
we formulate the optimal trade-off with respect to the delay violation probability pv (w) between
the code rate r and the error probability ε as a convex optimization problem. Thus, depending
on the requirements of the application, an optimal rate adaptation strategy for the transmitter
can be quickly determined. The biggest challenge in the characterization of that trade-off is the
fact that no closed-form representation of the problem exists in the literature, as ε is not given
in closed form. Thus, after summarizing stochastic network calculus and presenting the problem
in Sec. III-A, we derive in Sec. III-B and III-C a novel closed-form approximation of the error
probability ε. This analysis initially considers only imperfect CSI at the transmitter and yields
an approximation and upper bound for the outage probability εout , while ignoring the effects of
finite length coding. The approximation is extended to channel codes with finite blocklength in
Sec. III-C. This allows us then in Sec. III-D to address the optimal trade-off between the rate r
and error probability ε as convex problem based on the previously developed approximations.
In addition, we show in Sec. III-E how our approximation could be applied to scenarios
different from the one considered in this work.
12
A. Queueing Analysis
This subsection summarizes previous results on how the delay performance, specifically the
delay violation probability pv (w) given in (13), can be analyzed through stochastic network
calculus [13], [22]. While the delay violation probability can also be analyzed using effective
capacity [12], which has been successfully applied in numerous works, e.g., [16], [23], effective
capacity only provides an approximation of pv (w) in the tail of the delay distribution, i.e.,
for relatively large delays. Contrary to that, stochastic network calculus [22], which has been
recently extended to wireless network analysis in a transform domain [13] and which has also
been applied in various scenarios [14], [17], [24], [25], provides a strict upper bound on the
delay violation probability pv (w), even at low delay. This is beneficial for ultra-reliable low
latency systems in an industrial context: The modeled system will violate certain delay bounds
with an even lower probability than shown by the analysis. Parts of the following summary are
taken from our previous work in [17].
The delay W (t) in (12) is defined in terms of the arrival and departure processes. However,
for finding the statistical distribution of the delay, it is easier to use only the arrival and service
processes. The authors in [13] characterized these processes in the exponential domain, also
referred to as SNR domain. The main benefit of this approach is the elimination of the logarithm
in the channel capacity. Instead of describing the cumulative service and arrival S(τ, t) and
A(τ, t) in the bit domain, they are converted to the SNR domain (denoted by calligraphic letters)
as follows:
∆
A(τ, t) = eA(τ,t) ,
∆
∆
S(τ, t) = eS(τ,t) .
(14)
∆
Similarly, we define Ai = eAi and Si = eSi . According to the system model, the service process
Si is independent and identically distributed (i.i.d.) between time slots. We also require the arrival
process Ai to be i.i.d. in this work.
An upper bound on the delay violation probability pv (w) can then be computed in terms of the
Mellin transforms of Ai and Si , where we can drop the subscript i due to the i.i.d. assumption.
The Mellin transform MX (θ) of a nonnegative random variable X is defined as [13]
∆
MX (θ) = E X θ−1
(15)
13
for a parameter θ ∈ R. For the analysis, we always choose θ > 0 and first check whether the
stability condition MA (1 + θ)MS (1 − θ) < 1 holds. If it holds, define the kernel [13], [17]
∆
K (θ, w) = lim
t→∞
=
t
X
MA (1 + θ)t−u · MS (1 − θ)t+w−u
(16)
u=0
MS (1 − θ)w
.
1 − MA (1 + θ)MS (1 − θ)
(17)
This kernel is an upper bound for the delay violation probability, which holds for any time slot
t, including the limit t → ∞ (steady-state):
pv (w) ≤ inf {K (θ, w)} .
θ>0
(18)
For any parameter θ > 0, the kernel K (θ, w) provides an upper bound on the probability pv (w)
that the delay exceeds the target delay w. In order to find the tightest upper bound, one must
find the parameter θ > 0 that minimizes K (θ, w).
The kernel K (θ, w) depends on the Mellin transforms of A and S, i.e., of the arrival and
service processes in the SNR domain. For simplicity, we assume that the arrival process is
constant, i.e., in each time slot of length nslot symbols, a data packet with a constant size of
ᾱ · nslot bits arrives at the transmitter, thus MA (θ) = eᾱnslot (θ−1) . The service process describes
the number of bits that are successfully transmitted to the receiver. The random service S in (10)
can be described as S = nr(Γ̂) · Z, where r(Γ̂) is the code rate adapted to the measured SNR Γ̂
and Z is a Bernoulli random variable, which is zero in case of error, i.e., with probability ε, and
one in case of successful transmission, i.e. with probability (1 − ε). Thus, the Mellin transform
MS (θ) of the service process in the SNR domain S = eS is given as [17]
h
i
MS (θ) = E S θ−1 = EΓ̂,Z enr(Γ̂)·Z·(θ−1)
Z∞
=
(1 − ε)enr(γ̂)(θ−1) + ε fΓ̂ (γ̂)dγ̂ .
(19)
(20)
0
Note that, in general, the error probability ε is not constant but varies based on the estimated
SNR γ̂ and on the selected code rate r(γ̂). When MA (θ) and MS (θ) are known2 , the kernel
K (θ, w) and the upper bound (18) on the delay violation probability pv (w) can be computed.
2
In order to evaluate MS (θ) numerically, the integration range of (20) can be split into a finite set of intervals. As both the
error probability and fΓ̂ (γ̂) decrease in γ̂, replacing γ̂ in each interval with the lower limit of that interval will result in an
upper bound on MS (θ). The delay bound (18) remains valid when MS (θ) is replaced by its upper bound.
14
In this work, we seek to characterize the optimal trade-off with respect to the delay violation
probability pv (w) between training length m and code length n, as well as between the selected
rate r(γ̂) and the resulting error probability ε. The analytical bound (18) on pv (w) can be used
to solve this problem. First, in order to obtain the delay bound (18), one must iterate over
different parameters θ > 0, and compute the kernel K (θ, w) for each θ. The kernel K (θ, w) is
monotonically increasing in MS (1 − θ). Therefore, for each particular value of θ > 0, finding
the parameters m and r(γ̂) (with the corresponding values of n = nslot − m and ε) that minimize
MS (1 − θ) in (20) yields the desired minimum on the delay bound.
Concerning the optimal training length m, one can simply iterate over all possible choices
of m and pick the value that provides the lowest value of MS (1 − θ). However, finding the
optimal rates r(γ̂) that minimize (20) is hard, because the error probability ε in (7) can only be
evaluated numerically and depends on the training length m, on the estimated SNR γ̂, and on the
rate r(γ̂). In order to solve this problem, we develop a closed-form approximation for ε in the
following two sections. We then show in Sec. III-D that with this closed-form approximation,
finding the rates r(γ̂) that minimize (20) becomes a convex optimization problem.
B. Outage Probability Approximation for Imperfect CSI
When the blocklength of the code tends to infinity, i.e., when the effects of channel coding
at finite blocklength are ignored, then the error probability ε in (7) converges to the outage
probability εout in (4). Conditioned on the channel estimate, the outage probability can be
computed using (6), i.e., in terms of the Marcum Q-function. In this section, we provide an
upper bound for the outage probability based on the Gaussian Q-function.
Lemma 1. Given an imperfect estimate of the channel γ̂ and a rate r, the outage probability is
bounded by
εout ≤ Q
γ̂ − (2r − 1)
σICSI
,
(21)
∆
2
2
with σICSI
= 2σN
γ̄γ̂.
Proof. Given a known measurement ĥ, the random variable H is given in terms of the random
variable Z according to (2). Thus:
Γ = γ̄|H|2 = γ̄(ĥ + Z)(ĥ∗ + Z ∗ )
(22)
15
n
o
= γ̂ + 2γ̄< ĥ Z + γ̄ |Z|2
2
= γ̂ + 2γ̄|ĥ|< Z̄ + γ̄ Z̄
∗
(23)
(24)
where Z̄ = e−j∠(ĥ) Z is just a phase-rotated version of Z. The distribution and the magnitude
of a circularly symmetric random variable stay constant under phase rotation, and thus the real
2
part < Z̄ has Gaussian distribution N (0, σN
/2). It follows that the SNR Γ = γ̄ |H|2 is given
as
Γ = γ̂ + Γ̃G + Γ̃δ = γ̂ + Γ̃ ,
(25)
2
i.e. the estimation error Γ̃ = Γ − γ̂ is the sum of a Gaussian error Γ̃G ∼ N (0, σICSI
) and some
2
Γ̃δ = γ̄ Z̄ . The outage probability εout can then be bounded as:
n
o
εout = P Γ < 2r − 1 Γ̂ = γ̂
n
o
≤ P γ̂ + Γ̃G < 2r − 1
)
(
γ̂ − (2r − 1)
Γ̃G
>
=P −
σICSI
σICSI
(26)
(27)
(28)
where the inequality holds because Γ̃δ ≥ 0.
4
and thus Γ̃δ becomes very small
We observe that the variance of Γ̃δ is proportional to σN
relative to Γ̃G as the channel estimates become more accurate. In that case, (21) becomes a tight
upper bound on the outage probability.
Given a target outage probability of e.g. ε0out = 10−3 , it is difficult to find the exact rate r
for which the outage probability εout is exactly ε0out , as the Marcum Q-function cannot be easily
inverted. However, the approximation (21) is invertible:
Corollary 1. Given an imperfect channel estimate γ̂ and a target outage probability ε0out with
Q(γ̂/σICSI ) < ε0out < 1/2, the actual outage probability εout is less than or equal ε0out if the
transmitter chooses the rate
rICSI (γ̂, ε0out ) = log2 1 + γ̂ − σICSI Q−1 (ε0out ) .
(29)
Proof. Let the right-hand side of (21) be equal to ε0out and solve for the rate r. The condition
ε0out > Q(γ̂/σICSI ) ensures that the rate is positive.
16
Therefore, when the transmitter selects the rate rICSI (γ̂, ε0out ), the actual outage probability
εout is not exactly equal to the target outage probability ε0out , but εout is smaller than ε0out . This
upper bound on the outage probability allows for a worst-case performance analysis.
C. Combined Analysis of Imperfect CSI and Finite Blocklength
When analyzing the physical layer using the finite blocklength model, we focus first on the case
where the channel state information is perfect. If, in a specific time slot, the fading coefficient
h and the SNR γ = γ̄|h|2 are perfectly known at the transmitter and receiver, then (7) can be
computed easily, as the expected value is taken with respect to the constant h. In that case, (7)
can be solved for the achievable rate r given the error probability ε:
r
V(γ) −1
Q (ε) .
rFBL (γ, n, ε) ≈ log2 (1 + γ) −
n
(30)
This result corresponds to the approximation for the achievable rate in AWGN channels by
Polyanskiy et al. [1, Thm. 54]. We can obtain the same result from a different interpretation:
For a fixed capacity c = log2 (1 + γ), define the random blocklength-equivalent capacity
r
V(γ)
∆
· UFBL
Cb (γ) = log2 (1 + γ) −
n
(31)
with UFBL ∼ N (0, 1) and assume that errors occur if and only if an outage occurs, i.e. iff
Cb (γ) < rFBL (γ, n, ε). This means that rFBL is interpreted as the outage capacity for a channel
with random capacity Cb (γ), which simplifies the comparison and combination of imperfect CSI
and finite blocklength effects. However, while finite blocklength effects are modeled as Gaussian
variation UFBL in the capacity, we observed in Sec. III-B that channel estimation errors can be
approximated by Gaussian variations in the SNR. In order to analyze the combined impact of
both effects, we approximate the finite blocklength variations UFBL as variation in the SNR.
This is done using the first-order Taylor approximation of ln(x) around the point x0 , which has
gradient
1
.
x0
Due to the concavity of the ln-function, this linear approximation is larger than the
function itself:
ln(x0 ) −
1
(x0 a) ≥ ln (x0 − x0 a) .
x0
(32)
Due to ln(x) being continuous and monotonically increasing, this means that for some δ ≥ 0
and b = a log2 (e):
log2 (x0 ) − b = log2 x0 − x0
b
+δ
log2 (e)
.
(33)
17
By applying this result to (31) around x0 = 1 + γ, (31) can be rewritten as
Cb (γ) = log2 (1 + γ − σFBL (γ) · UFBL + Uδ )
with
1+γ
σFBL (γ) =
log2 (e)
∆
r
V(γ)
,
n
(34)
(35)
UFBL ∼ N (0, 1) and some random Uδ ≥ 0. Thus, we convert the Gaussian error in the rate
to a Gaussian error in the SNR, plus an unknown Uδ which can later be ignored because it is
non-negative.
As a next step, imperfect CSI is taken into account. When the transmitter has imperfect channel
state information, the error probability is given by ε defined in (7). The following lemma allows
an easier notation and interpretation of our results.
Lemma 2. The approximate error probability ε by Yang et al. [4] given in (7) is equal to the
blocklength-equivalent outage probability, i.e.
!
#
"
n
o
log2 (1 + Γ) − r
p
Γ̂ = γ̂ = P Cb (Γ) < r Γ̂ = γ̂ ,
E Q
V(Γ)/n
(36)
with Cb defined by (31) now depending on the random SNR Γ (conditioned on γ̂) and on the
random variable UFBL .
Proof. For a fixed SNR Γ = γ, the Q-function on the inside of the expectation in (7) is equal
to P {Cb (γ) < r} by definition of Cb (γ). The claim follows by taking the expectation over the
distribution of Γ (conditioned on the measurement γ̂) on both sides.
When the SNR is not perfectly known at the transmitter, the fixed value γ needs to be replaced
by the random Γ = γ̂ + Γ̃. We have seen before that the estimation error Γ̃ in the SNR Γ can
also be approximated as a Gaussian error: Γ̃ = Γ̃G + Γ̃δ with Γ̃δ ≥ 0. Thus, (34) becomes
Cb (Γ) = log2 1 + γ̂ + Γ̃ − σFBL (γ̂ + Γ̃)UFBL + Uδ
(37)
≥ log2 1 + γ̂ + Γ̃G − σFBL (γ̂ + Γ̃)UFBL .
(38)
When defining the right side of (38) as Cb,lower (Γ), the error probability ε can be bounded as
n
o
ε ≤ P Cb,lower (Γ) < r Γ̂ = γ̂ .
(39)
Naturally, the channel measurement error Γ̃ and its Gaussian approximation Γ̃G do not depend on
the noise in the data transmission phase. In addition, we assumed that the decoding performance
18
is not affected by imperfect CSI at the receiver. Therefore, Γ̃G and UFBL are considered to be
independent variables. To simplify the analysis, we make the following assumption:
Assumption 1. Inequality (39) holds when σFBL (γ̂ + Γ̃) is replaced by σFBL (γ̂), i.e.
n
o
ε ≤ P log2 1 + γ̂ + Γ̃G − σFBL (γ̂)UFBL < r
(40)
is assumed to hold for all parameters.
Motivation: When the estimated SNR γ̂ is larger than the actual SNR, then the variance is
replaced by a larger term, i.e., the finite blocklength effects are overestimated. Overestimating
the negative effects of finite blocklength coding should generally lead to an overestimation of
the error probability. On the other hand, when the estimated SNR γ̂ is smaller than the actual
SNR, then the channel is already better than predicted, and there is a high margin between the
actual capacity and the rate, so errors in this regime are very rare.
Lemma 3. If Assumption 1 holds, and if the estimated SNR γ̂ and the average SNR γ̄ are known,
then the error probability ε for a code with rate r is bounded as
γ̂ − (2r − 1) ∆ 0
ε≤Q
=ε ,
σIC,F (γ̂)
(41)
with
2
2
2
σIC,F
(γ̂) = σICSI
+ σFBL
(γ̂) .
(42)
2
Proof. The random variables Γ̃G ∼ N (0, σICSI
) and UFBL ∼ N (0, 1) are independent. Thus, the
2
2
difference Γ̃G − σFBL (γ̂)UFBL can be described by UIC,F ∼ N (0, σIC,F
(γ̂)), where σIC,F
(γ̂) is
the sum of the two variances. Then, starting from (40) we obtain:
n
o
ε ≤ P γ̂ + Γ̃G − σFBL (γ̂)UFBL < 2r − 1
= P {γ̂ + UIC,F < 2r − 1}
γ̂ − (2r − 1)
UIC,F
>
.
=P −
σIC,F (γ̂)
σIC,F (γ̂)
While all our numerical results confirmed that Assumption 1 holds, we are mainly interested in
an approximation of the error probability ε. The expression in (41) still provides an approximation
for ε even if Assumption 1 does not hold for some parameters. However, an upper bound for
19
ε can be very useful in the context of ultra-reliable low latency systems with rate adaptation,
specifically, when the transmitter wants to select a rate such that the error probability ε is below
a target error probability ε0 :
Corollary 2. Given an imperfect channel estimate γ̂ and a target error probability ε0 with
Q(γ̂/σIC,F (γ̂)) < ε0 < 1/2, the actual error probability ε is less than or equal ε0 if Assumption 1
holds and if the transmitter chooses the rate
rIC,F (γ̂, n, ε0 ) = log2 1 + γ̂ − σIC,F (γ̂)Q−1 (ε0 ) .
(43)
Proof. The proof follows by solving (41) for r, with ε0 > Q(γ̂/σIC,F (γ̂)) ensuring that r > 0.
D. Optimal Rate Adaptation
The problem addressed in this paper is finding the parameters of training sequence length m
and rates r(γ̂) (with corresponding values of n and ε) that minimize the upper bound (18) on
the delay violation probability pv (w). For the rate adaptation, we will show in this section that
Lemma 3 and Corollary 2 can be used to determine a nearly optimal solution.
As we already observed in Sec. III-A, an optimal rate adaptation strategy minimizes MS (1−θ)
for some parameter θ > 0, or equivalently, minimizes MS (θ) for some θ < 1. An approximate
solution can be found with Lemma 3, which provides an upper bound ε0 on the error probability
ε, resulting in an upper bound on MS (θ). However, instead of choosing the rates r(γ̂) and then
bounding the error probability, Corollary 2 allows choosing a target error probability ε0 for each
γ̂ and then computing the achievable rate rIC,F (γ̂, n, ε0 ), resulting in
MS (θ) ≤
Z∞
0
|
0
n·rIC,F (γ̂,n,ε0 )(θ−1)
(1 − ε )e
{z
g(γ̂,ε0 )
0
+ ε fΓ̂ (γ̂)dγ̂ .
}
(44)
The inequality is due to ε ≤ ε0 as established by Lemma 3. The optimal values of ε0 are the ones
which minimize the right-hand side of (44) for some parameter θ < 1. They can be found by
minimizing the term g(γ̂, ε0 ) over ε0 individually for each discretized value γ̂. Using observations
from [16], we find:
Lemma 4. g(γ̂, ε0 ) is convex in ε0 for Q(γ̂/σIC,F (γ̂)) < ε0 < 1/2 and θ < 1.
Proof. See the Appendix.
20
The convexity property establishes that the optimal ε0 for each γ̂ is unique and can be found
efficiently. The value of ε0 then directly provides the optimal rate rIC,F (γ̂, n, ε0 ) that should be
chosen by the transmitter.
Even though our numerical studies found no case where the approximation ε0 from Lemma 3
was below ε in (7), Lemma 3 relies on Assumption 1. Thus, after the optimal rate rIC,F (γ̂, n, ε0 )
has been found, the error probability ε can be computed from (7).
E. Further Uses of the Approximation
Beyond our own analysis, Lemma 3 could also be used for analyzing further system properties
and devising other system features:
1) Outdated CSI: Assume that the transmitter has an outdated observation Ĥold of a Rayleigh
fading channel, which is related to the current value H as Ĥold = ρH + Z with known ρ, H ∼
CN (0, 1), Z ∼ CN (0, 1−ρ2 ), and H and Z mutually independent. Furthermore, assume that the
receiver has perfect knowledge of the current channel H and that is knows (e.g. from additional
headers) which coding scheme was chosen by the transmitter. Then, the MMSE estimate of the
2
with
channel is given as Ĥold,MMSE = ρĤold . Replacing Ĥ in Sec. II with Ĥold,MMSE and σN
(1 − ρ2 ) leads to the same results, i.e., the error probability ε for this case can be computed
through (7) and approximated through Lemma 3.
2) Power Allocation: In certain cases, e.g., in battery-powered devices, the transmitter may
need to adapt the transmit power to the channel. Power control in the finite blocklength regime
was analyzed in [9] for perfect CSI and without considering the queueing performance. We
assume now for simplicity that the channel training is always performed with the same transmit
power, and that the SNR Γ and estimated SNR γ̂ are related to this training power. During the
data transmission phase, the transmitted signal power is scaled by a factor φ > 0 and thus the
SNR during the data transmission phase is changed to φΓ = φγ̂ + φΓ̃G + φΓ̃δ . We start again
by assuming, as in Assumption 1, that
n
o
ε ≤ P log2 1 + φγ̂ + φΓ̃G − σFBL (φγ̂)UFBL < r .
Following the same steps as in Sec. III-C, the error probability can be bounded as
φγ̂ − (2r − 1)
,
ε≤Q
σPA
(45)
(46)
21
10 0
ε for m = 10
ε′ for m = 10
ε for m = 25
ε′ for m = 25
Error Probability
10 -1
r = 0.95 ĉ
10 -2
10 -3
r = 0.9 ĉ
r = 0.75 ĉ
10 -4
4
6
8
10
12
14
16
18
20
Estimated SNR γ̂ (dB)
Fig. 2. Error probability ε and approximation/upper bound ε0 from Lemma 3 vs. estimated SNR γ̂, when the r(γ̂) = κ log2 (1+γ̂)
with κ ∈ {0.75, 0.9, 0.95}. Two choices of training length m ∈ {10, 25} are considered. n = 200, γ̄ = 15 dB.
2
2
2
(φγ̂). Although there is no closed-form solution for the minimal power
+σFBL
= φ2 σICSI
with σPA
scaling φ such that the error probability ε is below a certain target, one can quickly compute
(46) for different values of φ in order to determine the minimum required transmit power.
IV. N UMERICAL E VALUATION
For numerical evaluation, Sec. IV-A addresses the accuracy of the error probability approximation from Lemma 3, especially with respect to its use in rate adaptation. In Sec. IV-B, we
validate the accuracy of the system model itself. In Sec. IV-C, we compare the system performance without delay constraints with the performance under strict delay constraints. Finally,
in Sec. IV-D, we investigate the trade-off between training and data transmission time under
varying delay constraints.
A. Validation
In Fig. 2, we compare the error probability ε in (7) with its upper bound ε0 from Lemma 3.
The length of the training sequence is m ∈ {10, 25}, the average SNR γ̄ is 15 dB, and the length
of the data transmission phase is fixed at n = 200. In this example, the rate r(γ̂) is simply chosen
as a fraction κ of the estimated capacity ĉ = log2 (1 + γ̂), with κ ∈ {0.75, 0.9, 0.95}. First of
all, we confirm in all cases that ε0 is indeed an upper bound on ε, as expected from Lemma 3.
Second, even though we observe that this bound is not tight, especially when ε0 < 10−2 , it can
be seen that the upper bound ε0 predicts quite accurately how much the error probability ε will
22
change when reducing the rate or when increasing the number of training symbols m. As a
result, the bound/approximation ε0 may be accurate enough to decide how the rate r(γ̂) should
be adapted to the estimated SNR γ̂ and what training sequence length to choose for optimal
performance.
Specifically, the optimal delay performance can be reached by iterating over the parameter θ
and finding the rate adaptation which minimizes MS (θ) for each value of θ. For γ̄ = 15 dB,
n = 200, m = 25, and a specific choice3 of θ = 0.99, we compare in Fig. 3 the rate adaptation
based on the Corollary 2 as described in Sec. III-D (red dashed curve), with a rate adaptation
scheme that directly tries to minimize (20) by computing ε numerically for many different
values of r (black solid curve, labelled as perfectRA). We find that our proposed approxRA
scheme always selects a slightly lower rate than the perfectRA scheme. This is due to ε0 being
an upper bound to ε: the approximate rate adaptation scheme always overestimates the error
probability and then chooses a lower rate in order to avoid too many errors. However, as our
system model may not be accurate when ε becomes extremely small (see Sec. II-B), we will
from now on always restrict the approxRA scheme to choose only values ε0 ≥ 10−3 . The resulting
rates are shown in the dashed blue curve. We find that the difference between the perfectRA
and approxRA schemes is very small: the value of MS (θ) increases only slightly from 0.0291
to 0.0294. Furthermore, restricting approxRA to ε0 ≥ 10−3 is only relevant for γ̂ > 9 dB, and
has almost no effect on the value of MS (θ), as the Mellin transform depends mostly on the
behavior at low values of γ̂, where the error probabilities are much higher than 10−3 and the
data rates are small. In conclusion, even though our approximation is not tight, it can provide a
nearly optimal solution to the rate adaptation problem.
Additionally, Fig. 3 shows two suboptimal rate adaptation schemes. The green dash-dotted
curve shows the rate r that would be chosen when the transmitter always keeps the error
probability at a fixed value ε = 0.003 for all values of γ̂. The value of MS (θ) increases to
0.0394 for this scheme4 . The second suboptimal rate adaptation scheme (violet dotted curve)
is one that does not take the delay requirements into account, but optimizes the parameters to
achieve the maximum expected goodput r in (9). This scheme favors high data rates over high
3
For these parameters and an arrival rate ᾱ = 1.4 bits/symb., the bound K (θ1 , w) for target delay w = 5 was minimal at
θ1 ≈ 0.010, thus MS (θ) must be evaluated at θ = 1 − θ1 ≈ 0.99.
4
it is even higher for ε ∈ {0.0001, 0.001, 0.002, 0.01} and all other values we tested
23
Selected rate r in percent of ĉ
100%
RA without
adapting to delay
M S (θ) ≈ 0.04379
90%
80%
70%
approxRA with
ε′ ≥ 0.001
M S (θ) ≈ 0.02943
perfectRA
M S (θ) ≈ 0.02914
60%
approxRA
M S (θ) ≈ 0.02938
fixed ε = 0.003
M S (θ) ≈ 0.03935
50%
0
2
4
6
8
10
12
14
16
18
20
Estimated SNR γ̂ [dB]
Fig. 3. Choice of r(γ̂) in percent of ĉ = log2 (1 + γ̂) (with resulting values of MS (θ)) for different rate adaptation schemes.
m = 25, n = 200, γ̄ = 15 dB, θ = 0.99.
reliability, and causes MS (θ) to increase to 0.0438. Due to the massive increases in MS (θ), we
suspect that the delay performance will deteriorate with both suboptimal schemes.
This suspicion is confirmed by Fig. 4. It shows the delay violation probability pv (w), which
can be obtained by simulating the queueing system with random instances of the service and
arrival process, and its analytical upper bound (18), versus the target delay w for those different
rate adaptation schemes. We first note that while the upper bound (18) on pv (w) is not tight,
which was also observed in similar works [14], [17], the upper bound is very useful, as it
not only predicts the slope of pv (w) correctly, but also predicts which parameters (here: which
rate adaptation schemes) are optimal with respect to pv (w). We observe that the delay bounds
for the perfectRA (solid black curve) and approxRA (dashed red) are almost indistinguishable,
which is in line with the results in Fig. 3. The difference between the two schemes in pv (w) as
obtained from simulations is also not noticeable. Contrary to that, when using the suboptimal
schemes, which either use fixed ε = 0.003 or do not adapt the rate to the delay constraints,
the delay violation probability pv (w) at w = 4 degrades by nearly an order of magnitude, and
this degradation is correctly predicted by the analytical bounds. This suggests that it is quite
important to solve the rate adaptation problem optimally, taking the delay requirements into
account.
B. Validation of the System Model
While we assumed throughout the paper that the decoding error probability is exactly equal
to ε in (7), ε is itself only an approximation, and depends on the assumption of perfect CSIR.
Using [3, Cor. 3], we can numerically compute a strict lower bound on the achievable rate for a
24
Delay violation probability pv (w)
10 -1
Delay bound for perfectRA
Delay bound for approxRA
Delay bound for RA not adapted to delay
Delay bound for fixed ε = 0.003
pv (w) sim. for perfectRA
pv (w) sim. for approxRA
pv (w) sim. for RA not adapted to delay
pv (w) sim. for fixed ε = 0.003
10 -2
10 -3
10 -4
10 -5
10 -6
10 -7
10 -8
1
2
3
4
5
6
7
Target Delay w
Fig. 4. Target delay violation probability pv (w) (obtained from simulations over 1011 time steps) and its respective upper bound
(18) vs. target delay w. m = 25, n = 200, γ̄ = 15 dB, arrival rate ᾱ = 1.4 bits/symbol.
given error probability. For the error probability, we set the optimal value ε0 that was found with
the rate adaptation using Lemma 4. In Fig. 5, we compare the delay performance of a system
with this achievable rate to the delay performance of the original system model. In particular, we
compare the maximum possible size of arriving data packets per time slot such that the system
can still guarantee different quality-of-service constraints, versus the training sequence length
m. In order to guarantee a deadline of only w = 5 slots with pv (w) < 10−8 , only 200-300
bits (depending on the training length) should arrive in each time slot. The performance (i.e.,
the maximum arrival size) that can be guaranteed through [3, Cor. 3] can be 10% below the
performance predicted from our system model. Even as the training length m increases to 104 ,
which results in nearly perfect CSI at transmitter and receiver, this performance gap remains.
However, in case the CSI becomes perfect, we know that our system model converges to earlier
results by Polyanskiy et al. [1, Thm. 54], which were shown to be quite accurate. Therefore,
while [3, Cor. 3] provides a strict lower bound on the performance, it is presumably not a tight
bound when the CSI is nearly perfect. The following results assume again that the decoding
error probability is given by (7).
C. Performance under Delay Constraints
Next, we show the impact on the performance due to imperfect CSI and finite blocklength
under different delay constraints. When the system has to operate under strict delay constraints,
the transmitter should generally try to achieve high reliability, which may require spending more
time on channel training. In Fig. 6a we show the expected goodput r, i.e. the performance when
25
800
700
Maximum arrivals (bits/slot)
w = 20
600
500
w=7
400
300
w=5
200
100
0
10 1
System model
Ach. rate lower bound
10 2
10 3
10 4
Training length m
Fig. 5. Maximum arrivals A in bits per time slot vs. training length m for n = 200, γ̄ = 15 dB such that for target delay
w ∈ {5, 7, 20} slots, the delay constraint pv (w) < 10−8 is still satisfied.
there are no delay constraints, next to Fig. 6b, which shows the maximum supported arrival rate
ᾱ such that the system still meets strict delay constraints. In both cases, we show the performance
against the average SNR γ̄, for different channel models and different parameters. In all cases,
the total length nslot = m + n of one time slot is 250 symbols. First of all, Fig. 6a shows the
expected goodput. The black curve (“PCSI,IBL”) shows the performance for a simplified channel
model where the CSI is assumed to be perfect and transmissions are error-free at rate r equal
to the capacity. The black curve marked with ‘+’ (“PCSI,FBL”) shows the performance when
the CSI is still assumed to be perfect, but finite blocklength effects are taken into account. The
expected goodput decreases because some transmissions fail and also because the transmitter
must back off from the capacity and choose a smaller rate to get a low probability of error.
When the effects of imperfect channel state information are taken into account as well (two
blue curves, “ICSI,FBL”), the performance is even lower. This is again because a backoff is
required and because transmissions fail with a higher probability when the channel is unknown.
Another reason is that m symbols are used for channel estimation and thus fewer symbols are
used for data transmissions, which leads to lower normalized goodput. In this scenario, i.e. when
there are no delay constraints, we observe that the system performs better with m = 5 training
symbols than with m = 50 over a wide range of the average SNR γ̄. Here, the benefits of better
channel estimation at m = 50 cannot compensate for the reduced length of the data transmission
phase. Nevertheless, even when m = 50 symbols (20% of the time slot) are spent on training,
the performance is still better than the performance of a system that does not measure the
26
8
8
PCSI, IBL (n = 250)
PCSI, FBL (n = 250)
ICSI, FBL (m = 5, n = 245) approx. RA
ICSI, FBL (m = 50, n = 200) approx. RA
NoCSI, FBL (n = 250)
6
PCSI, IBL (n = 250)
PCSI, FBL (n = 250)
ICSI, FBL (m = 5, n = 245) approx. RA
ICSI, FBL (m = 50, n = 200) approx. RA
NoCSI, FBL (n = 250)
7
Maximum Arrival Rate (bits/symbol)
Expected Goodput (bits/symbol)
7
5
4
3
2
1
6
5
4
3
2
1
0
0
5
10
15
20
25
5
Average SNR γ̄ (dB)
(a)
10
15
20
25
Average SNR γ̄ (dB)
(b)
Fig. 6. (a) Expected goodput r vs. average SNR γ̄ and total slot length nslot = 250. (b) Maximum arrival rate ᾱ vs. average
SNR for nslot = 250 such that for target delay w = 5 slots, pv (w) < 10−8 .
channel (m = 0) and transmits at a fixed rate (red curve, “NoCSI,FBL”). However, note that
the differences are small, so this trend might change under different assumptions, for example
when the feedback of CSI is not instantaneous and error-free.
Fig. 6b shows the maximum arrival rate ᾱ per symbol if the upper bound on the delay violation
probability pv (w) for a target delay of w = 5 slots should not be higher than 10−8 , for the same
system parameters as in Fig. 6a. It can be seen that it is very difficult for the system to meet
these target requirements with imperfect CSI when the average SNR is low. The requirements
can only be satisfied by choosing an arrival rate ᾱ that is significantly lower than the expected
goodput. At low SNR, we also see that using m = 50 training symbols now leads to better
performance than m = 5. This means that under strict delay requirements, the higher reliability
gained through better channel estimation is more beneficial than additional data transmission
time. The trade-off between m and n depends on the delay constraints. Furthermore, we observe
that the fixed rate transmission scheme performs significantly worse than the schemes adapting
the rate to the imperfect measurement. Thus, rate adaptation seems to be beneficial especially
under tight delay constraints.
D. How much time to spend on training?
The previous Fig. 6b implies that with nslot = 250 and γ̄ = 15 dB, it is possible to meet the
QoS target pv (w = 5) < 10−8 when ᾱ ≈ 1.0 bits per symbol and m = 50. For m = 5, the
performance is worse, but what is the optimal value of m? Fig. 7 shows the delay bound for
27
10 -3
Delay violation probability
10 -4
ᾱ = 2.4
10 -5
ᾱ = 2.0
10
-6
ᾱ = 1.6
10 -7
ᾱ = 1.3
10
-8
10
-9
ᾱ = 1.0
0
10
20
30
40
50
60
70
80
Number of training symbols m
Fig. 7. Bounds on the delay violation probability pv (w) vs. number of training symbols m for target delay w = 5, average
SNR γ̄ = 15, nslot = 250 and different arrival rates (in bits per symbol). The minimum point (optimal m) is marked with ‘x’.
nslot = 250, γ̄ = 15 dB and different values of ᾱ versus the training length m. For ᾱ = 1.0, the
smallest delay violation probability is obtained at m = 33 training symbols, but the performance
remains similar for m between 20 and 50. On the other hand, when the arrival rate is increased,
fewer training symbols should be used; the delay violation probability easily increases by an
order of magnitude when the transmitter chooses too many (e.g. m = 50) training symbols.
Finally, Fig. 8 shows the optimal number of training symbols m for different delay requirements and different SNR levels. Here, we require that for all values of the target delay w, the
bound on the delay violation probability is pv (w) < 10−8 . The optimal m is then defined as
the value of m for which the system can support the highest arrival rate ᾱ while still satisfying
the delay constraints. We observe that when the delay requirements become very strict, a large
fraction of the available resources must be spent on training. On the other hand, when the
delay requirements become more relaxed, the optimal value of m is much smaller. Lastly, when
the SNR decreases, channel estimation becomes less reliable, and more training symbols are
required.
V. C ONCLUSIONS AND F UTURE W ORK
In this work, we studied the joint impact of rate adaptation with imperfect CSI at the transmitter
and finite blocklength channel coding on the delay performance of a wireless communication
system. Based on stochastic network calculus, we found that the transmitter must adapt the
rate not only to the channel estimate, but must also consider the delay requirements. In order
to find the optimal rate adaptation, we developed a closed-form approximation for the error
28
60
γ̄ = 10 dB
γ̄ = 15 dB
γ̄ = 20 dB
Training Length m
50
40
30
20
10
0
4
6
8
10
12
14
16
18
20
Target Delay w
Fig. 8. Optimal value of training length m when nslot = 250 against the target delay w, for different average SNR γ̄ ∈
{10, 15, 20} dB.
probability due to imperfect CSI and finite blocklength, which is invertible and can be used
beyond this specific analysis. Using this approximation, the optimal rate selection becomes a
convex problem. After validating various aspects of our work, we showed numerically that rate
adaptation typically outperforms approaches that are channel-agnostic (i.e. do not rely on CSI
at the transmitter). Furthermore, it could be shown that the optimal choice of training length
depends strongly on the average SNR as well as on the delay and reliability requirements of the
application. Making an optimal choice may reduce the delay violation probability of the system
by an order of magnitude. A possible extension of this work relates to multi-antenna systems,
which can offer higher reliability at the cost of even more channel estimation, while CSI at the
transmitter could also be used for beamforming.
A PPENDIX
To show convexity, we show that for fixed γ̂, the second derivative of g1 (ε0 ) = g(γ̂, ε0 ) is
strictly positive for θ < 1:
0
g1 (ε0 ) = (1 − ε0 )en·rIC,F (γ̂,n,ε )(θ−1) + ε0
(47)
n
= (1 − ε0 )(1 + γ̂ − σIC,F (γ̂)Q−1 (ε0 )) ln 2 (θ−1) + ε0
(48)
= (1 − ε0 )(a − bQ−1 (ε0 ))c + ε0
(49)
with constants a, b > 0 and c < 0. Due to ε0 > Q(γ̂/σIC,F (γ̂)), we have rIC,F (γ̂, n, ε0 ) > 0 and
(a − bQ−1 (ε0 )) > 1. The first derivative is given by
ġ1 (ε0 ) =(1 − ε0 )c(a − bQ−1 (ε0 ))c−1 (−bQ̇−1 (ε0 )) − (a − bQ−1 (ε0 ))c + 1.
(50)
29
The second derivative is given by
g̈1 (ε0 ) =(1 − ε0 )c(a − bQ−1 (ε0 ))c−1 (−bQ̈−1 (ε0 ))
+ (1 − ε0 )c(c − 1)(a − bQ−1 (ε0 ))c−2 (−bQ̇−1 (ε0 ))2
− 2c(a − bQ−1 (ε0 ))c−1 (−bQ̇−1 (ε0 )).
(51)
From [16], the derivatives of the inverse Q-function are:
√
Q−1 (ε0 )2
Q̇−1 (ε0 ) = − 2πe 2
−1 (ε0 )2
Q̈−1 (ε0 ) = 2πQ−1 (ε0 )eQ
(52)
(53)
Thus, for ε0 < 1/2, Q̇−1 (ε0 ) < 0 and Q̈−1 (ε0 ) > 0, and therefore g̈1 (ε0 ) > 0.
R EFERENCES
[1] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory,
vol. 56, no. 5, pp. 2307–2359, May 2010.
[2] W. Yang, G. Durisi, T. Koch, and Y. Polyanskiy, “Diversity versus channel knowledge at finite block-length,” in Proc.
IEEE Information Theory Workshop (ITW), Sep. 2012, pp. 572–576.
[3] ——, “Quasi-static SIMO fading channels at finite blocklength,” in Proc. IEEE Int. Symp. on Information Theory (ISIT),
Jul. 2013, pp. 1531–1535.
[4] ——, “Quasi-static multiple-antenna fading channels at finite blocklength,” IEEE Trans. Inf. Theory, vol. 60, no. 7, pp.
4232–4243, Jul. 2014.
[5] M. Medard, “The effect upon channel capacity in wireless communications of perfect and imperfect knowledge of the
channel,” IEEE Trans. Inf. Theory, vol. 46, no. 3, pp. 933–946, 2000.
[6] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?” IEEE Trans. Inf.
Theory, vol. 49, no. 4, pp. 951–963, 2003.
[7] J. Meng and E.-H. Yang, “Constellation and rate selection in adaptive modulation and coding based on finite blocklength
analysis and its application to lte,” IEEE Transactions on Wireless Communications, vol. 13, no. 10, pp. 5496–5508, 2014.
[8] E.-H. Yang and J. Meng, “New nonasymptotic channel coding theorems for structured codes,” IEEE Transactions on
Information Theory, vol. 61, no. 9, pp. 4534–4553, 2015.
[9] W. Yang, G. Caire, G. Durisi, and Y. Polyanskiy, “Optimum power control at finite blocklength,” IEEE Transactions on
Information Theory, vol. 61, no. 9, pp. 4598–4615, 2015.
[10] A. Lim and V. K. N. Lau, “On the fundamental tradeoff of spatial diversity and spatial multiplexing of MISO/SIMO links
with imperfect CSIT,” IEEE Trans. Wireless Commun., vol. 7, no. 1, pp. 110–117, 2008.
[11] V. K. N. Lau, M. Jiang, and Y. Liu, “Cross layer design of uplink multi-antenna wireless systems with outdated CSI,”
IEEE Trans. Wireless Commun., vol. 5, no. 6, pp. 1250–1253, 2006.
[12] D. Wu and R. Negi, “Effective capacity: a wireless link model for support of quality of service,” IEEE Trans. Wireless
Commun., vol. 2, no. 4, pp. 630–643, Jul. 2003.
[13] H. Al-Zubaidy, J. Liebeherr, and A. Burchard, “Network-layer performance analysis of multihop fading channels,”
IEEE/ACM Trans. Netw., vol. 24, no. 1, pp. 204–217, Feb. 2016.
30
[14] N. Petreska, H. Al-Zubaidy, R. Knorr, and J. Gross, “On the recursive nature of end-to-end delay bound for heterogenous
networks,” in IEEE Int. Conf. on Communications (ICC), Jun. 2015.
[15] P. Wu and N. Jindal, “Coding versus ARQ in fading channels: how reliable should the PHY be?” IEEE Trans. Commun.,
vol. 59, no. 12, pp. 3363–3374, Dec. 2011.
[16] M. C. Gursoy, “Throughput analysis of buffer-constrained wireless systems in the finite blocklength regime,” EURASIP
Journal on Wireless Communications and Networking, Dec. 2013.
[17] S. Schiessl, J. Gross, and H. Al-Zubaidy, “Delay analysis for wireless fading channels with finite blocklength channel
coding,” in Proc. 18th ACM Int. Conf. on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM).
ACM, 2015, pp. 13–22.
[18] J. Gross, “Scheduling with outdated CSI: Effective service capacities of optimistic vs. pessimistic policies,” in Proc. 20th
IEEE Int. Workshop on Quality of Service (IWQoS).
IEEE, 2012, pp. 1–9.
[19] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO achievable rates with downlink training and
channel state feedback,” IEEE Transactions on Information Theory, vol. 56, no. 6, pp. 2845–2866, 2010.
[20] A. H. Nuttall, “Some integrals involving the Q-M function,” IEEE Trans. Inf. Theory, vol. 21, no. 1, pp. 95–96, 1975.
[21] Y. Polyanskiy, “SPECTRE: short-packet communication toolbox,” 2016. [Online]. Available: https://github.com/yp-mit/
spectre
[22] M. Fidler, “A network calculus approach to probabilistic quality of service analysis of fading channels,” in Proc. IEEE
Global Telecommunications Conf. (GLOBECOM), Nov. 2006, pp. 1–6.
[23] M. Ozmen and M. C. Gursoy, “Throughput regions of multiple-access fading channels with Markov arrivals and QoS
constraints.” IEEE Wireless Commun. Letters, vol. 2, no. 5, pp. 499–502, 2013.
[24] S. Schiessl, F. Naghibi, H. Al-Zubaidy, M. Fidler, and J. Gross, “On the delay performance of interference channels,” in
IFIP Networking Conference, May 2016, pp. 216–224.
[25] H. Al-Zubaidy, G. Dán, and V. Fodor, “Performance of in-network processing for visual analysis in wireless sensor
networks,” in IFIP Networking Conference.
IEEE, 2015, pp. 1–9.
| 7cs.IT
|
arXiv:1704.04926v1 [math.ST] 17 Apr 2017
EXACT TESTS TO COMPARE CONTINGENCY TABLES
UNDER QUASI-INDEPENDENCE AND QUASI-SYMMETRY
C. BOCCI AND F. RAPALLO
Abstract. In this work we define log-linear models to compare several square
contingency tables under the quasi-independence or the quasi-symmetry model,
and the relevant Markov bases are theoretically characterized. Through Markov
bases, an exact test to evaluate if two or more tables fit a common model is
introduced. Two real-data examples illustrate the use of these models in different fields of applications.
1. Introduction
Complex models for contingency tables have received an increasing interest in
the last decades from researchers and practitioners in different fields, from Biology
to Medicine, from Economics to Social Science. For a general introduction to the
statistical models for contingency tables see for instance [1], [4] and [11]. Quasisymmetry and quasi-independence models are well known log-linear models for
square contingency tables. Starting from Caussinus in [6], several authors have
considered such models from the point of view of both theory and applications,
and it is impossible to give a complete account on all the papers where quasiindependence and quasi-symmetry are studied or used in data analysis. In the next
section we will recall the basic facts on the quasi-independence and quasi-symmetry
models, while for a full presentation and an historical overview the reader can refer
to [4] and [9]. Quasi-symmetry is also the topic of a special issue of the Annales de
la Faculté des Sciences de Toulouse, edited in 2002 by S. Fienberg and P. G. M.
van der Heijden [8].
Within Algebraic Statistics, quasi-independence and quasi-symmetry are very
important models for contingency tables, for several reasons. We briefly review why
the synergy between Algebraic Statistics and quasi-independence has been fruitful.
Firstly, Algebraic Statistics provides and exact goodness-of-fit test based on the
Diaconis-Sturmfels algorithm. Such test is very flexible when applied to complex
models, and it allows us to make exact inference also outside the basic independence
model, where the classical Fisher’s exact test is available. When the sample size
is small, the use of the asymptotic tests based on the chi-square approximation
of the test statistics may lead to wrong conclusions, and this fact is even more
relevant in this kind of models, where the asymptotics fails also with moderately
large sample sizes, see an example in [13]. Secondly, under quasi-independence it
is possible to fix the diagonal counts, or even to analyze incomplete tables where
the diagonal counts (or an arbitrary subset of cells) are undefined or unavailable.
2010 Mathematics Subject Classification. 62H17.
Key words and phrases. Algebraic Statistics, Markov bases, MCMC algorithms, Rater agreement, Social mobility tables.
1
2
C. BOCCI AND F. RAPALLO
To include structural zeros in the analysis, the notion of toric statistical model is a
generalization of log-linear model that permits to study also the boundary. Toric
models are described by non-linear polynomials, but in several cases it is possible
to describe the geometry of such models, or at least it is possible to write their
invariants through Computer Algebra systems. Quasi-independence and quasisymmetry from the point of view of Algebraic Statistics can be found in [13], [3],
[7], and [2]. Applications of quasi-symmetry to the problem of rater agreement in
biomedical experiments are presented in [14].
In this paper, we use classical techniques from Algebraic Statistics in order
to compare several contingency tables under the quasi-independence and quasisymmetry models. This is accomplished by the construction of a three-way table
and by the definition of suitable log-linear models in order to determine if two or
more tables fit a common quasi-independence (resp., quasi-symmetry) model, versus the alternative hypothesis that each table follows a specific quasi-independence
(resp., quasi-symmetry) model with its own parameters. A third model is also introduced, as its matrix representation is a well known object in Combinatorics, namely
the Lawrence lifting of a matrix. For the first two models, the relevant Markov bases
are computed theoretically using a distance-reducing argument, while for the third
model the Markov bases are characterized only in the case of two tables, and some
advices are presented to efficiently run the exact test in the general case. We show
two applications of these models on datasets coming from different areas: the first
example comes from a rater agreement problem in a biomedical experiments, while
the second one concerns the analysis of social mobility tables.
This research suggests several the future directions. From the point of view of
Algebra, it would be interesting to study such models when the starting model on
each layer is different from the quasi-independence and quasi-symmetry models, also
including the study of their ideals. From the point of view of Statistics, it would be
interesting to study the use of this technique to make inference on other measures of
mobility based on log-linear models, also including one-sided tests and their semialgebraic characterization. For an introductory overview of these measures, with
several examples from surveys in European countries, refer to [20] and [5].
The paper is organized as follows. In Sect. 2 we recall some definitions and basic
results about log-linear models and toric models, with special attention to quasiindependence and quasi-symmetry. In particular, we collect here several scattered
results on the Markov bases for these models. In Sect. 3 we show how to define
suitable log-linear models to compare two or more square tables. Given a base
log-linear model, we define new log-linear models through the specification of their
model matrices. For quasi-independence and quasi-symmetry on several tables, the
Markov bases are theoretically computed. Sect. 4 is devoted to the illustration of
two real-data examples.
2. Markov bases for quasi-independence and quasi-symmetry
In this section we recall some basic definitions and properties of log-linear models,
with special attention to quasi-independence and quasi-symmetry for square twoway tables. A probability distribution on a finite sample space X with K elements
is a normalized vector of K non-negative real numbers. Thus, the most general
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
probability model is the simplex
(
∆=
K
X
(p1 , . . . , pK ) : pk ≥ 0 ,
3
)
pk = 1
.
k=1
A statistical model is therefore a subset of ∆.
A classical example of finite sample space is the case of a multi-way contingency
table where the cells are the joint counts of two or more random variables with
a finite number of levels each. In the case of square two-way contingency tables,
where the sample space is usually written as a cartesian product of the form X =
{1, . . . , I} × {1, . . . , I} we will use the notation pi,j to ease the readability. In such
case, the two categorical variables are denoted with X and Y .
In the classical theory of log-linear models, under the Poisson sampling scheme
the cell counts are independent and identically distributed Poisson random variables
with expected values N p1 , . . . , N pK , where N is the sample size, and the statistical
model is constraints on the raw parameters p1 , . . . , pK . A model is log-linear if
the log-probabilities lie in an affine subspace of the vector space RK . Given d real
parameters α1 , . . . , αd , a log-linear model is described, apart from normalization,
through the equations:
(1)
log(pk ) =
d
X
Ak,r αr
r=1
for k = 1, . . . , K, where A is the model matrix (or design matrix, [12]). Exponentiating Eq. (1), we obtain the expression of the corresponding toric model
(2)
pk =
d
Y
Ak,r
ζr
r=1
for k = 1, . . . , K, where ζr = exp(αr ), r = 1, . . . , d, are new non-negative parameters. Allowing the ζr ’s to be non-negative (instead of strictly positive) leads us to
consider also the boundary of the models, with points having some entries equal to
zero. It follows that the model matrix A is also the matrix representation of the
minimal sufficient statistic of the model. The matrix representation of the toric
models as in Eq. (2) is widely discussed in, e.g., [15] and [7]. It is easy to see from
Eq. (1) that different model matrices with the same image as vector space generate
the same log-linear model.
In the two-way case, the simplest (and widely studied) log-linear model is the
independence model, which models the stochastic independence between the two
categorical variables X and Y . Its log-linear form is
(3)
(X)
log(pi,j ) = µ + αi
(Y )
+ βj
with the constraints
(4)
I
X
i=1
(X)
αi
= 0,
I
X
(Y )
βj
= 0.
j=1
Quasi-independence and quasi-symmetry are both derived from the independence model adding constraints on given subsets of cells (typically, the cells on the
main diagonal) and constraints on the symmetry of the table. Although quasiindependence can be defined for general rectangular tables with fixed counts on an
arbitrary subset of X , see [2], here we restrict our attention to the case of square
4
C. BOCCI AND F. RAPALLO
tables with fixed counts on the main diagonal. The log-linear form of the quasiindependence model is
(X)
(5)
log(pi,j ) = µ + αi
(Y )
+ βj
+ γi δi,j
where δi,j is the Kronecker delta. Also in the quasi-independence model the constraints in Eq. (4) hold.
The log-linear form of the quasi-symmetry model is
(X)
(6)
log(pi,j ) = µ + αi
(Y )
+ βj
+ γi,j
with the constraints
I
X
(X)
αi
= 0,
i=1
I
X
(Y )
βj
= 0,
γi,j = γj,i , i, j = 1, . . . , I .
j=1
(X)
(Y )
In Eq. (6), the αi
are the parameters of the row effect, the βj are the parameters of the column effect, while the parameters γi,j force the quasi-symmetry.
Comparing Equations (1) and (6) it is easy to explicitly write the model matrix Aqs
for the quasi-symmetry model. The first non-trivial example of quasi-independence
and quasi-symmetry models is the 3 × 3 case, and in this first case the two models
coincide as log-linear models. A possible choice is reported in Fig. 1.
Atqs
=
1
1
0
0
1
0
0
0
0
0
1
0
0
1
1
0
0
0
1
0
1
0
0
0
0
0
1
1
0
0
0
0
1
0
1
0
0
0
0
1
0
1
0
1
0
0
1
0
0
0
0
0
1
0
1
0
0
1
0
0
0
0
0
1
0
1
0
1
0
0
0
1
0
0
1
0
0
0
1
0
0
1
1
0
0
0
1
0
0
0
0
1
0
0
1
0
1
0
0
0
1
0
0
0
1
0
0
1
0
0
1
0
0
0
0
0
1
.
Figure 1. The model matrix of the quasi-symmetry model for I = 3.
Notice that in Atqs each column represents a cell of the table (the cells are ordered
lexicographically for convenience), and each row represents a parameter. Analyzing the structure of Atqs , the first 7 rows of Atqs form the model matrix of the
independence model Atind , while the last three rows of Atqs define one real parameter for each diagonal cell, and hence force the diagonal cells to be fitted exactly.
This parametrization is redundant, since the model has 1 degree of freedom, and
therefore 8 parameters are sufficient to describe the model. The ideal of the independence model with model matrix Aind is the set of all 2 × 2 minors of the table
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
5
of probabilities, i.e.,
(7)
IAind = Ideal(p1,1 p2,2 − p1,2 p2,1 , p1,1 p2,3 − p1,3 p2,1 , p1,1 p3,2 − p1,2 p3,1 ,
p1,1 p3,3 − p1,3 p3,1 , p1,2 p2,3 − p1,3 p2,2 , p1,2 p3,3 − p3,2 p2,3 ,
p2,1 p3,2 − p3,1 p2,2 , p2,1 p3,3 − p3,1 p2,3 , p2,2 p3,3 − p3,2 p2,3 ) ,
while for the quasi-independence model and for the quasi-symmetry model from
the matrix Aqs we have only one binomial:
IAqind = IAqs = Ideal(p1,2 p2,3 p3,1 − p1,3 p3,2 p2,1 ) .
From the ideals above one can easily derive the corresponding Markov bases. Given
a model matrix A, recall that a move is a table m with integer entries such that
Am = 0, and that a set of moves MA is a Markov basis if all fibers
FA,b = {f ∈ Nk : At f = b}
are connected. Following [7], from the point of view of computations, the easiest way
to build a Markov basis is to compute the binomials in a system of generators of the
+
−
toric ideal IA of A and to transform such binomials through the logs: xm −xm 7→
±m = ±(m+ − m− ). For instance, the ideal IAqs yields a Markov basis with only
two moves:
0 +1 −1
m = ± −1 0 +1 .
+1 −1 0
To conclude this section, we collect some results on Markov bases for quasiindependence and quasi-symmetry models to be found in [3], [2], [7], and [16].
A loop of degree r on X is an I × I a move m = ±mr (i1 , . . . , ir ; j1 , ..., jr ) for
1 ≤ i1 , . . . , ir ≤ I, 1 ≤ j1 , . . . , jr ≤ I, where mr (i1 , . . . , ir ; j1 , . . . , jr ) has entries
mi1 ,j1 = mi2 ,j2 = . . . = mir−1 ,jr−1 = mir ,jr = 1,
mi1 ,j2 = mi2 ,j3 = . . . = mir−1 ,jr = mir ,j1 = −1,
and all other elements are zero. The indices i1 , i2 , . . ., are all distinct, as well as the
indices j1 , j2 , . . ., i.e.
im 6= in and jm 6= jn for all m 6= n ,
A loop of degree 2, M2 (i1 , i2 ; j1 , j2 ), is called a basic move, and a loop mr is
called df 1 if its support does not contain the support of any other loop. A loop
mr is called a symmetric loop if {i1 , . . . , ir } = {j1 , . . . , jr }.
The first result concerns quasi-independence with possible structural zeros. Let
S be the set of structural zeros.
Proposition 2.1. The set of df 1 loops of degree 2, . . . , I with support on X \ S
forms a unique minimal Markov basis for I × I contingency tables under the quasiindependence model with possible structural zeros. When the fixed cells of the table
are located only on the main diagonal, the minimal Markov basis is formed by the
df 1 loops of degree 2 and 3.
For the quasi-symmetry model we have the following
Proposition 2.2. The set of symmetric loops of degree 3, . . . , I with support outside
the main diagonal form a unique minimal Markov basis for I × I contingency tables
with structural zeros under the quasi-symmetry model. Such set of moves is also a
Graver basis.
6
C. BOCCI AND F. RAPALLO
3. Comparison of several tables under quasi-independence and
quasi-symmetry
As outlined in the Introduction, in this section we define three log-linear models to compare two or more square tables under quasi-independence and quasisymmetry, and we study the corresponding Markov bases. Let us consider H tables (H ≥ 2) and define a three-way contingency table T by stacking the H tables.
Conversely, each original table is a layer of the table T . Let K 0 = HI 2 be the
number of cells of T . Since the definition of the new models can be done starting
form a generic log-linear model on the two-way table, we present the models is a
general context, and then we write the explicit log-linear representation in the case
of quasi-independence and quasi-symmetry in order to highlight the meaning of
such new models in our cases.
Definition 3.1. Let A be the model matrix of a log-linear model. We define three
log-linear models for T :
• Under the model M0 we assume that all the layers follow a common model
with model matrix A;
• Under the model M1 we assume that each layer of the table T follows
a model with model matrix A with its own parameters, without further
constraints;
• Under the model M2 we assume that each layer of the table T follows a
model with model matrix A with its own parameters, and with the additional constraint of fixed marginal sums over the layers.
The model matrices of M0 , M1 and M2 have a simple block structure. In fact:
At
1K
=
AtM0
···
..
At
.
1K
AtM1
=
t
A
..
.
AtM2
At
At
=
IK
..
.
..
.
t
A
IK
where 1K is a row vector of 1’s with length K, IK is a the identity matrix with
dimensions K × K and each empty block means a block filled with 0’s. The matrix
AtM2 is the H-th order Lawrence lifting of At and its properties in terms of Markov
and Graver bases have been studied in [18] and [17].
Writing explicitly the log-linear form of the three models in the case of quasiindependence we have the equations below. For M0 :
(X)
(8)
(Y )
(M0 )
log(pi,j,h ) = µ + µh + αi + βj + γi δi,j
PH
(X)
(Y )
with the constraint h=1 µh = 0 in addition to the constraints on αi and βj
naturally derived from the basic quasi-independence model in Eq. (5). The second
model M1 is defined by
(9)
(M1 )
(X)
(Y )
log(pi,j,h ) = µ + µh + αh,i + βh,j + γi,h δi,j
PH
(X)
(Y )
with the constraint h=1 µh = 0 in addition to the constraints on αh,i and βh,j
derived from the basic quasi-independence model in Eq. (5) and valid on each layer
of the table T . The third model M2 is defined by
(10)
(M2 )
(X)
(Y )
log(pi,j,h ) = µ + µh + µi,j + αi,h + βj,h + γi,h δi,j
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
7
PI
with the same constraints as in M1 plus the additional constraints
i=1 µi,j =
PI
0, j = 1, . . . , I and j=1 µi,j = 0, i = 1, . . . , I.
In the case of quasi-symmetry we obtain the expressions below. For M0 :
(X)
(11)
(Y )
(M0 )
log(pi,j,h ) = µ + µh + αi + βj + γi,j
PH
(X)
(Y )
with the constraint h=1 µh = 0 in addition to the constraints on αi , βj and
γi,j naturally derived from the basic quasi-symmetry model in Eq. (6). The second
model M1 is defined by
(X)
(12)
(Y )
(M1 )
log(pi,j,h ) = µ + µh + αh,i + βh,j + γi,j,h
PH
(X)
(Y )
with the constraint h=1 µh = 0 in addition to the constraints on αh,i , βh,j and
γi,j,h derived from the basic quasi-symmetry model in Eq. (6) and valid on each
layer of the table T . The third model M2 is defined by
(13)
(M2 )
(X)
(Y )
log(pi,j,h ) = µ + µh + µi,j + αi,h + βj,h + γi,j,h
PI
with the same constraints as in M1 plus the additional constraints
i=1 µi,j =
PI
0, j = 1, . . . , I and j=1 µi,j = 0, i = 1, . . . , I.
With a simple linear algebra argument, it easy to see that M0 ⊂ M1 ⊂ M2 .
As a consequence, these models can be used in two ways. They can be applied
separately, to define goodness-of-fit test to contrast the observed table with a given
model, or the can be used to define a test for nested models, see e.g. [1] where
several examples are introduced and discussed. In the next section we will focus on
tests for nested models.
Now, we compute the Markov bases for the models M0 , M1 and M2 defined
above. Let M be a Markov basis for quasi-independence (or quasi-symmetry). In
the proofs below we will make use of a distance-reducing argument, introduced in
[19].
For the model M0 , define the following two types of moves b:
type 1: fix a move m ∈ M and split it on the different layers with the condition
that with the condition that each row of m belongs to one layer.
type 2: choose integers 1 ≤ i1 < i2 ≤ I and 1 ≤ j1 < j2 ≤ I and define the moves
±b where b has zero coordinate except for
bi1 ,j1 ,h1
bi2 ,j2 ,h1
bi1 ,j1 ,h2
bi2 ,j2 ,h2
=1
= −1
= −1
=1
for 1 ≤ h1 < h2 ≤ H.
Consider as a set of moves
M 0 = B1 ∪ B 2
where Bi is the set of moves of type i.
Proposition 3.2. The set M0 above is a Markov basis for the model M0 .
Proof. Let v ∈ ZI×I×H . Then v ∈ KerZ (AtM0 ) if and only
P
PH
H
t
i) At
h=1 v•,•,h = 0 that is
h=1 v•,•,h ∈ KerZ (A );
PI
ii)
i,j=1 vi,j,h = 0 for all 1 ≤ h ≤ H
8
C. BOCCI AND F. RAPALLO
where ii) follows directly looking at the vectors of ones 1K in the definition of AtM0 .
Let u, v vectors with same value of the sufficient statistic, i.e. AtM0 u = AtM0 v.
We want to prove that there exists b ∈ B such that
u+b≥0
and
||u + b − v||1 < ||u − v||1 .
Since u and v are distinct with AtM0 u = AtM0 v then there exists a positive
entry in u − v, say ui1 ,j1 ,h1 − vi1 ,j1 ,h1 > 0. Suppose that such entry belongs to
the main diagonal, i.e., i1 = j1 . Then there exists another layer h2 such that
ui1 ,j1 ,h1 − vi1 ,j1 ,h1 < 0. Moreover, by the condition ii), there exists a positive entry
of u − v in the layer h2 and a negative entry in the layer h1 : ui2 ,j2 ,h1 − vi2 ,j2 ,h1 < 0
and ui3 ,j3 ,h2 − vi3 ,j3 ,h2 > 0 for some indices i2 , i3 , j2 , j3 . Now consider the move
b ∈ B2 defined by
bi1 ,j1 ,h1 = −1
bi1 ,j1 ,h2 = +1
bi3 ,j3 ,h2 = −1
bi3 ,j3 ,h1 = +1
that satisfies ||u − v||1 > ||u + b − v||1 .
Thus, we can consider only the case where u − v is zero on the main diagonal
of all layers. Let U and V be the sum over the H layers of u and v, respectively.
By condition i), U − V ∈ KerZ (At ). By Propositions 2.1 and 2.2, there exists a
distance reducing move m which is a df 1 loop (in the case of quasi-independence)
or a symmetric loop (in the case of quasi-symmetry):
||U + m − V ||1 < ||U − V ||1
Let I = {(i1 , j1 ), (i1 , j2 ), . . . , (it , jt ), (it , j1 )} be the set of the indices where m has
nonzero entries. Without loss of generality we can suppose that mi1 ,j1 = −1. The
move m is defined such that
mα,β = +1
mα,β = −1
if and only if
if and only if
Uα,β − Vα,β < 0
Uα,β − Vα,β > 0
for all (α, β) ∈ I, except at most for the last index (it , j1 ), where mit ,j1 = +1.
Now, we split m in the H layers. For this aim, notice that Uα,β − Vα,β < 0
implies that there exists a layer h such that uα,β,h − vα,β,h < 0, and similarly
Uα,β − Vα,β > 0 implies that there exists a layer h such that uα,β,h − vα,β,h > 0. We
then split m by putting the +1 in the layer with a negative entry of u − v and −1 in
the layer with a positive entry of u − v. This can be done for all (α, β) ∈ I \ (it , j1 ).
The last +1 can be assigned to an arbitrary layer. Thus we have a matrix b defined
by a sequence of indices
I 0 = {(i1 , j1 , h11 ), (i1 , j2 , h12 ), . . . , (it , jt , htt ), (it , j1 , ht1 )}
Now, we arrange the entries in such a way that the entries in each row belong to
the same layer. This will ensures that the split move b satisfies the condition ii)
above.
Let us consider the first row and the corresponding indices (i1 , j1 , h11 ), (i1 , j2 , h12 )
where b is nonzero. By construction, ui1 ,j1 ,h11 − vi1 ,j1 ,h11 > 0 and ui1 ,j2 ,h12 −
vi1 ,j2 ,h12 < 0. If h11 = h22 we have concluded, otherwise there is an entry (α, β, h12 )
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
9
in the layer h12 such that uα,β,h12 − vα,β,h12 > 0. Now, consider the preliminary
move b(p) ∈ B2 defined by
(p)
bα,β,h12 = −1
(p)
bi1 ,j1 ,h12 = +1
(p)
bi1 ,j1 ,h11 = −1
(p)
bα,β,h11 = +1.
If ui1 ,j1 ,h12 − vi1 ,j1 ,h12 < 0, then the move b(p) is distance-reducing and we have
(p)
concluded. Otherwise, ||u + b(p) − v||1 = ||u − v||1 and ui1 ,j1 ,h12 + bi1 ,j1 ,h12 −
vi1 ,j1 ,h12 > 0, and therefore we can move the −1 into the layer h12 . Now, the
remaining +1’s in b can be simply moved, row by row, in the same layer of the
corresponding −1 and this is enough to conclude: there is a move b ∈ B1 such that
||u + b(p) + b − v||1 < ||u − v||1 .
Example 3.3. If I = 3 and H = 2 an example of move of type B1 , for both the
quasi-independence and the quasi-symmetry models, is drawn in Fig. 2.
0
0
0
0
±
1
0
0
0
0
1
1
1
0
1
0
1
0
0
Figure 2. An example of split move of type B1 for the model M0 .
We remark that, while in the quasi-independence model the diagonal cells are
fixed, this is no longer true when two or more tables are compared under the model
M0 . In fact, the moves in B2 act also on the diagonal cells.
The model M1 is easy to analyze, and the moves are given by the following
proposition.
Proposition 3.4. The set of moves
M1 = {(m, 0, . . . , 0), (0, m, 0, . . . , 0), . . . , (0, . . . , 0, m) : m ∈ M}
is a Markov basis for M1 .
Example 3.5. If I = 3 and H = 2 the moves for the model M1 for quasi1
independence and quasi-symmetry are drawn in Fig. 3.
For the model M2 , we restrict to the case of two layers. Let us consider a Graver
basis G for the model matrix At , and consider the set of moves
L = {(m, −m) : m ∈ G} .
0
±
1
0
0
0
10
0
0
1
0
0
0
±
0
1
0
1
0
1
0
0
0
1
0
0
0
0
0
1
1
1
1
1
0
0
0
1
0
1
1
0
±
0
1
0
C. BOCCI AND F. RAPALLO
1
0
0
0
1
0
0
0
1
Figure 3. The moves for the model M1 .
0
0
0
Remark 1. For quasi-symmetry, the Markov basis from Proposition 2.2 is also a
1
0
Graver basis, while 1in the case
of quasi-independence
we need to consider the set
of all df 1 loops
of
degree
2,
.
.
.
,
I.
0
0
± 0
Proposition 3.6. The set L is a Graver basis (and thus also a Markov basis) for
2
the model M2 .
1
1
0
Proof. This follows
from
7.1 in [18].
0
0 Theorem
0
Example 3.7. For example, if I = 3 the are only two moves for the model M2 for
both the quasi-independence and the quasi-symmetry models. They are drawn in
Fig. 4.
1
0
0
1
1
1
2
1
±
1
0
1
1
1
1
0
1
1
0
0
Figure 4. The moves for the model M2 .
For more than 2 layers (i.e., for higher Lawrence configurations), the Markov
basis for M2 may be computed through 4ti2 [10], but the number of moves increases
rapidly with I and H. A valid alternative in this case is to run the Markov chain of
the Diaconis-Sturmfels algorithm without a Markov basis, as described in Chapter
16 of [2].
1
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
11
4. Examples
In this section we present two numerical examples where the comparison of two
quasi-symmetry tables may be used. Before introducing the two numerical examples, we briefly recall the Diaconis-Sturmfels algorithm, by adapting the notation
to the case of a test for nested models to contrast model M0 inside the model M1 .
For details on the Diaconis-Sturmfels algorithm, see [7].
Let f be the observed table of counts, and write f as a vector of length K 0
according to the row labels of AM0 . Moreover, let p be the vector of probabilities
associated to f . The test for nested models with null hypothesis H0 : p ∈ M0 ⊂ M1
versus H1 : p ∈ M1 can be done using the log-likelihood ratio statistic
!
K0
X
ˆ0k
f
,
G2 = 2
fk log
fˆ1k
k=1
where fˆ0k and fˆ1k are the maximum likelihood estimates of the expected cell counts
under the models M0 and M1 respectively. In the asymptotic theory, the value of
G2 must be compared with the quantiles of the chi-square distribution with the
appropriate number of degrees of freedom, depending on the dimensions of the
table.
We use here the Diaconis-Sturmfels algorithm based on a Markov basis M0 of
the model M0 . Given the observed table f , its reference set under M0 is
n
o
0
F(f ) = f 0 ∈ NK : AtM0 f 0 = AtM0 f
and it is connected using the algorithm below. At each step:
(1)
(2)
(3)
(4)
let f be the current table;
choose with uniform probability a move m ∈ M0 ;
define the candidate table as f+ = f + m;
generate a random number u with uniform distribution over [0, 1]. If f+ ≥ 0
and
H(f+ )
min 1,
>u
H(f )
then move the chain in f+ ; otherwise stay at f . Here H denotes the hypergeometric distribution on F(t).
After an appropriate burn-in-period and taking only tables at fixed times to reduce
correlation between the sampled tables, the proportion of sampled tables with test
statistics greater than or equal to the test statistic of the observed one is the Monte
Carlo approximation of p-value of the log-likelihood ratio test. The results in this
section are based on Monte Carlo samples of size 10, 000, and the corresponding
asymptotic p-values are displayed for comparison.
4.1. Rater agreement data. The data in Tab. 1 summarize the results of a medical experiment involving the evaluation of agreement among raters. Two independent raters are asked to assign a set of medical images to 5 different stages of a
disease (levels 1 to 5 in increasing order of severity of the disease). To check the
relevance of a thorough training of the raters, a first set of images has been classified
before the training session, while a second set has been classified after the training
session.
12
C. BOCCI AND F. RAPALLO
Table 1. Rater agreement data. Columns represent the grading
assigned by the first rater, rows represent the grading assigned by
the second rater. The data in the left panel have been collected
before the training session, the data in the right panel have been
collected after the training session.
1
2
3
4
5
1
2
3
4
5
10
4
0
1
0
2 1
4
8 4 1
0 10 3
1 4 11
1 3 3
0
1
1
0
10
1
2
3
4
5
1
2
3
4
5
16
0
1
0
0
5
15
3
2
2
1
2
14
0
0
0
1
1
14
3
0
0
1
3
14
We use the quasi-symmetry model, and the Markov basis M0 for the model M0
consists of 2 · 1004 moves (i.e., 1004 moves, each of them with the two signs). First,
we run two exact tests for the goodness-of-fit of the two tables separately under
quasi-symmetry, and we obtain exact p-values equal to 0.912 and 0.791 respectively
(G2 = 1.578 and G2 = 4.303 respectively, with 7 df). Running the test for a unique
quasi-symmetry model described in the previous section, the exact test gives a
p-value equal to 0.029 (G2 = 30.589 with 18 df, corresponding to an asymptotic
p-value equal to 0.014). While both the layers fit a quasi-symmetry model very
well, they do not fit a common quasi-symmetry model. This means that there are
significant differences in the classification of the medical images before and after
the training.
4.2. Social mobility data. As a second numerical example we consider the data
reported in Tab. 2 (adapted from [5] and originally collected during the “Italian
Household Longitudinal Survey”) where the inter-generational social mobility has
been recorded on a sample of 4, 343 Italian workers in 1997. The data take into
account the gender, and thus we have separate tables for men and women. There
are 4 categories of workers. A: “High level professionals”; B: “Employees and
commerce”; C: “Skilled working class and artisans”; D: “Unskilled working class”.
In [5] these data are analyzed extensively with a thorough presentation of a lot of
models to describe special patterns of mobility. Here we merely use the simplified
version displayed in Tab. 2 to show the practical applicability of the methodology
introduced in Sect. 3 also in this context.
Table 2. Table of social mobility in Italy (1997). Columns represent the father’s occupation, rows represent the son’s (or daughter’s) occupation. Male respondents in the left panel, female respondents in the right panel.
A
A
B
C
D
172
108
174
225
B
C
D
31
31 28
49
24 46
84 301 272
148 236 664
A
A
B
C
D
B
C
137 52 29
78 46 14
142 100 124
164 181 141
D
15
23
145
35
EXACT TESTS TO COMPARE CONTINGENCY TABLES UNDER QI AND QS
13
We use again the quasi-symmetry model, and the Markov basis M0 for the model
M0 is formed by 2 · 200 moves. If we consider the two layers separately, we obtain
exact p-values are equal to 0.051 and 0.088 respectively (G2 = 6.703 and G2 =
8.279 respectively, with 3 df). Although this fit may appear weak, nevertheless
the situation dramatically changes when considering the test for nested models.
Running the test for a unique quasi-symmetry model, the exact test produces a
p-value equal to 0 (G2 = 112.687 with 12 df, corresponding to an asymptotic
p-value less than 10−15 ), meaning that there is a strong departure from the null
hypothesis. Combining these results, one can conclude that the two genders present
strong differences in terms of patterns of mobility.
References
[1] Alan Agresti. Categorical Data Analysis. Wiley, New York, 3 edition, 2013.
[2] Satoshi Aoki, Hisayuki Hara, and Akimichi Takemura. Markov Bases in Algebraic Statistics.
Springer, New York, 2012.
[3] Satoshi Aoki and Akimichi Takemura. Markov chain Monte Carlo exact tests for incomplete
two-way contingency tables. J. Stat. Comput. Simul., 75(10):787–812, 2005.
[4] Yvonne M. Bishop, Stephen Fienberg, and Paul W. Holland. Discrete multivariate analysis:
Theory and practice. MIT Press, Cambridge, 1975.
[5] Richard Breen. Social Mobility in Europe. Oxford University Press, Oxford, 2007.
[6] Henri Caussinus. Contribution à l’analyse statistique des tableaux de corrélation. Ann. Fac.
Sci. Toulouse Math. (4), 29:77–183, 1965.
[7] Mathias Drton, Bernd Sturmfels, and Seth Sullivant. Lectures on Algebraic Statistics.
Birkhauser, Basel, 2009.
[8] Stephen E. Fienberg and Peter G. M. Van der Heijden. Introduction to special issue on
quasi-symmetry and categorical data analysis. Ann. Fac. Sci. Toulouse Math. (6), 11:439–
441, 2002.
[9] Leo A. Goodman. Contributions to the statistical analysis of contingency tables: Notes on
quasi-symmetry, quasi-independence, log-linear models, log-bilinear models, and correspondence analysis models. Ann. Fac. Sci. Toulouse, 11(4):525–540, 2002.
[10] Ralf Hemmecke, Raymond Hemmecke, and Peter Malkin. 4ti2 version 1.2—computation of
Hilbert bases, Graver bases, toric Gröbner bases, and more. Available at www.4ti2.de, 2005.
[11] Maria Kateri. Contingency Table Analysis. Methods and Implementation Using R. Springer,
New York, 2014.
[12] Giovanni Pistone, Eva Riccomagno, and Henry P. Wynn. Algebraic Statistics: Computational
Commutative Algebra in Statistics. Chapman&Hall/CRC, Boca Raton, 2001.
[13] Fabio Rapallo. Algebraic Markov bases and MCMC for two-way contingency tables. Scand.
J. Statist., 30(2):385–397, 2003.
[14] Fabio Rapallo. Algebraic exact inference for rater agreement models. Stat. Methods Appl.,
14(1):45–66, 2005.
[15] Fabio Rapallo. Toric statistical models: Parametric and binomial representations. Ann. Inst.
Statist. Math., 59(4):727–740, 2007.
[16] Fabio Rapallo and Ruriko Yoshida. Markov bases and subbases for bounded contingency
tables. Ann. Inst. Stat. Math., 62(4):785–805, 2010.
[17] Francisco Santos and Bernd Sturmfels. Higher Lawrence configurations. J. Combin. Theory,
Ser. A, 103(1):151–164, 2003.
[18] Bernd Sturmfels. Gröbner Bases and Convex Polytopes. Memoirs of the American Mathematical Society. American Mathematical Society, Providence, RI, 1996.
[19] Akimichi Takemura and Satoshi Aoki. Distance-reducing markov bases for sampling from a
discrete sample space. Bernoulli, 11(5):793–813, 2005.
[20] Yu Xie. The log-multiplicative layer effect model for comparing mobility tables. American
Sociological Review, 57:380–395, 1992.
14
C. BOCCI AND F. RAPALLO
Department of Information Engineering and Mathematics, University of Siena, Siena,
Italy
E-mail address: [email protected]
Department of Science and Technological Innovation, University of Piemonte Orientale, Alessandria, Italy
E-mail address: [email protected]
| 10math.ST
|
arXiv:1711.05376v2 [cs.CV] 16 Nov 2017
Sliced Wasserstein Distance for Learning Gaussian Mixture Models
Soheil Kolouri
HRL Laboratories, LLC
Gustavo K. Rohde
University of Virginia
Heiko Hoffmann
HRL Laboratories, LLC
[email protected]
[email protected]
[email protected]
Abstract
Moreover, GMMs could be used to approximate any density
defined on Rd with a large enough number of mixture components. To fit a finite GMM to the observed data, one is
required to answer the following questions: 1) how to estimate the number of mixture components needed to represent
the data, and 2) how to estimate the parameters of the mixture components. Several techniques have been introduced
to provide an answer for the first question [37]. The focus of
this paper in on the latter question.
Gaussian mixture models (GMM) are powerful parametric tools with many applications in machine learning and
computer vision. Expectation maximization (EM) is the
most popular algorithm for estimating the GMM parameters.
However, EM guarantees only convergence to a stationary
point of the log-likelihood function, which could be arbitrarily worse than the optimal solution. Inspired by the
relationship between the negative log-likelihood function
and the Kullback-Leibler (KL) divergence, we propose an
alternative formulation for estimating the GMM parameters
using the sliced Wasserstein distance, which gives rise to
a new algorithm. Specifically, we propose minimizing the
sliced-Wasserstein distance between the mixture model and
the data distribution with respect to the GMM parameters.
In contrast to the KL-divergence, the energy landscape for
the sliced-Wasserstein distance is more well-behaved and
therefore more suitable for a stochastic gradient descent
scheme to obtain the optimal GMM parameters. We show
that our formulation results in parameter estimates that are
more robust to random initializations and demonstrate that
it can estimate high-dimensional data distributions more
faithfully than the EM algorithm.
The existing methods to estimate the GMM parameters
are based on minimizing the negative log-likelihood (NLL)
of the data with respect to the parameters [51]. The Expectation Maximization (EM) algorithm [15] is the prominent
way of minimizing the NLL (though, see, e.g., as an alternative [38, 22]). While EM remains the most popular method
for estimating GMMs, it only guarantees convergence to
a stationary point of the likelihood function. On the other
hand, various studies have shown that the likelihood function has bad local maxima that can have arbitrarily worse
log-likelihood values compared to any of the global maxima
[22, 25, 2]. More importantly, Jin et al. [24] proved that with
random initialization, the EM algorithm will converge to a
bad critical point with high probability. This issue turns the
EM algorithm sensitive to the choice of initial parameters.
In the limit (i.e. having infinite i.i.d samples), minimizing
the NLL function is equivalent to minimizing the KullbackLeibler divergence between the data distribution and the
GMM, with respect to the GMM parameters. Here, we propose, alternatively, to minimize the p-Wasserstein distance,
and more specifically the sliced p-Wasserstein distance [28],
between the data distribution and the GMM. The Wasserstein
distance and its variations have attracted a lot of attention
from the Machine Learning (ML) and signal processing communities lately [28, 3, 16]. It has been shown that optimizing
with respect to the Wasserstein loss has various practical benefits over the KL-divergence loss [44, 16, 39, 3, 20]. Importantly, unlike the KL-divergence and its related dissimilarity
measures (e.g. Jensen-Shannon divergence), the Wasserstein distance can provide a meaningful notion of closeness
(i.e. distance) for distributions supported on non-overlapping
low dimensional manifolds. This motivates our proposed
1. Introduction
Finite Gaussian Mixture Models (GMMs), also called
Mixture of Gaussians (MoG), are powerful, parametric, and
probabilistic tools that are widely used as flexible models for
multivariate density estimation in various applications concerning machine learning, computer vision, and signal/image
analysis. GMMs have been utilized for: image representation [5, 17] to generate feature signatures, point set registration [24], adaptive contrast enhancement [9], inverse
problems including super-resolution and deblurring [19, 55],
time series classification [8], texture segmentation [43], and
robotic visuomotor transformations [23] among many others.
As a special case of general latent variable models, finite GMM parameters could serve as a concise embedding
[40], which provide a compressed representation of the data.
1
formulation for estimating GMMs.
To overcome the computational burden of the Wasserstein minimization for high-dimensional distributions, we
propose to use the sliced Wasserstein distance [6, 30, 28].
Our method slices the high-dimensional data distribution via
random projections and minimizes the Wasserstein distance
between the projected one-dimensional distributions with
respect to the GMM parameters. We note that the idea of
characterizing a high-dimensional distribution via its random
projections has been studied in various work before [52, 26].
The work in [26], for instance, minimizes the L1 norm between the slices of the data distribution and the GMM with
respect to the parameters. This method, however, suffers
from the same shortcomings as the KL-divergence based
methods.
The p-Wasserstein distances and more generally the optimal mass transportation problem have recently gained plenty
of attention from the computer vision and machine learning
communities [28, 49, 42, 29, 54, 47, 3]. We note that the
p-Wasserstein distances have also been used in regard to
GMMs, however, as a distance metric to compare various
GMM models [11, 34, 45]. Our proposed method, on the
other hand, is an alternative framework for fitting a GMM to
data via sliced p-Wasserstein distances.
In what follows, we first formulate the p-Wasserstein distance, the Radon transform, and the Sliced p-Wasserstein
distance in Section 2. In Section 3, we reiterate the connection between the K-means problem and the Wasserstein
means problem [21], extend it to GMMs, and formulate the
Sliced Wasserstein means problem. Our numerical experiments are presented in Section 4. Finally, we conclude our
paper in Section 5.
2. Preliminary
2.1. p-Wasserstein distance:
In this section we review the preliminary concepts and
formulations needed to develop our framework. Let Pp (Ω)
be the set of Borel probability measures with finite p’th
moment defined on a given metric space (Ω, d), and let ρ ∈
Pp (X) and ν ∈ Pp (Y ) be probability measures defined on
X, Y ⊆ Ω with corresponding probability density functions
Ix and Iy , dρ(x) = Ix (x)dx and dν(y) = Iy (y)dy. The
p-Wasserstein distance for p ∈ [1, ∞) between ρ and ν is
defined as the optimal mass transportation (OMT) problem
[53] with cost function c(x, y) = dp (x, y), such that:
p1
Z
p
Wp (ρ, ν) =
inf
d (x, y)dγ(x, y)
, (1)
γ∈Γ(ρ,ν)
X×Y
where Γ(ρ, ν) is the set of all transportation plans, γ ∈
Γ(ρ, ν), and satisfy the following:
γ(A × Y ) = ρ(A) for any Borel subset A ⊆ X
.
γ(X × B) = ν(B) for any Borel subset B ⊆ Y
Due to Brenier’s theorem [7], for absolutely continuous
probability measures ρ and ν (with respect to Lebesgue measure) the p-Wasserstein distance can be equivalently obtained
from,
Z
1
Wp (ρ, ν) = (inf f ∈M P (ρ,ν)
dp (f (x), x)dρ(x)) p
(2)
X
where, M P (ρ, ν) = {f : X → Y | f# ρ = ν} and f# ρ
represents the pushforward of measure ρ,
Z
Z
dρ(x) =
dν(y) for any Borel subset A ⊆ Y.
f −1 (A)
A
When a transport map exists, the transport plan and the
transport map are related via, γ = (Id × f )# ρ. Note that in
most engineering and computer science applications Ω is a
compact subset of Rd and d(x, y) = |x − y| is the Euclidean
distance. By abuse of notation we will use Wp (ρ, ν) and
Wp (Ix , Iy ) interchangeably throughout the manuscript. For
a more detailed explanation of the Wasserstein distances
and the optimal mass transport problem, we refer the reader
to the recent review article by Kolouri et al. [28] and the
references there in.
One-dimensional distributions: The case of onedimensional continuous probability measures is specifically
interesting as the p-Wasserstein distance has a closed form
solution. More precisely, for one-dimensional probability
measures there exists a unique monotonically increasing
transport map that pushesR one measure into another. Let
x
Jx (x) = ρ((−∞, x]) = −∞ Ix (τ )dτ be the cumulative
distribution function (CDF) for Ix and define Jy to be the
CDF of Iy . The transport map is then uniquely defined as,
f (x) = Jy−1 (Jx (x)) and consequently the p-Wasserstein
distance is calculated as:
Z
p1
p
−1
Wp (ρ, ν) =
d (Jy (Jx (x)), x)dρ(x)
X
Z
1
=
d
p
(Jy−1 (z), Jx−1 (z))dz
p1
(3)
0
where in the second line we used the change of variable
Jx (x) = z. The closed form solution of the p-Wasserstein is
an attractive property that gives rise to the Sliced-Wasserstein
(SW) distances. Next we review the Radon transform, which
enables the definition the Sliced p-Wasserstein distance.
2.2. Radon transform
The d-dimensional Radon transform, R, maps a functionR I ∈ L1 (Rd ) where L1 (Rd ) := {I : Rd →
R| Rd |I(x)|dx ≤ ∞} to the set of its integrals over the
hyperplanes of Rd and is defined as,
Z
RI(t, θ) :=
I(x)δ(t − x · θ)dx
(4)
Rd
For all θ ∈ Sd−1 where Sd−1 is the unit sphere in Rd . Note
that R : L1 (Rd ) → L1 (R × Sd−1 ). For the sake of completeness, we note that the Radon transform is an invertible,
linear transform and we denote its inverse as R−1 , which is
also known as the filtered back projection algorithm and is
defined as:
I(x)
=
=
R−1 (RI(t, θ))
Z
(RI(., θ) ∗ h(.)) ◦ (x · θ)dθ
(5)
Sd−1
where h(.) is a one-dimensional filter with corresponding
Fourier transform Fh(ω) = c|ω|d−1 (it appears due to the
Fourier slice theorem, see [41] for more details) and ‘∗’ denotes convolution. Radon transform and its inverse are extensively used in Computerized Axial Tomography (CAT) scans
in the field of medical imaging, where X-ray measurements
integrate the tissue-absorption levels along 2D hyper-planes
to provide a tomographic image of the internal organs. Note
that in practice acquiring infinite number of projections is
not feasible therefore the integration in Equation (5) is replaced with a finite summation over projection angles. A
formal measure theoretic definition of Radon transform for
probability measures could be find in [6].
Radon transform of empirical PDFs: The Radon transform of Ix simply follows Equation (4). However, in most
machine learning applications we do not have access to the
distribution Ix but to its samples, xn . Kernel density estimation could be used in such scenarios to approximate Ix from
its samples,
Nρ
1 X
Ix (x) ≈
φ(x − xn )
Nρ n=1
R
where φ : Rd → R+ is a density kernel where Rd φ(x)dx =
1 (e.g. Gaussian kernel). The Radon transform of Ix can
then be approximated from its samples via:
Nρ
1 X
RIx (t, θ) ≈
Rφ(t − xn · θ, θ)
Nρ n=1
(6)
Note that certain density kernels have analytic Radon
transformation. For instance when φ(x) = δ(x) the Radon
transform Rφ(t, θ) = δ(t).
Radon transform of multivariate Gaussians: Let
φ(x) = Nd (µ, Σ) be a d-dimensional multivariate Gaussian
distribution with mean µ ∈ Rd and covariance Σ ∈ Rd×d .
A slice/projection of the Radon transform of φ is then a onedimensional normal distribution Rφ(·, θ) = N1 (θ·x, θT Σθ).
Given the linearity of the Radon transform, this indicates
that a slice of a d-dimensional GMM is a one-dimensional
GMM with component means θ · µi and variance θT Σi θ.
2.3. Sliced p-Wasserstein Distance
The idea behind the sliced p-Wasserstein distance is to
first obtain a family of marginal distributions (i.e. onedimensional distributions) for a higher-dimensional probability distribution through linear projections (via Radon
transform), and then calculate the distance between two input
distributions as a functional on the p-Wasserstein distance of
their marginal distributions. In this sense, the distance is obtained by solving several one-dimensional optimal transport
problems, which have closed-form solutions. More precisely,
the Sliced Wasserstein distance between Ix and Iy is defined
as,
Z
1
SWp (Ix , Iy ) = (
Wpp (RIx (., θ), RIy (., θ))dθ) p (7)
Sd−1
The Sliced p-Wasserstein distance as defined above is symmetric, and it satisfies sub-additivity and coincidence axioms,
and hence it is a true metric [30].
The sliced p-Wasserstein distance is especially useful
when one only has access to samples of a high-dimensional
PDFs and kernel density estimation is required to estimate
I. One dimensional kernel density estimation of PDF slices
is a much simpler task compared to direct estimation of I
from its samples. The catch, however, is that as the dimensionality grows one requires larger number of projections to
estimate I from RI(., θ). In short, if a reasonably smooth
two-dimensional distribution can be approximated by its L
projections (up to an acceptable reconstruction error, ), then
one would require O(Ld−1 ) number of projections to approximate a similarly smooth d-dimensional distribution (for
d ≥ 2). In later sections we show that the projections could
be randomized in a stochastic Gradient descent fashion for
learning Gaussian mixture models.
3. Sliced Wasserstein Means and Gaussian
Mixture Models
Here we first reiterate the connection between the Kmeans clustering algorithm and the Wasserstein means problem, and then extend this connection to GMMs and state the
need for the sliced Wasserstein distance. Let yn ∈ Rd for
n = 1, ..., N be N samples and Y = [y1 , ..., yN ] ∈ Rd×N .
The K-means clustering algorithm seeks the best K points,
xk ∈ Rd for k = 1, ..., K and X = [x1 , ..., xK ] ∈ Rd×K ,
that represent Y . Formally,
1
kY − XC T k2
N
s.t. C1K = 1N , ci,j ∈ {0, 1}
inf C,X
(8)
×K
where C ∈ RN P
contains the one-hot labels of Y .
N
1
Let Iy = N n=1 φ(y−yn ) be the empirical distribution
of Y , where φ is a kernel density estimator (e.g. radial
basis function kernel or the Dirac delta function in its limit).
Then, the K-means problem can be alternatively solved by
minimizing a statistical distance/divergence between Iy and
PK
1
Ix = K
k=1 φ(x − xk ). A common choice for such
distance/divergence is the Kullback-Leibler divergence (KLdivergence) [4, 10]. Alternatively, the p-Wasserstein distance
could be used to estimate the parameters of Ix ,
inf Ix Wpp (Ix , Iy )
(9)
We discuss the benefits of the p-Wasserstein distance
over the KL-divergence in the next sub-section. Above
minimization is known as the Wasserstein means problem
and is closely related to the Wasserstein Barycenter problem [1, 46, 14, 21]. The main difference being in that in
these works the
Pgoal is to find a measure ν∗ such that
ν∗ = arg inf ν k Wpp (νk , ν), where νk are sets of given
low dimensional distributions (2 or 3D images or point
clouds). The strategy in [1, 46, 14] could also be extended
into a clustering problem, though the two formulations are
still significantly different given the inputs into the wasserstein distance being very different. Note also that K-means
is equivalent to a variational EM approximation of a GMM
with isotropic Gaussians [36], therefore, a natural extension
of the Wasserstein means problem could be formulated to fit
a general GMM to Iy . To do so, we let distribution Ix to be
the parametric GMM as follows:
Ix (x) =
X
d
k
(2π) 2
αk
1
exp(− (x − µk )T Σ−1
p
k (x − µk ))
2
det(Σk )
Figure 1. The corresponding energy landscapes for the negative
log-likelihood and the Wasserstein Means problem for scenario 1
(a) and scenario 2 (b). The energy landscapes are scaled and shifted
for visualization purposes.
P
where k αk = 1 and Equation (9) is solved to find µk s,
Σk s, and αk s. Next we describe the benefits of using the
Wasserstein distance in Equation (9) to fit a general GMM to
the observed data compared to the common log-likelihood
maximization schemes.
3.1. Wasserstein Means vs.
Likelihood
Maximum Log-
General GMMs are often fitted to the observed data points,
yn s, via maximizing the log-likelihood of samples with respect to Ix . Formally, this is written as:
max
µk ,Σk ,αk
N
1 X
log(Ix (yn ))
N n=1
(10)
It is straightforward to show that in the limit and as the
number of samples grows to infinity, N → ∞, the maximum
log-likelihood becomes equivalent to minimizing the KLdivergence between Ix and Iy (See supplementary material
for a proof):
lim
max
N →∞ µk ,Σk ,αk
N
1 X
log(Ix (yn )) = min KL(Ix , Iy )
µk ,Σk ,αk
N n=1
The p-Wasserstein distance has been shown to have certain benefits over the commonly used KL-divergence and its
related distances/divergences (i.e., other examples of Bregman divergences including the Jensen-Shannon (JS) distance
and Itakura-Saito distance) [3]. For a general GMM, the
model Ix is continuous and smooth (i.e. infinitely differentiable) in its parameters and is locally Lipschitz; therefore,
Wp (Ix , Iy ) is continuous and differentiable everywhere,
while this is not true for the KL-divergence. In addition,
in scenarios where the distributions are supported by low
dimensional manifolds, KL-divergence and other Bregman
divergences may be difficult cost functions to optimize given
their limited capture range. This limitation is due to their
‘Eulerian’ nature, in the sense that the distributions are compared at fixed spatial coordinates (i.e., bin-to-bin comparison
in discrete measures) as opposed to the p-Wasserstein distance, which is ‘Lagrangian’, and morphs one distribution
to match another by finding correspondences in the domain
of these distributions (i.e., Wasserstein distances perform
cross-bin comparisons).
To get a practical sense of the benefits of the Wasserstein
means problem over the maximum log-likelihood estimation, we study two simple scenarios. In the first scenario, we
generate N one-dimensional samples, yn , from a normal dis-
∂f (t,θ)
∂t RIy (f (t, θ), θ)
Figure 2. Illustration of the high-level approach for the SlicedWasserstein Means of GMMs.
= RIx (t, θ). Note that, since Ix is
an absolutely continuous PDF, an optimal transport map will
exist even when Iy is not an absolutely continuous PDF (e.g.
when φ(y) = δ(y)) . Moreover, since the slices/projections
are one-dimensional the transport map, f (., θ), is uniquely
defined and the minimization on f has a closed form solution
and is calculated from Equation (3). The Radon transformations in Equation (11) are:
PN
1
RIy (t, θ) ≈ N n=1 Rφ(t − yn · θ, θ)
(12)
2
P
k ·θ)
RIx (t, θ) =
√ αk
)
exp(− (t−µ
k
2θ T Σk θ
T
2πθ Σk θ
tribution N (0, σ) where we assume known σ and visualize
the negative log-likelihood (NLL) and the Wasserstein means
(WM) problem as a function of µ. Figure 1 (a) shows the first
scenario and the corresponding energy landscapes for the
negative log-likelihood and the Wasserstein means problem.
It can be seen that while the global optimum is the same for
both problems, the Wasserstein means landscape is less sensitive to the initial point, hence a gradient descent approach
would easily converge to the optimal point regardless of the
starting point. In the second scenario, we generated N samples, yn , from a mixture of two one-dimensional Gaussian
distributions. Next, we assumed that the mixture coefficients
αk s and the standard deviations σk s, for k ∈ {0, 1}, are
known and visualized the corresponding energy landscapes
for NLL and WM as a function of µk s (See Figure 1 (b)). It
can be clearly seen that although the global optimum of both
problems is the same, but the energy landscape of the Wasserstein means problem does not suffer from local minima and
is much smoother.
The Wasserstein means problem, however, suffers from
the fact that the W22 (., .) is computationally expensive to
calculate for high-dimensional Ix and Iy . This is true even
using very efficient OMT solvers, including the ones introduced by Cuturi [13], Solomon et al. [48], and Levy [32].
3.2. Sliced Wasserstein Means
We propose to use an approximation of the p-Wasserstein
distance between Ix and Iy using the SW distance. Substituting the Wasserstein distance in Equation (9) with the
SW distance leads to the Sliced p-Wasserstein Means (SWM)
problem,
inf
µk ,Σk ,αk
SWpp (Ix , Iy )
Z
=
Wpp (RIx (., θ), RIy (., θ))dθ,
Sd−1
which can be written as:
Z
Z
inf
inf f (.,θ) |f (t, θ) − t|p RIx (t, θ)dtdθ
µk ,Σk ,αk
Sd−1
The new formulation avoids the optimization for calculating the Wasserstein distance and enables an efficient
implementation for clustering high-dimensional data. Figure 2 demonstrates the high-level idea behind slicing
high-dimensional PDFs Ix and Iy and minimizing the pWasserstein distance between these slices over GMM components. Moreover, given the high-dimensional nature of the
problem estimating density Iy in Rd requires large number
of samples, however, the projections of Iy , RIy (., θ), are
one dimensional and therefore it may not be critical to have
large number of samples to estimate these one-dimensional
densities.
Optimization scheme: To obtain a numerical optimization scheme, which minimizes the problem in Equation (11)
we first discretize the set of directions/projections. This
corresponds to the use of a finite set Θ ∈ Sd−1 , and a minimization of the following energy function,
|Θ|
inf
µk ,Σk ,αk
1 X
|Θ|
l=1
Z
|f (t, θl ) − t|p RIx (t, θl )dt
(13)
R
A fine sampling of Sd−1 is required for Equation (13) to
be a good approximation of (11). Such sampling, however,
becomes prohibitively expensive for high-dimensional data.
Alternatively, following the approach presented in [6] we
utilize random samples of Sd−1 at each minimization step
to approximate the Equation (11). This leads to a stochastic
gradient descent scheme where instead of random sampling
of the input data, we random sample the projection angles.
Finally, the GMM parameters are updated through an EMlike approach where for fixed GMM parameters we calculate
the optimal transport map f between random slices of Ix
and Iy , followed by updating Ix for fixed transport maps
f (., θ)s. Below we describe these steps:
1. Generate L random samples from S(d−1) , {θ1 , ..., θL }.
R
(11)
where for a fixed θ, f (., θ) is the optimal transport
map between RIx (., θ) and RIy (., θ), and satisfies
2. Fix the GMM, Ix , and calculate the optimal transport
map between slices RIx (·, θl ) and RIy (·, θl ) via:
f (t, θl ) = RJy−1 (RJx (t, θl ), θl )
where RJx(y) (·, θl ) is the CDF of RIx(y) (·, θl ).
(14)
Figure 3. Results of 100 runs of EM-GMM and SW-GMM fitting a model with 10 modes to the ring-line-square dataset, showing four
samples of random initializations (Top) and histograms across all 100 runs for the negative log-likelihood of the fitted model and the
sliced-Wasserstein distance between the fitted model and the data distribution (Bottom).
3. For fixed transportmaps, f (·, θl )s, update the GMM
parameters by differentiating Equation (11):
∂SWpp
∂αk
∂SWpp
∂µk
∂SWpp
∂Σk
L Z
X
|f (t, θl ) − t|p
(t − µk · θl )2
)dt
exp(−
q
2θlT Σk θl
2πθlT Σk θl
l=1 R
Z
L
X
αk |f (t, θl ) − t|p
(t − µk · θl )2
=
exp(−
)
q
2θlT Σk θl
T
R
2πθl Σk θl
l=1
!
(t − µk · θl )
dt
θl
θlT Σk θl
Z
L
X
αk |f (t, θl ) − t|p (t − µk · θl )2
=
[
− 1]
q
θlT Σk θl
R
8π(θ T Σ θ )3
l=1
=
l
k l
exp(−
!
(t − µk · θl )2
T
)dt (θl θl )
T
2θl Σk θl
where the summation is over L random projections
θl ∈ Sd−1 . We use the RMSProp optimizer [50], which
provides an adaptive learning rate, to update the parameters of the GMM according to the gradients
4. Project the updated Σk s onto the positive
P semidefinite
cone, and renormalize αk s to satisfy k αk = 1.
Notice that the derivative with respect to the k’th component of the mixture model in Equation (15) is independent
of other components. In addition, the transport map for each
projection, f (·, θ), in Equation (14) is calculated independent of the other projections. Therefore the optimization can
be heavily parallelized in each iteration. We note that, we
have also experimented with the Adam optimizer [27] but
did not see any improvements over RMSProp. The detailed
update equations are included in the Supplementary materials. In what follows we show the SWM solver for estimating
GMM parameters in action.
4. Numerical Experiments
We ran various experiments on three datasets to test our
proposed method for learning GMM parameters. The first
dataset is a two-dimensional data-point distribution consisting a ring, a square, and a connecting line (See Figure 3). To
show the applicability of our method on higher-dimensional
datasets we also used the MNIST dataset [31] and the CelebFaces Attributes Dataset (CelebA) [35].
4.1. Robustness to initialization
We started by running a simple experiment to demonstrate
the robustness of our proposed formulation to different initializations. In this test, we used a two-dimensional dataset
consisting of a ring, a square, and a line connecting them.
For a fixed number of modes, K = 10 in our experiment,
we randomly initialized the GMM. Next, for each initialization, we optimized the GMM parameters using the EM
algorithm as well as our proposed method. We repeated this
experiment 100 times.
Figure 4. Qualitative performance comparison on the MNIST dataset between our method, SW-GMM, and EM-GMM, showing decoded
samples for each mode (Right). Modes with bad samples are shown in red. The GMM was applied to a 128-dimensional embedding space
(Left).
Figure 3 shows sample results of the fitted GMM models
for both algorithms (Top Row). Moreover, we calculated the
histograms of the negative log-likelihood of the fitted GMM
and the sliced-Wasserstein distance between the fitted GMM
and the empirical data distribution (bottom). It can be seen
that our proposed formulation provides a consistent model
regardless of the initialization. In 100% of initializations,
our method achieved the optimal negative log-likelihood,
compared to only 29% for EM-GMM. In addition, both the
negative log-likelihood and the sliced-Wasserstein distance
for our method are smaller than those of the EM algorithm,
indicating that our solution is closer to the global optimum
(up to permutations of the modes).
4.2. High-dimensional datasets
We analyzed the performance of our proposed method
in modeling high-dimensional data distributions, here, using the MNIST dataset [31] and the CelebA dataset [35].
To capture the nonlinearity of the image data and boost
the applicability of GMMs, we trained an adversarial deep
convolutional autoencoder (Figure 4, Left) on the image
data. Next, we modeled the distribution of the data in the
embedded space via a GMM. The GMM was then used to
generate samples in the embedding space, which were consequently decoded to generate synthetic (i.e. ’fake’) images.
In learning the GMM, we compared the EM algorithm with
our proposed method, SW-GMM. We note that the entire
pipeline is in an unsupervised learning setting. Figure 4
demonstrates the steps of our experiment (Left) and provides
a qualitative measure of the generated samples (Right) for
the MNIST dataset. It can be seen that the SW-GMM model
leads to more visually appealing samples compared to the
EM-GMM. In addition, we trained a CNN classifier on the
MNIST training data. We then generated 10,000 samples
from each GMM component and classified these samples
to measure the fidelity/pureness of each component. Ideally, each component should only be assigned to a single
digit. We found out that for EM-GMM the components were
in average 80.48% pure, compared to 86.98% pureness of
SW-GMM components.
Similarly, a deep convolutional autoencoder was learned
for the CelebA face dataset, and a GMM was trained in
the embedding space. Figure 5 shows samples generated
from GMM components learned by EM and by our proposed
method (The samples generated from all components is attached in the Supplementary materials). We note that, Figures 4 and 5 only provide qualitative measures of how well
the GMM is fitting the dataset. Next we provide quantitative
measures for the fitness of the GMMs for both methods.
We used adversarial training of neural networks [18, 33]
to provide a goodness of fitness of the GMM to the data
distribution. In short, we use success in fooling an adversary network as an evaluation metric for goodness of fit of a
GMM. A deep discriminator/classifier was trained to distinguish whether a data point was sampled from the actual data
distribution or from the GMM. The fooling rate (i.e. error
rate) of such a discriminator is a good measure of fitness for
the GMM, as a higher error rate translates to a better fit to
Fooling rate
EM-GMM
SW-GMM
Ring-Square-Line
46.83% ± 1.14%
47.56% ± 0.86%
MNIST
24.87% ± 8.39%
41.91% ± 2.35%
CelebA
10.37% ± 3.22%
31.83% ± 1.24%
Figure 6. A deep discriminator is learned to classify whether an
input is sampled from the true distribution of the data or via the
GMM. The fooling rate of such a discriminator corresponds to the
fitness score of the GMM.
datasets (i.e. the MNIST and CelebA datasets).
The details of the architectures used in our experiments
are included in the supplementary material.
5. Discussion
Figure 5. Qualitative performance comparison between our method,
SW-GMM (Bottom), and EM-GMM (Top), showing decoded samples for several GMM components. The images are contrast enhanced for visualization purposes.
the distribution of the data. Figure 6 shows the idea behind
this experiment, and reports the fooling rates for all three
datasets used in our experiments. Note that the SW-GMM
consistently provides a higher fooling rate, indicating a better fit to the datasets. Moreover, we point out that while for
the low-dimensional ring-square-line dataset both methods
provide reasonable models for the dataset, the SW-GMM
significantly outperforms EM-GMM for higher-dimensional
In this paper, we proposed a novel algorithm for estimating the parameters of a GMM via minimization of the
sliced p-Wasserstein distance. In each iteration, our method
projects the high-dimensional data distribution into a small
set of one-dimensional distributions utilizing random projections/slices of the Radon transform and estimates the
GMM parameters from these one-dimensional projections.
While we did not provide a theoretical guarantee that the
new method is convex, or that it has fewer local minima,
the empirical results suggest that our method is more robust compared to KL-divergence-based methods, including the EM algorithm, for maximizing the log-likelihood
function. Consistent with this finding, we showed that the
p-Wasserstein metrics result in more well-behaved energy
landscapes. We demonstrated the robustness of our method
on three datasets: a two-dimensional ring-square-line distribution and the high-dimensional MNIST and CelebA face
datasets. Finally, while we used deep convolutional encoders
to provide embeddings for two of the datasets and learned
GMMs in these embeddings, we emphasize that our method
could be applied to other embeddings including the original
data space.
6. Acknowledgement
This work was partially supported by NSF (CCF
1421502). The authors gratefully appreciate countless fruitful conversations with Drs. Charles H. Martin and Dejan
Slepćev.
[16]
[17]
References
[1] M. Agueh and G. Carlier. Barycenters in the Wasserstein
space. SIAM Journal on Mathematical Analysis, 43(2):904–
924, 2011. 4
[2] C. Améndola, M. Drton, and B. Sturmfels. Maximum likelihood estimates for gaussian mixtures are transcendental. In
International Conference on Mathematical Aspects of Computer and Information Sciences, pages 579–590. Springer,
2015. 1
[3] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein generative adversarial networks. In International Conference on
Machine Learning, pages 214–223, 2017. 1, 2, 4
[4] A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh. Clustering with bregman divergences. Journal of machine learning
research, 6(Oct):1705–1749, 2005. 4
[5] C. Beecks, A. M. Ivanescu, S. Kirchhoff, and T. Seidl. Modeling image similarity by gaussian mixture models and the
signature quadratic form distance. In Computer Vision (ICCV),
2011 IEEE International Conference On, pages 1754–1761.
IEEE, 2011. 1
[6] N. Bonneel, J. Rabin, G. Peyré, and H. Pfister. Sliced and
Radon Wasserstein barycenters of measures. Journal of Mathematical Imaging and Vision, 51(1):22–45, 2015. 2, 3, 5
[7] Y. Brenier. Polar factorization and monotone rearrangement
of vector-valued functions. Communications on pure and
applied mathematics, 44(4):375–417, 1991. 2
[8] W. M. Campbell, D. E. Sturim, and D. A. Reynolds. Support
vector machines using gmm supervectors for speaker verification. IEEE signal processing letters, 13(5):308–311, 2006.
1
[9] T. Celik and T. Tjahjadi. Automatic image equalization and
contrast enhancement using gaussian mixture modeling. IEEE
Transactions on Image Processing, 21(1):145–156, 2012. 1
[10] K. Chaudhuri and A. McGregor. Finding metric structure in
information theoretic clustering. In COLT, volume 8, page 10,
2008. 4
[11] Y. Chen, T. T. Georgiou, and A. Tannenbaum. Optimal
transport for gaussian mixture models. arXiv preprint
arXiv:1710.07876, 2017. 2
[12] F. Chollet et al. Keras. https://github.com/
fchollet/keras, 2015. 11
[13] M. Cuturi. Sinkhorn distances: Lightspeed computation of
optimal transport. In Advances in Neural Information Processing Systems, pages 2292–2300, 2013. 5
[14] M. Cuturi and A. Doucet. Fast computation of wasserstein
barycenters. In International Conference on Machine Learning, pages 685–693, 2014. 4
[15] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum
likelihood from incomplete data via the em algorithm. Journal
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
of the royal statistical society. Series B (methodological),
pages 1–38, 1977. 1
C. Frogner, C. Zhang, H. Mobahi, M. Araya, and T. A. Poggio.
Learning with a wasserstein loss. In Advances in Neural
Information Processing Systems, pages 2053–2061, 2015. 1
J. Goldberger, S. Gordon, and H. Greenspan. An efficient
image similarity measure based on approximations of kldivergence between two gaussian mixtures. In null, page
487. IEEE, 2003. 1
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. Courville, and Y. Bengio. Generative
adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 7
J. A. Guerrero-Colón, L. Mancera, and J. Portilla. Image
restoration using space-variant gaussian scale mixtures in
overcomplete pyramids. IEEE Transactions on Image Processing, 17(1):27–41, 2008. 1
I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and
A. Courville. Improved training of wasserstein gans. arXiv
preprint arXiv:1704.00028, 2017. 1
N. Ho, X. Nguyen, M. Yurochkin, H. H. Bui, V. Huynh, and
D. Phung. Multilevel clustering via wasserstein means. arXiv
preprint arXiv:1706.03883, 2017. 2, 4
H. Hoffmann. Unsupervised Learning of Visuomotor Associations, volume 11 of MPI Series in Biological Cybernetics.
Logos Verlag Berlin, 2005. 1
H. Hoffmann, W. Schenck, and R. Möller. Learning visuomotor transformations for gaze-control and grasping. Biological
Cybernetics, 93:119–130, 2005. 1
B. Jian and B. C. Vemuri. Robust point set registration using
gaussian mixture models. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 33(8):1633–1645, 2011.
1
C. Jin, Y. Zhang, S. Balakrishnan, M. J. Wainwright, and M. I.
Jordan. Local maxima in the likelihood of gaussian mixture
models: Structural results and algorithmic consequences. In
Advances in Neural Information Processing Systems, pages
4116–4124, 2016. 1
A. T. Kalai, A. Moitra, and G. Valiant. Disentangling gaussians. Communications of the ACM, 55(2):113–120, 2012.
2
D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6
S. Kolouri, S. R. Park, M. Thorpe, D. Slepcev, and G. K. Rohde. Optimal mass transport: Signal processing and machinelearning applications. IEEE Signal Processing Magazine,
34(4):43–59, 2017. 1, 2
S. Kolouri and G. K. Rohde. Transport-based single frame
super resolution of very low resolution face images. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 4876–4884, 2015. 2
S. Kolouri, Y. Zou, and G. K. Rohde. Sliced-Wasserstein
kernels for probability distributions. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 4876–4884, 2016. 2, 3
Y. LeCun. The mnist database of handwritten digits.
http://yann. lecun. com/exdb/mnist/. 6, 7
[32] B. Lévy. A numerical algorithm for L2 semi-discrete optimal transport in 3D. ESAIM Math. Model. Numer. Anal.,
49(6):1693–1715, 2015. 5
[33] J. Li, W. Monroe, T. Shi, A. Ritter, and D. Jurafsky. Adversarial learning for neural dialogue generation. arXiv preprint
arXiv:1701.06547, 2017. 7
[34] P. Li, Q. Wang, and L. Zhang. A novel earth mover’s distance
methodology for image matching with gaussian mixture models. In Proceedings of the IEEE International Conference on
Computer Vision, pages 1689–1696, 2013. 2
[35] Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face
attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 6, 7
[36] J. Lücke and D. Forster. k-means is a variational em approximation of gaussian mixture models. arXiv preprint
arXiv:1704.04812, 2017. 4
[37] G. J. McLachlan and S. Rathnayake. On the number of components in a gaussian mixture model. Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery, 4(5):341–
355, 2014. 1
[38] R. Möller and H. Hoffmann. An extension of neural gas to
local PCA. Neurocomputing, 62:305–326, 2004. 1
[39] G. Montavon, K.-R. Müller, and M. Cuturi. Wasserstein
training of restricted boltzmann machines. In Advances in
Neural Information Processing Systems, pages 3718–3726,
2016. 1
[40] K. P. Murphy. Machine learning: a probabilistic perspective.
MIT press, 2012. 1
[41] F. Natterer. The mathematics of computerized tomography,
volume 32. Siam, 1986. 3
[42] S. R. Park, S. Kolouri, S. Kundu, and G. K. Rohde. The cumulative distribution transform and linear pattern classification.
Applied and Computational Harmonic Analysis, 2017. 2
[43] H. Permuter, J. Francos, and I. Jermyn. A study of gaussian
mixture models of color and texture features for image classification and segmentation. Pattern Recognition, 39(4):695–706,
2006. 1
[44] G. Peyré, J. Fadili, and J. Rabin. Wasserstein active contours.
In Image Processing (ICIP), 2012 19th IEEE International
Conference on, pages 2541–2544. IEEE, 2012. 1
[45] Y. Qian, E. Vazquez, and B. Sengupta. Deep geometric retrieval. arXiv preprint arXiv:1702.06383, 2017. 2
[46] J. Rabin, G. Peyré, J. Delon, and M. Bernot. Wasserstein
barycenter and its application to texture mixing. In Scale
Space and Variational Methods in Computer Vision, pages
435–446. Springer, 2012. 4
[47] A. Rolet, M. Cuturi, and G. Peyré. Fast dictionary learning
with a smoothed wasserstein loss. In Artificial Intelligence
and Statistics, pages 630–638, 2016. 2
[48] J. Solomon, F. de Goes, P. A. Studios, G. Peyré, M. Cuturi,
A. Butscher, A. Nguyen, T. Du, and L. Guibas. Convolutional
Wasserstein distances: Efficient optimal transportation on
geometric domains. ACM Transactions on Graphics (Proc.
SIGGRAPH 2015), to appear, 2015. 5
[49] M. Thorpe, S. Park, S. Kolouri, G. K. Rohde, and D. Slepčev.
A transportation lˆ p distance for signal analysis. Journal of
Mathematical Imaging and Vision, 59(2):187–210, 2017. 2
[50] T. Tieleman and G. Hinton. Lecture 6.5-rmsprop: Divide
the gradient by a running average of its recent magnitude.
COURSERA: Neural networks for machine learning, 4(2):26–
31, 2012. 6, 11
[51] M. E. Tipping and C. M. Bishop. Mixtures of probabilistic
principal component analyzers. Neural Computation, 11:443–
482, 1999. 1
[52] S. S. Vempala. The random projection method, volume 65.
American Mathematical Soc., 2005. 2
[53] C. Villani. Optimal transport: old and new, volume 338.
Springer Science & Business Media, 2008. 2
[54] S. Xiao, M. Farajtabar, X. Ye, J. Yan, L. Song, and H. Zha.
Wasserstein learning of deep generative point process models.
arXiv preprint arXiv:1705.08051, 2017. 2
[55] G. Yu, G. Sapiro, and S. Mallat. Solving inverse problems
with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE Transactions on Image Processing, 21(5):2481–2499, 2012. 1
7. Supplementary material
7.3. Experimental Details
7.1. Maximum log-likelihood and KL-divergence
Here we provide the detailed implementation and architecture of the adversarial autoencoders we used in our paper.
Figure 8 shows the detailed architectures of the deep adversarial autoencoder for MNIST and CelebA datasets. The
architecture of the deep binary classifiers used for scoring
the fitness of the GMMs are shown in Figure 7. We used
Keras [12] for implementation of our experiments.
For the loss function of the autoencoder we used the mean
absolute error between the input image and the decoded image together with the adversarial loss of the decoded image
(equally weighted). The loss functions for training the adversarial networks and the binary classifiers were chosen to be
cross entropy. Finally, we used RMSProp [50] as the default
optimizer for all the models, and trained the models over
100 Epochs, with batch size of 250.
The KL-divergence between Ix and Iy is defined as:
Z
KL(Ix , Iy ) =
Iy (ρ)log(
Rd
Iy (ρ)
)dρ
Ix (ρ)
For the maximum log-likelihood and in the limit and as the
number of samples grows to infinity, N → ∞, we have:
lim argmax
N →∞ µk ,Σk ,αk
N
1 X
log(Ix (yn )) =
N n=1
Z
argmax
µk ,Σk ,αk
Iy (ρ)log(Ix (ρ))dρ =
Rd
Z
argmin −
Iy (ρ)log(Ix (ρ))dρ =
µk ,Σk ,αk
Rd
Z
argmin
Iy (ρ)log(Iy (ρ))dρ −
µk ,Σk ,αk Rd
Z
Iy (ρ)log(Ix (ρ))dρ =
Rd
Z
Iy (ρ)
)dρ =
argmin
Iy (ρ)log(
Ix (ρ)
µk ,Σk ,αk Rd
7.4. CelebA Generated Images
Figure 9 shows all the GMM components learned by EM
and our SWM formulation.
argmin KL(Ix , Iy )
µk ,Σk ,αk
7.2. RMSProp update equations
SGD often suffers from oscillatory behavior across the
slopes of a ravine while only making incremental progress
towards the optimal point. Various momentum based methods have been introduced to adaptively change the learning
rate of SGD and dampen this oscillatory behavior. In our
work, we used RMSProp, introduced by Tieleman and Hinton [50], which is a momentum based technique for SGD.
Let α be the learning rate, γ ∈ (0, 1) be the decay parameter,
and κ be the momentum parameter. The update equation for
a GMM parameter, here µk , is then calculated from:
(i)
∂SWpp (Ix ,Iy )
(i−1)
mk = γmk
+ (1 − γ)
∂µk
∂SWpp (Ix ,Iy ) 2
(i)
(i−1)
+ (1 − γ)(
)
∂µk
gk = γgk
∂SWpp (Ix ,Iy )
(i)
(i−1)
vk = κvk
− q (i) α
∂µk
gk −(mik )2 +
(i)
(i−1)
(i)
µk = µk
+ vk
(15)
Where mk and gk are the first and second moments
(uncentered variance) of ∂SWpp (Ix , Iy )/∂µk , respectively.
Similar update equations hold for Σk and αk .
Figure 7. Details of the convolutional autoencoders learned for the MNIST and CelebA face dataset
Figure 8. Details of the deep binary classifiers used for scoring the fitness of GMMs.
(a)
(b)
Figure 9. GMM Samples Generated from the GMM learned from EM-GMM (a), and from SW-GMM (b). Each column depicts random
samples from a single component of the GMM.
| 1cs.CV
|
Spike and Tyke, the Quantized Neuron Model
M. A. El-Dosuky1, M. Z. Rashad1, T. T. Hamza1, and A.H. EL-Bassiouny2
1
Department of Computer Sciences, Faculty of Computers and Information sciences, Mansoura University,
Egypt
[email protected]
[email protected]
[email protected]
2
Department of Mathematics, Faculty of Sciences, Mansoura University, Egypt
[email protected]
Abstract
Modeling spike firing assumes that spiking statistics is Poisson, but real data violates this
assumption. To capture non-Poissonian features, in order to fix the inevitable inherent
irregularity, researchers rescale the time axis with tedious computational overhead instead of
searching for another distribution! Spikes or action potentials are precisely-timed changes in
the ionic transport through synapses adjusting the synaptic weight, successfully modeled and
developed as a memristor. Memristance value is multiples of initial resistance. This reminds us
with the foundations of quantum mechanics. We try to quantize potential and resistance, as
done with energy.
After reviewing Planck curve for blackbody radiation, we propose the quantization equations.
We introduce and prove a theorem that quantizes the resistance. Then we define the tyke
showing its basic characteristics. Finally we give the basic transformations to model spiking and
link an energy quantum to a tyke. Investigation shows how this perfectly models the neuron
spiking., with over 97% match.
Keywords:
Spike Timing Dependent Plasticity, STDP, synaptic weights, Spike, Tyke
1. Introduction
For a long time, Hebb rule provides the basic conceptualization and learning algorithm for
altering connection weights in neural network models [1]. Hebbian learning is only sensitive to
the spatial correlations between pre- and postsynaptic neurons ([2], [3]). Spike-timing
dependent plasticity (STDP) is temporally asymmetric form of Hebbian learning ([4], [5], [6]).
STDP allows the demonstration of associative memory [9] and the construction of powerful
brain-like circuitry ([7], [8]). There are many models for neuron spiking ([9], [10], [11], [12], [13]).
Spikes or action potentials are precisely-timed changes in the ionic transport through synapses
adjusting the synaptic weight [14]. A biological synapse successfully modeled and developed
as a memristor in HP labs [15].
The Lagrange formalism of memory circuit elements shows how to approach the quantization of
a general memory element circuit [16]. Memristors can be used to generate higher harmonics,
as an example of the advantage of memristor circuits over circuits without memory[17].
The memristor is a state-based device, i.e., a pulse response depends on its starting point.
Memristance at time t, denoted as MT, is proved to be a function of flux φ [18],
2η∆Rφ (t )
M T (t ) = R0 1 −
(1)
Q0 R02
where, R0 is the initial resistance, η is the bias sign, ∆R is the difference between the maximum
and the minimum resistance. Q0 is the charge required to pass through the memristor.
According to the STDP rule, synaptic weight wij between a pre-synaptic neuron j and a neuron i
after a time step ∆t is, ([19],[ 20]):
wij (t + ∆t ) = wij (t ) + ∆wij (t )
(2),
where
∆wij = µ sgn(tˆi − tˆ j ) exp( − | tˆi − tˆ j | / τ d )
(3)
where tˆ j and tˆi are the relative timing of the pre- and post-synaptic spikes respectively, sgn(·) is
the signum function, |·| is the absolute value, and τd is a reference time delay.
The problem statement of this research can be summarized as follows. Spike train generation
assumes that spiking statistics is Poisson, but real data violates this assumption [21]. To
capture non-Poissonian features, in order to fix the inevitable inherent irregularity, researchers
rescale the time axis with tedious computational overhead ([22], [23]) instead of searching for
another distribution!
We claim that we get the best model of spike train modeling and generation based on the
following notice. From equation (1) we notice that memristance value is multiples of initial
resistance. This reminds us with the foundations of quantum mechanics, when Planck found
that atoms in a heated solid had energies that were multiples of a fixed amount, called a
quantum [24]. Bohr explained the line spectrum of hydrogen, as in the energy diagram [24],
shown in figure 1.
E↑
0
E4
E3
E2
E1
...
n=∞
n=4
n=3
n=2
n=1
Excited states
Ground state
Figure 1 Energy Diagram
We try to quantize resistance, as done with energy. First we review Planck curve for blackbody
radiation. Then we propose the quantization equations. Investigation shows how this perfectly
models the neuron spiking, with the introduction of a new term called Tyke.
2. Planck curve for blackbody radiation
Planck investigated black-body radiation that hot objects glow. He derived an expression that
perfectly describes the relation between the wavelength of radiation and the emitted power [25]:
2πhc 2
Pλ = 5 hc / λkT
(4)
λ (e
− 1)
where Pλ is power per m² area per m wavelength, h is Planck's constant (6.626 x 10-34 J.s) , c
is the speed of Light (3 x 108 m/s) , λ is the wavelength (m) , k is Boltzmann Constant (1.38 x
10-23 J/K), T is temperature (K). The curve is shown in Figure 2.
Planck assumed that atoms in the black body act as electromagnetic oscillators that emit
radiation with a frequency ν. These oscillators can only have energies given by [25]
E = nhυ , n = 1, 2, 3… (5)
Initially, at n =1, the smallest possible energy (hν) is the quantum of energy.
Figure 2, Relation between Wavelength and Intensity
The code listing that regenerates this curve is given below.
Code listing 1: Plank Curve
h = 6.6261*10^-34; % Planck's constant J. s
c = 2.9979*10^8; % speed of light m/s
k = 1.3807*10^-23; % Boltzmann's constant J/K
lambda = 1e-9:10e-9:3000e-9;
A1=(h.*c)./(k.* 4500.*lambda);
A2=(h.*c)./(k.* 6000.*lambda);
A3=(h.*c)./(k.* 7500.*lambda);
quantum =(8.*pi.*h.*c)./lambda.^5;
curve1= quantum.*(1./(exp(A1)-1));
curve2= quantum.*(1./(exp(A2)-1));
curve3= quantum.*(1./(exp(A3)-1));
plot(lambda, curve1,'-.b',lambda, curve2, 'g--',lambda, curve3, '-ro')
xlabel('Wavelength (\lambda)','FontSize',16)
ylabel('Intensity','FontSize',16)
title('\it{Relation Between Wavelength and Intensity}','FontSize',16)
hleg1 = legend('T_1 = 4500','T_2 = 6000', 'T_3 = 7500');
3. Spike and Tyke Model
Let us first introduce and prove the following theorem that quantizes the resistance. Then we
define the tyke showing its basic characteristics. Finally we give the basic transformations to
model spiking and link an energy quantum to a tyke.
nh
Theorem 1: At any time, the resistance is R = 2 where n = 1, 2, 3, …., an integer.
Q
Proof: The power is the rate of generating or consuming energy.
E
(6)
P=
t
Power relates to current and voltage as it is equal to their multiplication at any time.
P=I×V
(7)
Using Ohm's Law: V = I × R and substituting in equation (7) to get:
P = I² R
(8)
From equations (6) and (8) we get,
E = I² R t
(9)
From equations (5) and (9) we get,
nhυ
(10)
R= 2
I t
Current is the rate of charge flow, and frequency is the reciprocal of the period denoted by T, so
nht 2
nht
nh t
R= 2 = 2 = 2 ⋅
(11)
Q tT Q T Q T
t
t
where normalizes time with period time. Angular frequency is 2πυ, the angle θ is 2πυ t = 2π
T
T
t
θ
, so =
. For θ = 0, t = 0. For θ = 2π, t = 1/υ. If 0 ≤θ ≤ 2π, then 0 ≤ t ≤ 1/υ .This can be seen
T 2π
as making the relative time or ranking the time over the phase. We omit this normalization or
t
nh
ranking part , as the numbering variable n can undertake its role, so R = 2 ■
T
Q
Definition 1: Tyke
Tyke is the potential corresponding to the smallest possible resistance.
nh
Tyke can be calculated in different ways. Substituting n =1 in theorem 1, R = 2 , yields the
Q
h
smallest possible resistance ( 2 ).Multiplying by the current to gain the tyke potential. We
Q
believe that tyke goes beyond measurement, but the measurable spike is actually the
accumulation of tyke potentials over time.
Transformation from Planck curve to Spike and Tyke Model
Transforming the space of plank curve to the space of neuron potential is straight foreward, by
defining two transformation functions.
x
fx: wavelength Æ time,
fx (x) =
(12)
c
A
fy: Intensity Æ potential,
fy (y) = y
(13)
I
1
c
Function fx is based on
= υ = , so the time is gained by dividing the wavelength by c, the
T
λ
IV
speed of light. Function fy is based on Intensity =
, so the potential is gained by dividing the
A
Intensity by I and multiplying by the area A. Successful generation of spike train is shown in
figure 3. The code listing that regenerates this curve is given below.
Code listing 2: Spiking potential
h = 6.6261*10^-34;
c = 2.9979*10^8;
k = 1.3807*10^-23;
lambda_tyke = 1e-9:10e-9: 3000e-9;
lambda = 1e-9:10e-9: 8* 3000e-9;
time= lambda ./ c;
qu=(8.*pi.*h.*c)./lambda_tyke.^5;
figure(1),clf
spike=zeros([1, 300]);
for ii= 4500: 500 :7500
Ai=(h.*c)./(k.*ii.*lambda_tyke);
sgn=qu.*(1./(exp(Ai)-1))
spike = cat( 2, spike, sgn);
end
plot(t,spike, '-g', 'LineWidth',2)
hold on
xlabel('Time','FontSize',16)
ylabel('Potential','FontSize',16)
Figure 3, generated spike train, showing relation between time and potential
4. Investigation and Future work
To evaluate the proposed model, we run a simulation, with time between t0 = 3.3357* 10-18 sec
to tmax = 9.9770* 10-15 sec, and step size ∆t = 3.3357* 10-17 sec. That is about 300 points. We
use a simple performance measure of counting the matched points. On average we have 280
matched points. This is over 97% match of the curve points, as shown in figure 4.
Figure 4, evaluating the proposed model by counting matched points.
Future work has many dimensions. One possible direction is answering the question: What is
the relation between the class of problems solved by the current model and the class of
problems solved by chaotic stochastic models? If the two classes are equivalent or there is a
mapping between them, there would be no need to hybridize neural network with fuzzy sets,
rough sets, or chaos. Another future work direction is investigating how to exploit other quantum
physics equations, such as Wien's. Wien proposed an approximation for the spectrum of the
object, which was correct at high frequencies (short wavelength) but not at low frequencies
(long wavelength) [25]. Finally, a direction is to seek an optimal learning rule the spiking neuron
model, to enhance operation of artificial neural networks.
References
[1] Trappenberg, T. P., Fundamental of computational neuroscience. Oxford: Oxford University
Press. 2002
[2] G. Bi and M. Poo. Synaptic modification of correlated activity: Hebb’s postulate revisited.
Annu. Rev. Neurosci., 24:139–166, 2001.
[3] W. Gerstner and W. M. Kistler. Mathematical formulations of hebbian learning. Biol. Cybern.,
87:404–415, 2002.
[4] R. Kempter, W. Gerstner, and J. L. van Hemmen. Hebbian learning and spiking neurons.
Phys. Rev. E, 59:4498–4514, 1999.
[5] R. Gütig, R. Aharonov, S. Rotter, and H. Sompolinsky. Learning input correlations through
nonlinear temporally asymmetric hebbian plasticity. J. Neurosci., 23(9):3697–3714, 2003.
[6]
Pillow, J., Shlens, J., Paninski, L., Sher, A., Litke, A., Chichilnisky, E. & Simoncelli, E.
“Spatiotemporal correlations and visual signaling in a complete neuronal population,” Nature,
Vol. 454, pp. 995-999, 2008.
[7]
Y. V. Pershin, M. Di Ventra, “Experimental Demonstration of Associative Memory with
Memristive Neural Networks,” Neural Networks, No. 23, pp. 881-886, 2010.
[8] Katz, L.C., and Shatz, C.J., Synaptic activity and the construction of cortical circuits. Science
274, 1133-1138, 1996
[9] W. Gerstner and W. M. Kistler. Spiking Neuron Models. Cambridge University Press, 2002.
[10] E. M. Izhikevich, “Which model to use for cortical spiking neurons?” IEEE Transactions
on Neural Networks, Vol. 15, No. 5, pp. 1063–70, 2004.
[11] A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and
application to conduction and excitation in nerve,” J. Physiol., Vol. 117, pp. 500–544, 1954.
[12] E. M. Izhikevich, “Simple model of spiking neurons,” IEEE Transactions on Neural
Networks, Vol. 14, No. 6, pp. 1569–72, 2003.
[13] Gamble, E. and Koch, C, The dynamics of free calcium in dendritic spines in response to
repetitive synaptic input. Science, 236:1311-1315. 1987
[14] Y. Tang, J. R. Nyengaard, D. M. G. De Groot, and H. J. G. Gundersen, “Total regional
and global number of synapses in the human brain neocortex,” Synapse, No. 41, pp. 258–273,
2001.
[15] D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, “The missing memristor
found,” Nature, No. 453, pp. 80–83, 2008.
[16] G. Z. Cohen, Y. V. Pershin and M. Di Ventra,"Lagrange formalism of memory circuit
elements: classical and quantum formulations", Phys. Rev. B 85, 165428 , 2012
[17] G. Z. Cohen, Y. V. Pershin and M. Di Ventra "Second and higher harmonics generation
with memristive systems", Appl. Phys. Lett. 100, 133109 , 2012
[18] Y. N. Joglekar, and S. J. Wolf, “The elusive memristor: properties of basic electrical
circuits,” European Journal of Physics, Vol. 30, No. 4, pp. 661-675, 2009.
[19] S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumder, and W. Lu, “Nanoscale
Memristor Device as Synapse in Neuromorphic Systems,” Nano Letters, vol. 10, no. 4, pp.
1297-1301, 2010.
[20] I. Ebong, and P. Mazumder, “STDP CMOS and Memristor Based Neural Network Design
for Position Detection,” Proc. IEEE, Special Issue on Memristor Technology, Egypt, Dec 2010.
[21] Grün, Sonja; Rotter, Stefan (Eds.) , Analysis of Parallel Spike Trains. Series: Springer
Series in Computational Neuroscience, Vol. 7. 1st Edition., 2010
[22] Takeaki Shimokawa, Shinsuke Koyama, Shigeru Shinomoto: A characterization of the timerescaled gamma process as a model for spike trains. Journal of Computational Neuroscience
29(1-2): 183-191, 2010
[23] Takahiro Omi, Shigeru Shinomoto: Optimizing Time Histograms for Non-Poissonian Spike
Trains. Neural Computation 23(12): 3125-3144, 2011
[24] A. C. Phillips, Introduction to Quantum Mechanics (Manchester Physics Series), WileyBlackwell, 2003
[25] F. Mandl "Statistical Physics", 2nd Edition, pages 250-256, John Wiley ,1998
| 9cs.NE
|
Pure Gaussian quantum states from passive Hamiltonians and an
active local dissipative process
arXiv:1608.02698v1 [quant-ph] 9 Aug 2016
Shan Ma, Matthew J. Woolley, Ian R. Petersen, and Naoki Yamamoto
Abstract—We investigate the problem of preparing a pure
Gaussian state via reservoir engineering. In particular, we consider a linear quantum system with a passive Hamiltonian and
with a single reservoir which acts only on a single site of the
system. We then give a full parametrization of the pure Gaussian
states that can be prepared by this type of quantum system.
I. I NTRODUCTION
Recently, the problem of deterministically preparing a pure
Gaussian state in a linear quantum system has been studied in
the literature [1]–[6]. The main idea is to construct coherent
and dissipative processes such that the quantum system is
strictly stable and driven into a desired target pure Gaussian
state. This approach is often referred to as reservoir engineering [7], [8]. Let us consider a linear open quantum system that
obeys a Markovian Lindblad master equation [9]:
K
1 ∗
1 ∗
d
∗
ρ̂ = −i[Ĥ, ρ̂] + ∑ ĉ j ρ̂ ĉ j − ĉ j ĉ j ρ̂ − ρ̂ ĉ j ĉ j , (1)
dt
2
2
j=1
where ρ̂ is the density operator, Ĥ = Ĥ ∗ represents the system
Hamiltonian, and ĉ j is a Lindblad operator that represents a
system–reservoir interaction. Let L̂ , [ĉ1 ĉ2 · · · ĉK ]> be the
vector of Lindblad operators and for convenience, we call L̂ the
coupling vector. The Lindblad master equation (1) describes
the dynamics of a quantum system that interacts with many
degrees of freedom in a dissipative environment. Under some
conditions, the Lindblad master equation (1) can be strictly
stable and has a unique steady state limt→∞ ρ̂(t) = ρ̂(∞). Based
on this fact, it was shown in [1] that any pure Gaussian state
can be prepared in the above dissipative
way by selecting a
suitable pair of operators Ĥ, L̂ . However, for some pure
Gaussian states, the quantum systems that generate them
are hard to implement
experimentally, mainly because the
operators Ĥ, L̂ have a complex structure.
This paper complements our previous work [6]. In this
paper, we consider linear quantum systems subject to the
following two constraints. (i) The Hamiltonian Ĥ is of the form
N
N
Ĥ = ∑ ∑ g jk (q̂ j q̂k + p̂ j p̂k ), where g jk ∈ R, 1 ≤ j ≤ k ≤ N.
j=1 k= j
This type of Hamiltonian describes a set of passive beamsplitter-like interactions. (ii) The system is locally coupled
This work was supported by the Australian Research Council (ARC), the
Australian Academy of Science, and the Japan Society for the Promotion of
Science (JSPS).
S. Ma, M. J. Woolley and I. R. Petersen are with the School of
Engineering and Information Technology, University of New South Wales
at the Australian Defence Force Academy, Canberra, ACT 2600, Australia.
[email protected] [email protected]
[email protected]
N. Yamamoto is with the Department of Applied Physics and
Physico-Informatics, Keio University, Yokohama 223-8522, Japan.
[email protected]
to a single reservoir. In other words, the coupling vector L̂
consists of only one Lindblad operator which acts only on a
single site of the system. It is relatively simple to implement
a quantum system subject to the two constraints (i) and (ii).
We parametrize the class of pure Gaussian states that can be
prepared by this type of quantum system.
Notation. R denotes the set of real numbers. Rm×n denotes
the set of real m × n matrices. C denotes the set of complex
numbers. Cm×n denotes the set of complex-entried m × n
matrices. In denotes the n × n identity matrix. 0m×n denotes
the m × n zero matrix. The superscript ∗ denotes either the
complex conjugate of a complex number or the adjoint of an
operator. For a matrix A = [A jk ] whose entries A jk are complex
numbers or operators, we define A> = [Ak j ], A† = [A∗k j ]. For a
matrix A = A> ∈ Rn×n , A > 0 means that A is positive definite.
diag[A1 , · · · , An ] denotes a block diagonal matrix with diagonal
blocks A j , j = 1, 2, · · · , n. det(A) denotes the determinant of
the matrix A.
II. P RELIMINARIES
Consider an N-mode continuous-variable quantum system.
Let q̂ j and p̂ j be the position and momentum operators for
the jth mode, respectively. Then they satisfy the following
commutation relations (h̄ = 1)
[q̂ j , p̂k ] = iδ jk ,
[q̂ j , q̂k ] = 0,
and [ p̂ j , p̂k ] = 0.
Let us define a column vector of operators
x̂ = [q̂1 · · · q̂N p̂1 · · · p̂N ]> . Then we have
h
i
>
0
IN
>
>
>
x̂, x̂ = x̂x̂ − x̂x̂
= iΣ, Σ ,
. (2)
−IN 0
Let ρ̂ be the density operator of the system. Then
the mean value of the vector x̂ is given by hx̂i =
[tr(q̂1 ρ̂) · · · tr(q̂N ρ̂) tr( p̂1 ρ̂) · · · tr( p̂N ρ̂)]> and the covariance
matrix is given by V = 12 h4x̂4x̂> + (4x̂4x̂> )> i, where
4x̂ = x̂−hx̂i. A Gaussian state is entirely specified by its mean
vector hx̂i and its covariance matrix V . Since the mean vector
hx̂i contains no information about noise and entanglement, we
restrict our attention to zero-mean Gaussian states.
p The purity
of a Gaussian state is defined by p = tr(ρ̂ 2 ) = 1/ 22N det(V ).
A Gaussian state with covariance matrix V is pure if and only
if det(V ) = 2−2N . In fact, when a Gaussian state is pure, its
covariance matrix V always has the following decomposition
1 Y −1
Y −1 X
V=
,
(3)
2 XY −1 XY −1 X +Y
where X = X > ∈ RN×N , Y = Y > ∈ RN×N and Y > 0 [10]. For
the N-mode vacuum state, we have X = 0 and Y = IN . Let
Z , X +iY . Given Z, a covariance matrix V can be constructed
from it using (3). Thus, the matrix Z fully characterizes a
pure Gaussian state. We refer to Z = X + iY as a Gaussian
graph matrix [10]. Note that, a basic property of Z is Re(Z) =
Re(Z)> and Im(Z) = Im(Z)> > 0. This property guarantees
physicality of the corresponding state.
Suppose that the system Hamiltonian in (1) is a quadratic
function of x̂, i.e., Ĥ = 21 x̂> Gx̂, with G = G> ∈ R2N×2N , the
coupling vector is a linear function of x̂, i.e., L̂ = Cx̂, with
C ∈ CK×2N , and the dynamics of the density operator ρ̂ obey
the Lindblad master equation (1). Then the corresponding
dynamics of the mean vector hx̂(t)i and the covariance matrix
V (t) can be described by
dhx̂(t)i
= A hx̂(t)i,
(4)
dt
dV (t) = A V (t) +V (t)A > + D,
(5)
dt
where A = Σ G + Im(C†C) , D = Σ Re(C†C)Σ> [9, Chapter
6]. The linearity of the dynamics indicates that if the initial
system state ρ̂(0) is Gaussian, then the system will always be
Gaussian, with mean vector hx̂(t)i and covariance matrix V (t)
evolving according to (4) and (5), respectively. We shall be
particularly interested in the steady state of the system with the
covariance matrix V (∞). Recently, a necessary and sufficient
condition has been derived in [1], [2] for preparing a pure
Gaussian steady state via reservoir engineering. The result is
summarized as follows.
Lemma 1 ( [1], [2]). Let Z = X + iY be the Gaussian graph
matrix of an N-mode pure Gaussian state. Then this pure Gaussian state is generated by the Lindblad master equation (1) if
and only if
XRX +Y RY − ΓY −1 X − XY −1 Γ> −XR + ΓY −1
G=
,
−RX +Y −1 Γ>
R
(6)
and
C = P> [−Z IN ] ,
(7)
consider linear quantum systems subject to the following two
constraints.
¬ The Hamiltonian Ĥ is of the form Ĥ =
N N
∑ ∑ g jk (q̂ j q̂k + p̂ j p̂k ), where g jk ∈ R, 1 ≤ j ≤ k ≤ N.
j=1 k= j
The system is locally coupled to a single reservoir. That
is, the coupling vector L̂ is of the form L̂ = c1 q̂` + c2 p̂` ,
where c1 ∈ C, c2 ∈ C and ` ∈ {1, 2, · · · , N}.
Remark 2. The Hamiltonian in ¬ can be rewritten in terms
of the annihilation and creation operators as
N
Ĥ =
N
â∗j âk + â j â∗k ,
∑ ∑ g jk
1 ≤ j ≤ k ≤ N,
q̂ −i p̂
q̂ +i p̂
where â j = j√2 j and â∗j = j√2 j are the annihilation and creation operators for the jth mode, respectively. A Hamiltonian
Ĥ of this form is called passive and describes beam-splitterlike interactions [12], [13].
Remark 3. The constraint implies two crucial features of
the system. First, the system is coupled to a single reservoir,
i.e., K = 1 in (1). Second, the corresponding Lindblad operator
acts only on a single mode of the system.
To illustrate this type of linear quantum system, we consider
two examples.
Example 1. We consider a ring of three quantum harmonic
oscillators (N = 3), as depicted in Fig. 1. The quantum harmonic oscillators are labelled 1 to 3. Suppose the Hamiltonian
Ĥ is given by
3
Ĥ =
q̂2j + p̂2j + g12 (q̂1 q̂2 + p̂1 p̂2 )
∑ gjj
j=1
+ g13 (q̂1 q̂3 + p̂1 p̂3 ) + g23 (q̂2 q̂3 + p̂2 p̂3 ) ,
where g jk ∈ R, 1 ≤ j ≤ k ≤ 3. In addition, the system-reservoir
coupling vector L̂ consists of only one Lindblad operator
which acts on the third mode of the system. It is given by
L̂ = c1 q̂3 +c2 p̂3 , where c1 ∈ C and c2 ∈ C. This linear quantum
system satisfies the constraints ¬ and .
1
where R = R> ∈ RN×N , Γ = −Γ> ∈ RN×N , and P ∈ CN×K are
free matrices satisfying the following rank condition
rank [P QP · · · QN−1 P] = N, Q , −iRY +Y −1 Γ. (8)
Remark 1. A pair of matrices (A1 , A2 ), where A1 ∈
Cn×n and A2 ∈ Cn×m , is said to be controllable if
rank [A2 A1 A2 · · · An−1
1 A2 ] = n [11, Theorem 6.1]. Therefore, the rank constraint (8) is equivalent to saying that (Q, P)
is controllable.
g jk ∈ R,
j=1 k= j
2
3
Reservoir
Fig. 1. A ring of linearly coupled quantum harmonic oscillators. The systemreservoir coupling vector L̂ consists of only one Lindblad operator which acts
on the third mode of the system.
III. C ONSTRAINTS
According to Lemma 1, for some pure Gaussian states, the
quantum systems that generate them are hard to implement
experimentally because the operators Ĥ = 21 x̂> Gx̂ and L̂ = Cx̂
have a complex structure. Here we discuss another route. We
restrict our attention to a class of linear quantum systems that
are relatively simple to implement. Then we see which pure
Gaussian states can be prepared by this type of system. We
Example 2. We consider a chain of three quantum harmonic
oscillators (N = 3), as depicted in Fig. 2. The quantum harmonic oscillators are labelled 1 to 3. Suppose the Hamiltonian
Ĥ is given by
3
Ĥ =
∑ gjj
j=1
q̂2j + p̂2j + g12 (q̂1 q̂2 + p̂1 p̂2 )
+ g23 (q̂2 q̂3 + p̂2 p̂3 ) ,
•
where g11 , g22 , g33 , g12 , g23 ∈ R. In addition, the systemreservoir coupling vector L̂ consists of only one Lindblad
operator which acts on the second mode of the system. It
is given by L̂ = c1 q̂2 + c2 p̂2 , where c1 ∈ C and c2 ∈ C. This
linear quantum system satisfies the constraints ¬ and .
1
2
3
Reservoir
Fig. 2.
A chain of linearly coupled quantum harmonic oscillators. The
system-reservoir coupling vector L̂ consists of only one Lindblad operator
which acts on the second mode of the system.
The class of linear quantum systems subject to ¬ and
can be split into three disjoint subclasses (I, II, and III).
The subclass I consists of all the unstable linear quantum
systems subject to the constraints ¬ and . For any system
in the subclass I, there does not exist a steady state. The
subclass II consists of all the linear quantum systems with the
constraints ¬ and that are strictly stable, and that evolve
toward a mixed Gaussian steady state. The subclass III consists
of all the linear quantum systems with the constraints ¬ and
that are strictly stable, and that evolve toward a pure Gaussian
steady state.
In the following section, we are particularly interested in
the subclass III. Our objective is to characterize the class of
pure Gaussian states that can be generated by linear quantum
systems subject to ¬ and . In other words, we characterize all
of the pure Gaussian states for which there exist a Hamiltonian
Ĥ of the form ¬ and a system–reservoir coupling vector L̂ of
the form such that the state is the unique steady state of
the corresponding linear quantum system (1).
IV. PARAMETRIZATION
We give the main result. The following theorem provides
a full mathematical parametrization of the pure Gaussian
states that can be generated by systems subject to the two
constraints ¬ and . For clarity, we distinguish two cases: N
is odd and N is even.
Theorem 1.
• An N-mode pure Gaussian state (where N is odd) can be
generated by a linear quantum system subject to the two
constraints ¬ and if and only if its Gaussian graph
matrix Z can be written as
z̄
01×(N−1)
Z = P>
P,
(9)
0(N−1)×1 Q > Z̄Q
Z̄ = diag[Z̃1 , · · · , Z̃ N−1 ],
2
where P ∈
is a permutation matrix, Q ∈
R(N−1)×(N−1) is a real orthogonal matrix, z̄ ∈ Λ, and
Z̃ j ∈ ∆, 1 ≤ j ≤ N−1
2 .
RN×N
An N-mode pure Gaussian state (where N is even) can
be generated by a linear quantum system subject to the
two constraints ¬ and if and only if its Gaussian graph
matrix Z can be written as
z̄
01×(N−1)
>
P,
(10)
Z=P
0(N−1)×1 Q > Z̄Q
Z̄ = diag[Z̃1 , · · · , Z̃ N ],
2
where P ∈ RN×N is a permutation matrix, Q ∈
R(N−1)×(N−1) is a real orthogonal matrix, z̄ ∈ Λ, Z̃1 = − 1z̄ ,
and Z̃ j ∈ ∆, 2 ≤ j ≤ N2 .
Here Λ , {z z ∈ C and Im(z) > 0}, and ∆ , Z Z = Z > ∈
2×2 , Im (Z) > 0, diag[1, −1]Z 2 = −I , and det(Z + 1 I ) =
C
2
z̄ 2
0 .
The proof of Theorem 1 is provided in the Appendix. We
give some remarks to explain Theorem 1.
Remark 4. Multiplying a Gaussian graph matrix Z on the left
by a permutation matrix P > and on the right by P simply
corresponds to a relabeling of the N modes. Therefore, the
role that the matrix P plays in Equations (9) and (10) is not
important from a practical point of view. It simply amounts to
a relabeling of the N modes so that the first mode is the one
that is coupled to the reservoir.
Remark 5. If z̄ = i in Equations (9) and (10), then from
the condition that Z̃ j ∈ ∆, we can deduce that Z̃ j = iI2 . In
this case, the whole state is a trivial pure Gaussian state,
i.e., the N-mode vacuum state. It then follows from (7) that
the system-reservoir
i.e.,
√ coupling vector L̂ must be passive,
+i p̂`
L̂ = c (q̂` + i p̂` ) = 2câ` , where c ∈ C and â` = q̂`√
is
an
2
annihilation operator.
Remark 6. The constraint Im(z) > 0 in the definition of
Λ must be satisfied for the corresponding Z to be a valid
Gaussian graph matrix. For the same reason, the constraints
Z = Z > ∈ C2×2 and Im (Z) > 0 in the definition of ∆ must be
maintained. See the definition of the Gaussian graph matrix Z
in Section II for details.
Remark 7. Theorem 1 tells us that by choosing a permutation matrix P ∈ RN×N , a real orthogonal matrix Q ∈
R(N−1)×(N−1) , a complex number z̄ ∈ Λ and several complex
matrices Z̃ j ∈ ∆, we can construct a pure Gaussian state that
can be prepared by a system subject to the two constraints ¬
and .
Remark 8. As an extreme case, let us consider a one-mode
pure Gaussian state (N = 1). According to Theorem 1, a
one-mode pure Gaussian state can be generated by a linear
quantum system subject to the two constraints ¬ and if and
only if its Gaussian graph matrix Z satisfies Z = P > z̄P = z̄,
where z̄ ∈ Λ. Since the set Λ characterizes all of the one-mode
pure Gaussian states, so we conclude that all of the one-mode
pure Gaussian states can be generated by a linear quantum
system subject to the two constraints ¬ and .
Remark 9. Let us consider the two-mode case (N = 2).
According to Theorem 1, a two-mode pure Gaussian state can
be generated by a linear quantum system subject to the two
constraints ¬ and if and only
if its Gaussian graph matrix
z̄ 0
>
P, where z̄ ∈ Λ. Since
Z can be written as Z = P
0 − 1z̄
the role of the permutation matrix P is not important here, it
follows that the Gaussian
graph
matrix of the two-mode state
z̄ 0
is of the form Z =
, where z̄ ∈ Λ. The structure of the
0 − 1z̄
system that generates this state is shown in Fig. 3. It follows
from the form of the matrix Z that the two modes are not
entangled [5].
1
2
matrix Z can be written as
z̄
Z = P>
0(N−1)×1
01×(N−1)
P,
Q > Z̄Q
(11)
Z̄ = diag[Z̃1 , · · · , Z̃ N−1 ],
2
where P ∈
is a permutation matrix, Q ∈
R(N−1)×(N−1)
is a# real" orthogonal matrix,
z̄ ∈ Λ, and
" 2
#
z̄ −1
z̄2 +1
z̄2 −1
z̄2 +1
−
2z̄
2z̄
Z̃ j = z̄22z̄+1 z̄22z̄−1 or
, 1 ≤ j ≤ N−1
2 .
z̄2 +1
z̄2 −1
−
2z̄
2z̄
2z̄
2z̄
An N-mode pure Gaussian state (where N is even) can
be generated by a linear quantum system subject to the
two constraints ¬ and if and only if its Gaussian graph
matrix Z can be written as
z̄
01×(N−1)
>
P,
(12)
Z=P
0(N−1)×1 Q > Z̄Q
RN×N
•
Z̄ = diag[Z̃1 , · · · , Z̃ N ],
2
Reservoir
Fig. 3.
The structure of a quantum system that generates
the two-mode
pure Gaussian state with Gaussian graph matrix
Z=
z̄
0
0
.
− 1z̄
Remark 10. Let us consider the three-mode case (N = 3). According to Theorem 1, a three-mode pure Gaussian state can be
generated by a linear quantum system subject to the two constraints ¬ and if andonly if its Gaussian
graph matrix Z can
z̄
01×2
>
be written as Z = P
P, where z̄ ∈ Λ, and
02×1 Q > Z̃1 Q
Z̃1 ∈ ∆. As mentioned, the constraint defining Λ, i.e., Im(z) >
0, comes from the definition
of a Gaussian graph matrix. Simi
lar constraints (i.e., Z = Z > ∈ C2×2 and Im (Z) > 0 ) apply
to the set ∆. Let us analyze the other two constraints in ∆. One
2
is diag[1, −1]Z = −I2 , i.e., Z diag[1, −1]Z = − diag[1, −1],
which is similar to what we have obtained in Theorem 1 of [6].
This should not be surprising, since we can think of the mode
coupled to the reservoir as an auxiliary mode. Then the other
two modes are coupled to a single reservoir. This is indeed
the case considered in [6]. The other constraint defining ∆,
i.e., det(Z + 1z̄ I2 ) = 0, is equivalent to saying that − 1z̄ is an
eigenvalue of the matrix Z̃ j . This constraint comes from the
locality requirement on the original system–reservoir coupling.
Lemma 2. Given z̄ ∈ Λ, then the set ∆ contains only two
matrices, i.e.,
(" 2
# " 2
#)
2
z̄ −1
z̄2 +1
z̄ −1
− z̄ 2z̄+1
2z̄
2z̄
2z̄
∆=
,
.
2
z̄2 +1
z̄2 −1
z̄2 −1
− z̄ 2z̄+1
2z̄
2z̄
2z̄
The proof of Lemma 2 is provided in the Appendix. Using
Lemma 2, we have an equivalent version of Theorem 1.
Theorem 2.
•
is a permutation matrix, Q ∈
where P ∈
1
R(N−1)×(N−1)
is
a
real
orthogonal
" 2
# " 2 matrix,2 z̄ ∈#Λ, Z̃1 = − z̄ ,
z̄ −1
z̄2 +1
z̄ −1
z̄ +1
− 2z̄
2z̄
and Z̃ j = z̄22z̄+1 z̄22z̄−1 or
, 2 ≤ j ≤ N2 .
z̄2 +1
z̄2 −1
−
2z̄
2z̄
2z̄
2z̄
RN×N
An N-mode pure Gaussian state (where N is odd) can be
generated by a linear quantum system subject to the two
constraints ¬ and if and only if its Gaussian graph
Remark 11. Once we obtain a pure Gaussian state using
Theorem 2, we can immediately construct a linear quantum
system that generates such a state and also satisfies the
constraints ¬ and . The construction method is given as
follows.
" 2
#
2
Case
1
(N
is
odd).
If
Z̃ j =
z̄ −1
2z̄
z̄2 +1
2z̄
z̄ +1
2z̄
z̄2 −1
2z̄
,
τ̄ j
choose R̃21, j =
, where τ̄ j ∈ R and τ̄ j 6= 0.
−τ̄ j
τ̄ j
Otherwise, choose R̃21, j =
, where τ̄ j ∈ R
τ̄ j
i>
h >
>
and τ̄ j 6= 0. Let R̄21 = R̃21,1 · · · R̃21, (N−1) . Let
2 i
h
R̄22 = diag r1 , −r1 , r2 , −r2 , · · · , r N−1 , −r N−1 , where
2
2
r j ∈ R, r j 6= 0 and |r j| 6= |rk | whenever j 6=
k. In Lemma 1, let
0
R̄>
21 Q
us choose R = P >
P, Γ = XRY and
>
Q R̄21 Q > R̄22 Q
>
P = P > τ p 01×(N−1) , where τ p ∈ C and τ p 6= 0. Then
calculate the matrices G and C using (6) and (7), respectively.
The resulting linear quantum system with Hamiltonian
Ĥ = 12 x̂> Gx̂ and coupling vector L̂ = Cx̂ is strictly stable and
generates the pure Gaussian state. It can be shown that this
quantum system also satisfies the two constraints ¬ and .
Case 2 (N is "even). Choose
# R̃21,1 = τ̄1 , where τ̄1 ∈R and
z̄2 −1
z̄2 +1
τ̄ j
2z̄
2z̄
τ̄1 6= 0. If Z̃ j = z̄2 +1 z̄2 −1 , j ≥ 2, choose R̃21, j =
,
−τ̄ j
2z̄
2z̄
τ̄
where τ̄ j ∈ R and τ̄ j 6= 0. Otherwise, choose R̃21, j = j ,
τ̄ j
h
i>
>
>
where τ̄ j ∈ R and τ̄ j 6= 0. Let R̄21 = R̃21,1 · · · R̃21, N .
2
h
i
Let R̄22 = diag 0, r2 , −r2 , · · · , r N , −r N , where r j ∈ R,
2
2
r j 6= 0, j ≥ 2, and |r j |6= |rk | whenever j
6= k. In Lemma 1,
0
R̄>
21 Q
let us choose R = P >
P, Γ = XRY and
>
Q R̄21 Q > R̄22 Q
>
P = P > τ p 01×(N−1) , where τ p ∈ C and τ p 6= 0. Then calculate the matrices G and C using (6) and (7), respectively. The
resulting linear quantum system with Hamiltonian Ĥ = 12 x̂> Gx̂
and coupling vector L̂ = Cx̂ is strictly stable and generates the
pure Gaussian state. It can be shown that this system also
satisfies the two constraints ¬ and .
and the coupling vector
V. E XAMPLE
We consider the five-mode pure Gaussian state generated
in [14]. The Gaussian graph matrix of this pure Gaussian state
is given by
The coupling operator L̂ represents a standard dissipative
reservoir that acts only on the third mode. The eigenstate
corresponding to the zero eigenvalue of L̂ is a squeezed
state. The third system mode is then prepared in a squeezed
state, while other modes are coupled via passive interactions.
This leads to entanglement across the system, in a similar
way to which the interference of a squeezed optical mode
with a vacuum at a beam splitter results in entangled output
modes [15]. The structure of the system is shown in Fig. 4.
It is a chain of quantum harmonic oscillators with nearest–
neighbour Hamiltonian interactions. Only the central oscillator
is coupled to the reservoir.
Z=
cosh(2α)
0
0
i
0
sinh(2α)
0
cosh(2α)
0
− sinh(2α)
0
where α ∈
0
1
P =
0
0
0
0
0
cosh(2α) + sinh(2α)
0
0
0
− sinh(2α)
0
cosh(2α)
0
sinh(2α)
0
0
, (13)
0
cosh(2α)
Let us choose a permutation matrix
1 0 0
0 0 0
0 0 1
and
a
real
orthogonal
0 0 0
0 1 0
−1 0 −1 0
√
0 −1 0 1
.
matrix
Q = 22
Then
we
0
1
0 1
−1 0 1 0
z̄
01×4
P, Z̄ = diag[Z̃1 , Z̃2 ],
have Z = P >
04×1 Q > Z̄Q
where
z̄ = i (cosh(2α) + sinh(2α)),
and
Z̃ j =
cosh(2α)
(−1) j−1 sinh(2α)
i
, j = 1, 2. It
(−1) j−1 sinh(2α)
cosh(2α)
can be verified that z̄ ∈ Λ and Z̃ j ∈ ∆, j = 1, 2.
According to Theorem 1, this pure Gaussian state can
be generated by a linear quantum system satisfying
the two constraints ¬ and √
. Next, we construct
>
such a system. Let R̄21 = 2 −1 1 1 1
and
R̄22 = diag[1, −1, 3, −3]. Then in Lemma
1,
let
us
choose
−1 2 0 0 0
2 −1 2 0 0
0
R̄>
Q
>
21
0
2 0 2 0
R=P
P =
>
>
,
Q R̄21 Q R̄22 Q
0
0 2 1 2
0
0 0 2 1
>
cosh(α)−sinh(α)
√
0 0 1 0 0 . It can
Γ = 05×5 , and P = i
2
be verified that Y RY = R and
rank [P QP Q2 P Q3 P Q4 P]
= rank [P − iRY P − R2 P iR3Y P R4 P]
R.
0
0
0
1
0
=5.
Therefore, the resulting linear quantum system is strictly stable
and generates the target pure Gaussian state given in (13).
Substituting R, Γ and P into (6) and (7), we obtain the
Hamiltonian of the system
1
1
Ĥ = − q̂21 + p̂21 + q̂22 + p̂22 + q̂24 + p̂24 + q̂25 + p̂25
2
2
+ 2 (q̂1 q̂2 + p̂1 p̂2 + q̂2 q̂3 + p̂2 p̂3 )
+ 2 (q̂3 q̂4 + p̂3 p̂4 + q̂4 q̂5 + p̂4 p̂5 ) ,
cosh(α) − sinh(α)
√
[−i((cosh(2α) + sinh(2α)) q̂3 + p̂3 ]
2
cosh(α) + sinh(α)
cosh(α) − sinh(α)
√
√
=
q̂3 + i
p̂3
2
2
= cosh(α)â3 + sinh(α)â∗3 .
L̂ = i
1
2
3
4
5
Reservoir
Fig. 4. The pure Gaussian state given by (13) can be generated in a chain
of linearly coupled quantum harmonic oscillators with nearest–neighbour
Hamiltonian interactions. Only the central oscillator is coupled to the reservoir.
The oscillators of the system are entangled in pairs. The
first oscillator is entangled with the fifth oscillator and the
second oscillator is entangled with the fourth oscillator. The
central oscillator is not entangled with the other oscillators.
The amount of entanglement can be quantified using the
logarithmic negativity E [16]–[18]. The values are given by
E(1,5) = E(2,4) = 2|α|. For a more detailed discussion of this
example, we refer the reader to [14].
VI. C ONCLUSION
In this paper, we consider linear quantum systems subject
to constraints. First, we assume that the Hamiltonian Ĥ is
N
N
of the form Ĥ = ∑ ∑ g jk (q̂ j q̂k + p̂ j p̂k ), where g jk ∈ R,
j=1 k= j
1 ≤ j ≤ k ≤ N. Second, we assume that the system is locally
coupled to a single reservoir. Then we give a full mathematical parametrization of the pure Gaussian states that can be
prepared using this type of quantum system.
VII. A PPENDIX
In this section, we provide the proof of Theorem 1. First,
we provide some preliminary results which will be used in the
proof of Theorem 1.
Lemma 3 ( [5]). Suppose that an N-mode pure Gaussian state
is generated in a linear quantum system with a single dissipative process and that the corresponding Lindblad operator acts
only on the `th mode of the system, then
Z(`, j) = Z( j,`) = 0,
∀ j 6= `,
where Z(`, j) denotes the (`, j) element of the Gaussian graph
matrix Z. In addition, the `th mode is not entangled with the
other modes.
Lemma 4 ( [11]). If a pair of matrices (A1 , A2 ) is controllable,
where A1 ∈ Cn×n and A2 ∈ Cn×m , then F −1 A1 F, F −1 A2 is
controllable where F ∈ Cn×n is a non-singular matrix.
A
A12
Lemma 5. Suppose A = 11
, where A11 ∈ C, A12 ∈
A21 A22
C1×(n−1) , A21 ∈ C(n−1)×1 , and A22 ∈ C(n−1)×(n−1) , and ξ =
>
τ 01×(n−1) ∈ Cn×1 , where τ ∈ C and τ 6= 0. Then the
pair (A, ξ ) is controllable if and only if the pair (A22 , A21 ) is
controllable.
Proof: If (A, ξ ) is controllable, then the matrix
[A − λ In ξ ] has full
row rank for all λ ∈ C [19, Theorem
A11 − λ
A12
τ
3.1]. That is, rank
=n
A21
A22 − λ In−1 0(n−1)×1
for all λ ∈ C. It follows that rank ([A21 A22 − λ In−1 ]) = n − 1
for all λ ∈ C. Therefore, (A22 , A21 ) is controllable [19,
Theorem 3.1]. Conversely, if (A22 , A21 ) is controllable, then
rank ([A22 − λ In−1
A21 ]) = n − 1 for all λ ∈ C. Since
τ 6= 0,
A11 − λ
A12
τ
=n
it follows that rank
A21
A22 − λ In−1 0(n−1)×1
for all λ ∈ C. That is, the matrix [A − λ In ξ ] has full row rank
for all λ ∈ C. Therefore, (A, ξ ) is controllable.
Lemma 6. Suppose A = diag [A11 , A22 , · · · , Amm ], where
>
A j j ∈ Cn j ×n j , 1 ≤ j ≤ m, and ξ = ξ1> ξ2> · · · ξm> ,
where ξ j ∈ Cn j ×1 , 1 ≤ j ≤ m. Then the pair (A, ξ ) is
controllable if and only if all the pairs (A j j , ξ j ), 1 ≤ j ≤ m,
are controllable and A j j and Akk , j 6= k, have no common
eigenvalues.
Proof: hNecessity. If theipair (A, ξ ) is controllable, then
the matrix A − λ I∑mj=1 n j ξ has full row rank for all λ ∈
C [19, Theorem 3.1]. That is,
A − λ I
n1
11
rank
A22 − λ In2
..
.
Amm − λ Inm
ξ1
ξ2
..
.
ξm
Amm − λ Inm 0
has full
row rank. As a consequence, it can be shown
A11 − λ In1
ξ1
A
−
λ
I
ξ
n
22
2
2
has
that
.
..
..
.
Amm − λ Inm ξm
hfull row rank. i Therefore, we conclude that the matrix
A − λ I∑mj=1 n j ξ has full row rank for any eigenvalue λ of
A. Hence (A, ξ ) is controllable.
Proof of Theorem 1
Proof: Necessity. Suppose an N-mode pure Gaussian
state is generated by a linear quantum system subject to
the two constraints ¬ and , and Z is the corresponding
Gaussian graph matrix for this pure Gaussian state. We
will show that the Gaussian graph matrix Z of this pure
Gaussian state can be written in the form of Equation (9)
or Equation (10). According to ¬, the Hamiltonian
of the
R
0N×N
1 >
linear quantum system is Ĥ = 2 x̂
x̂, where
0N×N
R
2g11 g12 · · · g1N
g12 2g22 · · · g2N
R= .
..
.. . Using Lemma 1, we have
..
..
.
.
.
g1N
g2N
m
= ∑ n j.
Sufficiency. To
h show (A, ξ ) isicontrollable, we need to prove
that the matrix A − λ I∑mj=1 n j ξ has full row rank for all λ ∈
h
i
C. That is, we need to prove that the matrix A − λ I∑mj=1 n j ξ
has full row rank for any eigenvalue λ of A. Now suppose λ
is an arbitrary eigenvalue of A. Since A j j and Akk , j 6= k, have
no common eigenvalues, it follows that λ is an eigenvalue of
only one block. Without loss of generality, we assume that
λ is an eigenvalue of A11 . Then we have det(A11 − λ In1 ) = 0
and det(A j j − λ In j ) 6= 0, 2 ≤ j ≤ m. Since (A11 , ξ1 ) is controllable, it follows
that [A11 − λ In1 ξ1 ] has full row rank. Then
A11 − λ In1
ξ1
A22 − λ In2
0
we have
..
..
.
.
(14)
j=1
Then it follows that the matrix A j j − λ In j ξ j has full row
rank for all λ ∈ C. Therefore, all the pairs (A j j , ξ j ), 1 ≤ j ≤ m,
are controllable [19, Theorem 3.1].
To prove the second part of necessity, without loss of generality, we assume thatA11 and A22 share a common eigenvalue
A11
0
λ . Then the matrix
is a derogatory matrix [20],
0 A22
A11
0
ξ
[21]. Using Lemma 6 in [6], the pair
, 1
0 A22
ξ2
cannot be controllable. But we already know from the previous
result
that if the
pair (A, ξ ) is controllable then the pair
A11
0
ξ1
,
must be controllable. Therefore, we
0 A22
ξ2
reach a contradiction. We conclude that A j j and Akk , j 6= k,
have no common eigenvalues.
···
2gNN
XRX +Y RY − ΓY −1 X − XY −1 Γ> = R,
−XR + ΓY
−1
= 0.
(15)
(16)
It follows from (16) that Γ = XRY . Substituting this into (15)
gives
Y RY − XRX = R.
Γ + Γ>
Recall from Lemma 1 that
Combining this with (17) gives
(17)
= 0, i.e., XRY +Y RX = 0.
ZRZ = −R.
(18)
Suppose that the `th mode of the linear quantum system is
locally coupled to the reservoir. Then using Lemma 3, we
have Z(`, j) = Z( j,`) = 0, ∀ j 6= `. This fact implies that there
exists a permutation matrix P1 ∈ RN×N such that
z̄
01×(N−1)
>
Z = P1
P1 ,
0(N−1)×1
Z̆
where z̄ = Z(`,`) and Z̆ ∈ C(N−1)×(N−1) . Obviously, z̄ ∈ Λ. Let
R̆ , P1 RP1> . Then Equation (18) is transformed into
z̄
01×(N−1)
z̄
01×(N−1)
= −R̆. (19)
R̆
Z̆
0(N−1)×1
Z̆
0(N−1)×1
R̆
R̆>
21 , where
If we write R̆ in block form as R̆ = 11
R̆21 R̆22
(N−1)×(N−1) , and R̆ ∈ R(N−1)×1 , then
R̆11 ∈ R, R̆22 = R̆>
∈
R
21
22
Equation (19) becomes
z̄R̆11 z̄ = −R̆11 ,
Z̆ R̆22 Z̆ = −R̆22 ,
(20)
Z̆ R̆21 z̄ = −R̆21 .
(21)
Since R̆22 = R̆>
22 , it can be diagonalized by a real orthogonal
matrix Q1 ∈ R(N−1)×(N−1) . That is, R̆22 = Q1> R{ 22 Q1 , where
R{ 22 is a real diagonal matrix. Let Z{ , Q1 Z̆Q1> , and R{ 21 ,
Q1 R̆21 . Then the equations (20) and (21) are transformed into
{{ {
ZR22 Z = −R{ 22 ,
(22)
Z{ R{ 21 z̄ = −R{ 21 .
(23)
Since the `th mode of the linear quantum system is locally coupled to the reservoir, it follows from Lemma 1
that
the matrix P in (7) must be of the form P =
>
01×(`−1) τ p 01×(N−`) , where τ p ∈ C and τ p 6= 0. The
matrix Q in (8) is given by Q = −iRY + Y −1 Γ = −iRY
+
z̄
01×(N−1)
−1
>
P1 .
Y (−Y RX) = −RZ = −P1 R̆
0(N−1)×1
Z̆
From (8), we observe that
the pair
Using
(Q, P) is controllable.
z̄
01×(N−1)
, P1 P
Lemma 4, it follows that −R̆
0(N−1)×1
Z̆
>
τ
0
is controllable.
Note
that
P
P
=
and
p 1×(N−1)
1
z̄
01×(N−1)
R̆11 z̄ R̆>
Z̆
21
=−
−R̆
. It follows from
0(N−1)×1
Z̆
R̆21 z̄ R̆22 Z̆
Lemma 5 that −R̆22 Z̆, −R̆21 z̄ is controllable. That is,
{
>{
−Q1> R{ 22 ZQ
is controllable.
1 , −Q1 R21 z̄
Again using
{
Lemma 4, it follows that −R{ 22 Z,
−R{ 21 z̄ is controllable.
Since −R{ 21 z̄ ∈ C(n−1)×1 , by Lemma 6 in [6], the matrix −R{ 22 Z{
is a non-derogatory matrix. Then following similar arguments
as in the proof of Theorem 1 in [6], we have the following
preliminary result.
* If N is odd, then
Z{ = P2> Z̄P2 ,
Z̄ = diag[Z̃1 , · · · , Z̃ (N−1) ],
2
is a permutation matrix, Z̃1 ∈
where P2 ∈
(Π ∪ Ξ), and Z̃ j ∈ Ξ, 2 ≤ j ≤ (N−1)
2 .
* If N is even, then
R(N−1)×(N−1)
Z{ = P2> Z̄P2 ,
Z̄ = diag[Z̃1 , · · · , Z̃ N ],
2
where P2 ∈
is a permutation matrix, Z̃1 ∈
Λ, and Z̃ j ∈ Ξ, 2 ≤ j ≤ N2 .
Here
Π , {diag[z, i]
z ∈ C and Im(z) > 0}, and Ξ ,
2
Z Z = Z > ∈ C2×2 , Im (Z) > 0, and diag[1, −1]Z =
−I2 .
R(N−1)×(N−1)
Let R̄22 , P2 R{ 22 P2> and R̄21 , P2 R{ 21 . Then the equations (22) and (23) are transformed into
Z̄ R̄22 Z̄ = −R̄22 ,
(24)
Z̄ R̄21 z̄ = −R̄21 .
(25)
is
controllable,
i.e.,
Since
−R22 Z, −R21 z̄
is controllable, it follows
−P2> R̄22 Z̄P2 , −P2> R̄21 z̄
from Lemma 4 that (−R̄22 Z̄, −R̄21 z̄) is controllable.
First, if N is odd, we partition R̄22
and R̄21 as R̄22i =
h >
>
>
R̃
diag[R̃22,1 , · · · , R̃ (N−1) ] and R̄21 =
,
21,1 · · · R̃ (N−1)
{
{
22,
{
2
where R̃22, j ∈ R2×2 and R̃21, j ∈ R2×1 , 1 ≤ j ≤
the equations (24) and (25) become
Z̃ j R̃22, j Z̃ j = −R̃22, j ,
21, 2
(N−1)
2 . Then
(26)
Z̃ j R̃21, j z̄ = −R̃21, j ,
(27)
where 1 ≤ j ≤ (N−1)
2 . Since (−R̄22 Z̄, −R̄21 z̄) is control
lable, it follows from Lemma 6 that −R̃22, j Z̃ j , −R̃21, j z̄ ,
1 ≤ j ≤ (N−1)
2 , is controllable. Note that R̃22,1 is a real
diagonal matrix.
If Z̃1 ∈ Π and Z̃1 6= iI2 , solving(26) gives
0 0
0
0
.
R̃22,1 =
, where τ1 ∈ R. Then −R̃22,1 Z̃1 =
0 τ1
0 −τ1 i
Since −R̃22,1 Z̃1 , −R̃21,1 z̄ is controllable, it follows that
>
R̃21,1 = τ2 τ3 , with τ2 τ3 6= 0. Substituting this into (27)
yields z̄ = i and Z̃1 = iI2 . This contradicts the assumption that
Z̃1 6= iI2 . Therefore, we know that if Z̃1 ∈ Π, then Z̃1 = iI2.
Note that iI2 ∈ Ξ. Hence Z̃1 ∈ Ξ. Since −R̃22, j Z̃ j , −R̃21, j z̄
is controllable, it follows that R̃21, j 6= 02×1 , 1 ≤ j ≤ (N−1)
2 .
Then by (27), − 1z̄ is an eigenvalue of Z̃ j , i.e., det Z̃ j + 1z̄ = 0,
1 ≤ j ≤ (N−1)
2 . Combining this fact with Z̃ j ∈ Ξ, we conclude
that Z̃ j ∈ ∆, 1 ≤ j ≤ (N−1)
2 . Let P = P1 , and Q = P2 Q1
in (9). Obviously, Q is an orthogonal matrix and Equation (9)
holds. This completes the first part of the necessity proof.
Second, if N is even, we partition
h R̄22 and R̄21i>as R̄22 =
>
>
diag[R̃22,1 , · · · , R̃ N ] and R̄21 = R̃21,1 · · · R̃21, N , where
22, 2
2
R̃22,1 ∈ R, R̃21,1 ∈ R, R̃22, j ∈ R2×2 and R̃21, j ∈ R2×1 , 2 ≤ j ≤ N2 .
Then the equations (24) and (25) become
Z̃ j R̃22, j Z̃ j = −R̃22, j ,
(28)
Z̃ j R̃21, j z̄ = −R̃21, j ,
(29)
where 1 ≤ j ≤ N2 . Recall that (−R̄22 Z̄, −R̄21 z̄) is controllable.
Then it follows from Lemma 6 that −R̃22, j Z̃ j , −R̃21, j z̄ , 1 ≤
j ≤ N2 , is controllable. Hence we have R̃21,1 6= 0 and R̃21, j 6=
02×1 , 2 ≤ j ≤ N2 . Then by (29), − 1z̄ is an eigenvalue of Z̃ j ,
i.e., det Z̃ j + 1z̄ = 0, 1 ≤ j ≤ N2 . Therefore, we obtain Z̃1 = − 1z̄
and Z̃ j ∈ ∆, 2 ≤ j ≤ N2 . Let P = P1 , and Q = P2 Q1 in (10).
Obviously, Q is an orthogonal matrix and Equation (10) holds.
This completes the second part of the necessity proof.
Sufficiency. Suppose the graph matrix Z of an N-mode pure
Gaussian state (where N is odd) satisfies Equation (9). We
now construct an N-mode linear quantum system subject to
the two constraints ¬ and , such that the state is obtained
as the unique steady state of this system. We see from
Equation (9) that Z̃ j ∈ ∆, 1 ≤ j ≤ (N−1)
. Using Theorem 2
2>
>
in [6], it can be shown that ξv1 = 1 1 and ξv2 = 1 −1
2×2
are two eigenvectors of Z̃ j , 1 ≤ j ≤ (N−1)
2 . Since Z̃ j ∈ C
1
1
and − z̄ is an eigenvalue of Z̃ j , we have Z̃ j ξv1 = − z̄ ξv1 or
>
Z̃ j ξv2 = − 1z̄ ξv2 . Using this fact, we choose R̃21, j = τ̄ j 1 1 ,
>
>
and
where τ̄ j ∈ R and τ̄ j 6= 0, if Z̃ j 1 1 = − 1z̄ 1 1
>
R̃21, j = τ̄ j 1 −1
otherwise. Then we have Z̃ j R̃21, j =
i>
h >
>
(N−1)
1
− z̄ R̃21, j , 1 ≤ j ≤ 2 . Let R̄21 = R̃21,1 · · · R̃21, (N−1)
and
2 i
h
let R̄22 = diag r1 , −r1 , r2 , −r2 , · · · , r N−1 , −r N−1 , where
2
2
r j ∈ R, r j 6= 0, and |r j | 6= |rk | whenever j 6= k. Then it can be
verified that Z̄ R̄22 Z̄ = −
R̄22 , and Z̄ R̄21>z̄ = −R̄21 . In Lemma 1,
0
R̄21 Q
>
let us choose R = P
P, Γ = XRY and
Q > R̄21 Q > R̄22 Q
>
P = P > τ p 01×(N−1) where τ p ∈ C and τ p 6= 0. Then it
can be verified that R = R> ∈ RN×N and ZRZ = −R. It then
follows that Y RY − XRX = R and XRY + Y RX = 0. Hence
Γ + Γ> = 0, i.e., Γ is a skew symmetric matrix. Substituting R
and Γ into (6) gives G = diag[R, R]. Therefore, the Hamiltonian is Ĥ = 12 x̂> Gx̂, which satisfies the first constraint ¬. We
have
Q = −RZ
z̄
0
R̄>
Q
21
= −P
Q > R̄21 Q > R̄22 Q 0(N−1)×1
0
R̄>
Z̄Q
>
21
= −P
P.
Q > R̄21 z̄ Q > R̄22 Z̄Q
>
01×(N−1)
P
Q > Z̄Q
We now show that (Q, P) is controllable. Let R̃22, j =
diag[r j , −r j ]. We have
rank R̃21, j z̄ R̃22, j Z̃ j R̃21, j z̄
= rank R̃21, j z̄ − R̃22, j R̃21, j
= rank R̃21, j R̃22, j R̃21, j
τ̄ j r j τ̄ j
τ̄ j r j τ̄ j
= rank
or rank
τ̄ j −r j τ̄ j
−τ̄ j r j τ̄ j
(N − 1)
.
=2, 1 ≤ j ≤
2
Then it follows that R̃22, j Z̃ j , R̃21, j z̄ , 1 ≤ j ≤ (N−1)
2 , is controllable. In addition, we have (R̃22, j Z̃ j )2 = −R̃222, j = −r2j I2 .
Using Lemma 2 in [6], it follows that the matrix R̃22, j Z̃ j
is diagonalizable and its eigenvalues are either r j i or −r j i.
Hence R̃22, j Z̃ j , 1 ≤ j ≤ (N−1)
2 , have no common eigenvalues. By Lemma 6, we have (R̄22 Z̄, R̄21 z̄) is controllable. By
> R̄ Z̄Q, Q > R̄ z̄ is controllable. According
Lemma 4, Q
22
21
τp
0
R̄>
Z̄Q
21
to Lemma 5,
,
is con0(N−1)×1
Q > R̄21 z̄ Q > R̄22 Z̄Q
trollable. Based on Lemma 4, we conclude that (Q, P) is
controllable. Therefore, by Lemma 1, the resulting linear
quantum system is strictly stable and generates the desired
target pure Gaussian state specified by Equation (9). The
system-reservoir coupling vector L̂ is given by
L̂ = Cx̂ = P> [−Z IN ] x̂
>
= − τ p z̄ 01×(N−1) P q̂1 q̂2 · · · q̂N
>
+ τ p 01×(N−1) P p̂1 p̂2 · · · p̂N
= −τ p z̄q̂` + τ p p̂` ,
>
where ` = 1 01×(N−1) P 1 2 · · · N . It can be seen
that the system-reservoir coupling vector L̂ consists of only
one Lindblad operator which acts on the `th mode of the system. Hence it satisfies the second constraint . This completes
the first part of the sufficiency proof.
Suppose the graph matrix Z of an N-mode pure Gaussian
state (where N is even) satisfies Equation (10). We now
construct an N-mode linear quantum system subject to the
two constraints ¬ and , such that the state is obtained
as the unique steady state of this system. We see from
Equation (10) that Z̃1 = − 1z̄ and Z̃ j ∈ ∆, 2 ≤ j ≤ N2 . Using
>
Theorem 2 in [6], it can be shown that ξv1 = 1 1 and
>
ξv2 = 1 −1
are two eigenvectors of Z̃ j , 2 ≤ j ≤ N2 .
2×2
Since Z̃ j ∈ C
and − 1z̄ is an eigenvalue of Z̃ j , we have
1
Z̃ j ξv1 = − z̄ ξv1 or Z̃ j ξv2 = − 1z̄ ξv2 , 2 ≤ j ≤ N2 . Let R̃21,1 = τ̄1 ,
>
where τ̄1 ∈ R and τ̄1 6= 0. For 2 ≤ j ≤ N2 , let R̃21, j = τ̄ j 1 1 ,
>
>
and
where τ̄ j ∈ R and τ̄ j 6= 0, if Z̃ j 1 1 = − 1z̄ 1 1
>
R̃21, j = τ̄ j 1 −1
otherwise. Then we have Z̃ j R̃21, j =
h
i>
>
>
− 1z̄ R̃21, j , 1 ≤ j ≤ N2 . Let R̄21 = R̃21,1 · · · R̃21, N
and
2
i
h
let R̄22 = diag 0, r2 , −r2 , · · · , r N , −r N , where r j ∈ R,
2
2
r j 6= 0, j ≥ 2, and |r j | =
6 |rk | whenever j 6= k. Then it
can be verified
that Z̄ R̄22 Z̄> = −R̄22 and Z̄ R̄21 z̄ = −R̄21 .
0
R̄21 Q
Let R = P >
P, Γ = XRY and P =
Q > R̄21 Q > R̄22 Q
>
P > τ p 01×(N−1) where τ p ∈ C and τ p 6= 0. Then it can be
verified that R = R> ∈ RN×N and ZRZ = −R. It then follows
that Y RY − XRX = R and XRY +Y RX = 0. Hence Γ + Γ> = 0,
i.e., Γ is a skew symmetric matrix. Substituting R and Γ
into (6) gives G = diag[R, R]. Therefore, the Hamiltonian is
Ĥ = 12 x̂> Gx̂, which satisfies the first constraint ¬. We have
Q = −RZ
z̄
0
R̄>
Q
21
= −P
Q > R̄21 Q > R̄22 Q 0(N−1)×1
0
R̄>
Z̄Q
>
21
= −P
P.
Q > R̄21 z̄ Q > R̄22 Z̄Q
>
01×(N−1)
P
Q > Z̄Q
We now show that (Q, P) is controllable. Let R̃22,1 = 0 and
R̃22, j = diag[r j , −r j ], 2 ≤ j ≤ N2 . First, it can be seen that
R̃22,1 Z̃1 , R̃21,1 z̄ is controllable. We also have
rank R̃21, j z̄ R̃22, j Z̃ j R̃21, j z̄
= rank R̃21, j z̄ − R̃22, j R̃21, j
= rank R̃21, j R̃22, j R̃21, j
τ̄ j r j τ̄ j
τ̄ j r j τ̄ j
= rank
or rank
τ̄ j −r j τ̄ j
−τ̄ j r j τ̄ j
N
=2, 2 ≤ j ≤ .
2
Then it follows that R̃22, j Z̃ j , R̃21, j z̄ , 2 ≤ j ≤ N2 , is controllable. In addition, we have (R̃22, j Z̃ j )2 = −R̃222, j = −r2j I2 , 2 ≤
j ≤ N2 . Using Lemma 2 in [6], it follows that the matrix R̃22, j Z̃ j
is diagonalizable and its eigenvalues are either r j i or −r j i,
2 ≤ j ≤ N2 . Bearing in mind R̃22,1 Z̃1 = 0, it follows that R̃22, j Z̃ j ,
1 ≤ j ≤ N2 , have no common eigenvalues. By Lemma 6,
(R̄22 Z̄, R̄21 z̄) is controllable. It then follows from Lemma 4
>
Q > R̄22 Z̄Q,
It follows from
Q R̄21 z̄ is controllable.
>
Z̄Q
0
R̄>
21
τ
0
,
Lemma 5 that
p
1×(N−1)
Q > R̄21 z̄ Q > R̄22 Z̄Q
is controllable. According to Lemma 4, we conclude that
(Q, P) is controllable. Therefore, by Lemma 1, the resulting
linear quantum system is strictly stable and generates the
desired target pure Gaussian state specified by Equation (10).
The system-reservoir coupling vector L̂ is given by
that
L̂ = Cx̂ = P> [−Z IN ] x̂
>
= − τ p z̄ 01×(N−1) P q̂1 q̂2 · · · q̂N
>
+ τ p 01×(N−1) P p̂1 p̂2 · · · p̂N
= −τ p z̄q̂` + τ p p̂` ,
>
where ` = 1 01×(N−1) P 1 2 · · · N . It can be seen
that the system-reservoir coupling vector L̂ consists of only
one Lindblad operator which acts on the `th mode of the system. Hence it satisfies the second constraint . This completes
the second part of the sufficiency proof.
Proof of Lemma 2
Proof: Suppose Z ∈ ∆. Then
it follows
from Theorem 2
z11 z12
in [6] that Z has the form Z =
, where z11 ∈ C and
z12 z11
2
z12 ∈ C. Substituting this into the equation diag[1, −1]Z =
−I2 gives
z211 − z212 = −1.
(30)
The constraint det(Z + 1z̄ I2 ) = 0 is equivalent to
1 2 2
z11 +
− z12 = 0.
z̄
z̄2 −1
2z̄
z̄2 −1
2z̄
2
− z̄ 2z̄+1
Combining (30) and
# z11 "=
" 2 (31) yields
2
Therefore, Z =
z̄ −1
2z̄
z̄2 +1
2z̄
z̄ +1
2z̄
z̄2 −1
2z̄
or
(31)
2
z̄ +1
and z12 =
# ± 2z̄ .
2
− z̄ 2z̄+1
. Since
z̄2 −1
2z̄
Z = Z > , it can be easily seen that Re(Z) = Re(Z)> and
Im(Z) = Im(Z)> . To prove Im(Z) > 0, we have to show
Im(z11 ) > 0 and (Im(z11 ))2 − (Im(z12 ))2 > 0. Suppose z̄ =
x + iy, x ∈ R, y ∈ R, and y > 0. Then
z̄2 − 1 1
1
1
x − iy
z11 =
=
z̄ −
=
x + iy − 2
.
2z̄
2
z̄
2
x + y2
y
Hence Im(z11 ) = 21 y + x2 +y
> 0. Since
2
z̄2 + 1
1
1
1
x − iy
z12 = ±
=±
z̄ +
=±
x + iy + 2
,
2z̄
2
z̄
2
x + y2
we have
(Im(z11 ))2 − (Im(z12 ))2
2
2
1
y
1
y
=
y+ 2
−
y
−
4
x + y2
4
x2 + y2
y2
=
> 0.
(x2 + y2 )2
Therefore,
we
have
Im(Z) > 0.
Combining
the
results # above,
we
conclude
that
∆ =
("
"
#)
z̄2 −1
z̄2 +1
z̄2 −1
z̄2 +1
−
2z̄
2z̄
2z̄
2z̄
,
. This completes
2
z̄2 +1
z̄2 −1
z̄2 −1
− z̄ 2z̄+1
2z̄
2z̄
2z̄
the proof.
R EFERENCES
[1] K. Koga and N. Yamamoto, “Dissipation-induced pure Gaussian state,”
Physical Review A, vol. 85, no. 2, p. 022103, 2012.
[2] N. Yamamoto, “Pure Gaussian state generation via dissipation: a quantum stochastic differential equation approach,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering
Sciences, vol. 370, no. 1979, pp. 5324–5337, 2012.
[3] Y. Ikeda and N. Yamamoto, “Deterministic generation of Gaussian pure
states in a quasilocal dissipative system,” Physical Review A, vol. 87,
no. 3, p. 033802, 2013.
[4] S. Ma, M. J. Woolley, I. R. Petersen, and N. Yamamoto, “Preparation
of pure Gaussian states via cascaded quantum systems,” in Proceedings
of IEEE Conference on Control Applications (CCA), October 2014, pp.
1970–1975.
[5] S. Ma, M. J. Woolley, I. R. Petersen, and N. Yamamoto, “Cascade
and locally dissipative realizations of linear quantum systems for
pure Gaussian state covariance assignment,” arXiv:1604.03182, 2016.
[Online]. Available: http://arxiv.org/abs/1604.03182
[6] S. Ma, I. R. Petersen, and M. J. Woolley, “Linear quantum systems
with diagonal passive Hamiltonian and a single dissipative channel,”
arXiv:1511.04929, 2015. [Online]. Available: http://arxiv.org/abs/1511.
04929
[7] J. F. Poyatos, J. I. Cirac, and P. Zoller, “Quantum reservoir engineering
with laser cooled trapped ions,” Physical Review Letters, vol. 77, no. 23,
pp. 4728–4731, 1996.
[8] M. J. Woolley and A. A. Clerk, “Two-mode squeezed states in cavity
optomechanics via engineering of a single reservoir,” Physical Review
A, vol. 89, no. 6, p. 063805, 2014.
[9] H. M. Wiseman and G. J. Milburn, Quantum Measurement and Control.
Cambridge University Press, 2010.
[10] N. C. Menicucci, S. T. Flammia, and P. van Loock, “Graphical calculus
for Gaussian pure states,” Physical Review A, vol. 83, no. 4, p. 042335,
2011.
[11] C. T. Chen, Linear System Theory and Design, 3rd ed.
Oxford
University Press, 1999.
[12] R. Simon, N. Mukunda, and B. Dutta, “Quantum-noise matrix for
multimode systems: U(n) invariance, squeezing, and normal forms,”
Physical Review A, vol. 49, no. 3, pp. 1567–1583, 1994.
[13] N. C. Gabay and N. C. Menicucci, “Passive interferometric symmetries
of multimode Gaussian pure states,” Physical Review A, vol. 93, no. 5,
p. 052326, 2016.
[14] S. Zippilli, J. Li, and D. Vitali, “Steady-state nested entanglement
structures in harmonic chains with single-site squeezing manipulation,”
Physical Review A, vol. 92, no. 3, p. 032319, 2015.
[15] M. S. Kim, W. Son, V. Bužek, and P. L. Knight, “Entanglement by
a beam splitter: nonclassicality as a prerequisite for entanglement,”
Physical Review A, vol. 65, no. 3, p. 032323, 2002.
[16] G. Vidal and R. F. Werner, “Computable measure of entanglement,”
Physical Review A, vol. 65, no. 3, p. 032314, 2002.
[17] M. B. Plenio, “The logarithmic negativity: a full entanglement monotone
that is not convex,” Physical Review Letters, vol. 95, no. 9, p. 090503,
2005.
[18] G. Adesso, A. Serafini, and F. Illuminati, “Extremal entanglement and
mixedness in continuous variable systems,” Physical Review A, vol. 70,
no. 2, p. 022318, 2004.
[19] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control.
Prentice Hall, 1996.
[20] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. Cambridge
University Press, 2012.
[21] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas,
2nd ed. Princeton University Press, 2009.
| 3cs.SY
|
Extremum Seeking-based Iterative Learning Model Predictive Control
(ESILC-MPC)
arXiv:1512.02627v1 [cs.SY] 8 Dec 2015
Anantharaman Subbaraman, Mouhacine Benosman
Abstract
In this paper, we study a tracking control problem for linear time-invariant systems, with model parametric
uncertainties, under input and states constraints. We apply the idea of modular design introduced in [1], to solve
this problem in the model predictive control (MPC) framework. We propose to design an MPC with input-to-state
stability (ISS) guarantee, and complement it with an extremum seeking (ES) algorithm to iteratively learn the
model uncertainties. The obtained MPC algorithms can be classified as iterative learning control (ILC)-MPC.
I. I NTRODUCTION
Model predictive control (MPC), e.g., [2], is a model-based framework for optimal control of constrained multi-variable systems. MPC is based on the repeated, receding horizon solution of a finite-time
optimal control problem formulated from the system dynamics, constraints on system states, inputs,
outputs, and a cost function describing the control objective. However, since MPC is a model-based
controller, its performance inevitably depends on the quality of the prediction model used in the optimal
control computation.
In contrast, extremum seeking (ES) control is a well known approach where the extremum of a cost
function associated with a given process performance (under some conditions) is found without the need
for detailed modelling information, see, e.g., [3], [4], [5]. Several ES algorithms (and associated stability
analysis) have been proposed, [6], [4], [7], [5], [7], [3], [8], [9], and many applications of ES have been
reported [10], [11], [12], [13], [14].
The idea that we want to theoretically analyze in this paper, is that the performance of a modelbased MPC controller can be combined with the robustness of a model-free ES learning algorithm for
simultaneous identification and control of linear time-invariant systems with structural uncertainties. We
refer the reader to [1], [14], [15] where this idea of learning-based modular adaptive control has been
introduced in a more general setting of nonlinear dynamics
A. Subbaraman ([email protected]) is with the Department of Electrical Engineering, University of California at Santa
Barbara, USA. M. Benosman (m [email protected]) is with Mitsubishi Electric Research Laboratories (MERL), Cambridge, MA 02139,
USA.
We aim at proposing an alternative approach to realize an iterative learning-based adaptive MPC. We
introduce an approach for an ES-based iterative learning MPC that merges a model-based linear MPC
algorithm with a model-free ES algorithm to realize an iterative learning MPC that adapts to structured
model uncertainties. Due to the iterative nature of the learning model improvement, we want here to
compare the proposed approach to some existing Iterative learning control (ILC)-MPC methods. Indeed,
ILC method introduced in [16] is a control technique which focuses on improving tracking performance
of processes that repeatedly execute the same operation over time. It is of particular importance in
robotics and in chemical process control of batch processes. We refer the reader to [17], [18] and [19]
for more details on ILC and its applications.
At the intersection of learning based control and constrained control is the ILC-MPC concept. For
instance, ILC-MPC for chemical batch processes are studied in [20], [21], and [22]. As noted in [21]
one of the shortcomings of the current literature is a rigorous justification of feasibility, and Lyapunovbased stability analysis for ILC-MPC . For example, in [20] the goal is to reduce the error between
the reference and the output over multiple trials while satisfying only input constraints. However, the
reference signals is arbitrary and the MPC scheme for tracking such signals is not rigorously justified.
Furthermore, the MPC problem does not have any stabilizing conditions (terminal cost or terminal
constraint set). The ILC update law is an addition of the MPC signal of the current trial to the MPC
signal of the previous trial. In [21], an ILC-MPC scheme for a general class of nonlinear systems with
disturbances is proposed. The proof is presented only for MPC without constraints. In [22], the ILC
update law is designed using MPC. State constraints are not considered in [22]. In [23] a batch MPC
(BMPC) is proposed, which integrates conventional MPC scheme with an iterative learning scheme. A
simplified static input-output map is considered in the paper as opposed to a dynamical system.
In summary, we think that there is a need for more rigorous theoretical justification attempted in
this paper. Furthermore, to the best of our knowledge, the literature on ILC-MPC schemes do not
consider state constraints, do not treat robust feasibility issues in the MPC tracking problem, rigorous
justification of reference tracking proofs for the MPC is not present in the literature and stability proofs
for the combination of the ILC and MPC schemes are not established in a systematic manner. Finally,
we want to cite the work of [24], [25], [26], where similar control objectives as the one targeted in
this paper, have been studied using a learning-based MPC approach. The main differences are in the
control/learning design methodology and the proof techniques.
The main contribution of this work is to present a rigorous proof of an ILC-MPC scheme using existing
Lyapunov function based stability analysis established in [27] and extremum seeking algorithms in
[28], to justify the modular design method for ILC-MPC proposed in [29], where an ES-based modular
approach to design ILC-MPC schemes for a class of constrained linear systems is proposed.
The rest of the paper is organized as follows. Section II contains some useful notations and definitions.
The MPC control problem formulation is presented in Section III. Section IV is dedicated to a rigorous
analysis of the proposed ES-based ILC-MPC. Finally, some concluding comments are presented in
Section V.
II. N OTATION
AND BASIC DEFINITIONS
Throughout this paper, R denotes the set of real numbers and Z denotes the set of integers. State
constraints and input constraints are represented by X ⊂ Rn and U ⊂ Rm respectively. The optimization
horizon for MPC is denoted by N ∈ Z≥1 . The feasible region for the MPC optimization problem is
denoted by XN . A continuous function α : R≥0 → R≥0 with α(0) = 0 belongs to class K if it is
increasing and bounded. A function β belongs to class K∞ if it belongs to class K and is unbounded.
A function β(s, t) ∈ KL if β(·, t) ∈ K and limt→∞ β(s, t) = 0. Given two sets A and B, such that
A ⊂ Rn , B ⊂ Rn , the Minkowski sum is defined as A ⊕ B := {a + b|a ∈ A, b ∈ B}. The Pontryagin
set difference is defined as A ⊖ B := {x|x ⊕ B ∈ A}. Given a matrix M ∈ Rm×n , the set MA ⊂ Rm ,
is defined as MA , {Ma : a ∈ A}. A positive definite matrix is denoted by P > 0. The standard
√
Euclidean norm is represented as |x| for x ∈ Rn , |x|P := xT P x for a positive definite matrix P ,
|x|A := inf y∈A |x − y| for a closed set A ⊂ Rn and kAk represents an appropriate matrix norm where
A is a matrix. B represents the closed unit ball in the Euclidean space. Also, a matrix M ∈ Rn×n is
said to be Schur iff all its eigenvalues are inside the unitary disk.
III. P ROBLEM
FORMULATION
In this section we will describe in detail the problem studied in this paper. We consider linear systems
of the form
x(k + 1) = (A + ∆A)x(k) + (B + ∆B)u(k),
y(k) = Cx(k) + Du(k),
(1)
(2)
where ∆A and ∆B represent the uncertainty in the system model. We will assume that the uncertainties
are bounded as follows:
Assumption 1: The uncertainties k∆Ak ≤ ℓA and k∆Bk ≤ ℓB for some ℓA , ℓB > 0.
Next, we impose some assumptions on the reference signal r.
Assumption 2: The reference signal r : [0, T ] → R is a piecewise constant trajectory for some T > 0.
Under Assumptions 1 and 2, the goal is to design a control scheme guarantying tracking with
sufficiently small errors by learning the uncertain parameters of the system. Next, we will explain
in detail the optimization problem associated with the MPC based controller. The results stated here are
from [27]. We exploit the analysis results in [27] to establish that the closed-loop system has an ISS
property with respect to the parameter estimation error.
Since the value of ∆A and ∆B are not known a priori, the MPC uses a model of the plant based on
ˆ and ∆B.
ˆ
the current estimate ∆A
We will now formulate the MPC problem with a given estimate of the uncertainty for a particular
iteration of the learning process. We will rewrite the system dynamics as
x(k + 1) = f (x, u) + g(x, u, ∆) = F (x, u, ∆),
(3)
where f (x, u) = Ax + Bu and g(x, u, ∆) = ∆Ax + ∆Bu.
Assumption 3: The state constraint set X ⊂ Rn and control constraint set U ⊂ Rm are compact,
convex polyhedral sets.
ˆ ∆B
ˆ and is expressed as
The MPC model is generated using an estimate ∆A,
ˆ = F (x, u, ∆).
ˆ
x(k + 1) = f (x, u) + g(x, u, ∆)
(4)
We can now rewrite the actual model as
ˆ + (∆A − ∆A)x
ˆ
ˆ
x(k + 1) = f (x, u) + g(x, u, ∆)
+ (∆B − ∆B)u.
(5)
This system can now be compared to the model in [27]. So we have
ˆ + w(k),
x(k + 1) = F (x(k), u(k), ∆)
(6)
ˆ
ˆ
w(k) = (∆A − ∆A)x(k)
+ (∆B − ∆B)u(k),
(7)
where
and x(k) ∈ X , u(k) ∈ U. The following assumption will be justified in the next section.
ˆ
ˆ
Assumption 4: The estimates of the uncertain parameters are bounded with k∆Ak
≤ ℓA and k∆Bk
≤
ℓB for all iterations of the extremum seeking algorithm.
We now impose certain conditions on the disturbance w(k) and system matrices in accordance with
[27, Assumption 1].
ˆ B + ∆B)
ˆ
ˆ and ∆B.
ˆ
Assumption 5: The pair (A + ∆A,
is controllable for every realization of ∆A
We will denote the actual model using (x, u) and the MPC model through (x̄, ū). Hence we have
ˆ + w,
x(k + 1) = F (x, u, ∆)
ˆ
x̄(k + 1) = F (x̄, ū, ∆).
A. Robust positive invariant sets
We denote the error between the states of the true model and MPC model by e(k) = x(k) − x̄(k).
We want the error to be bounded during tracking. The error dynamics is then given by
ˆ + (B + ∆B)K)e(k)
ˆ
e(k + 1) = (A + ∆A
+ w(k),
(8)
ˆ + (B + ∆B)K)
ˆ
where u = ū + Ke and the matrix K is such that AK := (A + ∆A
is Schur.
We first recall the definition of a robust positive invariant set (RPI), e.g., [27].
Definition 1: A set ΦK is called an RPI set for the uncertain dynamics (8), if AK Φk ⊕ W ⊆ ΦK .
So, we let ΦK be an RPI set associated with the error dynamics (8), i.e., AK ΦK ⊕ W ⊆ ΦK .
B. Tightening the constraints
Now we follow [27] and tighten the constraints for the MPC model so that we achieve robust constraint
satisfaction for the actual model with uncertainties. Let X1 = X ⊖ΦK and U1 = U ⊖KΦK . The following
result is from [30, Proposition 1, Theorem 1 and Corollary 1 ].
Proposition 1: Let ΦK be RPI for the error dynamics. If e(0) ∈ ΦK , then x(k) ∈ x̄(k) ⊕ ΦK for all
k ≥ 0 and w(k) ∈ W. If in addition, x̄(k) ∈ X1 and ū(k) ∈ U1 then with the control law u = ū + Ke,
x(k) ∈ X and u(k) ∈ U for all k ≥ 0.
C. Invariant set for tracking
As in [30] and [27], we will characterize the set of nominal steady states and inputs so that we can
relate them later to the tracking problem. Let zs = (x̄s , ūs ) be the steady state for the MPC model.
Then,
ˆ
ˆ
A + ∆A − I B + ∆B
x̄
0
s = .
C
D
ūs
ȳs
(9)
From the controllability assumption on the system matrices, the admissible steady states can be characterized by a single parameter θ̄ as
z̄s = Mθ θ̄,
(10)
ȳs = Nθ θ̄,
(11)
for some θ̄ and matrices Mθ and Nθ = [C D]Mθ . We let Xs , Us denote the set of admissible steady
states that are contained in X1 , U1 and satisfy (9). Ys denotes the set of admissible output steady states.
Now we will define an invariant set for tracking which will be utilized as a terminal constraint for the
optimization problem.
Definition 2: [27, Definition 2] An invariant set for tracking for the MPC model is the set of initial
conditions, steady states and inputs (characterized by θ̄) that can be stabilized by the control law ū =
K̄ x̄ + Lθ̄ with L := [−K̄ I]Mθ while (x̄(k), ū(k)) ∈ X1 × U1 for all k ≥ 0.
ˆ
ˆ K̄) is Schur. We refer the reader to [30]
+ ∆B)
We choose the matrix K̄ such that AK̄ := (A+ ∆A+(B
and [27] for more details on computing the invariant set for tracking. We will refer to the invariant set for
tracking as ΩK̄ . We say a point (x̄(0), θ̄) ∈ ΩK̄ if with the control law u = K̄(x̄ − x̄s ) + ūs = K̄ x̄ + Lθ̄,
the solutions of the MPC model from x̄(0) satisfy x̄(k) ∈ Projx (ΩK̄ ) for all k ≥ 0. As stated in [27]
the set can be taken to be a polyhedral.
D. MPC Optimization problem
Now we will define the optimization problem that will be solved at every instant to determine the
control law for the actual plant dynamics. For a given target setpoint yt and initial condition x, the
optimization problem PN (x, yt ) is defined as,
min VN (x, yt , x̄(0), θ̄, ū)
x̄(0),θ̄,ū
s.t x̄(0) ∈ x ⊕ (−ΦK )
ˆ
ˆ
x̄(k + 1) = (A + ∆A)x̄(k)
+ (B + ∆B)ū(k)
x̄s = Mθ θ̄
ȳs = Nθ θ̄
(x̄(k), ū(k)) ∈ X1 × U1 , k ∈ Z≤N −1
(x̄(N), θ̄) ∈ ΩK̄ ,
where the cost function is defined as follows
VN (x, yt , x̄(0), θ̄, ū) =
+|ū(k) −
ūs |2R
+ |x̄(N) −
N
−1
X
k=0
x̄s |2P
|x̄(k) − x̄s |2Q̃
+ |ȳs − yt |2T .
(12)
Such cost function is frequently used in MPC literature for tracking except for the additional term in
the end which penalizes the difference between the artificial stead state and the actual target value. We
refer the reader to [31], [30] and [27] for more details.
Assumption 6: The following conditions are satisfied by the optimization problem
1) The matrices Q̃ > 0, R > 0, T > 0.
ˆ + (B + ∆B)K)
ˆ
2) (A + ∆A
is Schur matrix, ΦK is a RPI set for the error dynamics, and X1 , U1 are
non-empty.
ˆ + (B + ∆B)
ˆ K̄ is Schur and P > 0 satisfies:
3) The matrix K̄ is such that A + ∆A
ˆ + (B + ∆B)
ˆ K̄)T P (A + ∆A
ˆ + (B + ∆B)
ˆ K̄) =
P − (A + ∆A
Q̃ + K̄ T RK̄.
4) The set ΩK̄ is an invariant set for tracking subject to the tightened constraints X1 , U1 .
As noted in [27], the feasible set XN does not vary with the set points yt and the optimization problem
is a Quadratic programming (QP) problem. The optimal values are given by x̄∗s , ū∗ (0), x̄∗ . The MPC
control law writes then as: u = κN (x) = K(x − x̄∗ ) + ū∗ (0). The MPC law κN implicitly depends on
ˆ Also it follows from the results in [32] that the control law
the current estimate of the uncertainty ∆.
for the MPC problem is continuous1.
IV. DIRECT
EXTREMUM SEEKING - BASED ITERATIVE LEARNING
MPC
A. DIRECT-based iterative learning MPC
In this section we will explain the assumptions regarding the learning cost function2 used for identifying the true parameters of the uncertain system via nonlinear programming based extremum seeking
called DIRECT, e.g., [33]. Let ∆ be a vector that contains the entries in ∆A and ∆B. Similarly the
ˆ Then ∆, ∆
ˆ ∈ Rn(n+m) .
estimate will be denoted by ∆.
Since we do not impose the presence of attractors for the closed-loop system as in [34] or [35], the
cost function that we utilize Q : Rn(n+m) → R≥0 depends on x0 . For iterative learning methods, the
ˆ as
same initial condition x0 is used to learn the uncertain parameters and hence we refer to Q(x0 , ∆)
ˆ since x0 is fixed.
only Q(∆)
Assumption 7: The learning cost function Q : Rn(n+m) → R≥0 is
1) Lipshitz in the compact set of uncertain parameters
ˆ for all ∆
ˆ =
2) The true parameter ∆ is such that Q(∆) < Q(∆)
6 ∆.
One example of a learning cost functions is identification-type cost function, where the error between
outputs measurements from the system are compared to the MPC model outputs. Another example of a
learning cost function, can be a performance-type cost function, where a measured output of the system
is directly compared to a desired reference trajectory.
We then use the DIRECT optimization algorithm introduced in [33] for finding the global minimum
of a Lipschitz function without knowledge of the Lipschitz constant. The algorithm is implemented
in MATLAB using [36]. We will utilize a modified termination criterion introduced in [35] for the
1
2
The authors would like to thank Dr. S. Di Cairano for pointing out to us the paper [32].
Not to be confused with the MPC cost function.
DIRECT algorithm to make it more suitable for extremum seeking applications. As we will mention in
later sections, the DIRECT algorithm has nice convergence properties which will be used to establish
our main results.
B. Main results: Proof of the MPC ISS and the learning convergence
We will now present the main results of this paper, namely the stability analysis of the proposed
ILC-MPC algorithm, using the existing results for MPC tracking and DIRECT algorithm established in
[27] and [28], respectively.
First, we define the value function
VN∗ (x, yt ) = min VN (x, yt , x̄(0), θ, ū)
x̄(0),θ,ū
for a fixed target yt . Also, we let θ̃ := arg minθ̄ |Nθ θ̄ − yt |, (x̃s , ũs ) = Mθ θ̃ and ỹs = C x̃s + Dũs . If
the target steady state yt is not admissible, the MPC tracking scheme drives the output to converge to
the point ỹs which is a steady state output that is admissible and also minimizes the error with the
target steady state, i.e., graceful target degradation principle, e.g., [37]. The proof of the following result
follows from [27, Theorem 1] and classically uses VN∗ (x, yt ) as a Lyapunov function for the closed-loop
system.
Proposition 2: Let yt be given. For all x(0) ∈ XN , the MPC problem is recursively feasible. The
state x(k) converges to x̃s ⊕ ΦK and the output y(k) converges to ỹs ⊕ (C + DK)ΦK .
The next result states the convergence properties of the modified DIRECT algorithm, which we will
used in establishing the main result. This result is stated as [28, Assumption 7] and it follows from the
analysis of the modified DIRECT algorithm in [35].
ˆ t , t = 1, 2, ... from the modified DIRECT algorithm
Proposition 3: For any sequence of updates ∆
ˆ t | ≤ ε for t ≥ N.
and ε > 0, there exists a N > 0 such that |∆ − ∆
Remark 1: Note that the results in [35] also include a robustness aspect of the DIRECT algorithm.
This can be used to account for measurement noises and computational error associated with the learning
cost Q.
We now state the main result of the section that combines the ISS MPC formulation and the extremum
seeking algorithm.
Theorem 1: Under Assumptions 1-7, given an initial condition x0 , an output target yt , such that yt
is constant over [0, T ∗] for some T ∗ sufficiently large. Then, for every ε > 0, there exists N1 and N2
such that |y(k) − ỹs | <= ε for k ∈ [N1 , T ∗] after N2 iterations of the ILC-MPC scheme.
Proof 1: It can observed that since the size of ΦK grows with the size of W and ΦK = {0} for the
case without disturbances that without loss of generality ΦK ⊆ Γ(w ∗)B, where Γ ∈ K and w ∗ = k∆A −
∗
∗
ˆ
ˆ
∆AkX
+ k∆B − ∆BkU
, where X ∗ = maxx∈X |x| and U ∗ = maxu∈U |u|. Here X ∗ , U ∗ are fixed over
both regular time and learning iteration number, but the uncertainties vary over iterations because of the
modified DIRECT algorithm updates. Since the worst case disturbance depends directly on the estimation
ˆ
ˆ
error, without loss of generality we have that ΦK ⊆ γ(|∆ − ∆|)B
and (C + DK)ΦK ⊆ γ ∗ (|∆ − ∆|)B
for some γ, γ ∗ ∈ K . It follows from Proposition 2 that limk→∞ |x(k)|x̃s ⊕ΦK = 0. Then,
lim |x(k) − x̃s | ≤ max |x|
k→∞
x∈ΦK
ˆ
≤ γ(|∆ − ∆|).
We observe that the above set of equations state that the closed-loop system with the MPC controller
has the asymptotic gain property and it is upper bounded by the size of the parameter estimation error.
ˆ is constant for a particular iteration of the process. Also, for the case of no
Note that the estimate ∆
uncertainties we have 0−stability (Lyapunov stability for the case of zero uncertainty). This can be proven
by using the cost function VN∗ (x, yt ) as the Lyapunov function, such that VN∗ (x(k +1), yt) ≤ VN∗ (x(k), yt )
and λmin (Q̃)|x − x̃s |2 ≤ VN∗ (x, yt ) ≤ λmax (P )|x − x̃s |2 , see [38]. Furthermore, here the stability and
asymptotic gain property can be interpreted with respect to the compact set A := {x̃s }.
Since the MPC control law is continuous, the closed-loop system for a particular iteration of the
ILC-MPC scheme is also continuous with respect to the state. Then, from [39, Theorem 3.1] we can
conclude that the closed-loop system is ISS with respect to the parameter estimation error and hence
satisfies,
ˆ
|x(k) − x̃s | ≤ β(|x(0) − x̃s |, k) + γ̂(|∆ − ∆|),
where β ∈ KL and γ̂ ∈ K. Now, let ε1 > 0 be small enough such that γ̂(ε1 ) ≤ ε/2. From Proposition
ˆ t | ≤ ε1 for t ≥ N2 , where t is the iteration
3, it follows that there exists a N2 > 0 such that |∆ − ∆
number of the ILC-MPC scheme. Hence there exists N1 > 0 such that |β(|x(0) − x̃s |, k)| ≤ ε/2 for
k ≥ N1 . We choose T ∗ such that T ∗ > N1 . Then, we have that for k ∈ [N1 , T ∗ ] and for t ≥ N2 ,
|x(k) − x̃s | ≤ ε.
Similarly, using the linearity dependence between y and x, we can also establish that, ∃ ε̃(ε), such
that for k ∈ [N1 , T ∗ ] and for t ≥ N2
|y(k) − ỹs | ≤ ε̃(ε).
V. C ONCLUSION
In this paper, we have reported some results about extremum seeking-based ILC-MPC algorithms. We
have argued that it is possible to merge together a model-based linear MPC algorithm with a model-free
ES algorithm to iteratively learn structural model uncertainties and thus improve the overall performance
of the MPC controller. We have presented the stability analysis of this modular design technique for
ES-based ILC-MPC. where we addressed both feasibility and tracking performance. Future work can
include extending this method to a wider class of nonlinear systems, tracking a more richer class of
signals, employing different non-smooth optimization techniques for the extremum seeking algorithm,
etc.
VI. REFERENCES
[1] M. Benosman, “Learning-based adaptive control for nonlinear systems,” in European Control
Conference, 2014, pp. 920–925.
[2] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Constrained model predictive
control: Stability and optimality,” Automatica, vol. 36, pp. 789–814, 2000.
[3] K. B. Ariyur and M. Krstic, Real-time optimization by extremum-seeking control. John Wiley &
Sons, 2003.
[4] ——, “Multivariable extremum seeking feedback: Analysis and design,” in Proc. of the Mathematical Theory of Networks and Systems, South Bend, IN, August 2002.
[5] D. Nesic, “Extremum seeking control: Convergence analysis,” European Journal of Control, vol. 15,
no. 34, pp. 331 – 347, 2009.
[6] M. Krstic, “Performance improvement and limitations in extremum seeking,” Systems & Control
Letters, vol. 39, pp. 313–326, 2000.
[7] Y. Tan, D. Nesic, and I. Mareels, “On non-local stability properties of extremum seeking control,”
Automatica, no. 42, pp. 889–903, 2006.
[8] M. Rotea, “Analysis of multivariable extremum seeking algorithms,” in Proceedings of the American
Control Conference, vol. 1, no. 6. IEEE, 2000, pp. 433–437.
[9] M. Guay, S. Dhaliwal, and D. Dochain, “A time-varying extremum-seeking control approach,” in
American Control Conference, 2013.
[10] T. Zhang, M. Guay, and D. Dochain, “Adaptive extremum seeking control of continuous stirred-tank
bioreactors,” AIChE J., no. 49, p. 113123., 2003.
[11] N. Hudon, M. Guay, M. Perrier, and D. Dochain, “Adaptive extremum-seeking control of
convection-reaction distributed reactor with limited actuation,” Computers & Chemical Engineering,
vol. 32, no. 12, pp. 2994 – 3001, 2008.
[12] C. Zhang and R. Ordez, Extremum-Seeking Control and Applications. Springer-Verlag, 2012.
[13] M. Benosman and G. Atinc, “Multi-parametric extremum seeking-based learning control for
electromagnetic actuators,” in American Control Conference, 2013.
[14] ——, “Nonlinear learning-based adaptive control for electromagnetic actuators,” in European
Control Conference, 2013.
[15] G. Atinc and M. Benosman, “Nonlinear learning-based adaptive control for electromagnetic
actuators with proof of stability,” in IEEE, Conference on Decision and Control, 2013.
[16] S. Arimoto, “Robustness of learning control for robot manipulators,” in Proceedings of the IEEE
International Conference on Robotics and Automation. IEEE, 1990, pp. 1528–1533.
[17] Y. Wang, F. Gao, and F. J. Doyle III, “Survey on iterative learning control, repetitive control, and
run-to-run control,” Journal of Process Control, vol. 19, no. 10, pp. 1589–1600, 2009.
[18] K. L. Moore, “Iterative learning control: an expository overview,” in Applied and computational
control, signals, and circuits. Springer, 1999, pp. 151–214.
[19] H.-S. Ahn, Y. Chen, and K. L. Moore, “Iterative learning control: brief survey and categorization,”
IEEE Transactions on Systems Man and Cybernetics, vol. 37, no. 6, p. 1099, 2007.
[20] Y. Wang, D. Zhou, and F. Gao, “Iterative learning model predictive control for multi-phase batch
processes,” Journal of Process Control, vol. 18, no. 6, pp. 543–557, 2008.
[21] J. R. Cueli and C. Bordons, “Iterative nonlinear model predictive control. stability, robustness and
applications,” Control Engineering Practice, vol. 16, no. 9, pp. 1023–1034, 2008.
[22] J. Shi, F. Gao, and T.-J. Wu, “Single-cycle and multi-cycle generalized 2D model predictive iterative
learning control (2D-GPILC) schemes for batch processes,” Journal of Process Control, vol. 17,
no. 9, pp. 715–727, 2007.
[23] K. S. Lee, I.-S. Chin, H. J. Lee, and J. H. Lee, “Model predictive control technique combined with
iterative learning for batch processes,” AIChE Journal, vol. 45, no. 10, pp. 2175–2187, 1999.
[24] A. Aswani, H. Gonzalez, S. S. Sastry, and C. Tomlin, “Provably safe and robust learning-based
model predictive control,” Automatica, vol. 49, no. 5, pp. 1216–1226, 2013.
[25] A. Aswani, P. Bouffard, and C. Tomlin, “Extensions of learning-based model predictive control
for real-time application to a quadrotor helicopter,” in American Control Conference (ACC), 2012.
IEEE, 2012, pp. 4661–4666.
[26] A. Aswani, N. Master, J. Taneja, D. Culler, and C. Tomlin, “Reducing transient and steady state
electricity consumption in hvac using learning-based model-predictive control,” Proceedings of the
IEEE, vol. 100, no. 1, pp. 240–253, 2012.
[27] D. Limon, I. Alvarado, T. Alamo, and E. Camacho, “Robust tube-based MPC for tracking of
constrained linear systems with additive disturbances,” Journal of Process Control, vol. 20, no. 3,
pp. 248–260, 2010.
[28] S. Z. Khong, D. Nešić, Y. Tan, and C. Manzie, “Unified frameworks for sampled-data extremum
seeking control: Global optimisation and multi-unit systems,” Automatica, vol. 49, no. 9, pp. 2720–
2733, 2013.
[29] M. Benosman, S. D. Cairano, and A. Weiss, “Extremum seeking-based iterative learning linear
MPC,” in IEEE Multi-conference on Systems and Control, 2014.
[30] I. Alvarado, D. Limon, T. Alamo, M. Fiacchini, and E. Camacho, “Robust tube based MPC for
tracking of piece-wise constant references,” in Decision and Control, 2007 46th IEEE Conference
on. IEEE, 2007, pp. 1820–1825.
[31] I. Alvarado, D. Limon, T. Alamo, and E. Camacho, “Output feedback robust tube based MPC
for tracking of piece-wise constant references,” in Proceedings of the 46th IEEE Conference on
Decision and Control. IEEE, 2007, pp. 2175–2180.
[32] A. Bemporad, M. Morari, V. Dua, and E. N. Pistikopoulos, “The explicit linear quadratic regulator
for constrained systems,” Automatica, vol. 38, no. 1, pp. 3–20, 2002.
[33] D. R. Jones, C. D. Perttunen, and B. E. Stuckman, “Lipschitzian optimization without the Lipschitz
constant,” Journal of Optimization Theory and Applications, vol. 79, no. 1, pp. 157–181, 1993.
[34] D. Popovic, M. Jankovic, S. Magner, and A. R. Teel, “Extremum seeking methods for optimization
of variable cam timing engine operation,” IEEE Transactions on Control Systems Technology,
vol. 14, no. 3, pp. 398–407, 2006.
[35] S. Z. Khong, D. Nešić, C. Manzie, and Y. Tan, “Multidimensional global extremum seeking via
the DIRECT optimisation algorithm,” Automatica, vol. 49, no. 7, pp. 1970–1978, 2013.
[36] D. E. Finkel, “DIRECT optimization algorithm user guide,” Center for Research in Scientific
Computation, North Carolina State University, vol. 2, 2003.
[37] M. Benosman and K.-Y. Lum, “On-line references reshaping and control re-allocation for nonlinear
fault tolerant control,” IEEE, Transactions on Control Systems Technology, vol. 17, no. 2, pp. 366–
379, March 2009.
[38] D. Limón, I. Alvarado, T. Alamo, and E. F. Camacho, “MPC for tracking piecewise constant
references for constrained linear systems,” Automatica, vol. 44, no. 9, pp. 2382–2387, 2008.
[39] C. Cai and A. R. Teel, “Characterizations of input-to-state stability for hybrid systems,” Systems
& Control Letters, vol. 58, no. 1, pp. 47–53, 2009.
| 3cs.SY
|
Reconfigurable Antenna Multiple Access for 5G
mmWave Systems
Mojtaba Ahmadi Almasi∗ , Hani Mehrpouyan∗, David Matolak∗∗ , Cunhua Pan†, Maged Elkashlan†
∗ Department of Electrical and Computer Engineering, Boise State University, {mojtabaahmadialm,hanimehrpouyan}@boisestate.edu
∗∗ Department of Electrical Engineering, University of South Carolina, {matolak}@cec.sc.edu
arXiv:1803.09918v1 [cs.IT] 27 Mar 2018
† School of Electronic Engineering and Computer Science, Queen Mary University of London, {c.pan,maged.elkashlan}@qmul.ac.uk
Abstract—This paper aims to realize a new multiple access
technique based on recently proposed millimeter-wave reconfigurable antenna architectures. To this end, first we show that integration of the existing reconfigurable antenna systems with the
well-known non-orthogonal multiple access (NOMA) technique
causes a significant degradation in sum rate due to the inevitable
power division in reconfigurable antennas. To circumvent this
fundamental limit, a new multiple access technique is proposed.
The technique which is called reconfigurable antenna multiple
access (RAMA) transmits only each user’s intended signal at
the same time/frequency/code, which makes RAMA an inter-user
interference-free technique. Two different cases are considered,
i.e., RAMA with partial and full channel state information (CSI).
In the first case, CSI is not required and only the direction
of arrival for a specific user is used. Our analytical results
indicate that with partial CSI and for symmetric channels,
RAMA outperforms NOMA in terms of sum rate. Further, the
analytical result indicates that RAMA for asymmetric channels
achieves better sum rate than NOMA when less power is assigned
to users that experience better channel quality. In the second
case, RAMA with full CSI allocates optimal power to each user
which leads to higher achievable rates compared to NOMA
for both symmetric and asymmetric channels. The numerical
computations demonstrate the analytical findings.
I. I NTRODUCTION
The rapid growth of global mobile data traffic is expected
to be satisfied by exploiting a plethora of new technologies,
deemed as 5th generation (5G) networks. To meet this demand,
millimeter-wave (mmWave) communications operating in the
30 − 300 GHz range is emerging as one of the most promising solutions [1]. The existence of a large communication
bandwidth at mmWave frequencies represents the potential for
significant throughput gains. Indeed, the shorter wavelengths
at the mmWave band allow for the deployment of a large
number of antenna elements in a small area, which enables
mmWave systems to potentially support more higher degrees
of beamforming gain and multiplexing [1]. However, significant path loss, channel sparsity, and hardware limitations are
major obstacles for the deployment of mmWave systems. In
order to address these obstacles, several mmWave systems
have been proposed to date [2]–[4].
An analog beamforming mmWave system is designed in [2]
which uses one radio frequency (RF) chain and can support
only one data stream. Subsequent work considers a hybrid
beamforming mmWave system to transmit multiple streams
This project is supported in part by the NSF ERAS grant award numbers
1642865.
by exploiting several RF chains [3]. In [4], the concept of
beamspace multi-input multi-output (MIMO) is introduced
where several RF chains are connected to a lens antenna array
via switches. In addition to these systems, non-orthogonal
multiple access (NOMA) has been also considered as another
promising enabling technique for 5G to enhance spectral
efficiency in multi-user scenarios [5], [6]. Unlike orthogonal
multiple access (OMA) techniques that are realized in time,
frequency, or code domain, NOMA is realized in the power
domain [7]. In fact, NOMA performs superposition coding
(SC) in the power domain at the transmitter. Subsequently,
successive interference cancellation (SIC) is applied at the
receiver [8], [9].
In order to serve more users in 5G wireless communications, recently, the integration of NOMA in mmWave
systems, i.e., mmWave-NOMA, has been studied [10]–[14].
The work in [10] designs a random beamforming technique
for mmWave-NOMA systems. The base station (BS) randomly
radiates a directional beam toward paired users. In [11], it
is shown that mismatch between the users’ channel vector
and finite resolution analog beamforming1 simplifies utilizing
NOMA in mmWave-MIMO systems. The work in [12], proposes the combination of beamspace MIMO and NOMA, to
ensure more users can be served with a limited number of RF
chains. As a result, the number of served users is more than the
number of RF chains [12]. In [13], NOMA is studied for hybrid mmWave-MIMO systems. A power allocation algorithm
has been provided in order to maximize energy efficiency.
Newly, a joint power allocation and beamforming algorithm
for NOMA in the analog mmWave systems has been proposed
in [14]. Although the works [10]–[14] have all resulted in
maximizing bandwidth efficiency, this gain has come at the
costs of higher complexity at the receiver and the use of
multiple RF chains or one RF chain along with a large number
of phase shifters and power amplifiers which can be costly at
mmWave frequencies. Hence, in contrast to the prior works,
here, we will propose a new multiple access scheme that takes
advantage of reconfigurable antennas to outperform NOMA,
while at the same requiring one RF chain at the transmitter
and simple receiver structure.
Recently, there has been a new class of reconfigurable
antennas that can support the transmission of multiple radiation beams with one RF chain [15], [16]. The unique
1 Finite resolution analog beamforming is due to the use of a finite number
of phase shifters.
II. S YSTEM M ODEL AND NOMA
In this section, first we introduce the reconfigurable antenna
systems and their properties. Then, NOMA technique for a
BS and multiple users is described. Finally, NOMA for the
reconfigurable antennas is investigated.
A. Reconfigurable antenna systems
A reconfigurable antenna can support multiple reconfigurable orthogonal radiation beams. To accomplish this, a
spherical dielectric lens is fed with multiple tapered slot
antennas (TSAs), as shown in Fig. 1. The combination of each
TSA feed and the lens produces highly directive beams in far
field [15], [16]. That is, each TSA feed generates a beam in
a given direction in the far field. Therefore, a reconfigurable
antenna system is a multi-beam antenna capable of generating
M ≫ 1 independent beams where M is the number of TSA
feeds. Only the feed antennas that generate the beams in the
desired directions need to be excited. To this end, the output
of the RF chain is connected to a beam selection network (see
Fig. 1). The network has one input port that is connected to
the RF chain and M output ports that are connected to the M
TSA feeds [15]. For instance, in Fig. 1, the network selects
(a) RF Transceiver Chain
Spherical
Dielectric Lens
RF Transciever Chain
Mixer
PA
Duplexer
Mixer
LNA
Far-Field
Pencil Beams
Beam Selection
Network
property that distinguishes the system in [15], [16] from prior
art is that the proposed reconfigurable antenna architecture
can support multiple simultaneous orthogonal reconfigurable
beams via one RF chain. Inspired by this class of antennas, this
paper proposes a fresh multiple access technique for mmWave
reconfigurable antenna systems which is called reconfigurable
antenna multiple access (RAMA). We consider a scenario in
which a single BS is equipped with a mmWave reconfigurable
antenna and each beam of the antenna serves one user where
the users are not aligned with the same direction. Given
that the limitation on the RF circuitry of the antenna [15]
results in the division of the transmitted power amongst the
beams, the current state-of-the-art in mmWave-NOMA would
not operate efficiently in such a setting. To enhance the
performance of multiple access schemes in the mmWave band
and also overcome this fundamental limit for reconfigurable
antennas, unlike NOMA, RAMA aims to transmit only the
intended signal of each user. To accommodate this technique,
we will consider two cases, RAMA with partial channel state
information (CSI) and RAMA with full CSI. In the first case,
channel gain information is not required and only the direction
of arrival (DoA) for a specific user is used. Our results
indicate that with partial CSI and for symmetric channels,
RAMA outperforms NOMA in terms of sum rate. Further, the
analytical result indicates that RAMA for asymmetric channels
achieves better sum rate than NOMA when less power is
assigned to a user that experiences a better channel quality. In
the second case, RAMA with full CSI allocates optimal power
to each user which leads to higher achievable rates compared
to NOMA for both symmetric and asymmetric channels. Our
extensive numerical computations demonstrate the analytical
findings.
√
Notations: Hereafter, j = −1. Also, E[·] and | · | denote
the expected value and amplitude value of (·), respectively.
Tapered Slot
Antenna Feeds
Fig. 1. Schematic of the reconfigurable antenna steering multiple beams. The
antenna is composed of a spherical lens fed with a number of tapered slot
antenna feeds. Each feed generates a beam in a given direction in the far
field [15].
only five outputs that are connected to the input ports of five
TSA feeds. Accordingly, the network divides power of the
output signal of the RF chain equally or unequally amongst
five TSA feeds.
It is mentioned that the reconfigurable antennas steer reconfigurable independent beams. This steering is achieved
by selecting the appropriate TSA feed. Also, recall that we
assume that the transmitter has knowledge of the DoA of
the users. Accordingly, by appropriately steering the beams,
the reconfigurable antenna can manipulate the phase of the
received signal at the user terminal. Therefore, steering multiple reconfigurable independent beams and routing the power
amongst those beams are two properties of the reconfigurable
antennas.
B. Review of NOMA
In this paper, we consider the downlink of a single communication cell with a BS in the center that serves multiple users.
The BS and users are provided with a signal omnidirectional
antennas. For simplicity, the number of users is restricted to
two where the users are not aligned with the same direction,
i.e., there is an angle gap between the users.
Let the BS have signals si (i = 1, 2) for User i, where
E[|si |2 ] = 1, with power transmission pi . The sum of pi s, for
i = 1, 2, equals to p. According to the principle of the NOMA
downlink, at the transmitter, s1 and s2 are superposition coded
as
√
√
(1)
x = p 1 s1 + p 2 s2 .
Hence, the received signal at the ith user, for i = 1, 2, is
given by
√
√
yi = xhi + ni ≡ ( p1 s1 + p2 s2 )hi + ni ,
(2)
where hi is the complex channel gain between the BS and User
i, and ni denotes the additive white Gaussian noise with power
σi2 . At the receiver, each user performs the SIC process to
decode the desired signal. The optimal decoding order depends
on the channel gain. Without loss of generality, let us assume
that User 1 have better channel gain, i.e., |h1 |2 /σ12 ≥ |h2 |2 /σ22 ,
which gives p1 ≤ p2 .
After applying SIC, the achievable rate for NOMA for User
i can be determined as
RN = log (1 + p1 |h21 |2 ),
1
2
σ1
(3)
2
2
R2N = log2 (1 + p2 |h22| /σ2 2 ).
p1 |h2 | /σ +1
2
Allocated power to user 1
Power
Allocated power to user 2
Time/Frequency/Code
...,s2, s1
Beam Selection Network
This result indicates that power allocation greatly affects the
achievable rate for each user. For example, an improper power
allocation does now allow User 1 to decode s2 correctly, which
in turn does not allow for the interference from User 2 to be
successfully eliminated.
RF
Im
Re
C. NOMA for the reconfigurable antennas
8-PSK constellation
Phase
Detector
ej
User 1
0
s1
.5p
0.5p
s2
User 2
TSA feed
Suppose that a BS is equipped with the described recon2. Schematic of the BS for reconfigurable antenna multiple access
figurable antenna system and aims to simultaneously serve Fig.
technique with partial CSI and equal power division amongst the TSA feeds.
two users by using NOMA. The reconfigurable antenna steers It is assumed that the signals s1 and s2 are selected form 8-PSK constellation
two beams by feeding two TSAs. Each TSA serves one user
which is equipped with a single omnidirectional antenna. The
superposition coding of si with allocated power pi is defined A. RAMA with Partial CSI
in (1). Users 1 and 2 receive the following signals as
Let us assume that the BS utilizes a reconfigurable antenna
(
√ √
√
√
and
has partial CSI, i.e., knows the DoA of users. Thus, power
z1 = αxh1 + n1 ≡ α( p1 s1 + p2 s2 )h1 + n1 ,
√
√
√
√
is allocated equally for each user, i.e.,
z2 = 1 − αxh2 + n2 ≡ 1 − α( p1 s1 + p2 s2 )h2 + n2 ,
(4)
p1 = p2 = 0.5p.
(6)
respectively, where factor α ∈ (0, 1) is due to the power
division in the reconfigurable antennas. When α = 0.5,
it means the power is divided equally between two TSAs.
Thus, each user receives only a portion of the power of the
transmitted signal x.
Consider the same order for the channel gains, i.e,
|h1 |2 /σ12 ≥ |h2 |2 σ22 . An error-free SIC process results in the
achievable rate for each user as
RNR = log (1 + αp1 |h2 1 |2 ),
1
2
σ1
(5)
2
2
R2NR = log2 (1 + (1−α)p2 |h22| /σ2 2 ).
(1−α)p1 |h2 | /σ2 +1
Obviously, signal to noise ratio (SNR) for User 1 and signal
to interference plus noise ratio (SINR) for User 2 in (5)
is less than that of the NOMA given in (3). As a result,
combining NOMA with reconfigurable antennas will reduce
the achievable user rate when considering the same channel
gains, power allocation, and noise power. It is noteworthy
that for reconfigurable antenna-NOMA the definition of power
allocation is different from power division. Power allocation
is a strategy which is used in NOMA. While, power division
is one of the properties of the reconfigurable antennas that
divides power of the superposition coded signal among two
TSAs.
In brief, when users are not aligned in the same direction,
the reconfigurable antennas cannot harvest the benefits of
NOMA due to the power division property. Whereas, when
users are located on the same direction, the reconfigurable antenna steers only one beam to serve users which means power
division is not required. In this case, reconfigurable antennaNOMA and NOMA in Subsection II.B achieve identical sum
rate performance.
III. R ECONFIGURABLE A NTENNA M ULTIPLE ACCESS
In this section, a novel multiple access technique for the
reconfigurable antenna systems is proposed. The technique
which we call RAMA takes advantage of the reconfigurable
antenna and directional transmission in mmWave bands.
Our main objective is to suppress inter-user interference. To
this end, we aim to transmit only the intended signal for each
user at the same time/frequency/code blocks. The intended
signals for Users 1 and 2 are s1 and s2 , respectively. Let us
assume that si , for i = 1, 2, is drawn from a phase shift
keying (PSK) constellation and E[|si |2 ] = 1. Accordingly, s2
can be expressed in terms of s1 as
s2 = s1 ej∆θ ,
(7)
where ∆θ denotes the difference between the phases of s1 and
s2 . Similar to (1), the power of the transmitted signal, x, is
assumed to be p.
Unlike NOMA, in RAMA only one of the signals, say s1 ,
is upconverted by the RF chain block and the whole power
p is allocated to that signal before the power division step.
Therefore, x is given by
√
x = ps1 .
(8)
It is clear that the superposition coded signal in (1) and the
signal in (8) carry the same average power p. The proposed
multiple access technique for the signal in (8) is shown in
Fig. 2. The phase detector block calculates the phase difference
between s1 and s2 , ej∆θ . Moreover, as shown in Fig. 2, the
beam selection network selects two TSA feeds as highlighted
with black color based on the users’ DoA. For simplicity, we
call them TSA 1 and TSA 2. The network divides the power
equally between TSA√1 and TSA 2. That is, the signal of TSAs
1 and 2 is given by 0.5ps1 .
The signal intended for TSA 1 is the desired signal for
User 1. To transmit the signal intended for User 2 via TSA
2 we take advantage of the reconfigurable antennas. Thanks
to properties of reconfigurable antennas, signal at each TSA
can independently be rotated with an arbitrary angle. Using
this proper, the beam selection network multiplies the signal
j∆θ
in corresponding to TSA 2 with e√
. This results√ in the
transmitted signal via TSA 2 to be 0.5ps1 ej∆θ = 0.5ps2
which is the desired signal for User 2. Since transmission
is directional in mmWave bands, each user receives only its
intended signal as
(
√
z1 = 0.5ps1 h1 + n1 ,
(9)
√
z2 = 0.5ps2 h2 + n2 .
It is noteworthy that, here, we assume that there is no
interference that is imposed from the signal intended for User
1 on User 2 and vice versa. This assumption is very well
justified since the structure of the proposed lens based slotted
reconfigurable antenna results in very directional beams with
very limited sidelobes. Moreover, due to significant pathloss
and shadowing at mmWave frequencies we do not expect the
signals from the sidelobes to reach the unintended user2 . We
also highlight that in contrast to NOMA, full CSI and the
SIC process are not required at the receiver. Furthermore, in
RAMA power allocation and power division carry the same
concept, such that the routed power for TSA 1 (or 2) is the
same as the allocated power for User 1 (or 2).
The achievable rate of RAMA for each user under equal
power allocation is obtained as
RR,I = log (1 + p|h12|2 ),
1
2
2σ1
(10)
2
R2R,I = log2 (1 + p|h22| ).
2σ2
Let us denote RAMA with partial CSI by RAMA-I. It is
valuable to compare the sum rate of NOMA and RAMA-I. To
this end, we consider two extreme cases as follows. By definiN
tion, sum rate for NOMA and RAMA-I are Rsum
= R1N + R2N
R,I
R,I
R,I
and Rsum = R1 + R2 , respectively, where R1N and R2N
are defined in (3) and R1R,I and R2R,I are defined in (10).
Case I: p|h1 |2 /σ12 = p|h2 |2 /σ22 . In his paper is case is
called symmetric channel [5]. That is, the two users have the
same SNR. In this case, RAMA-I always achieves higher sum
rate than NOMA.
To show this, we calculate the sum rate for NOMA and
RAMA. For NOMA, the sum rate can be calculated as
p2 |h2 |2 /σ22
p1 |h1 |2
N
+ log2 1 +
Rsum
= log2 1 +
2
σ1
p1 |h2 |2 /σ22 + 1
p1 |h1 |2
p2 |h2 |2 /σ22
(a)
= log2 1 +
1
+
σ12
p1 |h2 |2 /σ22 + 1
2
p2 |h2 |2
p1 |h1 |
+
= log2 1 +
σ12
σ22
2
p|h|
(b)
= log2 1 + 2 .
(11)
σ
The (a) follows from log2 a + log2 b = log2 ab and the (b)
follows the assumption that |h|2 /σ 2 = |h1 |2 /σ12 = |h2 |2 /σ22
and p1 + p2 = p.
Also, for RAMA-I, we follow the same steps as in NOMA.
Hence, it is obtained as
R,I
Rsum
= log2 1 +
p2 |h|4
p|h|2
+
.
2
σ
4σ 4
(12)
R,I
N
Since p2 |h|4 /4σ 4 > 0, it gives Rsum
≤ Rsum
.
2 Detail analysis of the impact of sidelobes on inter-user interference in
RAMA is subject of future research.
Case II: p|h1 |2 /σ12 ≥ p|h2 |2 /σ22 . Here, this case is called
asymmetric channel [5]. That is, we assume that the channel
gain for User 1 is stronger than User 2. It can be shown that
for asymmetric channels, RAMA-I achieves higher sum rate
than NOMA when the power isproperly allocated for Users 1
and 23 . To proof this claim, we have
p2 |h2 |2 /σ22
p1 |h1 |2
) + log2 (1 +
)
2
σ1
p1 |h2 |2 /σ22 + 1
p1 |h1 |2
p2 |h2 |2 /σ22
(a)
= log2 (1 +
)(1 +
)
2
σ1
p1 |h2 |2 /σ22 + 1
(b)
p2 |h2 |2
p1 |h1 |2
)(1 +
) ,
(13)
≤ log2 (1 +
2
σ1
σ22
N
Rsum
= log2 (1 +
where p1 and p2 are allocated power for Users 1 and 2, respectively. Also, the (a) follows from log2 a + log2 b = log2 ab,
and the (b) is a result of p1 |h2 |2 /σ22 > 0 . Using (13) and
R,I
N
(10), the inequality Rsum
≤ Rsum
holds when the following
condition follows
(1 +
p1 |h1 |2
p2 |h2 |2
p|h1 |2
p|h2 |2
)(1
+
)
≤
(1
+
)(1
+
).
σ12
σ22
2σ12
2σ22
(14)
Obviously, for p1 /p ∈ (0, 0.5] the inequality holds which
indicates that User 1 should have lower power than User
2. Although this range is not tight, it gives a considerable
insight. This result implies that with proper power allocation in
NOMA, RAMA-I attains higher sum rate. However, RAMA-I
may not achieve user fairness when channel gain of one of
the users is significantly greater than that of other user. In this
case, the allocated power for user with strong channel gain
should be far less than other user and equal power allocation
would not lead to user fairness.
B. RAMA with full CSI
Assume that full CSI is available at the BS. Furthermore, the
BS can unequally allocate the power between two users. For
signals s1 and s2 that are chosen from a QAM constellation
and E[|si |2 ] = 1, the relationship between two arbitrary signals
selected from the constellation is given by
s2 = s1 s̄ej∆θ
(15)
where s̄ denotes |s2 |/|s1 |. For RAMA with full CSI, the
transmitted signal is defined as
p
p
(16)
x = p′ s1 ≡ ( p1 + p2 s̄2 )s1 ,
which obviously has the same average power as the signal in
(1). It is assumed that p1 and p2 are the allocated power for
User 1 and User 2, respectively. For simplicity, we consider
that our power allocation strategy is exactly the same as
NOMA. Fig. 3 depicts the schematic of the reconfigurable
antenna for the RAMA. The phase detector and s̄ calculator
blocks calculate ej∆θ and s̄, respectively. The beam selection
network first selects two suitable TSA feeds, TSA 1 and TSA
2 which are highlighted with black color in Fig. 3, based on
3 To achieve user fairness in NOMA, when |h |2 /σ 2 ≥ |h |2 /σ 2 we have
1
2
2
1
p1 ≤ p2 [5].
•
Power
Allocated power to User 1
Allocated power to User 2
...,s2, s1
Beam Selection Network
Time/Frequency/Code
RF
Im
Re
16-QAM constellation
Calculator
Phase
Detector
User 1
s1
√p 1
√p2 s
2
User 2
TSA feed
ejΔθ
RAMA technique is not an alternative for NOMA. When
users are aligned in the same direction, RAMA would
be combined with one of OMA or NOMA. Accordingly,
the users located on the same direction are considered
as a cluster. Each cluster will be served via RAMAOMA or RAMA-NOMA. Integration of RAMA with
other multiple access techniques is beyond the scope of
this paper.
Fig. 3. Schematic of the BS for reconfigurable antenna multiple access
technique regarding full CSI and unequal power allocation. It is assumed
that signals s1 and s2 are chosen from 16-QAM constellation.
σ2
respectively.
It is straightforward to show that RAMA with full CSI,
denoted by RAMA-II, achieves higher sum rate than NOMA
R,II
N
irrespective of their channel condition. That is, Rsum
≤ Rsum
R,II
R,II
R,II
where Rsum = R1 +R2 . Further, RAMA-II considers user
fairness the same as NOMA.
The following remarks are in order:
•
•
•
RAMA-I significantly reduces system overhead. In highly
dense networks where the number of users is much larger
than two, it is reasonable to serve users with RAMAI. This is because RAMA-I does not require the BS to
know full CSI and only users’ direction information is
necessary.
RAMA-I implemented by PSK constellation is very
simple in practice. The beam selection network divides
the power equally by using a simple power divider and
the received signal is interference-free. The later is also
preserved for RAMA-II. Indeed, RAMA-I with QAM
constellations is operational if power is divided properly.
RAMA-II needs an optimal power allocation strategy. In
Subsection III.B it is pointed out that power allocation
strategy for NOMA is adopted for RAMA-II. However, this strategy may not be efficient in reconfigurable
antenna systems with full CSI. It is because NOMA
considers the interference and the minimum achievable
rate for each user to allocate the power [5]. Whereas,
for RAMA-II the interference is removed which leads to
designing a better power allocation strategy.
20
20
NOMA
RAMA
10
5
0
-10
NOMA, p 1 /p = 1/4
NOMA, p 1 /p = 1/2
15
Sum rate (bits/s/Hz)
respectively. Accordingly, the achievable rate for user 1 and
user 2 is obtained as
RR,II = log (1 + p1 |h21 |2 ),
1
2
σ1
(18)
2
R2R,II = log2 (1 + p2 |h22 | ),
This section evaluates the performance of the proposed multiple access technique by using numerical computations, where
the analytical findings will also be verified. Transmission in
mmWave bands can be done through both line-of-sight and
non line-of-sight paths. Here, for the sake of simplicity, we
will consider Rayleigh fading channels.
Figure 4.(a) represents sum rate versus symmetric channel
plot for NOMA and RAMA-I. It is clear that RAMA-I
achieves better sum rate than NOMA. This is because the user
achievable rate for NOMA is limited due to the interference
from other users, which degrades the sum rate. At symmetric
channels, the interference in NOMA is severe and as a result
leads to a considerable sum rate gap compared to the RAMA
technique which is an inter-user interference-free technique.
This result verifies our claim in Case I Subsection III.A.
Figure 4.(b) illustrates sum rate performance versus asymmetric channel gains for RAMA-I and various power allocation schemes for NOMA. The aim of this simulation is to support our analytical finding in Case II Subsection III.A. When (p|h1 |2 /σ12 )/(p|h2 |2 /σ22 ) is not large
enough, RAMA-I outperforms NOMA for all power allocation
schemes since the channel is similar to a symmetric channel.
By increasing (p|h1 |2 /σ12 )/(p|h2 |2 /σ22 ) channel satisfies the
condition in Case II Subsection III.A. At high region of
(p|h1 |2 /σ12 )/(p|h2 |2 /σ22 ), for p1 /p ≤ 1/2, e.g., p1 /p = 1/2
and p1 /p = 1/4, sum rate of RAMA-I is always better than
NOMA which is consistent with Case II. For p1 /p = 3/4 and
large channel gain difference, NOMA has a little better sum
rate. This is because much more power is allocated to User 1
which can nearly achieve maximum sum rate. However, in this
condition NOMA does not consider user fairness. In contrast,
Sum rate (bits/s/Hz)
the users’ DoA information. Then, it divides the signal in (16)
√
√
into p1 s1 and p2 s1 s̄ regarding the obtained CSI and the
√
power allocation strategy. The signal for User 1, p1 s1 , is
ready to transmit from TSA 1. For User 2, the intended signal
√
√
is built by multiplying p2 s1 s̄ by ej∆θ which yields p2 s2 .
Hence, the interference-free received signals for Users 1 and
2 are attained as
(
√
z 1 = p 1 s1 + n1 ,
(17)
√
z 2 = p 2 s2 + n2 ,
IV. N UMERICAL C OMPUTATIONS
NOMA, p 1 /p = 3/4
15
RAMA-I, p 1 /p = 1/2
10
5
0
0
10
2
20
2
30
0
20
40
|h 1 | = |h 2 | (dB)
(p|h 1 | 2 / 21 )/(p|h 2 | 2 / 22 ) (dB)
(a)
(b)
60
Fig. 4. Sum rate comparison between NOMA and RAMA-I for (a) symmetric
channel and (b) symmetric channel.
4
3
2
OMA
NOMA
RAMA-II
1
0
0
2
4
Achievable rate for User 1 (bits/s/Hz)
Achievable rate for User 2 (bits/s/Hz)
Achievable rate for User 2 (bits/s/Hz)
5
1
0.8
0.6
0.4
0.2
OMA
NOMA
RAMA-II
0
0
5
10
Achievable rate for User 1 (bits/s/Hz)
(a)
(b)
Fig. 5. Achievable rate region of user 1 and 2 for (a) symmetric channel
with p|hi |2 /σi2 = 15 dB for i = 1, 2 and (b) asymmetric channel with
p|h1 |2 /σ12 = 30 dB and p|h2 |2 /σ22 = 0 dB.
equal power is allocated for the users in RAMA-I and they
cannot exploit maximum sum rate.
Figure 5 shows achievable rate region of two users for
OMA, NOMA, and RAMA-II. OMA in downlink transmission
is assumed to be implemented by OFDMA technique where
R1O = βlog2 (1 + p1 |h1 |2 /βσ12 ) and R2O = (1 − β)log2 (1 +
p2 |h2 |2 /(1 − β)σ22 ) are achievable rate for Users 1 and 2 with
the bandwidth of β Hz assigned to User 1 and (1 − β) Hz
assigned to User 2 [5].
In Fig. 5.(a), channel is assumed to be symmetric where its
gain is set to p|hi |2 /σi2 = 15 dB for i = 1, 2. The achievable
rate region for OMA and NOMA are identical. The region
for RAMA-II is much wider than that for OMA and NOMA
because RAMA-II neither suffers from inter-user interference
nor divides the bandwidth among users. For instance, when
User 2 achieves rate 2.5 bits/s/Hz, achievable rate of User 1
for RAMA-II with channel gain information is approximately
twice higher than that for OMA and NOMA. Notice that when
total power is allocated to either user, all three techniques are
able to achieve maximum rate for that user.
The achievable sum rate of asymmetric channel for
p|h1 |2 /σ12 = 30 dB and p|h2 |2 /σ22 = 0 dB has been
represented in Fig. 5.(b). From the figure it is clear that for
OMA, NOMA and RAMA-II only User 1 achieves maximum
sum rate when whole power is allocated to that user. However,
NOMA achieves wider rate region than OMA as expected.
Interestingly, the achievable rate region for RAMA-II is greater
than NOMA. As an example, when we want that User 1
to achieve 8 bits/s/Hz, User 1 can reach 0.68 bits/s/Hz and
0.8 bits/s/Hz with NOMA and RAMA-II, respectively. This is
because the achievable rate of User 2 in NOMA is affected by
the interference term from User 1 due to using superposition
coding at transmitter and SIC process at receiver. Whereas, in
RAMA-II, User 1 does not impart interference on User 2. In
other words, the gap between NOMA and RAMA-II reflects
the impact of inter-user interference on NOMA.
V. C ONCLUSION
In this paper, we propose a new multiple access technique
for mmWave reconfigurable antennas in order to simultaneously support two users by using a single BS in downlink.
First, we show that NOMA is not suitable technique for
serving the users for the reconfigurable antenna systems. Then,
by wisely using the properties of reconfigurable antennas, a
novel multiple access technique called RAMA is designed
by assuming partial CSI and full CSI. The proposed RAMA
provides mmWave reconfigurable antenna system with an
inter-user interference-free user serving. That is, the users with
higher allocated power do not required to decode signals of
other users. It is shown that for symmetric channels RAMAI outperforms NOMA for an arbitrary p1 in terms of sum
rate. Also, for asymmetric channels RAMA-I demonstrates
better sum rate performance if approximately more than half
of the power is allocated to User 2. Further, RAMA-II always
achieves higher sum rate than NOMA. Numerical computations support our analytical investigations.
R EFERENCES
[1] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N.
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave
mobile communications for 5G cellular: It will work!” IEEE Access,
vol. 1, pp. 335–349, 2013.
[2] J. Kim and I. Lee, “802.11 WLAN: history and new enabling MIMO
techniques for next generation standards,” IEEE Commun. Mag., vol. 53,
no. 3, pp. 134–140, 2015.
[3] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
“Spatially sparse precoding in millimeter wave mimo systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, 2014.
[4] J. Brady, N. Behdad, and A. M. Sayeed, “Beamspace MIMO for
millimeter-wave communications: System architecture, modeling, analysis, and measurements,” IEEE Trans. Antennas Propagat., vol. 61, no. 7,
pp. 3814–3827, 2013.
[5] K. Higuchi and A. Benjebbour, “Non-orthogonal multiple access
(NOMA) with successive interference cancellation for future radio
access,” IEICE Trans. Commun., vol. 98, no. 3, pp. 403–414, 2015.
[6] L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-Lin, and Z. Wang, “Nonorthogonal multiple access for 5G: solutions, challenges, opportunities,
and future research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp.
74–81, 2015.
[7] D. Tse and P. Viswanath, Fundamentals of wireless communication.
Cambridge university press, 2005.
[8] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “Systemlevel performance evaluation of downlink non-orthogonal multiple access (NOMA),” in Proc. IEEE Int. Symp. Pers., Indoor Mobile Radio
Commun. (PIMRC). IEEE, Sep. 2013, pp. 611–615.
[9] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and
K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular
future radio access,” in Proc. IEEE VTC Spring. IEEE, 2013, pp.
1–5.
[10] Z. Ding, P. Fan, and H. V. Poor, “Random beamforming in millimeterwave NOMA networks,” IEEE Access, 2017.
[11] Z. Ding, L. Dai, R. Schober, and H. V. Poor, “NOMA meets finite
resolution analog beamforming in massive MIMO and millimeter-wave
networks,” IEEE Commun. Lett., 2017.
[12] B. Wang, L. Dai, Z. Wang, N. Ge, and S. Zhou, “Spectrum and energyefficient beamspace MIMO-NOMA for millimeter-wave communications using lens antenna array,” IEEE J. Select. Areas Commun., vol. 35,
no. 10, pp. 2370–2382, 2017.
[13] W. Hao, M. Zeng, Z. Chu, and S. Yang, “Energy-efficient power allocation in millimeter wave massive MIMO with non-orthogonal multiple
access,” IEEE Wireless Commun. Lett., vol. 6, no. 6, pp. 782–785, 2017.
[14] Z. Xiao, L. Zhu, J. Choi, P. Xia, and X.-G. Xia, “Joint power allocation
and beamforming for non-orthogonal multiple access (NOMA) in 5G
millimeter-wave communications,” arXiv preprint arXiv:1711.01380,
2017.
[15] M. A. Almasi, H. Mehrpouyan, V. Vakilian, N. Behdad, and H. Jafarkhani, “Reconfigurable antennas in mmWave MIMO systems,” arXiv
preprint arXiv:1710.05111, 2017.
[16] B. Schoenlinner, X. Wu, J. P. Ebling, G. V. Eleftheriades, and G. M.
Rebeiz, “Wide-scan spherical-lens antennas for automotive radars,”
IEEE Trans. Microwave Theory Tech., vol. 50, no. 9, pp. 2166–2175,
2002.
| 7cs.IT
|
A method for constructing parity-check matrices
of non-binary quasi-cyclic LDPC codes
arXiv:1705.10717v1 [cs.IT] 30 May 2017
Stanislav Kruglik, Valeriya Potapova and Alexey Frolov
Skolkovo Institute of Science and Technology
Moscow, Russia
Institute for Information Transmission Problems
Russian Academy of Sciences
Moscow, Russia
[email protected], [email protected],
[email protected]
Abstract. An algorithm for constructing parity-check matrices of nonbinary quasi-cyclic low-density parity-check (NB QC-LDPC) codes is
proposed. The algorithm finds short cycles in the base matrix and tries
to eliminate them by selecting the circulants and the elements of GF(q).
Firstly the algorithm tries to eliminate the cycles with the smallest number edges going outside the cycle. The efficiency of the algorithm is
demonstrated by means of simulations. In particular, it was shown that
NB QC-LDPC codes constructed with use of our algorithm loose less
that 0.1 dB in comparison to the best NB LDPC codes.
Keywords: LDPC codes, non-binary, quasi-cyclic, cycle, ACE value
1
Introduction
In this paper we consider the problem of constructing parity-check matrices of NB QC-LDPC codes.
QC-LDPC codes were proposed in [1],[2]. These codes form an important subclass of LDPC codes [3], [4]. These codes also are a subclass of
protograph-based LDPC codes [5]. QC-LDPC codes can be easily stored
as their parity-check matrices can be easily described. Besides such codes
have efficient encoding and decoding algorithms [7]. All of these makes
the codes very popular in practical applications.
NB LDPC codes have advantages over binary LDPC codes. Davey
and MacKay [8] were first who used belief propagation (BP) to decode
NB LDPC codes. They showed that NB LDPC codes significantly outperform their binary counterparts. Moreover, non-binary LDPC codes are
especially good for the channels with burst errors and high-order modulations [9]. However we have to mention, that their decoding complexity
is still large in comparison to binary LDPC codes [10], [11], [12].
There are numerous methods for constructing parity-check matrices of
binary QC-LDPC codes [13], [14], [15]. In this paper we generalize them
to non-binary case.
Our contribution is as follows. An algorithm for constructing paritycheck matrices of non-binary quasi-cyclic low-density parity-check (NB
QC-LDPC) codes is proposed. The algorithm finds short cycles in the
base matrix and tries to eliminate them by selecting the circulants and the
elements of GF(q). Firstly the algorithm tries to eliminate the cycles with
the smallest number edges going outside the cycle. The efficiency of the
algorithm is demonstrated by means of simulations. In particular, it was
shown that NB QC-LDPC codes constructed with use of our algorithm
loose less that 0.1 dB in comparison to the best NB LDPC codes.
2
Preliminary results
Consider a binary matrix of size m × n
Hbase = [hi,j ] ∈ {0, 1}m×n .
In what follows the matrix will be referred to as the base matrix.
Let us construct a parity-check matrix H of size ms × ns of NB QCLDPC code C. For this purpose we extend the matrix Hbase with circulant
matrices (circulant) multiplied by non-zero elements of GF(q), i.e.
H=
α1,1 P1,1
α2,1 P2,1
..
.
α1,2 P1,2
α2,2 P2,2
..
.
···
···
..
.
α1,n P1,n
α2,n P2,n
..
.
αm,1 Pm,1 αm,2 Pm,2 · · · αm,n Pm,n
,
(1)
where Pi,j is a circulant over a binary field of size s × s and of weight1
hi,j , i = 1, . . . , m, j = 1, . . . , n, and αi,j ∈ GF (q)\{0}, i = 1, . . . , m,
j = 1, . . . , n.
Let us denote the length of the code C by N = ns, such inequality
follows for the rate of the code
R(C) ≥ 1 −
m
.
n
Let F be some field, by F [x] we denote the ring of all the polynomials
with coefficients in F . It is well-known that the ring of circulants of size
1
the weight of a circulant is a weight of its first row.
s × s over F is isomorphic to the factor ring F (s) [x] = F [x]/(xs − 1). Thus
with the parity-check matrix H we associate a polynomial parity-check
matrix H(x) of size m × n:
H(x) =
α1,1 p1,1 (x)
α2,1 p2,1 (x)
..
.
α1,2 p1,2 (x)
α2,2 p2,2 (x)
..
.
···
···
..
.
α1,n p1,n (x)
α2,n p2,n (x)
..
.
αm,1 pm,1 (x) αm,2 pm,2 (x) · · · αm,n pm,n (x)
Ps
,
(2)
where pi,j (x) = t=1 Pi,j (t, 1)xt−1 , by Pi,j (t, 1) we mean an element at
the intersection of the t-th row and the first column in the matrix Pi,j .
Example 1. Let us consider the following base matrix
Hbase =
"
0 1 1
1 0 1
#
We can extend it in such a way. Each one in the base matrix is replaced
with a polynomial of for βxz , where β ∈ GF (q)\{0} and z ∈ Zs , e.g. we
have
"
#
0 x2 2x
H(x) =
.
1 0 3x2
Note, that H(x) means the following parity-check matrix
H=
0 0 0
0 0 0
0 0 0
0 1 0
0 0 1
1 0 0
1 0 0
0 1 0
0 0 1
0 0 0
0 0 0
0 0 0
0 0 2
2 0 0
0 2 0
0 3 0
0 0 3
3 0 0
Remark 1. We note, that all elements of a circulant are multiplied by the
only non-binary value. This is very important for a practical realization of
the codes, as the parity-check matrix is very compact and can be stored
very efficiently. The proposed construction of the parity check matrix is
also good for parallel implementation of the decoding algorithm.
3
Description of the algorithm for constructing
parity-check matrices of NB QC-LDPC codes
In this section we provide a description of algorithm for constructing
parity-check matrices of NB QC-LDPC codes. The algorithm can be divided into two stages:
1. Use NB-EXIT analysis to select a base matrix (protograph)
2. Lift the graph using NB-ACE algorithm (proposed below).
NB-EXIT analysis can be done in a similar way as for binary LDPC
codes (see detailed description in [16], [17]). In what follows we propose
the lifting method.
It is well-known, that pseudo codewords are the reason of belief propagation algorithm bad behavior. Pseudo codewords always contain cycles,
so the aim of the algorithm is to eliminate short cycles. As an example
let us consider the elimination of a cycle of length 4 (see Fig. 1).
cycle
H_base =
1
1
1
1
Replace each
non-zero
position
with a pair
(offset, nb value)
(z1 , α1 )
H=
(z4 , α4 )
(z2 , α2 )
1
1
1
1
(z3 , α3 )
Fig. 1. Cycle elimination
Let I be subset of rows, containing a cycle, J be subset of columns,
containing a cycle. We say that the cycle is eliminated if
det(H(x)I,J ) 6= 0.
So the algorithm tries to eliminate as much cycles as possible. If we
cannot eliminate a cycle we count the number of edges going from variable
nodes in a cycle outside the cycle (so called ACE value of the cycle [13]).
As there are many cycles and they are of different lengths the algorithm
really deals with the ACE vector
ace = (e4 , e6 , . . .),
where e2j is a minimal number of edges going outside a cycle of length 2j
in a parity-check matrix. If there are no cycles of length 2j, then e2j = ∞.
We propose a greedy algorithm, which tries to find an optimal ACE
vector on each step. We compare ACE vectors lexicographically, this
means, that
(1, 2, 3) < (2, 2, 3), (1, 2, 3) < (1, 3, 3), . . .
All these stages are described in the algorithm below.
Algorithm 1 Algorithm to construct NB QC-LDPC codes
Input: Hbase , Q, s (size of circulant), depth (maximal cycle length)
for i ∈ {1, 2, . . . , n} do
for j ∈ ones in row i do
find all cycles in Hbase going through variable node j with length ≤ depth
for test ∈ {1, 2, . . . , 100} do
H(i, j) ← (randi(s − 1), randi(Q − 1))
calculate ace vector
if acemax < ace then acemax ← ace
else revert previous value of H(i, j)
end if
end for
end for
end for
Output: H
4
Simulation results
In this section we present the simulation results. The simulation was
carried out in AWGN channel with 64-QAM modulation. Let q = 64,
N = 4620 and K = 4060. At first we add the simulation results for
best NB LDPC codes over GF(64), which do not have quasi-cyclic form,
and binary Turbo codes. Let us construct two parity-check matrices of
NB QC-LDPC codes. In the first case the base matrix has size 4 × 33
(s = 140), in the second case the base matrix has size 8 × 66 (s = 70).
Obtained results are shown in Fig. 2.
Note, that a curve for base matrix with size 4 × 33 has a clear error
floor. Let us investigate these phenomenon. The reason is the bad minimal
code distance. According to [18] the minimal distance of QC-LDPC code
has the following upper bound:
D(C) ≤ ⌊ℓ⌋!ℓm−⌊ℓ⌋ (m + 1),
where ℓ is the number of ones in a column (in our case ℓ = 2) and m is
the height of base matrix.
For the matrix with size 4 × 33 this bound is as follows D(C) ≤ 40. At
the same time for the matrix with size 8 × 66 the bound is much better:
D(C) ≤ 1152. The error correcting capabilities of the latter matrix are
good, it looses just 0.1 dB at the level where block error rate (BLER) is
equal to = 10−3 .
0
10
−1
10
BLER
−2
10
−3
10
LTE TURBO
(4620, 4060) LDPC GF(64)
8x66 base matrix, factor 70 QC−LDPC GF(64)
4x33 base matrix, factor 140 QC−LDPC GF(64)
−4
10
−5
10
10
10.5
11
11.5
12
Eb/N0
Fig. 2. Simulation results. Parameters: channel with AWGN, 64-QAM modulation,
q = 64, code length in field symbols N = 4620, code dimension in field symbols
K = 4060
5
Conclusion
Greedy algorithm for constructing parity-check matrices of NB QC-LDPC
codes was proposed. The algorithm finds short cycles in the base matrix
and tries to eliminate them by selecting the circulants and the elements of
GF(q). Firstly the algorithm tries to eliminate the cycles with the smallest
number edges going outside the cycle. The efficiency of the algorithm is
demonstrated by means of simulations. In particular, it was shown that
NB QC-LDPC codes constructed with use of our algorithm loose less that
0.1 dB in comparison to the best NB LDPC codes.
Acknowledgments
The research of Stanislav Kruglik was supported by the Russian Foundation for Basic Research (project 16-01-00716).
References
1. R. M. Tanner. On quasi-cyclic repeat-accumulate codes. in Proc. 37th Allerton
Conf. Commun., Contr., Comput., Monticello, IL, Sep. 22–24, 1999, pp. 249–259,
Allerton House.
2. M. P. C. Fossorier. Quasi-cyclic low-density parity-check codes from circulant permutation matrices. IEEE Trans. Inf. Theory, vol. 50, no. 8, pp. 1788–1793, Aug.
2004.
3. R. G. Gallager. Low-Density Parity-Check Codes. Cambridge, MA: M.I.T. Press,
1963.
4. R. M. Tanner. A recursive approach to low-complexity codes. IEEE Trans. Inf.
Theory, vol. 27, no. 5, pp. 533–547, Sep. 1981.
5. J. Thorpe. Low-density parity-check (LDPC) codes constructed from protographs.
JPL, IPN Progress Rep., Aug. 2003, vol. 42–154.
6. Z. Li, L. Chen, L. Zeng, S. Lin, and W. H. Fong. Efficient encoding of quasi-cyclic
low-density parity-check codes. IEEE Trans. Commun., vol. 54, no. 1, pp. 71–78,
Jan. 2006.
7. F. R. Kschischang, B. J. Frey, and H.-A. Loeliger. Factor graphs and the sumproduct algorithm. IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 498–519, Feb. 2001.
8. M. Davey and D. MacKay, Low-density parity check codes over GF(q), IEEE Commun. Lett., vol. 2, no. 6, pp. 165–167, Jun. 1998
9. H. Song, J. R. Cruz, Reduced-Complexity Decoding of Q-ary LDPC Codes for
Magnetic Recording, IEEE Trans. on Magnetics, vol. 39, no. 2, March 2003
10. H. Wymeersch, H.Steendam, M.Moeneclaey, Log-domain decoding of LDPC codes
over GF(q), IEEE Int. Conf. on Communications, 2004, pp 772–776
11. L. Barnault and D. Declercq, Fast Decoding Algorithm for LDPC over GF(2q ),
The Proc. 2003 Inform. Theory Workshop, Paris, France, pp. 70–73, Mar. 2003
12. D.Declercq, M. Fossorier, Decoding Algorithms for Nonbinary LDPC Codes over
GF(q), IEEE Trans. On Communications, vol.55, no.4, April 2007, pp 633–643
13. D. Vukobratovic and V. Senk, Generalized ACE Constrained Progressive EdgeGrowth LDPC Code Design, in IEEE Communications Letters, vol. 12, no. 1, pp.
32-34, January 2008.
14. H.Xiao, A. H. Banihashemi, Improved Progressive-Edge-Growth (PEG) Construction of Irregular LDPC Codes, IEEE Comm. Letters, vol.8, no. 12, December 2004,
pp.715–717
15. X. Hu, E. Eleftheriou, D. Arnold, Irregular Progressive Edge-Growth (PEG) Tanner Graphs, ISIT 2002, Lausanne, Switzerland, June 30–July 5, 2002
16. G. Liva and M. Chiani. Protograph LDPC Codes Design Based on EXIT Analysis.
IEEE GLOBECOM 2007 - IEEE Global Telecommunications Conference, Washington, DC, 2007, pp. 3250–3254.
17. T. Y. Chen, K. Vakilinia, D. Divsalar and R. D. Wesel. Protograph-Based RaptorLike LDPC Codes. IEEE Trans. on Comm., vol. 63, no. 5, pp. 1522–1532, May
2015.
18. A. Frolov, Upper Bounds on the Minimum Distance of Quasi-Cyclic LDPC codes
// XIV International Workshop on Algebraic and Combinatorial Coding Theory
(ACCT 2014) September 7-13, 2014, Kaliningrad, Russia. P. 163–168.
| 7cs.IT
|
Parameterized Complexity of Diameter
Matthias Bentert and André Nichterlein
arXiv:1802.10048v1 [cs.DS] 27 Feb 2018
Institut für Softwaretechnik und Theoretische Informatik, TU Berlin, Germany,
{matthias.bentert,andre.nichterlein}@tu-berlin.de
February 28, 2018
Abstract
Diameter—the task of computing the length of a longest shortest path—is a fundamental
graph problem. Assuming the Strong Exponential Time Hypothesis, there is no O(n1.99 )time algorithm even in sparse graphs [Roditty and Williams, 2013]. To circumvent this lower
bound we aim for algorithms with running time f (k)(n + m) where k is a parameter and f is
a function as small as possible. For our choices of k we systematically explore a hierarchy of
structural graph parameters.
1
Introduction
The diameter is arguably among the most fundamental graph parameters. Most known algorithms
for determining the diameter first compute the shortest path between each pair of vertices (APSP:
All-Pairs Shortest Paths) and then return the maximum [1]. The
currently fastest algo√
rithms for APSP in weighted graphs have a running time of O(n3 /2Ω( log n) ) in dense graphs [8]
and O(nm + n2 log n) in sparse graphs [25], respectively. In this work, we focus on the unweighted
case. Formally, we study the following problem:
Diameter
Input: An undirected, connected, unweighted graph G = (V, E).
Task: Compute the length of a longest shortest path in G.
The (theoretically) fastest algorithms for Diameter run in O(n2.373 ) time and are based
on fast matrix multiplication [33]. This upper bound can (presumably) not be improved by
much as Roditty and Williams [32] showed that solving Diameter in O((n + m)2−ε ) time for
any ε > 0 breaks the SETH (Strong Exponential Time Hypothesis [23, 24]). Seeking for ways to
circumvent this lower bound, we follow the line of “parameterization for polynomial-time solvable
problems” [19] (also referred to as “FPT in P”). Given some parameter k we search for an algorithm
with a running time of f (k)(n + m)2−ε that solves Diameter. Starting FPT in P for Diameter,
Abboud et al. [1] proved that, unless the SETH fails, the function f has to be an exponential
function if k is the treewidth of the graph. We extend their research by systematically exploring
the parameter space looking for parameters where f can be a polynomial. If this is not possible (due
to conditional lower bounds), then we seek for matching upper bounds of the form f (k)(n + m)2−ε
where f is exponential.
In a second step, we combine parameters that are known to be small in many real world graphs.
We concentrate on social networks which often have special characteristics, including the “smallworld” property and a power-law degree distribution [26, 27, 28, 29, 30]. We therefore combine
parameters related to the diameter with parameters related to the h-index1 .
1 The
h-index of a graph G is the largest number ℓ such that G contains at least ℓ vertices of degree at least ℓ.
1
1.1
Related Work
Due to its importance, Diameter is extensively studied. The theoretically fastest algorithms
are based on matrix multiplication and run in O(n2.373 ) time [33] and any O((n + m)2−ε )-time
algorithm refutes the SETH [32]. The diameter can be computed in linear time in several special
graph classes including trees [22], outerplanar graphs [17], interval graphs [15, 31], ptolemaic
graphs [15], dually chordal graphs [5], distance-hereditary graphs [13, 14], benzenoid systems [10],
and planar triangulations and quadrangulations [11].
The following results on approximating Diameter are known: It is folklore that a simple
breadth-first search gives a linear-time 2-approximation. Aingworth et al. [2] improved
√ the approximation factor to 3/2 at the expense of the higher running time of O(n2 log n + m n log n).
The lower bound of Roditty and Williams [32] also implies that approximating Diameter within
a factor of 3/2 − δ in O(n2−ε ) time refutes the SETH. On planar graphs, there is an approximation
scheme with near linear running time [36]; the fastest exact algorithm for Diameter on planar
graphs runs in O(n1.667 ) time [18].
Concerning FPT in P, Diameter can be solved in 2O(k log k) n1+o(1) time where k is the
treewidth of the graph; however, an 2o(k) n2−ε -time algorithm refutes the SETH [1]. In fact,
the construction actually proves the same running time lower bound with k being the vertex
cover number. The reduction for the lower bound of Roditty and Williams [32] also implies that
the SETH is refuted by any f (k)(n + m)2−ε -time algorithm for Diameter for any computable
function f when k is either the the (vertex deletion) distance to chordal graphs or the combined
parameter h-index and domination number. Evald and Dahlgaard [16] adapted the reduction by
Roditty and Williams and proved that any f (k)(n + m)2−ε -time algorithm for Diameter parameterized by the maximum degree k for any computable function f refutes the SETH.
1.2
Our Contribution
We make progress towards systematically classifying the complexity of Diameter parameterized
by structural graph parameters. Figure 1 gives an overview of previously known and new results
and their implications (see Brandstädt et al. [6] for definitions of the parameters). In Section 2,
we follow the “distance from triviality parameterization” [21] aiming to extend known tractability
results for special graph classes. For example, Diameter is linear-time solvable on trees or cliques.
Then, for the (vertex deletion) distance k to clique we provide an O(k · (n + m))-time algorithm.
Similarly, for the parameter feedback edge number k (edge deletion number to trees) we obtain
an O(k · n) algorithm. However, these are our only k O(1) (n + m)-time algorithms in this section.
For the remaining parameters, it is already known that such algorithms refute the SETH. For the
parameter distance k to cographs we therefore provide an 2O(k) (n + m)-time algorithm. Finally,
for the parameter distance to bipartite graphs, we use the recently introduced notion of GeneralProblem-hardness [3] to show that even an f (k)n2.3 -time algorithm for any function f would imply
an O(n2.3 )-time algorithm for Diameter.
In Section 3, we investigate parameters that are usually small in social networks like the diameter and h-index. We prove that a k O(1) (n + m)-time algorithm where k is the combined parameter
diameter and maximum degree would refute the SETH. Complementing this lower bound, we provide an f (k)(n + m)-time algorithm where k is the combined parameter diameter and h-index.
1.3
Preliminaries
For ℓ ∈ N we set [ℓ] := {1, 2, . . . , ℓ}. We use mostly standard graph notation. Given a graph G =
(V, E) set n := |V | and m := |E|. A path P = v0 . . . va is a graph with vertex set {v0 , . . . , va } and
edge set {{vi , vi+1 } | 0 ≤ i < a}. For u, v ∈ V , we denote with distG (v, u) the distance between u
and v in G. If G is clear from the context, then we omit the subscript. For a vertex subset V ′ ⊆ V ,
we denote with G[V ′ ] the graph induced by V ′ . We set G − V ′ := G[V \ V ′ ]. We denote by dia(G)
the diameter of G.
2
Max. Degree
+ Dom. No.
Ob. 6
Vertex Cover
Number
h-index +
Dom. No.
[1]
Distance to
Clique
Feedback
Edge Number
Ob. 3
[32]
Th. 4
h-index +
Diameter
Maximum
Degree
Th. 7
Th. 6
[16]
Treewidth
Ac. Chrom. No.
+ Diameter
h-index
Distance to
Interval
[1]
Ob. 2
Domination
Number
Distance to
Cograph
Distance to
Chordal
[32]
Th. 2
[32]
Acyclic
Chromatic No.
Distance to
Bipartite
Average
Degree
Th. 1
Girth
Th. 5
Bisection
Width
Ac. Chrom. No.
+ Dom. No.
Th. 3
Max Diameter
of Components
Max Degree
+ Diameter
Ob. 4
Distance to
Perfect
Chromatic
Number
Th. 1
Minimum
Degree
Th. 4
Figure 1: Overview of the relation between the structural parameters and the respective results
for Diameter. An edge from a parameter α to a parameter β below of α means that β can
be upper-bounded in a polynomial (usually linear) function in α (see also [34]). The three small
boxes below each parameter indicate whether there exist (from left to right) an algorithm running
in f (k)n2 , f (k)(n log n + m), or k O(1) (n log n + m) time, respectively. If a small box is green,
then a corresponding algorithm exist and the box to the left is also green. Similarly, a red box
indicates that a corresponding algorithm is a major breakthrough. More precisely, if a middle box
(right box) is red, then an algorithm running in f (k) · (n + m)2−ε (in k O(1) · (n + m)2−ε ) time
refutes the SETH. If a left box is red, then an algorithm with running time f (k)n2 implies a faster
algorithm for Diameter in general. Hardness results for a parameter α imply the same hardness
results for the parameters below α. Similarly, algorithms for a parameter β imply algorithms for
the parameters above β.
Parameterized Complexity and GP-hardness. A language L ⊆ Σ∗ × N is a parameterized
problem over some finite alphabet Σ, where (x, k) ∈ Σ∗ × N denotes an instance of L and k is
the parameter. Then L is called fixed-parameter tractable if there is an algorithm that on input (x, k) decides whether (x, k) ∈ L in f (k) · |x|O(1) time, where f is some computable function
only depending on k and |x| denotes the size of x. For a parameterized problem L, the language L̂ = {x ∈ Σ∗ | ∃k : (x, k) ∈ L} is called the unparameterized problem associated to L. We
use the notion of general problem hardness [3].
Definition 1. Let P ⊆ Σ∗ × N be a parameterized problem, let P̂ ⊆ Σ∗ be the unparameterized
decision problem associated to P , and let f : N → N be a function. We call P ℓ-General-Problemhard(f ) (ℓ-GP-hard(f )) if there exists an algorithm A transforming any input instance I of P̂ into
a new instance (I ′ , k ′ ) of P such that
3
v1
v2
w1
u2
u1
w2
v3
v4
w3
u3
u4
w4
Figure 2: Example for the construction in the proof of Theorem 1. The input graph given on the
left side has diameter two and the constructed graph on the right side has diameter three. Longest
shortest paths are highlighted.
(G3) k ′ ≤ ℓ, and
(G4) |I ′ | ∈ O(|I|).
(G1) A runs in O(f (|I|)) time,
(G2) I ∈ P̂ ⇐⇒ (I ′ , k ′ ) ∈ P ,
We call P General-Problem-hard(f ) (GP-hard(f )) if there exists an integer k such that P is kGP-hard(f ). We omit the running time and call P k-General-Problem-hard (k-GP-hard) if f is a
linear function.
Showing GP-hardness for some parameter k allows to transfer lower bounds from the unparameterized problem to the parameterized problem as stated next.
Lemma 1 ([3]). Let f : N → N be a function, let P ⊆ Σ∗ × N be a parameterized problem that
is ℓ-GP-hard(f ), and let P̂ ⊆ Σ∗ be the unparameterized decision problem associated to P . If
there is an algorithm solving each instance (I, k) of P in O(g(k) · f (|I|)) time, then there is an
algorithm solving each instance I ′ of P̂ in O(f (|I ′ |)) time.
2
Deletion Distance to Special Graph Classes
In this section we investigate parameterizations that measure the distance to special graph classes.
The hope is that when Diameter can be solved efficiently in a special graph class Π, then
Diameter can be solved if the input graph is “almost” in Π. We study the following parameters
in this order: distance to bipartite graphs, distance to interval graphs, distance to cographs, distance
to clique, and feedback edge number. All these parameters except the last one measure the vertex
deletion distance to some graph class. Feedback edge number measures the edge deletion distance to
trees. Note that the lower bound of Abboud et al. [1] for the parameter vertex cover number already
implies for k being one of the first three parameters in our list, that there is no k O(1) (n+m)2−ε -time
algorithm unless the SETH breaks.
Distance to bipartite graphs. We show that Diameter parameterized by distance to bipartite
graphs and girth is 4-GP-hard. Consequently solving Diameter in f (k) · n2.3 for any computable
function f , implies an O(n2.3 )-time algorithm for Diameter which would be a breakthrough
result.
Theorem 1. Diameter is 4-GP-hard with respect to the combined parameter distance to bipartite
graphs and girth.
Proof. Let G = (V, E) be an arbitrary undirected graph where V = {v1 , v2 . . . vn }. We construct
a new graph G′ = (V ′ , E ′ ) as follows:
V ′ ..= {ui , wi | vi ∈ V }, and
E ′ ..= {{ui , wj }, {uj , wi } | {vi , vj } ∈ E} ∪ {{ui , wi } | vi ∈ V }.
An example of this construction can be seen in Figure 2. We will now prove that all prop-
4
erties of Definition 1 hold. It is easy to verify that the reduction can be implemented in linear
time and therefore the resulting instance is of linear size as well. Observe that {ui | vi ∈ V }
and {wi | vi ∈ V } are both independent sets and therefore G′ is bipartite. Notice further that
for any edge {vi , vj } ∈ E there is an induced cycle in G′ containing the vertices {ui , wi , uj , wj }.
Since G′ is bipartite there is no induced cycle of length three in G′ and thus the girth of G′ is four.
Lastly, we show that dia(G′ ) = dia(G)+1 by proving that if dist(vi , vj ) is odd, then dist(ui , wj ) =
dist(vi , vj ) and dist(ui , uj ) = dist(vi , vj ) + 1, and if dist(vi , vj ) is even, then dist(ui , uj ) =
dist(vi , vj ) and dist(ui , wj ) = dist(vi , vj ) + 1. Since dist(ui , wi ) = 1 and dist(ui , wj ) = dist(uj , wi ),
this will conclude the proof.
Let P = va0 va1 . . . vad be a shortest path from vi to vj where va0 = vi and vad = vj .
Let P ′ = ua0 wa1 ua2 wa3 . . . be a path in G′ . Clearly, P ′ is also a shortest path as there are
no edges {ui , wj } ∈ E ′ where {vi , vj } ∈
/ E.
If d is odd, then ua0 wa1 . . . wad is a path of length d from ui to wj and ua0 wa1 . . . wad uad is
a path of length d + 1 from ui to uj . If d is even, then ua0 wa1 . . . wad−1 uad is a path of length d
from ui to uj and ua0 wa1 . . . wad−1 uad wad is a path of length d + 1 from ui to wj . Notice that G′
is bipartite and thus dist(ui , uj ) must be even and dist(ui , wj ) must be odd.
Distance to interval graphs. Next, we show that Diameter can be solved in O(k · n2 ) time
provided that a set K of size O(k) is given such that G′ = G − K is in a graph class on which AllPairs Shortest Paths can be solved in O(n2 ) time. The idea is fairly simple: First compute G′
and solve All-Pairs Shortest Paths on it in O(n2 ) time. Next, compute a breadth-first search
from every vertex in K in O(k · m) time and store all distances found in a table. Lastly, compute
each pair a, c ∈ V \ K
distG (a, c) := min{distG′ (a, c), min{distG (a, b) + distG (b, c)}}.
b∈K
This algorithm takes altogether O(n + m + n2 + k · m + n2 · k) = O(k · n2 ) time.
Observation 1. Let Π be a graph class such that All-Pairs Shortest Paths can be solved
in O(n2 ) time on Π. If the (vertex) deletion set K to Π is given, then All-Pairs Shortest
Paths can be solved in O(|K| · n2 ) time.
It is known that All-Pairs Shortest Paths can be solved in O(n2 ) time on interval
graphs [9, 35]. Thus we obtain the following.
Observation 2. Diameter parameterized by the distance k to interval graphs is solvable in
O(k · n2 ) time provided that the deletion set is given.
Distance to cographs. Providing an algorithm that matches the lower bound of Abboud et al. [1],
we will show that Diameter parameterized by distance k to cographs can be solved in O(k · (n +
m) + 2O(k) ) time.
A graph is a cograph if and only if it is P4 -free. Given a graph G one can determine in linear
time whether G is a cograph and can return an induced P4 if this is not the case [7, 12]. This
implies that in O(k · (n + m)) time one can compute a set K ⊆ V with |K| ≤ 4k such that G − K
is a cograph: Iteratively add all four vertices of a returned P4 into the solution set and delete
those vertices from G until it is P4 -free. In the following, we hence assume that such a set K is
given. Notice that every cograph has diameter at most two as any graph with diameter at least
three contains an induced P4 .
Theorem 2. Diameter can be solved in O(k · (n + m) + 216k k) time when parameterized by
distance k to cographs.
Proof. Let G = (V, E) be the input graph with distance k to cograph. Let K = {x1 , . . . , xj } be a
set of vertices such that G′ = G − K is a cograph with j ≤ 4k. Recall that K can be computed
in O(k · (n + m)) time.
5
We first compute all connected components and their diameter in G′ in linear time and store
for each vertex the information in which connected component it is. Notice that we only need
to check for each connected component C, whether C induces a clique in G′ since otherwise its
diameter is two. In a second step, we perform in O(k · (n + m)) time a breadth-first search in G
from each vertex v ∈ K and store the distance between v and every other vertex w in a table.
Next we introduce some notation. The type of a vertex u ∈ V \ X is a vector of length d where
the ith entry describes the distance from u to xi with the addition that any value above three is
set to 4. We say a type is non-empty, if there is at least one vertex with this type. We compute
for each vertex u ∈ V \ X its type. Additionally we store for each non-empty type the connected
component its vertex is in or that there are at least two different connected components containing
a vertex of that type. This takes O(n · k) time and there are at most 4j many different types.
Lastly, we iterate over all pairs of types (including the pairs where both types are the same) and
compute the largest distance between vertices of these types. Since each vertex is either in K (and
therefore the distance to each vertex is known) or has some type, this concludes the algorithm.
Let y, z be the vertices of the respective types with maximum pairwise distance. We will first show
how to find y and z and then show how to correctly compute their distance in O(k) time. If both
types only appear in the same connected component, then the distance between the two vertices of
these types is at most two. If two types appear in different connected components, then a longest
shortest path between vertices of the respective type contain at least one vertex in K. Observe
that since each cograph has diameter at most two, each third vertex in any longest shortest path
must be in K. Thus a shortest y-z–path contains at least one vertex xi ∈ K with dist(xi , y) < 3.
By definition, each vertex with the same type as y has the same distance to xi and therefore the
same distance to z unless there is no shortest path from it to z that passes through xi , that is, it is
in a same connected component as z. Thus, we can choose two arbitrary vertices of the respective
types in different connected components. We can compute the distance between y and z in O(k)
time by computing minx∈K dist(y, x) + dist(x, z). Observe that the shortest path from y to z
contains xi and therefore dist(y, xi ) + dist(xi , z) = dist(y, z). We can compute the diameter of G
this way in O(k · (n + m) + 48k · k) time.
Distance to clique. We next show how to solve Diameter parameterized by distance k to
clique in O(k · (n + m)) time. Observe that distance to clique is equivalent to vertex cover number
of the complement graph and can therefore be 2-approximated in linear time.
Observation 3. Diameter parameterized by distance k to clique can be solved in O(k · (n + m))
time.
Proof. We first provide a linear-time 2-approximation to compute the distance set K.
Let G = (V, E) be the input graph and let k be its distance to clique Compute in linear time
the degree of each vertex and the number n = |V | of vertices. Iteratively check for each vertex v
whether its degree is n − 1. If deg(v) = n − 1, then v can be deleted as it is in every largest
clique and thus decrease n by one and the degree of each other vertex by one. If not, then we
can find a vertex w which is not adjacent to v in O(deg(v)) time. Put v and w in the solution
set, delete both vertices and all incident edges and adjust the number
P of vertices and their degree
accordingly. This can be done in O(deg(v) + deg(w)) time. Since v∈V deg(v) ∈ O(n + m) this
procedure takes O(n + m) time.
We use the algorithm described above to compute a set K such that G′ = G − K is a clique
and |K| ≤ 2k in linear time. Since G′ is a clique, its diameter is one if there are at least two
vertices in the clique. We therefore assume that there is at least one vertex in K. Compute for each
vertex v ∈ K a breadth-first search rooted in v in linear time and return the largest distance found.
The returned value is the diameter of G as each longest induced path is either of length one or
has at least one endpoint in K. The procedure takes O(|K| · (n + m) + n + m) = O(k · (n + m))
time.
Feedback edge number. We will prove that Diameter parameterized by feedback edge number k can be solved in O(k · n) time. One can compute a minimum feedback edge set K, k = |K|,
6
in linear time by taking all edges not in a spanning tree. We will first present a data reduction
rule that iteratively removes all vertices of degree one in linear time from the graph and therein
introduces a weight function pen on the remaining vertices. Intuitively, pen stores the length of a
longest path to a vertex in a “pending tree”, that is, to a vertex which is removed by the reduction
rule. Afterwards, we will compute the diameter of the remaining vertex-weighted graph in O(k · n)
time.
Theorem 3. Diameter parameterized by feedback edge number k can be solved in O(k · n) time.
Proof. Let G = (V, E) be the input graph with feedback edge number k and let K be a feedback
edge set with |K| = k. Compute in linear time the degree of each vertex and sort all vertices by
their degree using bucket-sort. Since we use a vertex-weighted problem variant, initialize pen(v) =
0 for each vertex v.
The main idea of our algorithm is as follows: We store in each vertex v the length of a longest
path P ending in v such that v is the only vertex of P that was not removed. When deleting
a degree-one vertex u set the pen-value of its unique neighbor v to max{pen(u) + 1, pen(v)}.
Store in an additional variable s the maximum of s (before the reduction rule) and pen(u) +
pen(v) + 1. This addresses the case that a longest shortest path has both its endpoints in pending
trees that are connected to the same vertex v. Initially, s is set to zero. We then compute the
maximum dist(u, v) + pen(u) + pen(v) over all pairs of remaining vertices u, v. We will prove that
this is the diameter of G.
Reduction Rule 1. Let u be a vertex of degree one and let v be its neighbor. Delete u and the
incident edge from G, set s = max{s, pen(u) + pen(v) + 1} and pen(v) = max{pen(u) + 1, pen(v)}.
We first prove the correctness of this data reduction rule and afterwards analyze its running
time. Observe that a degree-one vertex has only one neighbor so any path ending in the degreeone vertex must pass through the unique neighbor. If we delete u, then the pen-value of its
unique neighbor v is set to max{pen(u) + 1, pen(v)} since any longest path from a pending tree
that ends in u now ends in v and its length increases by one. Thus any shortest path that
ended in a pending tree connected to some vertex u and some other vertex w is now represented
by pen(u) + pen(w) + dist(u, w). Note that it might be the case that a longest shortest path
has both its endpoints in pending trees that are connected to the same vertex v. Let v1 , v2 ∈
N (v) be the roots of these pending trees, respectively. In this case, the diameter of the graph
is pen(v1 )+pen(v2 )+2. Note that one of v1 and v2 is deleted first. Let without loss of generality v1
be the vertex deleted first. Then, pen(v) = pen(v1 ) + 1. In the step where v2 is deleted, s is set
to pen(v2 ) + pen(v) + 1 = pen(v2 ) + pen(v1 ) + 2 = dia(G). Thus the reduction rule is correct.
We now analyze the running time. Notice that all vertices are already sorted by their degree.
Removing u and {u, v} and updating pen(v), deg(v) and s take constant time and the sorting can
clearly be adjusted in constant time as well by just moving v from one bucket to another. Since
each vertex is removed at most once, the overall running time to apply this rule exhaustively is
in O(n).
Since the resulting graph has no degree-one vertices we can partition the vertex set V into
vertices V =2 of degree exactly two and vertices V ≥3 of degree at least three. Using standard
argumentation we can show that |V ≥3 | ∈ O(min{k, n}) and all vertices in V =2 are either in pending
cycles or in maximal paths [4, Lemma 1]. A maximal path is an induced subgraph P = x0 x1 . . . xa
where {xi , xi+1 } ∈ E for all 0 ≤ i < a, x0 , xa ∈ V ≥3 , xi ∈ V =2 for all 0 < i < a, and x0 6= xa . A
pending cycle is basically the same except x0 = xa and deg(x0 ) may possibly be two. The set C
of all pending cycles and P of maximal paths can be computed in linear time [4, Lemma 2].
Since pending cycles are easier to deal with, we will start with them. Let C = x0 x1 . . . xa be
a pending cycle. We distinguish between two cases: the longest shortest path starts and ends
in C or it leaves C. We first start with those paths that start and end in C. To this end in O(a)
time we compute i such that dist(xi , x0 ) + pen(xi ) is maximized and set s = max{s, pen(x0 ) +
dist(xi , x0 ) + pen(xi )} if i 6= 0. (For i = 0 we do not update s.) Next we use a dynamic program
that given the furthest vertex xj from xℓ (with respect to dist(xj , xℓ ) + pen(xj )) computes the
7
furthest vertex xj ′ from xℓ+1 . To this end, we only consider the shortest paths xℓ+1 xℓ+2 . . . xr
for ℓ + 1 ≤ r ≤ ℓ + 1 + ⌊a/2⌋ mod a, that is, we only consider shortest paths to half of the
vertices in C. Observe that all paths that are ignored by this restriction will be considered in the
iteration where xj ′ is considered and xℓ+1 is the vertex furthest away. Notice that the furthest
vertex xj ′ from xℓ+1 is either the furthest vertex from xℓ or some vertex that is ignored by xℓ since
all distances that are considered are exactly one larger for xℓ than they are for xℓ+1 . The only
possible vertex that is ignored by xℓ but not by xℓ+1 is xℓ+1+⌊a/2⌋ mod a . Thus we can compute
the furthest vertex from xℓ+1 in constant time and update s accordingly. (Note that if the furthest
vertex is xℓ+1 , then we do not update s.) The whole pending cycle can be checked in O(a) time
in this way.
Notice that all paths that leave C contain x0 and therefore only the distance to x0 and the penvalue matter. Thus we can use i (recall, this is the index that maximizes dist(xi , x0 ) + pen(xi ),
possibly i = 0) to update pen(x0 ) = dist(xi , x0 ) + pen(xi ) and delete C from G.
We now have all ingredients to state the main algorithm. First we perform a breadth-first search
(BFS) from each vertex v ∈ V ≥3 in overall O(min{k, n}·(2n+k)) = O(k·n) time and store for each
vertex u ∈ V \ {v} the distance dist(v, u) and update s = max{s, pen(v) + pen(u) + dist(v, u)}.
This way we find all shortest paths that start or end in a vertex in V ≥3 (or a pendant tree
connected to such a vertex). It remains to analyze all shortest v-u-paths with v, u ∈ V =2 , that is,
both endpoints of the path are in some maximal paths. The case where v and u are in the same
maximal path is similar to the case of pending cycles. The only adjustment is the computation
of the “mid”-index, that is, the index that is considered by xℓ+1 but not by xℓ . For a maximal
path P = x0 x1 . . . xa the respective index for xℓ+1 is ℓ + 1 + ⌊(a + dist(x0 , xa ))/2⌋ if ℓ + 1 < a/2
and ℓ + 1 − ⌊(a + dist(x0 , xa ))/2⌋ otherwise. If this index is not within [1, a − 1], then it can safely
be ignored.
The last remaining case is where v and u are in two different maximal paths. Let P1 = x0 x1 . . . xa
and P2 = y0 y1 . . . yb be two different maximal paths. We compute the longest shortest path between v and u where v ∈ {x1 , . . . xa−1 } and u ∈ {y1 , . . . yb−1 } in O(a + b) time as follows: We
use two dynamic programs to sweep over P1 . One computes the furthest distance of each vertex
with the additional requirement, that a shortest path passes through xa and the other one with
the requirement that a shortest path passes through x0 . Since they are completely symmetric, we
will only describe the first one here.
To this end, we first compute the index i that maximizes pen(yi ) + dist(x1 , yi ) where a
shortest path passes through xa . This can be done by computing for each vertex yh the distances dist(xa , yh ) = min{dist(xa , y0 ) + h, dist(xa , yb ) + b − h}. For each vertex xℓ ∈ V (P1 ) there
are three cases: There exist shortest paths to all vertices in P2 that pass through x0 , there exists
shortest paths to all vertices in P2 that pass through xa or shortest paths to some vertices pass
through x0 and to other vertices they pass through xa . Notice that if the first case applies for some
vertex xh , then it also applies to all xg , g < h. Analogously, if the second case applies for some
vertex xh , then it also applies to all xg , g > h. Lastly, if the second case applies to two adjacent
vertices xh , xh+1 , a shortest path from xh to y0 passes through x0 and a shortest path from xh
to yb passes through xa , then shortest xh+1 -y0 paths also pass through x0 and shortest xh+1 -yb
paths also pass through xa . Since dist(x0 , y0 ), dist(x0 , yb ), dist(xa , y0 ) and dist(xa , yb ) are given,
we can test which case applies to each vertex in constant time per vertex and partition P1 into
three parts according to this in O(a) time. Observe that for all vertices to which the first case
applies, we already know that yi is the furthest vertex. Analogously, for all vertices to which the
second case applies, the anchor of the second dynamic program is the furthest vertex in P2 . Thus
we only consider the subpath to which the third case applies. Let xj be the first vertex in this
subpath. We compute in O(b) the furthest vertex yi′ in P2 such that a shortest path to it passes
through xa . The dynamic program then computes the furthest distance from xℓ+1 that passes
through xa given the furthest distance from xℓ through xa . The vertex yi′ is the anchor for our
dynamic program.
Now the main idea is still the same as for pending cycles—the furthest vertex either stays the
same or is ignored by xℓ . This “mid”-index for the dynamic program is the vertex yj such that
the distance from xℓ to yj through xa is larger than through x0 and for xℓ+1 its at most as large
8
as through x0 . There are two possible cases: The shortest path from xℓ to y0 passes through xa
or through x0 . Since both cases are symmetric, we will only consider the first one. This “mid”index j for xℓ+1 is then computed by j = ℓ + 1 + ⌊(b + dist(x0 , yb ) − a − dist(xa , y0 ))/2⌋ since this
is the index where the distance through xa becomes smaller than the distance through x0 . We
again also consider the pen-values and update s accordingly.
It remains to analyze the running time of this algorithm. All of the preprocessing steps can be
done in O(n + m) = O(n + k) time. To deal with vertices in V ≥3 we run a BFS from O(min{n, k})
vertices which takes O(min{n, k} · (n + k)) = O(n · k) time. The time to compute all the dynamicPprograms for the cases where u and v are in the same maximal path or pending cycle
length of longest shortest
is O( S∈C∪P |S|) ⊆ O(|V =2 |) ⊆ O(min{n, k}). The time
P to compute
P
path between vertices in pairs of maximal paths is in O( P1 ∈P P2 ∈P |P1 | + |P2 |) ⊆ O(|V =2 |2 ) ⊆
O(min{n, k} · min{n, k}) ⊆ O(n · k). This concludes the proof.
3
Parameters for Social Networks
Here, we study parameters that we expect to be small in social networks. Recall that social
networks have the “small-world” property and a power-law degree distribution [26, 27, 28, 29, 30].
The “small-world” property directly transfers to the diameter. We capture the power-law degree
distribution by the h-index as only few high-degree exist in the network. Thus, we investigate
parameters related to the diameter and to the h-index starting with degree-related parameters.
3.1
Degree Related Parameters
We investigate the parameters minimum degree and average degree. Unsurprisingly, the minimum
degree is not helpful for parameterized algorithms. In fact, we show that Diameter is 2-GP-hard
with respect to the combined parameter bisection width and minimum degree.
Theorem 4. Diameter is 2-GP-hard with respect to bisection width and minimum degree.
Proof. Let (G = (V, E), d) be an arbitrary input instance for Diameter where V = {v1 , v2 , . . . vn }.
We construct a new instance (G′ = (V ′ , E ′ ), d + 4) as follows: Let
V ′ = {si , ti , ui | i ∈ [n]} ∪ {wi | i ∈ [3n]} and
E ′ = T ∪ W ∪ E ′′ , where
T = {{si , ti }, {ti , ui } | i ∈ [n]},
W = {u1 , w1 } ∪ {{w1 , wi } | i ∈ ([3n] \ {1})}, and
E ′′ = {{ui , uj } | {vi , vj } ∈ E}.
An example of this construction can be seen in Figure 3. We will now prove that all properties
of Definition 1 hold. It is easy to verify that the reduction runs in linear time and that there
are 6n vertices and 5n + m edges in G′ . Notice that {si , ti , ui | i ∈ [n]} and {wi | i ∈ [3n]} are
both of size 3n and that there is only one edge ({u1 , w1 }) between these two sets of vertices. The
bisection width of G′ is therefore one and the minimum degree is also one as s1 is only adjacent
to t1 .
It remains to show that G has diameter d if and only if G′ has diameter d + 4. First, notice
that the subgraph of G′ induced by {ui | i ∈ [n]} is isomorphic to G. Note that dist(si , ui ) = 2
and thus dist(si , sj ) = dist(ui , uj ) + 4 = dist(vi , vj ) + 4 and therefore the diameter of G′ is at
least d + 4. Third, notice that for all vertices x ∈ V ′ \ {ti } it holds that dist(si , x) > dist(ti , x).
Lastly, observe that for all i ∈ [3n] and for all vertices x ∈ V ′ it holds that dist(wi , x) ≤
max{dist(s1 , x), 4}. Thus the longest shortest path in G′ is between two vertices si , sj and is
of distance dist(ui , uj ) + 4 = dist(vi , vj ) + 4 ≤ d + 4.
9
T
v1
v2
s1
t1
s2
t2
W
w11
u1
w10
w12
w9
u2
w8
w1
v3
s3
t3
u3
v4
s4
t4
u4
w2
w7
w3
w6
w4
w5
E ′′
Figure 3: Example for the construction in the proof of Theorem 4. The input graph given on the
left side has diameter 2 and the constructed graph on the right side has diameter 2 + 4 = 6.
We mention in passing, that the constructed graph in the proof of Theorem 4 contains the original graph as an induced subgraph and if the original graph is bipartite, then so is the constructed
graph. Thus, first applying the construction in the proof of Theorem 1 (see also Figure 2) and then
the construction in the proof of Theorem 4 proves that Diameter is GP-hard even parameterized
by the sum of girth, bisection width, minimum degree, and distance to bipartite graphs.
Average degree. Observe that it holds 2m = n · k and therefore the standard algorithm
that computes a breadth-first search from every vertex and returns the largest distance found
takes O(n · m) = O(kn2 ) time.
Observation 4. Diameter parameterized by the average degree k is solvable in O(k · n2 ) time.
3.2
Parameters related to both diameter and h-index
Here, we will study combinations of two parameters where the first one is related to diameter and
the second to h-index (see Figure 1 for an overview of closely related parameters). We start with
the combination maximum degree and diameter. Usually, this parameter is not interesting as the
graph size can be upper-bounded by this parameter and thus fixed-parameter tractability with
respect to this combined parameter is trivial. Diameter is no exception: Since we may assume
that the input graph only consists of one connected component, every vertex is found by any
breadth-first search. Any breadth-first search may only reach depth d, where d isPthe diameter of
d
the input graph, and as each vertex may only have ∆ neighbors there are at most i=0 ∆i < ∆d+1
vertices. Since m ≤ n · ∆ the O(n · m)-time algorithm implies the following.
Observation 5. Diameter parameterized by diameter d and maximum degree ∆ is solvable
in O(∆2d+3 ) time.
Interestingly, although the parameter is quite large, the algorithm above cannot be improved
by much as stated next.
Theorem 5. There is no (d + ∆)O(1) (n + m)2−ǫ -time algorithm that solves Diameter parameterized by maximum degree ∆ and diameter d unless the SETH is false.
√
c
for
Proof. We prove a slightly stronger statement excluding 2o( d+∆) · (n + m)2−ǫ -time algorithms
√
o( r d+∆)
some constant c. Assume towards a contradiction that for each constant r there is a 2
·
(n + m)2−ǫ -time algorithm that solves Diameter parameterized by maximum degree ∆ and diameter d. Evald and Dahlgaard [16] have shown a reduction from CNF-SAT to Diameter where the
resulting graph has maximum degree three such that for any constant ǫ > 0 an O((n+m)2−ǫ )-time
algorithm (for Diameter) would refute the SETH. A closer look reveals that there is some constant c such that the diameter d in their constructed graph is in O(log c (n + m)). By assumption
10
we can solve Diameter parameterized by maximum degree and diameter in 2o(
time. Observe that
2o(
= 2o(
√
c
d+∆)
√
c
√
c
d+∆)
· (n + m)2−ǫ
· (n + m)2−ǫ
logc (n+m))
· (n + m)2−ǫ
= 2o(log(n+m)) · (n + m)2−ǫ
′
= (n + m)o(1) · (n + m)2−ǫ ∈ O((n + m)2−ǫ ) for some ε′ > 0.
′
Since we constructed for some ǫ′ > 0 an O((n + m)2−ǫ )-time algorithm for Diameter
the SETH
√
c
falls and thus we reached a contradiction. Finally, notice that (d + ∆)O(1) ⊂ 2o( d+∆) for any
constant c.
Next, observe that for any graph of n vertices, domination number d, and maximum degree ∆
it holds that n ≤ d · (∆ + 1) as each vertex is in a dominating set or is a neighbor of at least one
vertex in it. The next observation follows from m ≤ n · ∆ and provides the only parameterized
algorithm in this section where the dependency on the parameter is polynomial.
Observation 6. Diameter parameterized by domination number d and maximum degree ∆ is
solvable in O(d2 · ∆3 ) time.
h-index and diameter. We next investigate in the combined parameter h-index and diameter.
Notice that the reduction by Roditty and Williams [32] has constant domination number and
logarithmic vertex cover number (in the input size). Since the diameter d is upper-bounded by the
domination number and the h-index h is upper-bounded by the vertex cover number, any algorithm
that solves Diameter parameterized by the combined parameter (d + h) in 2o(d+h) · (n + m)2−ǫ
time disproves the SETH. We will now present an algorithm for Diameter parameterized by
h-index and diameter that matches the lower bound.
Theorem 6. Diameter parameterized by diameter d and h-Index h is solvable in O(2d+h · h · (n +
m)) time.
Proof. Let H = {x1 , . . . , xh } be a set of vertices such that all vertices in V \ H have degree at
most h in G. Clearly, H can be computed in linear time. We will describe a two-phase algorithm
with the following basic idea: In the first phase it performs a breadth-first search from each
vertex v ∈ H, stores the distance to each other vertex and uses this to compute the “type” of
each vertex, that is, a characterization by the distance to each vertex in H, of each vertex. In
the second phase it iteratively increases a value e and verifies whether there is a vertex pair of
distance at least e. If at any point no vertex pair is found, then the diameter of G is e − 1.
The first phase is straight forward: Compute a BFS from each vertex v in H and store the
distance from v to every other vertex w in a table. Then iterate over each vertex w ∈ V \ H and
compute a vector of length h where the ith entry represents the distance from w to xi . Also store
the number of vertices of each type. Since the distance to any vertex is at most d, there are at
most dh different types. This first phase takes O(h · (m + n)) time.
For the second phase, we initialize e with the largest distance found so far, that is, the maximum
value stored in the table and compute G′ = G − H. Iteratively check whether there is a pair of
vertices in V \ H of distance at least e + 1 as follows. We check for each vertex v ∈ V \ H and
each type whether there is a path of length at most e from v to each vertex of this type through
a vertex in H. This can be done by computing the sum of the two type-vectors in O(h) time and
comparing the minimum entry in this sum with e. If all entries are larger than e, then no shortest
path from v to some vertex w of the respective type of length at most e can contain any vertex
in H. Thus we compute a BFS from v in G′ up to depth e and count the number of vertices of
the respective type we found. If this number equals the total number of vertices of the respective
type, then for all vertices w of this type it holds that dist(v, w) ≤ e. If the two numbers do not
11
match, then there is a vertex pair of distance at least e + 1 so we can increase e by one and start
the process again.
There are at most d iterations in which e is increased and the check is done. Recall that the
maximum degree in G′ is h and therefore each of these iteration takes O(n·dh ·(he +h)) time as each
BFS to depth e takes O(he ) time. Thus the overall running time is in O(h · (m + n) + 2h·d · h · n) =
O(2h·d · h · n) as e ≤ d + 1 and the h-index upper-bounds the average degree.
Acyclic chromatic number and domination number. We next analyze the parameterized
complexity of Diameter parameterized by acyclic chromatic number a and domination number d.
Since the acyclic chromatic number upper-bounds the average degree the standard O(n · m)-time
algorithm runs in O(n2 · a) time. We will show that this is essentially the best one can hope for as
we can exclude f (a, d) · (n + m)2−ε -time algorithms assuming SETH. Our result is based on the
reduction by Roditty and Williams [32] and is modified such that the acyclic chromatic number
and domination number are both four in the resulting graph.
Theorem 7. There is no f (a, d) · (n + m)2−ǫ -time algorithm for any computable function f that
solves Diameter parameterized by acyclic chromatic number a and domination number d unless
the SETH is false.
Proof. We provide a reduction from CNF-SAT to Diameter where the input instance has constant acyclic chromatic number and domination number and such that an O((n + m)2−ε )-time
algorithm refutes the SETH. Since the idea is the same as in Roditty and Williams [32] we refer the reader to their work for more details. Let φ be a CNF-SAT instance with variable
set W and clause set C. Assume without loss of generality that |W | is even. We construct an
instance (G = (V, E), k) for Diameter as follows:
Randomly partition W into two set W1 , W2 of equal size. Add three sets V1 , V2 and B of
vertices to G where each vertex in V1 (in V2 ) represents one of 2|W1 | = 2|W2 | possible assignments
of the variables in W1 (in W2 ) and each vertex in B represents a clause in C. Clearly |V1 | + |V2 | =
2 · 2|W |/2 and |B| = |C|. For each vi ∈ V1 and each uj ∈ B we add a new vertex sij and the
two edges {vi , sij } and {uj , sij } to G if the respective variable assignment does not satisfy the
respective clause. We call the set of all these newly introduced vertices S1 . Now repeat the
process for all vertices wi ∈ V2 and all uj in B and call the newly introduced vertices qij and
the set S2 . Finally we add four new vertices t1 , t2 , t3 , t4 and the following sets of edges to G:
{{t1 , v} | v ∈ V1 }, {{t2, s} | s ∈ S1 }, {{t3 , q} | q ∈ S2 }, {{t4 , w} | w ∈ V2 }, {{t2 , b}, {t3 , b} | b ∈ B},
and {{t1 , t2 }, {t2 , t3 }, {t3 , t4 }}. See Figure 4 for a schematic illustration of the construction.
We will first show that φ is satisfiable if and only if G has diameter five and then show that
both the domination number and acyclic chromatic number of G are four. First assume that φ
is satisfiable. Then there exists some assignment β of the variables such that all clauses are
satisfied, that is, the two assignment of β with respect to the variables in W1 and W2 satisfy all
clauses. Let v1 ∈ V1 and v2 ∈ V2 be the vertices corresponding to β. Thus for each b ∈ B we
have dist(v1 , b) + dist(v2 , b) ≥ 5. Observe that all paths from a vertex in V1 to a vertex in V2 that
do not pass a vertex in B pass through t2 and t3 . Since dist(v1 , t3 ) = 3 and dist(v2 , t2 ) = 3 it
follows that dist(v1 , v2 ) = 5. Observe that the diameter of G is at most five since each vertex is
connected to some vertex in {t1 , t2 , t3 , t4 } and these four are of pairwise distance at most three.
Next assume that the diameter of G is five. Clearly there is a shortest path between a vertex vi ∈ V1 and vj ∈ V2 of length five. Thus there is no path of the form vi sih uh qjh wj for
any uh ∈ B. This corresponds to the statement that the variable assignment of vi and wj satisfy
all clauses and therefore φ is satisfiable.
The domination number of G is four as each vertex is adjacent to at least one vertex in {t1 , t2 , t3 , t4 }.
The acyclic chromatic number of G is at most four as V1 ∪ V2 ∪ B, S1 ∪ S2 ∪ {t1 , t2 }, {t3 } and {t4 }
each induce an independent set.
Now assume that we have an O(f (k)·(n+m)2−ǫ )-time algorithm for Diameter parameterized
by domination number and acyclic chromatic number. Since the constructed graph has O(2|W |/2 ·|C|)
12
V1
S1
B
S2
V2
v1
s11
u1
q1m
w1
v2
s12
u2
q22
w2
v3
s22
u3
q2m
w3
v4
s3m
..
.
q31
w4
..
.
s41
um
q43
..
.
v2n/2
..
.
..
.
w2n/2
t1
t2
t3
t4
Figure 4: A schematic illustration of the construction in the proof of Theorem 7. Note that
the resulting graph has acyclic chromatic number four (see colors) and a dominating number four
({t1 , t2 , t3 , t4 }).
vertices and edges, this would imply an algorithm with running time
O(f (8) · (2|W |/2 · |C|)2−ǫ )
= O(2(|W |/2)(2−ǫ) · |C|(2−ǫ) )
= O(2|W |(1−ǫ/2) · |C|(2−ǫ) )
′
= 2|W |(1−ǫ ) · (|C| + |W |)O(1) for some ε′ > 0.
Hence, such an algorithm for Diameter would refute the SETH.
4
Conclusion
We have resolved the complexity status of Diameter for most of the parameters in the complexity
landscape shown in Figure 1. Still, several open questions remain. For example, is there an f (k)n2 time algorithm with respect to the parameter diameter? Of course, the list of parameters displayed
in Figure 1 is by no means exhaustive. Hence, the question arises which other parameters are small
in typical scenarios? For example, social networks also have special community structures [20].
How can this structure be exploited by parameterized algorithms? Finally, is there a generic
′
procedure to solve Diameter parameterized by distance to Π in O(f (k)(n+m)2−ε ) time provided
that Diameter can be solved in O((n + m)2−ε ) time on graphs in Π?
References
[1] Amir Abboud, Virginia Vassilevska Williams, and Joshua R. Wang. Approximation and fixed
parameter subquadratic algorithms for radius and diameter in sparse graphs. In Proceedings of
13
the 27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA ’16), pages 377–391.
SIAM, 2016. URL http://dx.doi.org/10.1137/1.9781611974331.ch28.
[2] Donald Aingworth, Chandra Chekuri, Piotr Indyk, and Rajeev Motwani. Fast estimation of
diameter and shortest paths (without matrix multiplication). SIAM Journal on Computing,
28(4):1167–1181, 1999. URL http://dx.doi.org/10.1137/S0097539796303421.
[3] Matthias Bentert, Till Fluschnik, André Nichterlein, and Rolf Niedermeier. Parameterized
aspects of triangle enumeration. In Proceedings of the 21st International Symposium on
Fundamentals of Computation Theory (FCT ’17), volume 10472 of LNCS, pages 96–110.
Springer, 2017. URL https://doi.org/10.1007/978-3-662-55751-8_9.
[4] Matthias Bentert, Alexander Dittmann, Leon Kellerhals, André Nichterlein, and Rolf
Niedermeier. Towards improving brandes’ algorithm for betweenness centrality. CoRR,
abs/1802.06701, 2018. URL http://arxiv.org/abs/1802.06701.
[5] Andreas Brandstädt, Victor Chepoi, and Feodor F. Dragan. The algorithmic use of hypertree
structure and maximum neighbourhood orderings. In Proceedings of the 20th International
Workshop on Graph-Theoretic Conecpts in Computer Science (WG ’94), volume 903 of LNCS,
pages 65–80. Springer, 1994. URL http://dx.doi.org/10.1007/3-540-59071-4_38.
[6] Andreas Brandstädt, Van Bang Le, and Jeremy P. Spinrad. Graph Classes: a Survey, volume 3
of SIAM Monographs on Discrete Mathematics and Applications. SIAM, 1999.
[7] Anna Bretscher, Derek G. Corneil, Michel Habib, and Christophe Paul. A simple linear
time lexbfs cograph recognition algorithm. SIAM Journal on Discrete Mathematics, 22(4):
1277–1296, 2008. URL http://dx.doi.org/10.1137/060664690.
[8] Timothy M. Chan and Ryan Williams. Deterministic apsp, orthogonal vectors, and more:
Quickly derandomizing razborov-smolensky. In Proceedings of the 27th Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA ’16), pages 1246–1255. SIAM, 2016. URL
http://dx.doi.org/10.1137/1.9781611974331.ch87.
[9] Danny Z. Chen, D. T. Lee, R. Sridhar, and Chandra N. Sekharan. Solving the all-pair shortest
path query problem on interval and circular-arc graphs. Networks, 31(4):249–258, 1998. URL
http://dx.doi.org/10.1002/(SICI)1097-0037(199807)31:4<249::AID-NET5>3.0.CO;2-D.
[10] Victor Chepoi. On distances in benzenoid systems. Journal of Chemical Information and
Computer Sciences, 36(6):1169–1172, 1996.
[11] Victor Chepoi, Feodor F. Dragan, and Yann Vaxès. Center and diameter problems in
plane triangulations and quadrangulations. In Proceedings of the 13th Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA ’02), pages 346–355. ACM/SIAM, 2002. URL
http://dl.acm.org/citation.cfm?id=545381.545427.
[12] Derek G. Corneil, Yehoshua Perl, and Lorna K. Stewart. A linear recognition algorithm for cographs.
SIAM Journal on Computing, 14(4):926–934, 1985.
URL
http://dx.doi.org/10.1137/0214065.
[13] Feodor F. Dragan. Dominating cliques in distance-hereditary graphs. In Proceedings of the
4th Scandinavian Workshop on Algorithm Theory (SWAT ’94), volume 824 of LNCS, pages
370–381. Springer, 1994. URL http://dx.doi.org/10.1007/3-540-58218-5_34.
[14] Feodor F. Dragan and Falk Nicolai. Lexbfs-orderings of distance-hereditary graphs with
application to the diametral pair problem. Discrete Applied Mathematics, 98(3):191–207,
2000. URL http://dx.doi.org/10.1016/S0166-218X(99)00157-2.
14
[15] Feodor F. Dragan, Falk Nicolai, and Andreas Brandstädt. Lexbfs-orderings and power of
graphs. In Proceedings of the 22nd International Workshop on Graph-Theoretic Conecpts in
Computer Science (WG ’96), volume 1197 of LNCS, pages 166–180. Springer, 1996. URL
http://dx.doi.org/10.1007/3-540-62559-3_15.
[16] Jacob Evald and Søren Dahlgaard.
Tight hardness results for distance and centrality problems in constant degree graphs.
CoRR, abs/1609.08403, 2016.
URL
http://arxiv.org/abs/1609.08403.
[17] Arthur M. Farley and Andrzej Proskurowski. Computation of the center and diameter of outerplanar graphs. Discrete Applied Mathematics, 2(3):185 – 191, 1980. URL
http://www.sciencedirect.com/science/article/pii/0166218X80900396.
[18] Pawel Gawrychowski, Haim Kaplan, Shay Mozes, Micha Sharir, and Oren Weimann. Voronoi
diagrams on planar graphs, and computing the diameter in deterministic Õ (n 5/3) time. In
Proceedings of the 29th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA ’18),
pages 495–514. SIAM, 2018. URL https://doi.org/10.1137/1.9781611975031.33.
[19] Archontia C. Giannopoulou, George B. Mertzios, and Rolf Niedermeier. Polynomial fixedparameter algorithms: A case study for longest path on interval graphs. Theoretical Computer
Science, 689:67–95, 2017. URL https://doi.org/10.1016/j.tcs.2017.05.017.
[20] M. Girvan and M. E. J. Newman. Community structure in social and biological networks.
Proceedings of the National Academy of Sciences, 99(12):7821–7826, 2002.
[21] Jiong Guo, Falk Hüffner, and Rolf Niedermeier. A structural view on parameterizing problems: Distance from triviality. In Proceedings of the 1st International Workshop on Parameterized and Exact Computation(IWPEC 04), volume 3162 of LNCS, pages 162–173. Springer,
2004. URL https://doi.org/10.1007/978-3-540-28639-4_15.
[22] Gabriel Y. Handler. Minimax location of a facility in an undirected tree graph. Transportation
Science, 7(3):287–293, 1973. URL http://dx.doi.org/10.1287/trsc.7.3.287.
[23] Russell Impagliazzo and Ramamohan Paturi.
sat.
Journal of Computer and System Sciences,
http://dx.doi.org/10.1006/jcss.2000.1727.
On the complexity
62(2):367–375, 2001.
of kURL
[24] Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly
exponential complexity? Journal of Computer and System Sciences, 63(4):512–530, 2001.
[25] Donald B. Johnson.
Efficient algorithms for shortest paths in sparse networks.
Journal of the ACM, 24(1):1–13, 1977.
doi:
10.1145/321992.321993.
URL
http://doi.acm.org/10.1145/321992.321993.
[26] Jure Leskovec and Eric Horvitz. Planetary-scale views on a large instant-messaging network.
In Proceedings of the 17th International World Wide Web Conference (WWW ’08), pages
915–924. ACM, 2008. ISBN 978-1-60558-085-2.
[27] Stanley Milgram. The small world problem. Psychology Today, 1:61–67, 1967.
[28] M. E. J. Newman. The structure and function of complex networks. SIAM Review, 45(2):
167–256, 2003.
[29] M. E. J. Newman. Networks: An Introduction. Oxford University Press, 2010.
[30] M. E. J. Newman and Juyong Park. Why social networks are different from other types of
networks. Physical Review E, 68(3):036122, 2003.
15
[31] Stephan Olariu. A simple linear-time algorithm for computing the center of an interval graph. International Journal of Computer Mathematics, 34(3-4):121–128, 1990. URL
http://dx.doi.org/10.1080/00207169008803870.
[32] Liam Roditty and Virginia Vassilevska Williams. Fast approximation algorithms for
the diameter and radius of sparse graphs.
In Proceedings of the 45th Symposium
on Theory of Computing Conference (STOC ’13), pages 515–524. ACM, 2013. URL
http://dx.doi.org/10.1145/2488608.2488673.
[33] Raimund Seidel.
On the all-pairs-shortest-path problem in unweighted undirected
graphs.
Journal of Computer and System Sciences, 51(3):400–403, 1995.
URL
http://dx.doi.org/10.1006/jcss.1995.1078.
[34] Manuel Sorge and Mathias Weller. The graph parameter hierarchy. Manuscript, 2013. URL
http://fpt.akt.tu-berlin.de/msorge/parameter-hierarchy.pdf.
[35] Alan P. Sprague and Tadao Takaoka. O(1) query time algorithm for all pairs shortest distances
on interval graphs. International Journal of Foundations of Computer Science, 10(4):465–472,
1999. URL http://dx.doi.org/10.1142/S0129054199000320.
[36] Oren Weimann and Raphael Yuster. Approximating the diameter of planar graphs in near
linear time. In Proceedings of the 40th International Colloquium on Automata, Languages,
and Programming (ICALP ’13), volume 7965 of LNCS, pages 828–839. Springer, 2013. URL
http://dx.doi.org/10.1007/978-3-642-39206-1_70.
16
| 8cs.DS
|
Instruction Sequence Notations
with Probabilistic Instructions
arXiv:0906.3083v2 [cs.PL] 1 Oct 2014
J.A. Bergstra and C.A. Middelburg
Informatics Institute, Faculty of Science, University of Amsterdam,
Science Park 107, 1098 XG Amsterdam, the Netherlands
[email protected],[email protected]
Abstract. This paper concerns instruction sequences that contain probabilistic instructions, i.e. instructions that are themselves probabilistic by
nature. We propose several kinds of probabilistic instructions, provide an
informal operational meaning for each of them, and discuss related work.
On purpose, we refrain from providing an ad hoc formal meaning for the
proposed kinds of instructions. We also discuss the approach of projection semantics, which was introduced in earlier work on instruction
sequences, in the light of probabilistic instruction sequences.
Keywords: instruction sequence, probabilistic instruction, projection semantics.
1998 ACM Computing Classification: D.1.4, F.1.1, F.1.2.
1
Introduction
In this paper, we take the first step on a new subject in a line of research whose
working hypothesis is that the notion of an instruction sequence is relevant
to diverse subjects in computer science (see e.g. [12,13,14,15]). In this line of
research, an instruction sequence under execution is considered to produce a
behaviour to be controlled by some execution environment: each step performed
actuates the processing of an instruction by the execution environment and a
reply returned at completion of the processing determines how the behaviour
proceeds. The term service is used for a component of a system that provides an
execution environment for instruction sequences, and a model of systems that
provide execution environments for instruction sequences is called an execution
architecture. This paper is concerned with probabilistic instruction sequences.
We use the term probabilistic instruction sequence for an instruction sequence that contains probabilistic instructions, i.e. instructions that are themselves probabilistic by nature, rather than an instruction sequence of which the
instructions are intended to be processed in a probabilistic way. We will propose
several kinds of probabilistic instructions, provide an informal operational meaning for each of them, and discuss related work. We will refrain from a formal
semantic analysis of the proposed kinds of probabilistic instructions. Moreover,
we will not claim any form of completeness for the proposed kinds of probabilistic instructions. Other convincing kinds might be found in the future. We
will leave unanalysed the topic of probabilistic instruction sequence processing,
which includes all phenomena concerning services and execution architectures
for which probabilistic analysis is necessary.
Viewed from the perspective of machine-execution, execution of a probabilistic instruction sequence using an execution architecture without probabilistic
features can only be a metaphor. Execution of a deterministic instruction sequence using an execution architecture with probabilistic features, i.e. an execution architecture that allows for probabilistic services, is far more plausible. Thus,
it looks to be that probabilistic instruction sequences find their true meaning
by translation into deterministic instruction sequences for execution architectures with probabilistic features. Indeed projection semantics, the approach to
define the meaning of programs which was first presented in [9], need not be
compromised when probabilistic instructions are taken into account.
This paper is organized as follows. First, we set out the scope of the paper
(Section 2) and review the special notation and terminology used in the paper
(Section 3). Next, we propose several kinds of probabilistic instructions (Sections 4 and 5). Following this, we formulate a thesis on the behaviours produced
by probabilistic instruction sequences under execution (Section 6) and discuss
the approach of projection semantics in the light of probabilistic instruction sequences (Section 7). We also discuss related work (Section 8) and make some
concluding remarks (Section 10).
In the current version of this paper, we moreover mention some outcomes of
a sequel to the work reported upon in this paper (Section 9).
2
On the Scope of this Paper
We go into the scope of the paper and clarify its restrictions by giving the
original motives. However, the first version of this paper triggered off work on the
behaviour of probabilistic instruction sequences under execution and outcomes
of that work invalidated some of the arguments used to motivate the restrictions.
The relevant outcomes of the work concerned will be mentioned in Section 9.
We will propose several kinds of probabilistic instructions, chosen because of
their superficial similarity with kinds of deterministic instructions known from
PGA (ProGram Algebra) [9], PGLD (ProGramming Language D) with indirect
jumps [10], C (Code) [19], and other similar notations and not because any computational intuition about them is known or assumed. For each of these kinds,
we will provide an informal operational meaning. Moreover, we will exemplify
the possibility that the proposed unbounded probabilistic jump instructions are
simulated by means of bounded probabilistic test instructions and bounded deterministic jump instructions. We will also refer to related work that introduces
something similar to what we call a probabilistic instruction and connect the proposed kinds of probabilistic instructions with similar features found in related
work.
2
We will refrain from a formal semantic analysis of the proposed kinds of
probabilistic instructions. When we started with the work reported upon in this
paper, the reasons for doing so were as follows:
– In the non-probabilistic case, the subject reduces to the semantics of PGA.
Although it seems obvious at first sight, different models, reflecting different
levels of abstraction, can and have been distinguished (see e.g. [9]). Probabilities introduce a further ramification.
– What we consider sensible is to analyse this double ramification fully. What
we consider less useful is to provide one specific collection of design decisions
and working out its details as a proof of concept.
– We notice that for process algebra the ramification of semantic options after
the incorporation of probabilistic features is remarkable, and even frustrating
(see e.g. [24,27]). There is no reason to expect that the situation is much
simpler here.
– Once that a semantic strategy is mainly judged on its preparedness for a
setting with multi-threading, the subject becomes intrinsically complex –
like the preparedness for a setting with arbitrary interleaving complicates
the semantic modeling of deterministic processes in process algebra.
– We believe that a choice for a catalogue of kinds of probabilistic instructions
can be made beforehand. Even if that choice will turn out to be wrong,
because prolonged forthcoming semantic analysis may give rise to new, more
natural, kinds of probabilistic instructions, it can at this stage best be driven
by direct intuitions.
In this paper, we will leave unanalysed the topic of probabilistic instruction
sequence processing, i.e. probabilistic processing of instruction sequences, which
includes all phenomena concerning services and concerning execution architectures for which probabilistic analysis is necessary. At the same time, we admit
that probabilistic instruction sequence processing is a much more substantial
topic than probabilistic instruction sequences, because of its machine-oriented
scope. We take the line that a probabilistic instruction sequence finds its operational meaning by translation into a deterministic instruction sequence and
execution using an execution architecture with probabilistic features.
3
Preliminaries
In the remainder of this paper, we will use the notation and terminology regarding instructions and instruction sequences from PGA. The mathematical
structure that we will use for quantities is a signed cancellation meadow. That
is why we briefly review PGA and signed cancellation meadows in this section.
In PGA, it is assumed that a fixed but arbitrary set of basic instructions has
been given. The primitive instructions of PGA are the basic instructions and in
addition:
– for each basic instruction a, a positive test instruction +a;
3
– for each basic instruction a, a negative test instruction −a;
– for each natural number l, a forward jump instruction #l;
– a termination instruction !.
The intuition is that the execution of a basic instruction a produces either T
or F at its completion. In the case of a positive test instruction +a, a is executed
and execution proceeds with the next primitive instruction if T is produced.
Otherwise, the next primitive instruction is skipped and execution proceeds with
the primitive instruction following the skipped one. If there is no next instruction to be executed, execution becomes inactive. In the case of a negative test
instruction −a, the role of the value produced is reversed. In the case of a plain
basic instruction a, execution always proceeds as if T is produced. The effect
of a forward jump instruction #l is that execution proceeds with the l-th next
instruction. If l equals 0 or the l-th next instruction does not exist, execution
becomes inactive. The effect of the termination instruction ! is that execution
terminates.
The constants of PGA are the primitive instructions and the operators of
PGA are:
– the binary concatenation operator ; ;
– the unary repetition operator ω .
Terms are built as usual. We use infix notation for the concatenation operator
and postfix notation for the repetition operator.
A closed PGA term is considered to denote a non-empty, finite or periodic
infinite sequence of primitive instructions.1 Closed PGA terms are considered
equal if they denote the same instruction sequence. The axioms for instruction
sequence equivalence are given in [9]. The unfolding equation X ω = X ; X ω is
derivable from those equations. Moreover, each closed PGA term is derivably
equal to one of the form P or P ; Qω , where P and Q are closed PGA terms in
which the repetition operator does not occur.
In [9], PGA is extended with a unit instruction operator u which turns
sequences of instructions into single instructions. The result is called PGAu .
In [35], the meaning of PGAu programs is described by a translation from PGAu
programs into PGA programs.
In the sequel, the following additional assumption is made: a fixed but arbitrary set of foci and a fixed but arbitrary set of methods have been given.
Moreover, we will use f.m, where f is a focus and m is a method, as a general
notation for basic instructions. In f.m, m is the instruction proper and f is the
name of the service that is designated to process m.
The signature of signed cancellation meadows consists of the following constants and operators:
– the constants 0 and 1;
– the binary addition operator + ;
1
A periodic infinite sequence is an infinite sequence with only finitely many distinct
suffixes.
4
–
–
–
–
the
the
the
the
binary multiplication operator · ;
unary additive inverse operator −;
unary multiplicative inverse operator
unary signum operator s.
−1
;
Terms are build as usual. We use infix notation for the binary operators
+ and · , prefix notation for the unary operator −, and postfix notation for
the unary operator −1 . We use the usual precedence convention to reduce the
need for parentheses. We introduce subtraction and division as abbreviations:
p − q abbreviates p + (−q) and p/q abbreviates p · (q −1 ). We use the notation n
for numerals and the notation pn for exponentiation with a natural number as
exponent. The term n is inductively defined as follows: 0 = 0 and n + 1 = n + 1.
The term pn is inductively defined as follows: p0 = 1 and pn+1 = pn · p.
The constants and operators from the signature of signed cancellation meadows are adopted from rational arithmetic, which gives an appropriate intuition
about these constants and operators. The equational theories of signed cancellation meadows is given in [8]. In signed cancellation meadows, the functions min
and max have a simple definition (see also [8]).
A signed cancellation meadow is a cancellation meadow expanded with a
signum operation. The prime example of cancellation meadows is the field of rational numbers with the multiplicative inverse operation made total by imposing
that the multiplicative inverse of zero is zero, see e.g. [20].
4
Probabilistic Basic and Test Instructions
In this section, we propose several kinds of probabilistic basic and test instructions. It is assumed that a fixed but arbitrary signed cancellation meadow M
has been given. Moreover, we write q̂, where q ∈ M, for max(0, min(1, q)).
We propose the following probabilistic basic instructions:
– %(), which produces T with probability 1/2 and F with probability 1/2;
– %(q), which produces T with probability q̂ and F with probability 1 − q̂,
for q ∈ M.
The probabilistic basic instructions have no side-effect on a state.
The basic instruction %() can be looked upon as a shorthand for %(1/2). We
distinguish between %() and %(1/2) for reason of putting the emphasis on the
fact that it is not necessary to bring in a notation for quantities ranging from 0
to 1 in order to design probabilistic instructions.
Once that probabilistic basic instructions of the form %(q) are chosen, an
unbounded ramification of options for the notation of quantities is opened up. We
will assume that closed terms over the signature of signed
√ cancellation meadows
are used to denote quantities. Instructions such as %( 1 + 1) are implicit in the
√
form %(q), assuming that it is known how to view
as a notational extension
of signed cancellation meadows (see e.g. [8]).
Like all basic instructions, each probabilistic basic instruction give rise to
two probabilistic test instructions:
5
– %() gives rise to +%() and −%();
– %(q) gives rise to +%(q) and −%(q).
Probabilistic primitive instructions of the form +%(q) and −%(q) can be considered probabilistic branch instructions where q is the probability that the branch
is not taken and taken, respectively, and likewise the probabilistic primitive instructions +%() and −%().
We find that the primitive instructions %() and %(q) can be replaced by #1
without loss of (intuitive) meaning. Of course, in a resource aware model, #1
may be much cheaper than %(q), especially if q is hard to compute. Suppose
that %(q) is realized at a lower level by means of %(), which is possible, and
suppose that q is a computable real number. The question arises whether the
expectation of the time to execute %(q) is finite.
To exemplify the possibility that %(q) is realized by means of %() in the case
where q is a rational number, we look at the following probabilistic instruction
sequences:
−%(2/3) ; #3 ; a ; ! ; b ; ! ,
(+%() ; #3 ; a ; ! ; +%() ; #3 ; b ; !)ω .
It is easy to see that these instruction sequences produce on execution the same
behaviour: with probability 2/3, first a is performed and then termination follows; and with probability 1/3, first b is performed and then termination follows.
In the case of computable real numbers other than rational numbers, use must
be made of a service that does duty for the tape of a Turing machine (such a
service is, for example, described in [18]).
Let q ∈ M, and let random(q) be a service with a method get whose reply is T with probability q̂ and F with probability 1 − q̂. Then a reasonable
view on the meaning of the probabilistic primitive instructions %(q), +%(q) and
−%(q) is that they are translated into the deterministic primitive instructions
random(q).get, +random(q).get and −random(q).get, respectively, and executed
using an execution architecture that provides the probabilistic service random(q).
Another option is possible here: instead of a different service random(q) for each
q ∈ [0, 1] and a single method get, we could have a single service random with
a different method get(q) for each q ∈ [0, 1]. In the latter case, %(q), +%(q)
and −%(q) would be translated into the deterministic primitive instructions
random.get(q), +random.get(q) and −random.get(q).
5
Probabilistic Jump Instructions
In this section, we propose several kinds of probabilistic jump instructions. It
is assumed that the signed cancellation meadow M has been expanded with an
operation N such that, for all q ∈ M, N(q) = 0 iff q = n for some n ∈ N.
We write l̄, where l ∈ M is such that N(l) = 0, for the unique n ∈ N such that
l = n.
We propose the following probabilistic jump instructions:
6
– #%U(k), having the same effect as #j with probability 1/k for j ∈ [1, k̄],
for k ∈ M with N(k) = 0;
– #%G(q)(k), having the same effect as #j with probability q̂ · (1 − q̂)j−1 for
j ∈ [1, k̄], for q ∈ M and k ∈ M with N(k) = 0;
– #%G(q)l, having the same effect as #l̄ · j with probability q̂ · (1 − q̂)j−1 for
j ∈ [1, ∞), for q ∈ M and l ∈ M with N(l) = 0.
The letter U in #%U(k) indicates a uniform probability distribution, and the
letter G in #%G(q)(k) and #%G(q)l indicates a geometric probability distribution. Instructions of the forms #%U(k) and #%G(q)(k) are bounded probabilistic jump instructions, whereas instructions of the form #%G(q)l are unbounded
probabilistic jump instructions.
Like in the case of the probabilistic basic instructions, we propose in addition
the following probabilistic jump instructions:
– #%G()(k) as the special case of #%G(q)(k) where q = 1/2;
– #%G()l as the special case of #%G(q)l where q = 1/2.
We believe that it must be possible to eliminate all probabilistic jump instructions. In particular, we believe that it must be possible to eliminate all
unbounded probabilistic jump instructions. This belief can be understood as the
judgement that it is reasonable to expect from a semantic model of probabilistic
instruction sequences that the following identity and similar ones hold:
+a ; #%G()2 ; (+b ; ! ; c)ω =
+a ; +%() ; #8 ; #10 ;
(+b ; #5 ; #10 ; +%() ; #8 ; #10 ;
! ; #5 ; #10 ; +%() ; #8 ; #10 ;
c ; #5 ; #10 ; +%() ; #8 ; #10)ω .
Taking this identity and similar ones as our point of departure, the question arises
what is the most simple model that justifies them. A more general question is
whether instruction sequences with unbounded probabilistic jump instructions
can be translated into ones without probabilistic jump instructions provided it
does not bother us that the instruction sequences may become much longer (e.g.
expectation of the length bounded, but worst case length unbounded).
6
The Probabilistic Process Algebra Thesis
In the absence of probabilistic instructions, threads as considered in BTA (Basic
Thread Algebra) [9] or its extension with thread-service interaction [17] can
be used to model the behaviours produced by instruction sequences under execution.2 Processes as considered in general process algebras such as ACP [6],
2
In [9], BTA is introduced under the name BPPA (Basic Polarized Process Algebra).
7
CCS [32] and CSP [26] can be used as well, but they give rise to a more complicated modeling of the behaviours of instruction sequences under execution (see
e.g. [13]).
In the presence of probabilistic instructions, we would need a probabilistic
thread algebra, i.e. a variant of BTA or its extension with thread-service interaction that covers probabilistic behaviours. When we started with the work
reported upon in this paper, it appeared that any probabilistic thread algebra is
inherently more complicated to such an extent that the advantage of not using a
general process algebra evaporates. Moreover, it appeared that any probabilistic
thread algebra requires justification by means of an appropriate probabilistic
process algebra. This led us to the following thesis:
Thesis. Modeling the behaviours produced by probabilistic instruction sequences
under execution is a matter of using directly processes as considered in some
probabilistic process algebra.
A probabilistic thread algebra has to cover the interaction between instruction sequence behaviours and services. Two mechanisms are involved in that.
They are called the use mechanism and the apply mechanism (see e.g. [17]).
The difference between them is a matter of perspective: the former is concerned
with the effect of services on behaviours of instruction sequences and therefore
produces behaviours, whereas the latter is concerned with the effect of instruction sequence behaviours on services and therefore produces services. When we
started with the work reported upon in this paper, it appeared that the use
mechanism would make the development of a probabilistic thread algebra very
intricate.
The first version of this paper triggered off work on the behaviour of probabilistic instruction sequences under execution by which the thesis stated above
is refuted. The ultimate point is that meanwhile an appropriate and relatively
simple probabilistic thread algebra has been devised (see [16]). Moreover, our
original expectations about probabilistic process algebras turned out to be too
high.
The first probabilistic process algebra is presented in [23] and the first probabilistic process algebra with an asynchronous parallel composition operator is
presented in [5]. A recent overview of the work on probabilistic process algebras
after that is given in [29]. This overview shows that the multitude of semantic
ideas applied and the multitude of variants of certain operators devised have
kept growing, and that convergence is far away. In other words, there is little well-established yet. In particular, for modeling the behaviours produced by
probabilistic instruction sequences, we need operators for probabilistic choice,
asynchronous parallel composition, and abstraction from internal actions. For
this case, the attempts by one research group during about ten years to develop a satisfactory ACP-like process algebra (see e.g. [1,2,3]) have finally led
to a promising process algebra. However, it is not yet clear whether the process
algebra concerned will become well-established.
8
All this means that a justification of the above-mentioned probabilistic thread
algebra by means of an appropriate probabilistic process algebra will be of a
provisional nature for the time being.
7
Discussion of Projectionism
Notice that once we move from deterministic instructions to probabilistic instructions, instruction sequence becomes an indispensable concept. Instruction
sequences cannot be replaced by threads or processes without taking potentially
premature design decisions. In preceding sections, however, we have outlined
how instruction sequences with the different kinds of probabilistic instructions
can be translated into instruction sequences without them. Therefore, it is a
reasonable to claim that, like for deterministic instruction sequence notations,
all probabilistic instruction sequence notations can be provided with a probabilistic semantics by translation of the instruction sequences concerned into
appropriate single-pass instruction sequences. Thus, we have made it plausible
that projectionism is feasible for probabilistic instruction sequences.
Projectionism is the point of view that:
– any instruction sequence P , and more general even any program P , first and
for all represents a single-pass instruction sequence as considered in PGA;
– this single-pass instruction sequence, found by a translation called a projection, represents in a natural and preferred way what is supposed to take
place on execution of P ;
– PGA provides the preferred notation for single-pass instruction sequences.
In a rigid form, as in [9], projectionism provides a definition of what constitutes
a program.
The fact that projectionism is feasible for probabilistic instruction sequences,
does not imply that it is uncomplicated. To give an idea of the complications
that may arise, we will sketch below found challenges for projectionism.
First, we introduce some special notation. Let N be a program notation. Then
we write N2pga for the projection function that gives, for each program P in N,
the closed PGA terms that denotes the single-pass instruction sequence that
produces on execution the same behaviour as P .
We have found the following challenges for projectionism:
– Explosion of size. If N2pga(P ) is much longer than P , then the requirement
that it represents in a natural way what is supposed to take place on execution of P is challenged. For example, if the primitive instructions of N include
instructions to set and test up to n Boolean registers, then the projection to
N2pga(P ) may give rise to a combinatorial explosion of size. In such cases,
the usual compromise is to permit single-pass instruction sequences to make
use of services (see e.g. [17]).
– Degradation of performance. If N2pga(P )’s natural execution is much slower
than P ’s execution, supposing a clear operational understanding of P , then
9
the requirement that it represents in a natural way what is supposed to
take place on execution of P is challenged. For example, if the primitive
instructions of N include indirect jump instructions, then the projection to
N2pga(P ) may give rise to a degradation of performance (see e.g. [10]).
– Incompatibility of services. If N2pga(P ) has to make use of services that are
not deterministic, then the requirement that it represents in a natural way
what is supposed to take place on execution of P is challenged. For example,
if the primitive instructions of N include instructions of the form +%(q) or
−%(q), then P cannot be projected to a single-pass instruction sequence
without the use of probabilistic services. In this case, either probabilistic
services must be permitted or probabilistic instruction sequences must not
be considered programs.
– Complexity of projection description. The description of N2pga may be so
complex that it defeats N2pga(P )’s purpose of being a natural explanation of
what is supposed to take place on execution of P . For example, the projection semantics given for recursion in [7] suffers from this kind of complexity
when compared with the conventional denotational semantics. In such cases,
projectionism may be maintained conceptually, but rejected pragmatically.
– Aesthetic degradation. In N2pga(P ), something elegant may have been replaced by nasty details. For example, if N provides guarded commands, then
N2pga(P ), which will be much more detailed, might be considered to exhibit
signs of aesthetic degradation. This challenge is probably the most serious
one, provided we accept that such elegant features belong to program notations. Of course, it may be decided to ignore aesthetic criteria altogether.
However, more often than not, they have both conceptual and pragmatic
importance.
One might be of the opinion that conceptual projectionism can accept explosion of size and/or degradation of performance. We do not share this opinion:
both challenges require a more drastic response than a mere shift from a pragmatic to a conceptual understanding of projectionism. This drastic response may
include viewing certain mechanisms as intrinsically indispensable for either execution performance or program compactness. For example, it is reasonable to
consider the basic instructions of the form %(q), where q is a computable real
number, indispensable if the expectations of the times to execute their realizations by means of %() are not all finite.
Nevertheless, projectionism looks to be reasonable for probabilistic programs:
they can be projected adequately to deterministic single-pass instruction sequences for an execution architecture with probabilistic services.
8
Related Work
In [38], a notation for probabilistic programs is introduced in which we can write,
for example, random(p·δ0 +q·δ1 ). In general, random(λ) produces a value according to the probability distribution λ. In this case, δi is the probability distribution
that gives probability 1 to i and probability 0 to other values. Thus, for p+q = 1,
10
p·δ0 +q·δ1 is the probability distribution that gives probability p to 0, probability
q to 1, and probability 0 to other values. Clearly, random(p·δ0 +q·δ1 ) corresponds
to %(p). Moreover, using this kind of notation, we could write #( k1 ·(δ1 +· · ·+δk̄ ))
for #%U(k) and #(q̂ · δ1 + q̂ · (1 − q̂) · δ2 + · · · + q̂ · (1 − q̂)k−1 · δk ) for #%G(q)(k).
In much work on probabilistic programming, see e.g. [25,30,33], we find the
binary probabilistic choice operator p ⊕ (for p ∈ [0, 1]). This operator chooses
between its operands, taking its left operand with probability p. Clearly, P p ⊕ Q
can be taken as abbreviations for +%(p);u(P ;#2);u(Q). This kind of primitives
dates back to [28] at least.
Quite related, but from a different perspective, is the toss primitive introduced in [21]. The intuition is that toss(bm, p) assigns to the Boolean memory
cell bm the value T with probability p̂ and the value F with probability 1− p̂. This
means that toss(bm, p) has a side-effect on a state, which we understand as making use of a service. In other words, toss(bm, p) corresponds to a deterministic
instruction intended to be processed by a probabilistic service.
Common in probabilistic programming are assignments of values randomly
chosen from some interval of natural numbers to program variables (see e.g. [37]).
Clearly, such random assignments correspond also to deterministic instructions
intended to be processed by probabilistic services. Suppose that x=i is a primitive instruction for assigning value i to program variable x. Then we can write:
#%U(k) ; u(x=1 ; #k) ; u(x=2 ; #k−1) ; . . . ; u(x=k ; #1). This is a realistic representation of the assignment to x of a value randomly chosen from {1, . . . , k}.
However, it is clear that this way of representing random assignments leads to
an exponential blow up in the size of any concrete instruction sequence representation, provided the concrete representation of k is its decimal representation.
The refinement oriented theory of programs uses demonic choice, usually
written ⊓, as a primitive (see e.g. [30,31]). A demonic choice can be regarded as a
probabilistic choice with unknown probabilities. Demonic choice could be written
+⊓ in a PGA-like notation. However, a primitive instruction corresponding to
demonic choice is not reasonable: no mechanism for the execution of +⊓ is
conceivable. Demonic choice exists in the world of specifications, but not in the
world of instruction sequences. This is definitely different with +%(p), because
a mechanism for its execution is conceivable.
Features similar to probabilistic jump instructions are not common in probabilistic programming. To our knowledge, the mention of probabilistic goto statements of the form pr goto {l1 , l2 } in [4] is the only mention of a similar feature
in the computer science literature. The intuition is that pr goto {l1 , l2 }, where
l1 and l2 are labels, has the same effect as goto l1 with probability 1/2 and has
the same effect as goto l2 with probability 1/2. Clearly, this corresponds to a
probabilistic jump instruction of the form #%U(1/2).
It appears that quantum computing has something to offer that cannot be
obtained by conventional computing: it makes a stateless generator of random
bits available (see e.g. [22,34]). By that quantum computing indeed provides a
justification of +%(1/2) as a probabilistic instruction.
11
9
On the Sequel to this Paper
The first version of this paper triggered off work on the behaviour of probabilistic
instruction sequences under execution and outcomes of that work invalidated
some of the arguments used to motivate the restricted scope of this paper. In
this section, we mention the relevant outcomes of that work.
After the first version of this paper (i.e. [11]) appeared, it was found that:
– The different levels of abstraction that have been distinguished in the nonprobabilistic case can be distinguished in the probabilistic case as well and
only the models at the level of the behaviour of instruction sequences under
execution are not essentially the same.
– The semantic options at the behavioural level after the incorporation of
probabilistic features are limited because of (a) the orientation towards behaviours of a special kind and (b) the semantic constraints induced by the
informal explanations of the enumerated kinds of probabilistic instructions
and the desired elimination property of all but one kind.
– For the same reasons, preparedness for a setting with multi-threading does
not really complicate matters.
The current state of affairs is as follows:
– BTA and its extension with thread-service interaction, which are used to
describe the behaviour of instruction sequences under execution in the nonprobabilistic case, have been extended with probabilistic features in [16].
– In [16], we have added the probabilistic basic and test instructions proposed in Section 4 to PGLB (ProGramming Language B), an instruction
sequence notation rooted in PGA and close to existing assembly languages,
and have given a formal definition of the behaviours produced by the instruction sequences from the resulting instruction sequence notation in terms of
non-probabilistic instructions and probabilistic services.
– The bounded probabilistic jump instructions proposed in Section 5 can be
given a behavioural semantics in the same vein. The unbounded probabilistic
jump instructions fail to have a natural behavioural semantics in the setting
of PGA because infinite instruction sequences are restricted to eventually
periodic ones.
– The extension of BTA with multi-threading, has been generalized to probabilistic multi-threading in [16] as well.
10
Conclusions
We have made a notational proposal of probabilistic instructions with an informal semantics. By that we have contrasted probabilistic instructions in an execution architecture with deterministic services with deterministic instructions in
an execution architecture with partly probabilistic services. The history of the
proposed kinds of instructions can be traced.
12
We have refrained from an ad hoc formal semantic analysis of the proposed
kinds of instructions. There are many solid semantic options, so many and so
complex that another more distant analysis is necessary in advance to create a
clear framework for the semantic analysis in question.
The grounds of this work are our conceptions of what a theory of probabilistic
instruction sequences and a complementary theory of probabilistic instruction
sequence processing (i.e. execution architectures with probabilistic services) will
lead to:
– comprehensible explanations of relevant probabilistic algorithms, such as the
Miller-Rabin probabilistic primality test [36], with precise descriptions of the
kinds of instructions and services involved in them;
– a solid account of pseudo-random Boolean values and pseudo-random numbers;
– a thorough exposition of the different semantic options for probabilistic instruction sequences;
– explanations of relevant quantum algorithms, such as Shor’s integer factorization algorithm [39], by first giving a clarifying analysis in terms of probabilistic instruction sequences or execution architectures with probabilistic
services and only then showing how certain services in principle can be realized very efficiently with quantum computing.
Projectionism looks to be reasonable for probabilistic programs: they can be
projected adequately to deterministic single-pass instruction sequences for an
execution architecture with appropriate probabilistic services. At present, it is
not entirely clear whether this extends to quantum programs.
References
1. Andova, S., Baeten, J.C.M.: Abstraction in probabilistic process algebra. In: Margaria, T., Yi, W. (eds.) TACAS 2001. Lecture Notes in Computer Science, vol.
2031, pp. 204–219. Springer-Verlag (2001)
2. Andova, S., Baeten, J.C.M., Willemse, T.A.C.: A complete axiomatisation of
branching bisimulation for probabilistic systems with an application in protocol
verification. In: Baier, C., Hermans, H. (eds.) CONCUR 2006. Lecture Notes in
Computer Science, vol. 4137, pp. 327–342. Springer-Verlag (2006)
3. Andova, S., Georgievska, S.: On compositionality, efficiency, and applicability of
abstraction in probabilistic systems. In: Nielsen, M., et al. (eds.) SOFSEM 2009.
Lecture Notes in Computer Science, vol. 5404, pp. 67–78. Springer-Verlag (2009)
4. Arons, T., Pnueli, A., Zuck, L.: Parameterized verification by probabilistic abstraction. In: Gordon, A.D. (ed.) FOSSACS 2003. Lecture Notes in Computer Science,
vol. 2620, pp. 87–102. Springer-Verlag (2003)
5. Baeten, J.C.M., Bergstra, J.A., Smolka, S.A.: Axiomatizing probabilistic processes:
ACP with generative probabilities. Information and Computation 121(2), 234–255
(1995)
6. Baeten, J.C.M., Weijland, W.P.: Process Algebra, Cambridge Tracts in Theoretical
Computer Science, vol. 18. Cambridge University Press, Cambridge (1990)
13
7. Bergstra, J.A., Bethke, I.: Predictable and reliable program code: Virtual machine
based projection semantics. In: Bergstra, J.A., Burgess, M. (eds.) Handbook of
Network and Systems Administration, pp. 653–685. Elsevier, Amsterdam (2007)
8. Bergstra, J.A., Bethke, I., Ponse, A.: Cancellation meadows: A generic basis theorem and some applications. Computer Journal 56(1), 3–14 (2013)
9. Bergstra, J.A., Loots, M.E.: Program algebra for sequential code. Journal of Logic
and Algebraic Programming 51(2), 125–156 (2002)
10. Bergstra, J.A., Middelburg, C.A.: Instruction sequences with indirect jumps. Scientific Annals of Computer Science 17, 19–46 (2007)
11. Bergstra, J.A., Middelburg, C.A.: Instruction sequence notations with probabilistic
instructions. arXiv:0906.3083v1 [cs.PL] (June 2009)
12. Bergstra, J.A., Middelburg, C.A.: Instruction sequence processing operators. Acta
Informatica 49(3), 139–172 (2012)
13. Bergstra, J.A., Middelburg, C.A.: On the behaviours produced by instruction sequences under execution. Fundamenta Informaticae 120(2), 111–144 (2012)
14. Bergstra, J.A., Middelburg, C.A.: On the expressiveness of single-pass instruction
sequences. Theory of Computing Systems 50(2), 313–328 (2012)
15. Bergstra, J.A., Middelburg, C.A.: Instruction sequence based non-uniform complexity classes. Scientific Annals of Computer Science 24(1), 47–89 (2014)
16. Bergstra, J.A., Middelburg, C.A.: A thread algebra with probabilistic features.
arXiv:1409.6873v1 [cs.LO] (September 2014)
17. Bergstra, J.A., Ponse, A.: Combining programs and state machines. Journal of
Logic and Algebraic Programming 51(2), 175–192 (2002)
18. Bergstra, J.A., Ponse, A.: Execution architectures for program algebra. Journal of
Applied Logic 5(1), 170–192 (2007)
19. Bergstra, J.A., Ponse, A.: An instruction sequence semigroup with involutive antiautomorphisms. Scientific Annals of Computer Science 19, 57–92 (2009)
20. Bergstra, J.A., Tucker, J.V.: The rational numbers as an abstract data type. Journal of the ACM 54(2), Article 7 (2007)
21. Chadha, R., Cruz-Filipe, L., Mateus, P., Sernadas, A.: Reasoning about probabilistic sequential programs. Theoretical Computer Science 379(1–2), 142–165 (2007)
22. Gay, S.J.: Quantum programming languages: Survey and bibliography. Mathematical Structures in Computer Science 16(4), 581–600 (2006)
23. Giacalone, A., Jou, C.C., Smolka, S.A.: Algebraic reasoning for probabilistic concurrent systems. In: Proceedings IFIP TC2 Working Conference on Programming
Concepts and Methods. pp. 443–458. North-Holland (1990)
24. van Glabbeek, R.J., Smolka, S.A., Steffen, B.: Reactive, generative and stratified models of probabilistic processes. Information and Computation 121(1), 59–80
(1995)
25. He Jifeng, Seidel, K., McIver, A.K.: Probabilistic models for the guarded command
language. Science of Computer Programming 28(2–3), 171–192 (1997)
26. Hoare, C.A.R.: Communicating Sequential Processes. Prentice-Hall, Englewood
Cliffs (1985)
27. Jonsson, B., Larsen, K.G., Yi, W.: Probabilistic extensions of process algebras. In:
Bergstra, J.A., Ponse, A., Smolka, S.A. (eds.) Handbook of Process Algebra, pp.
685–710. Elsevier, Amsterdam (2001)
28. Kozen, D.: A probabilistic PDL. Journal of Computer and System Sciences 30(2),
162–178 (1985)
29. López, N., Núñez, M.: An overview of probabilistic process algebras and their
equivalences. In: Baier, C., et al. (eds.) Validation of Stochastic Systems. Lecture
Notes in Computer Science, vol. 2925, pp. 89–123. Springer-Verlag (2004)
14
30. McIver, A.K., Morgan, C.C.: Demonic, angelic and unbounded probabilistic choices
in sequential programs. Acta Informatica 37(4–5), 329–354 (2001)
31. Meinicke, L., Solin, K.: Refinement algebra for probabilistic programs. Electronic
Notes in Theoretical Computer Science 201, 177–195 (2008)
32. Milner, R.: Communication and Concurrency. Prentice-Hall, Englewood Cliffs
(1989)
33. Morgan, C.C., McIver, A.K., Seidel, K.: Probabilistic predicate transformers. ACM
Transactions on Programming Languages and Systems 18(3), 325–353 (1996)
34. Perdrix, S., Jorrand, P.: Classically controlled quantum computation. Mathematical Structures in Computer Science 16(4), 601–620 (2006)
35. Ponse, A.: Program algebra with unit instruction operators. Journal of Logic and
Algebraic Programming 51(2), 157–174 (2002)
36. Rabin, M.O.: Probabilistic algorithms. In: Traub, J.F. (ed.) Algorithms and Complexity: New Directions and Recent Results, pp. 21–39. Academic Press, New York
(1976)
37. Schöning, U.: A probabilistic algorithm for k-SAT based on limited local search
and restart. Algorithmica 32(4), 615–623 (2002)
38. Sharir, M., Pnueli, A., Hart, S.: Verification of probabilistic programs. SIAM Journal of Computing 13(2), 292–314 (1984)
39. Shor, P.W.: Algorithms for quantum computation: Discrete logarithms and factoring. In: FOCS ’94. pp. 124–134. IEEE Computer Society Press (1994)
15
| 6cs.PL
|
Journal of Artificial Intelligence Research 1 (2015) 3-17
Submitted 6/91; published 9/91
Recursive Feature Generation for Knowledge-based Learning
Lior Friedman
Shaul Markovitch
LIORF @ CS . TECHNION . AC . IL
SHAULM @ CS . TECHNION . AC . IL
arXiv:1802.00050v1 [cs.AI] 31 Jan 2018
Technion-Israel Institute of Technology
Haifa 32000, Israel
Abstract
When humans perform inductive learning, they often enhance the process with background
knowledge. With the increasing availability of well-formed collaborative knowledge bases, the
performance of learning algorithms could be significantly enhanced if a way were found to exploit
these knowledge bases. In this work, we present a novel algorithm for injecting external knowledge
into induction algorithms using feature generation. Given a feature, the algorithm defines a new
learning task over its set of values, and uses the knowledge base to solve the constructed learning
task. The resulting classifier is then used as a new feature for the original problem. We have
applied our algorithm to the domain of text classification using large semantic knowledge bases. We
have shown that the generated features significantly improve the performance of existing learning
algorithms.
1. Introduction
In recent decades, machine learning techniques have become more prevalent in a wide variety of
fields. Most of these methods rely on the inductive approach: they attempt to locate a hypothesis that
is supported by given labeled examples. These methods have proven themselves successful when
the number of examples is sufficient, and a collection of good, distinguishing features is available.
In many real-world applications, however, the given set of features is insufficient for inducing a high
quality classifier (Levi & Weiss, 2004; Paulheim & Fümkranz, 2012).
One approach for overcoming this difficulty is to generate new features that are combinations
of the given ones. For example, the LFC algorithm (Ragavan, Rendell, Shaw, & Tessmer, 1993)
combines binary features through the use of logical operators such as ∧, ¬. Another example is
the LDMT algorithm (Brodley & Utgoff, 1995), which generates linear combinations of existing
features. Deep Learning methods combine basic and generated features using various activation
functions such as sigmoid or softmax. The FICUS framework (Markovitch & Rosenstein, 2002)
presents a general method for generating features using any given set of constructor functions.
The above approaches are limited in that they merely combine existing features to make the
representation more suitable for the learning algorithm. While this approach often suffices, there
are many cases where simply combining existing features is not sufficient. When people perform
inductive learning, they usually rely on a vast body of background knowledge to make the process
more effective (McNamara & Kintsch, 1996). For example, assume two positive examples of people
suffering from some genetic disorder, where the value of the country-of-origin feature is Poland and
Romania. Existing induction algorithms, including those generating features by combination, will
not be able to generalize over these two examples. Humans, on the other hand, can easily generalize
and generate a new feature, is located in Eastern Europe, based on their previously established
background knowledge.
c 2015 AI Access Foundation. All rights reserved.
In this work, we present a novel algorithm that uses a similar approach for enhancing inductive
learning with background knowledge through feature generation. Our method assumes that in addition to the labeled set of example it receives a body of external knowledge represented in relational
form. The algorithm treats feature values as objects and constructs new learning problems, using
the background relational knowledge as their features. The resulting classifiers are then used as
new generated features. For example, in the above very simple example, our algorithm would have
used the continent and region features of a country, inferred from a geographic knowledge base, to
create the new feature that enables us to generalize. One significant advantage of using background
knowledge in the form of generated features is that it allows us to utilize existing powerful learning
algorithms for the induction process.
We have implemented our algorithm in the domain of text classification, using Freebase and
YAGO2 as our background knowledge bases, and performed an extensive set of experiments to test
the effectiveness of our method. Results show that the use of background knowledge through our
methodology significantly improves the performance of existing learning algorithms.
2. Motivation
Before we delve into the detailed description of our algorithm, we would like to illustrate its main
ideas using an example. Suppose we are attempting to identify people with a high risk of suffering
from a genetic disorder. Assume that the target concept to be discovered is that those at risk are
women with ancestors originating from desert areas. To identify women at risk, we are given a
training sample of sick and healthy people, containing various features, including gender and their
full name. We call this learning problem T1 . Assuming we have no additional information, an
induction algorithm (a decision tree learner, in this example) would likely produce a result similar
to that shown in Figure 1. While such a classifier will achieve a low training error, the hundreds of
seemingly unrelated surnames will cause it to generalize poorly.
The above example illustrates a case where, without additional knowledge, an induction algorithm will yield a very poor result. However, if we assume access to a relational knowledge base
connecting surnames to common countries of origin, we can begin to apply our knowledge-based
feature generation techniques to the problem, as we can move from the domain of surnames to that
of countries. Our algorithm does so by creating a new learning problem T2 . The training objects
for learning problem T2 are surnames; surnames of people at risk are labeled as positive. The features for these new objects are extracted from the knowledge base. In this case, we have a single
feature: the country of origin. Solving the above learning problem through an induction algorithm
yields a classifier on surnames that distinguishes between surnames of patients with the disease and
surnames of healthy individuals. This classifier for T2 can then be used as a binary feature for the
original problem T1 by applying it to the feature value of surname. For example, it can be used as a
feature in the node corresponding a gender of female in Figure 1, yielding the tree seen in Figure 2.
This new feature gives us a better generalization over the baseline solution, as we now abstract
the long list of surnames to a short list of countries. This result also allows us to capture previously
unseen surnames from those countries. However, this is not a sufficient solution, as we have no way
of generalizing on previously unseen countries of origin.
If, however, we would have recursively applied our method for solving T2 , we could have obtained a better generalization. When learning to classify surnames, our method creates a new learning problem, T3 , with countries as its objects. Countries of surnames belonging to people with high
2
Gender
Surname
hundreds of surnames
su
e
rn
am
am
rn
e1
-
su
n
-
+
+
-
-
•••
Figure 1: A decision tree for the basic features
surname
Gender
dozens of
countries
Country of
Origin
1
Co
-
-
+
m
Co
un
try
try
un
-
+
+
-
•••
-
+
Figure 2: A constructed feature used within a decision tree
risk are labeled as positive. The knowledge base regarding countries is then used to extract features
for this new training set. Applying a standard learning algorithm to T3 will yield a classifier that
separates between countries of origin of people at risk and those not at risk. This classifier will do so
by looking at the properties of countries, and conclude that countries with high average temperature
and low precipitation, the characteristics of desert areas, are associated with people at high risk.
The result of this process, depicted in Figure 3, is a new feature for T2 , that is, a feature on
surnames. This feature is then used to construct a classifier for T2 , which is in turn used as a
feature for T1 , yielding a feature on patients. This new feature for patients will check whether their
3
Labeled
Countries
of Origin
country
(1)
population
avg. temp
precipit
••••
(2)
training set for recursive learning problem
+
surname
Gender
Average
Temperature
-
F
> 30°c
-
-
+
A classifier that
tries to separate
between
countries of
T
origin of people
with the disease
and those of
Precipitation
people without
the disease
F
Countries
of origin of
people
with the
disease
(4)
labeled
feature
vectors
Countries
of origin of
people
without the
disease
Induction
algorithm
(3)
T
< 250mm
-
+
+
-
Figure 3: Recursive construction of a learning problem on countries of origin. (1) Creating the objects for the new problem. (2) Creating features using the knowledge base. (3) Applying
an induction algorithm. (4) The resulting feature.
surname corresponds with a country of origin that has desert-like characteristics. We see that this
feature allows us to correctly capture the target concept.
3. Generating features through recursive induction
In the following sections, we formally define the feature generation problem, present a solution
in the form of a simple algorithm that generates features by using relational expansions, and then
proceed to describe our main recursive feature generation algorithm.
3.1 Problem definition
We begin our discussion with a standard definition of an induction problem. Let O be a set of
objects. Let Y = {0, 1} be a set of labels1 . Let C : O → Y be a target concept. Let S =
{(o1 , y1 ), . . . , (om , ym )} be a set of labeled examples such that oi ∈ O, yi ∈ Y, C(oi ) = yi . Let
F = {f1 , . . . , fn } be a feature map, a set consisting of feature functions fi : O → Ii where Ii
is the image of fi . This definition implies a training set represented by feature vectors: SF =
{(hf1 (oi ), . . . , fn (oi )i, yi )|(oi , yi ) ∈ S}. A learning algorithm L takes SF as inputs, and outputs a
classifier hSF : O → Y .
Definition 3.1. Let L(S, F ) = hSF be the classifier given as an output by L given hS, F i. Assuming
S ∼ D, the generalization error of a learning algorithm L is the probability P r(hSF (x) 6= y), where
(x, y) ∼ D.
Definition 3.2. A feature generation algorithm A is an algorithm that, given hS, F i, creates a new
feature map F 0 = {f10 , . . . , fk0 }, fi0 : O → Ii .
In order to evaluate the output of a feature generation algorithm A, we must define its utility.
Given hS, F i, A generates a feature set FA0 . Let S be a training set representative of the true
1. We assume binary labels for ease of discussion.
4
distribution of the target concept. Given S , a feature set F , a generated feature set FA0 and a
learning algorithm L, the utility of A is U (A(S, F )) = P r(hSF (x) 6= y) − P r(hSF 0 (x) 6= y),
A
where x ∈ O, y = C(x).
Thus, in order for the utility of A to be positive, the generated feature set FA0 must yield a lower
generalization error than the original feature map F .
In this work, we assume that, in addition to SF , A also has access to a set of binary2 relations
R = {R1 , . . . , Rt }, Rj : Dj × Dj 0 representing our knowledge base. For each individual relation
Rj , its set of departure is marked Dj and its co-domain is denoted as Dj 0 .
Definition 3.3. A knowledge-based feature-generation algorithm A is an algorithm that, given
hS, F, Ri, creates a new feature map FR = {f10 , . . . , fk0 }, fi0 : O → Ii .
3.2 Expansion-based feature generation
In this section, we present our first method for knowledge-based feature generation that ignores
the labels of the original learning problem. The algorithm extends each feature value by all of
the tuples it appears in. Let fi be an original feature. Let Rj : Dj × Dj 0 be a relation such that
Image(fi ) ⊆ Dj . We can generate a new feature function fi,j : O → Dj 0 by composing Rj onto
fi , yielding our new feature function fi,j (x) = Rj ◦ fi .
In the general case, composing Rj onto Fi yields a set of values, meaning fi,j (x) = {v ∈
0
Dj |(fi (x), v) ∈ Rj }. In our work, we preferred to work on singular values rather than set-based
features. To do so, we use aggregation functions. For the experiments described in this paper, we use
two aggregation function types: Majority and Any, but in general any other reasonable aggregation
function can be used instead. The Majority aggregator, for example, is defined as follows. For
each value v ∈ Dj 0 , we generate a binary feature function with value 1 only if v is the majority
value: M ajority v (X) = 1 ⇐⇒ majority(X) = v. The pseudo-code of this algorithm (called
Expander-FG) is listed in Algorithm 1.
Algorithm 1 Expander-FG
σ - An aggregation function family
function G ENERATE F EATURES(S,F , R)
generated = ∅
for fi ∈ F do
for Rj ∈ R such that Image(fi ) ⊆ Dj do
if Rj is a function then
add {fi,j = Rj ◦ fi } to generated
else
σ
add Fi,j
= {σ v (fi,j )|v ∈ Dj 0 } to generated
. Rj is a relation Rj : Dj × Dj 0
return generated
3.3 Recursive feature generation algorithm
One way to extend the Expander-FG algorithm described in the previous section is to apply it repeatedly to its own output. Extending the algorithm in this fashion, however, would yield an exponential
2. If our relations are not binary, we can use projection to create multiple binary relations instead.
5
increase in the number of generated features. To that end, we propose an alternative knowledgebased feature generation algorithm. Given an input hS, F, Ri, for each feature fi ∈ F, fi : O → Ii ,
our algorithm creates a recursive learning problem whose objects are the values of fi , where values
associated with positive examples in S are labeled positive. Features for this generated problem are
created using the relations in R. Once this new learning problem hSi0 , FR i is defined, an learning
algorithm is used to induce a classifier hi : Ii → Y . Finally, our algorithm outputs a single generated feature for fi , fi0 (x) = hi ◦ fi = hi (fi (x)), fi0 : O → Y . Note that during the induction process
of the newly created learning problem, we can apply a feature generation algorithm on hSi0 , FR , Ri.
In particular, we can apply the above method recursively to create additional features. We call this
algorithm FEAGURE (FEAture Generation Using REcursive induction).
Given a feature fi , we create a recursive learning problem hSi0 , FR i. Let vi (S) = {v|(o, y) ∈
S, fi (o) = v} be the set of feature values for fi in the example set S. We use vi (S) as our set of
objects. To label each v ∈ vi (S), we examine at the labels in the original problem. If there is a
single example o ∈ S such that fi (o) = v, then the label of v will be the label of o. Otherwise, we
take the majority label label(v) = majority({y|(o, y) ∈ S, fi (o) = v}).
To define our learning problem, we must specify a feature map over vi (S). Similarly to ExpanderFG, we use the relations in R on the elements in the new training set Si0 = {(v, label(v))|v ∈
vi (S)}. For each Rj ∈ R, if it is relevant to the problem domain, meaning that vi (S) ⊆ Dj , we
utilize it as a feature by applying it to v. If Rj (v) is a set, we use aggregators, as described in the
previous section. The result of this process is a generated feature map for Si0 , denoted as FR .
We now have a new induction problem hSi0 , FR i. We can further extend FR by recursively using
0 . The depth of recursion is controlled by a parameter
FEAGURE, yielding a new feature map FR
d, that will usually be set according to available learning resources. We proceed to use a learning
0 i in order to train a classifier, giving us h : I → Y . We can then use h on
algorithm3 on hSi0 , FR
i
i
i
objects in S as discussed above, giving us a new feature fi0 (x) = hi (fi (x)), fi0 : O → Y .
The full algorithm is listed in Algorithm 2. While the FEAGURE algorithm can be used as
described above, we found it more useful to use it in the context of a divide & conquer approach, in
a manner similar to the induction of decision trees. In this approach, the set of examples is given as
an input to a decision tree induction algorithm. The FEAGURE algorithm is applied at each node.
This allows us to generate features that are locally useful for a subset of examples. At the end of the
process the tree is discarded and the generated features are gathered as the final output.
3. For our experiments, we used a decision tree learner, but any induction algorithm can be used.
6
Algorithm 2 FEAGURE algorithm
function G ENERATE F EATURES(F , S, R, d)
for fi ∈ F do
Si0 , FR = C REATE N EW P ROBLEM(fi ,S,R,d)
hi = I NDUCTIONA LGORITHM(Si0 , FR )
add fi0 (x) = hi ◦ fi to generated features
return generated features
function C REATE N EW P ROBLEM(fi , S, R, d)
vi (S) = {v|(o, y) ∈ S, fi (o) = v}
Let s(v) = {o|(o, y) ∈ S, fi (o) = v}
Si0 = {(v, majority-label(s(v)))|v ∈ vi (S)}
FR = {Rj (v)|Rj ∈ R, vi (S) ⊆ Dj }
if d > 0 then
FR = FR ∪G ENERATE F EATURES(FR , Si0 , R, d − 1)
return Si0 , FR
3.4 Finding Locally Improving Features
Some generated features may prove very useful for separating only a subset of the training data but
may be difficult to identify in the context of the full training data. In the motivating example in
Section 2, for instance, examples of male individuals (who are not at risk) are irrelevant to the target
concept, and thus mask the usefulness of the candidates for feature generation. In this subsection,
we present an extension of our FEAGURE algorithm that evaluates features in local contexts using
the divide & conquer approach. This algorithm uses a decision tree induction method as the basis of
its divide & conquer process. At each node, the FEAGURE algorithm is applied to the given set of
features, yielding a set of generated features. Out of the expanded (base and generated) feature set,
the feature with the highest information gain measure (Quinlan, 1986) is selected, and the training
set is split based on the values of that feature. This feature may or may not be one of the generated
features. We continue to apply this approach recursively to the examples in each child node, using
the expanded feature set as a baseline. Once a stopping criterion has been reached, the generated
features at each node are gathered as the final output. In our case, we stop the process if the training
set is too small or if all examples have the same label. The decision tree is then discarded. In addition
to generating features that operate within localized contexts of the original induction problem, our
approach offers several advantages for feature generation:
1. Orthogonality: Because all examples with the same value for a given feature are grouped
together, any further splits must make use of different features. Due to this and the fact that
the features selected in each step have high IG, features chosen later in the process will be
mostly orthogonal to previously chosen features. This results in a larger variety of features
overall. The feature chosen as a split effectively prunes the search tree of possible features
and forces later splits to rely on other features and thus different domains.
2. Interpretability: Looking at the features used at each splitting point gives us an intuitive
understanding of the resulting subsets. Because of this, we can more easily understand why
certain features were picked over others, which domains are no longer relevant, and so on.
3. Iterative construction: The divide & conquer approach allows for an iterative search process, which can be interrupted if a sufficient number of features were generated, or when the
7
remaining training set is no longer sufficiently representative for drawing meaningful conclusions.
The above advantages give us a strong incentive to utilize this approach when attempting to
generate features using FEAGURE. We call this new algorithm Deep-FEAGURE, as it goes into
increasingly deeper local contexts. Pseudocode for this method is shown in Algorithm 3. Through
the use of a divide & conquer approach, we can better identify strong, locally useful features that
the FEAGURE algorithm may have difficulty generating otherwise. In our experiments, we used
this approach to generate features.
Algorithm 3 Deep FEAGURE- Divide & conquer feature generation
minSize: minimal size of a node.
generatedFeatures: A global list of all generated features.
SelectFeature: Method that selects a single feature, such as highest information gain.
function D EEP FEAGURE(S, F , R, d)
if all examples in S are of same class c then
return leaf(c)
if |S| <minSize then
return leaf(majority class of S)
localGeneratedFeatures=FEAGURE(S, F , R, d)
add localGeneratedFeatures to generatedFeatures
f = S ELECT F EATURE(S, F ∪localGeneratedFeatures)
children = ∅
for v ∈ Domain(f ) do
S(f ) = {(o, y) ∈ S|f (o) = v}
subTree= D EEP FEAGURE(S(f ), F ∪localGeneratedFeatures, R, d)
add subTree to children
return children
4. Empirical evaluation
We have applied our feature generation algorithm to the domain of text classification.
4.1 Application of FEAGURE to Text Classification
The text classification problem is defined by a set of texts O labeled by a set of categories Y 4 such
that we create S = {(oi , yi )|oi ∈ O, yi ∈ Y }. Given S, the learning problem is to find a hypothesis
h : O → Y that minimizes generalization error over all possible texts of the given categories. To
measure this error, a testing set is used as an approximation.
In recent years, we have seen the rise of Semantic Linked Data as a powerful semantic knowledge base for text-based entities, with large databases such as Google Knowledge Graph (Pelikánová, 2014), Wikidata (Vrandečić & Krötzsch, 2014) and YAGO2 (Hoffart et al., 2013) becoming common. These knowledge bases represent semantic knowledge through the use of relations,
mostly represented by triplets of various schema such as RDF, or in structures such as OWL and
XML. These structures conform to relationships between entities such as “born in” (hyponyms), as
well as type information (hypernyms).
4. We can assume Y = {0, 1} for ease of analysis.
8
To use FEAGURE for text classification, we use words as binary features and Freebase and
YAGO2 as our semantic knowledge bases. YAGO2 (Hoffart et al., 2013) is a large general knowledge base extracted automatically from Wikipedia, WordNet and GeoNames. YAGO2 contains over
10 million entities and 124 million relational facts, mostly dealing with individuals, countries and
events. Freebase (Bollacker, Evans, Paritosh, Sturge, & Taylor, 2008) has been described as “a massive, collaboratively edited database of cross-linked data.” Freebase is constructed as a combination
of data harvested from databases and data contributed by users. The result is a massive knowledge
base containing 1.9 billion facts.
To apply our approach to the domain of text classification, we perform a few minor adjustments
to the FEAGURE algorithm:
1. To enable linkage between the basic features and the semantic knowledge bases, we use entity linking software (Hoffart, Yosef, Bordino, Fürstenau, Pinkal, Spaniol, Taneva, Thater, &
Weikum, 2011; Milne & Witten, 2013) to transform these words into semantically meaningful
entities.
2. Once we have created a new classifier hi , we cannot simply compose it on fi , since every
example might contain multiple entities. To that end, we apply hi on each entity and take the
majority vote.
3. Since our features are binary, we use the entities extracted from the text as the set of values vi (S). We split vi (S) into several subsets according to relation domains and apply the
FEAGURE algorithm independently to each domain.
4.2 Methodology
We evaluated our performance using a total of 101 datasets from two dataset collections:
TechTC-100 (Davidov, Gabrilovich, & Markovitch, 2004) is a collection of 100 different binary
text classification problems of varying difficulty, extracted from the Open Dictionary project. We
used the training and testing sets defined in the original paper. As our knowledge base for this task,
we used YAGO2. For entity extraction, we used AIDA (Hoffart et al., 2011), a framework for entity
detection and disambiguation.
OHSUMED (Hersh, Buckley, Leone, & Hickam, 1994) is a large dataset of medical abstracts
from the MeSH categories of the year 1991. First, we took the first 20,000 documents, similarly
to Joachims (Joachims, 1998). We limited the texts further to medical documents that contain only
a title. Due to the relatively sparse size of most MeSH categories, we only used the two with the
most documents, C1 and C20. The result is a dataset of 850 documents of each category, for a total
of 1700 documents. We used ten-fold cross-validation to evaluate this dataset. Since the YAGO2
knowledge base does not contain many medical relations, we used Freebase instead. We used the
same data dump used by Bast, Bäurle, Buchhold, and Haußmann (Bast et al., 2014).
In our experiments, we generated features using the FEAGURE algorithm. We then proceeded
to use these new features alongside three learning algorithms: SVM (Cortes & Vapnik, 1995), K-NN
(Fix & Hodges Jr, 1951) and CART (Breiman, Friedman, Stone, & Olshen, 1984).
We compared the performance of a learning algorithm with the generated features to the baseline
of the same induction algorithm without the constructed features. In addition, since we could not
obtain the code of competitive approaches for relation-based feature generation (such as FeGeLOD,
9
SGLR), we instead compared our algorithm to Expander-FG, which we believe to be indicative of
the performance of these unsupervised approaches.
4.3 Results
Table 1 shows average accuracies across all 10 folds for OHSUMED, as well as the average accuracies for all 100 datasets in techTC-100. When the advantage of our method over the baseline was
found to be significant using a pairwise t-test (with p < 0.05), we marked the p-value. Best results
are marked in bold. For the TechTC-100 dataset, FEAGURE shows a significant improvement over
the baseline approach. Of particular note are the results for KNN and SVM, where the two-level
activation of FEAGURE (d=2) shows statistically significant improvement over Expander-FG as
well as the baseline accuracy (p < 0.05). One notable exception to our good results is the poor
performance of K-NN for the OHSUMED dataset. This is likely due to the sensitivity of K-NN
to the increase in dimension. For SVM as the external classifier, the FEAGURE algorithm showed
an improvement in accuracy for 87 of 100 datasets for d = 1, and 91 datasets for d = 2. Using a
Friedman test (Friedman, 1937), we see a significant improvement (p < 0.001) over the baseline.
Table 1: Average accuracy over all datasets. The columns specify feature generation approach, with
baseline being no feature generation. The rows specify the induction algorithm used on the
generated features for evaluation. Results marked with * are significant with p < 0.001.
Dataset
OHSUMED
TechTC-100
Classifier
KNN
SVM
CART
KNN
SVM
CART
Baseline
0.777
0.797
0.806
0.531
0.739
0.81
Expander-FG
0.756
0.804
0.814
0.702*
0.782*
0.815
FEAGURE(d=1)
0.769
0.816 (p < 0.05)
0.809
0.772*
0.796*
0.814
FEAGURE(d=2)
0.75
0.819 (p < 0.05)
0.829 (p < 0.05)
0.775*
0.807*
0.825 (p < 0.05)
Figures 4 and 5 show the accuracies for datasets in techTC-100 using a SVM classifier. The
x-axis represents the baseline accuracy without feature generation, and the y-axis represents the
accuracy using our new feature set generated using FEAGURE. Therefore, any dataset that falls
above the y = x line marks an improvement in accuracy. The results show a strong trend of
improvement, with high (> 10%) improvement being common. We see that for 8 of the datasets,
there is a degradation in accuracy. This can be a result of mistakes in the entity extraction and
linking process.
In their paper on TechTC-100, Davidov et al. (2004) define a metric called Maximal Achievable
Accuracy (MAA). This criterion attempts to assess the difficulty of the induction problem by the
maximal ten-fold accuracy over three very different induction algorithms (SVM, K-NN and CART).
Figure 6 shows the performance of FEAGURE on the 25 hardest datasets in TechTC-100, in terms of
the MAA criterion. We call this dataset collection “TechTC-25MAA.” Table 2 shows the accuracies
for “TechTC-25MAA.” These results show a much more pronounced increase in accuracy, and illustrate that we can, in general, rely on FEAGURE to yield positive features for difficult classification
problems.
10
1.0
0.9
0.9
FEAGURE 2-level Accuracy
FEAGURE Accuracy
1.0
0.8
0.7
0.6
10% difference
5% difference
y=x
0.5
0.4
0.5
0.6
0.7
0.8
Baseline Accuracy
0.8
0.7
0.6
y=x
5% difference
10% difference
0.5
0.9
0.5
Figure 4: Accuracy of baseline approach
compared to one-level activation
of FEAGURE (SVM). Each point
represents a dataset. The dotted
lines represent a 5 and 10 percent
difference in accuracy
0.6
0.7
0.8
Baseline Accuracy
0.9
Figure 5: Accuracy of baseline approach
compared to two-level activation
of FEAGURE (SVM). Each point
represents a dataset. The dotted
lines represent a 5 and 10 percent
difference in accuracy
1.0
FEAGURE Accuracy
0.9
0.8
0.7
0.6
y=x
5% difference
10% difference
0.5
0.5
0.6
0.7
0.8
Baseline Accuracy
0.9
1.0
Figure 6: Accuracy of the baseline approach compared to single-level activation of FEAGURE
(SVM). Displayed are the 25 hardest datasets (those with the lowest MAA)
11
Table 2: Average accuracy over the 25 hardest datasets in terms of MAA. The columns specify
the feature generation approach, with the baseline being no feature generation. The rows
specify the induction algorithm used on the generated features for evaluation. Best results
are marked in bold.
Dataset
TechTC-25MAA
TechTC-100
Classifier
KNN
SVM
CART
KNN
SVM
CART
Baseline
0.524
0.751
0.82
0.531
0.739
0.81
Expander-FG
0.723 (p < 0.001)
0.815 (p < 0.001)
0.839
0.702 (p < 0.001)
0.782 (p < 0.001)
0.815
FEAGURE
0.803 (p < 0.001)
0.817 (p < 0.001)
0.837
0.772 (p < 0.001)
0.796 (p < 0.001)
0.814
FEAGURE 2-level
0.795 (p < 0.001)
0.829 (p < 0.001)
0.849 (p < 0.05)
0.775 (p < 0.001)
0.807 (p < 0.001)
0.825 (p < 0.05)
As we have discussed in section 3.3, FEAGURE creates a generic learning problem as part of its
execution. For our main results we learned a decision tree classifier for this new induction problem.
We also tested the effects of using K-NN and SVM classifiers instead. This choice is orthogonal
to that of the learning algorithm used to evaluate the generated features. Our experiments showed
that in general, replacing the internal tree induction algorithm lowers the achievable accuracy of the
resulting feature map. The only exception to this trend is the case of an external K-NN classifier for
the OHSUMED dataset. In this case, an internal RBF-SVM induction algorithm yields an average
accuracy of 0.795 (across ten folds), a significant (p < 0.05) improvement over the baseline.
4.4 Quantitative Analysis
To better understand the behavior of the Deep-FEAGURE algorithm, we measured its output and
performed several aggregations over it. We first look at the average number of features considered
by Deep-FEAGURE at every node: for TechTC-100, out of an average of 36.7 partitions by type
that are considered for each feature, an average of 7.3 are expanded into new features by FEAGURE.
The rest of the generated problems are discarded due to the filtering criteria mentioned in Section
3. We note that there are 46 available relations, and thus we can expect to look at no more than 46
features (if a relation does not apply to a sub-problem, it is not counted in this average).
When the depth of the search tree increases, the number of generated features decreases rapidly,
as shown in Figure 7. This is unsurprising, as we know that increased depth will cause generated
problems to have a smaller training set, increasing the likelihood of those features to be filtered out.
Additionally, we see that the number of available features decreases in a roughly linear manner with
depth. This is again unsurprising, as features deeper in the tree cannot easily make use of the same
relation multiple times due to the orthogonality trait discussed in Section 3.4.
Let us now compare the relative size of recursive problems to the existing induction problems.
Figure 8 shows the ratio between the number of examples (size) in the newly constructed induction
problem and the size of the original learning problem at various depths of the divide & conquer
search process. As depth in the search tree increases, the size ratio increases as well. This is
somewhat surprising, as an increase in depth means a smaller induction problem, and thus we would
expect a similar size ratio to be maintained. Even more unexpected, however, is that the average
recursive problem size increases with problem depth. Intuitively, we would have expected the
12
number of features tried
number of features generated
40
# Features
30
20
10
0
0
5
10
Tree depth
15
20
Figure 7: Average number of features tried vs. average number of features generated per depth.
recursive problem size to decrease as problem depth increases. However, as we search smaller
problems deeper down the search process, the relations that result in a smaller or roughly similar
sized recursive induction problem have already been used, and thus relations that cause a larger size
ratio are required.
We also note that these recursive problems do not maintain the same label balance as the original
learning problems: one label set is much larger. This is expected, as the divide & conquer search
strategy aims to use the label set as a basis for separation. We may therefore wish to make use of
various known strategies to reduce the impact of label imbalance in induction problems when using
Deep-FEAGURE.
3.5
3.0
Mean Ratio
2.5
2.0
1.5
1.0
0.5
0
2
4
6
8
10
Tree depth
12
14
16
Figure 8: Mean size ratio of newly created recursive problem compared to the original learning
problem size (new problem training set size divided by original problem training set size).
Finally, we look at the local information gain of the newly created features. In Figure 9, we see
the mean information gain of generated features per depth, compared to the best information gain
13
achieved by a non-recursive feature for that depth. We see that our recursive features tend to have a
much higher information gain, especially in the beginning of the search process. We see a decrease
in both measures as depth increases, because it is then more difficult to find distinguishing features.
0.5
recursive tree ig
base ig
0.4
mean ig
0.3
0.2
0.1
0.0
0
2
4
6
8
10
Tree depth
12
14
16
Figure 9: Mean information gain of generated features compared to the best information gain for
the original learning problem.
4.5 FEAGURE Demonstration
To demonstrate FEAGURE, We selected one problem from TechTC-100. In this example, texts
refer either to locations in and around Texas, or to locations in and around New York. The extracted
entities are locations, with the “Located in” relation as our domain (Figure 10). Applicable relations
are used to then create a new induction problem. FEAGURE uses the “Located in” and “Happened
in” relations as features for this problem, as shown in Figure 11.
Texas
text (+)
New York
text (-)
Dallas(+)
Tagged Entities
Texas
text (+)
Located In
Galveston
(+)
New York
text (-)
Manhattan
(-)
Maryland
(-)
Figure 10: Entities are extracted from the text, and entities in the “Located in” relation are used as
labeled objects.
Since we chose d = 2 as the recursion depth parameter, the algorithm calls FEAGURE recursively to try and generate new features for the new induction problem. The values of the feature
14
Dallas(+)
Galveston
(+)
Manhattan
(-)
Maryland
(-)
Location
label
Is located in(x)
Happened in(x)
Dallas
+
{Dallas County,
Texas, USA}
{Assassination of J.F.K}
Galveston
+
{Texas, USA}
{Battle of Galveston,
American Civil War}
Manhattan
-
{New York City, USA}
{Manhattan Project}
Maryland
-
{USA}
{Battle of Antietam,
American Civil War}
Figure 11: Construction of a recursive learning problem based on the “Located in” relation. Applicable relations are used to create a feature set for the newly constructed example set.
“Happened in” are events. These events are used as objects for a recursive learning problem (Figure
12). We use the “Type” relation as a feature, relying on hypernyms to classify events (Figure 13).
The resulting classifier (a decision tree induction algorithm was used) is shown in Figure 14, and
can be interpreted as “is this event a battle or conflict?”.
Assassination(+)
Dallas(+)
Galveston
(+)
Manhattan
(-)
Happened In
Maryland
(-)
Battle of
Galveston
(+)
Battle of
Antietam
(-)
Manhattan
Project (-)
American
Civil War (+)
Figure 12: Construction of a second level recursive learning problem based on the “Happened in”
relation. Feature values are treated as objects and labeled according to the labels of the
problem on locations.
Once we have generated this classifier on events, we can use it as a binary feature. FEAGURE
uses this new feature to expand the constructed induction problem on locations, shown in Figure
11. This feature is applied to a location through a majority vote over events that happened in that
location. The result is a feature for locations representing the concept “were most notable events
in this location battles/conflicts?” Finally, a decision tree learner is used on the expanded feature
set to learn a classifier on locations to be used as a feature for our original learning problem. The
new classifier for locations is shown in Figure 15. It can be described as “is this location located in
Texas, or the site of battles or conflicts?”. Texts mentioning locations in and around Texas are more
likely to link to locations that correspond to the output of this classifier. We note that this feature
15
Battle of
Antietam
(-)
Assassination
(+)
Battle of
Galveston
(+)
Manhattan
Project (-)
Event
label
Type(x)
Assassination
of J.F.K
+
{wordnet_event, wordnet_murder}
Battle of
Galveston
+
{wordnet_event, wordnet_battle}
American Civil
War
American
Civil War
(+)
{wordnet_event, wordnet_conflict}
Manhattan
Project
-
{wordnet_event, wordnet_project}
Battle of
Antietam
-
{wordnet_event, wordnet_battle}
Figure 13: Construction of a recursive learning problem based on the “Happened in” relation. Once
the example set has been created, applicable relations are used to create a feature set for
the newly constructed induction problem.
Type (x,
wordnet_battle)
False
True
+
Type (x,
wordnet_conflict)
False
-
True
+
Figure 14: Recursive feature constructed by FEAGURE for entities in the “happened in” relation.
This feature operates on events, and can be used in a classifier on locations.
was generated by FEAGURE, and was later used by our external induction algorithm due to its high
information gain.
16
Is located in(x,
Texas)
False
True
+
Type (x,
wordnet_battle)
False
True
Type (x,
wordnet_conflict)
False
True
-
+
False
-
+
True
+
Figure 15: Final generated feature constructed by FEAGURE for entities in the “located in” relation. The feature in the left branch is the recursive feature constructed by applying
FEAGURE to the new learning problem.
5. Related work
Many feature generation methodologies have been developed to search for new features that better
represent the target concepts. There are three major approaches for feature generation: tailored
methods, combinational techniques, and algorithms utilizing external knowledge.
17
Tailored approaches (Sutton & Matheus, 1991; Hirsh & Japkowicz, 1994) are designed for
specific problem domains and rely on domain-specific techniques. One such example is the bootstrapping algorithm (Hirsh & Japkowicz, 1994), designed for the domain of molecular biology. The
algorithm represents features as nucleotide sequences whose structure is determined by existing
background knowledge. The algorithm uses an initial set of feature sequences, produced by human
experts, and uses a domain-specific set of operators to change them into new sequence features.
Such special-purpose algorithms may be effectively tailored for a given domain, but have proven
difficult to generalize to other domains and problems.
Combinational feature generation techniques are domain-independent methods for constructing
new features by combining existing features. The LDMT algorithm (Brodley & Utgoff, 1995)
performs feature construction in the course of building a decision-tree classifier. At each created
tree node, the algorithm constructs a hyperplane feature through linear combinations of existing
features in a way likely to produce concise, relevant hyperplanes. The LFC algorithm (Ragavan
et al., 1993) combines binary features through the use of logical operators such as ∧, ¬. The FICUS
algorithm (Markovitch & Rosenstein, 2002) allows the use of any combinational feature generation
technique, based on a given set of constructor functions. Recent work by Katz et al. (2016) uses a
similar approach.
Deep Learning (Rumelhart et al., 1986; LeCun et al., 1998) is another major class of combinational feature generation approaches. Here, the activation functions of the nodes can be viewed as
feature schemes, which are instantiated during the learning process by changing the weights.
One limitation of combinational approaches is that they merely combine existing features to
make the representation more suitable for the learning algorithm. Our FEAGURE algorithm belongs
to a third class of approaches that inject additional knowledge into the existing problem through the
feature generation process.
Propositionalization approaches (Kramer & Frank, 2000; Cheng et al., 2011) rely on relational
data to serve as external knowledge. They use several operators to create first-order logic predicates
connecting existing data and relational knowledge. Cheng et al. (2011) devised a generic propositionalization framework using linked data via relation-based queries. FeGeLOD (Paulheim &
Fümkranz, 2012) also uses linked data to automatically enrich existing data. FeGeLOD uses feature
values as entities and adds related knowledge to the example, thus creating additional features.
Unsupervised approaches allow us to utilize external knowledge, but they have a major issue:
Should we try to construct deep connections and relationships within the knowledge base, we would
experience an exponential increase in the number of generated features. To that end, FEAGURE
and other supervised approaches use the presence of labeled examples to better generate deeper
features. Most supervised methods can trace their source to Inductive Logic Programming (ILP)
(Muggleton, 1991; Quinlan, 1990), a supervised approach that induces a set of first-order logical
formulae to separate the different categories of examples in the training set. ILP methods start
from single relation formulae and add additional relational constraints using the knowledge base,
until formulae that separate the training set into positive and negative examples are found. To that
end, these approaches make use of a refinement operator. When applied on a relational formula,
this operator creates a more specialized case of that formula. For example, given the logical formula BornIn(X, Y ), where X is a person and Y is a city, one possible refinement is the formula
BornIn(X, Y ) ∧ CapitalOf (Y, Z), where Z is a country. The result is a logical formula that considers a more specific case. Additionally, we can look at a refinement that restricts by a constant,
18
turning BornIn(X, Y ) into, for example, BornIn(X, United States). This refinement process
continues until a sufficient set of consistent formulae is found.
An algorithm suggested by Terziev (2011) shows an interesting approach to supervised feature
generation. In his paper, Terziev (2011) suggests a decision tree based approach, where in each node
of the tree, an expansion of features is done similarly to FeGeLOD, with an entropy-based criterion
to decide whether further expansion is required. This technique bears several similarities to DeepFEAGURE (Algorithm 3). Unlike Deep-FEAGURE, the feature expansion process is unsupervised,
and the resulting feature must be a decision tree, restricting the generality of the approach.
The dynamic feature generation approach used by the SGLR algorithm (Popescul & Ungar,
2007) can be seen as the supervised equivalent of propositionalization methods. Feature generation is performed during the training phase, allowing for complex features to be considered by
performing a best-first search on possible candidates. This process allows SGLR to narrow the
exponential size of the feature space to a manageable number of candidates. While this supervised approach overcomes the exponential increase in features that unsupervised approaches suffer
from, the space of generated features that it searches is significantly less expressive than that of our
approach. Through the use of recursive induction algorithm, our approach automatically locates
relationships and combinations that we would not consider otherwise.
Since we have focused on the problem of text classification, we also discuss a few text-based
approaches for feature generation. Linguistic methods such as those described in a study by Moschitti and Basili (2004) attempt to use part-of-speech and grammar information to generate more
indicative features than the bag-of-words representation of texts. Concept-based approaches are the
most similar to our own. Two examples of such methods are Explicit Semantic Analysis (ESA)
(Gabrilovich & Markovitch, 2009), which generates explicit concepts from Wikipedia based on
similarity scores, and Word2Vec (Mikolov et al., 2013), which generates latent concepts based on a
large corpus. Both approaches provide us with a feature set of semantic concepts. These concepts
can be used alongside our approach, allowing us to induce over them.
6. Conclusions
When humans use inductive reasoning to draw conclusions from their experience, they use a vast
amount of general and specific knowledge. In this paper we introduced a novel methodology
for enhancing existing learning algorithms with background knowledge represented by relational
knowledge bases. The algorithm works by generating complex features induced using the available
knowledge base. It does so through the extraction of recursive learning problems based on existing
features and the knowledge base, that are then given as input to induction algorithms. The output of
this process is a collection of classifiers that are then turned into features for the original induction
problem.
An important strength of our approach is its generality. The features generated by FEAGURE
can be used by any induction algorithm, allowing us to inject external knowledge in a general, algorithm independent manner. One potential limitation of our approach is that it requires features
with meaningful values in order to operate. Despite this limitation, there is a wide range of problems where feature values have meaning. For these problems, we can apply one of several general
and domain-specific knowledge bases, depending on the problem. In recent years, we have seen an
increase in the number of available knowledge bases. These knowledge bases include both general
knowledge bases such as Freebase, YAGO, Wikidata and the Google Knowledge Graph, and more
19
domain-specific knowledge bases, from the British Geographical Survey (BGS) linked dataset, containing roughly one million geological facts regarding various geographical formations, through
biological databases such as Proteopedia, composed of thousands of pages regarding biological
proteins and molecules, to entertainment-focused databases such as IMDB, containing millions of
facts on movies, TV-series and known figures in the entertainment industry. With the recent surge
of well-formed relational knowledge bases, and the increase in use of strong learning algorithms
for a wide variety of tasks, we believe our approach can take the performance of existing machine
learning techniques to the next level.
20
References
Bast, H., Bäurle, F., Buchhold, B., & Haußmann, E. (2014). Easy access to the freebase dataset. In
Proceedings of the companion publication of the 23rd international conference on World wide
web companion, pp. 95–98. International World Wide Web Conferences Steering Committee.
Bollacker, K. D., Evans, C., Paritosh, P., Sturge, T., & Taylor, J. (2008). Freebase: a collaboratively
created graph database for structuring human knowledge. In Proceedings of the 2008 ACM
SIGMOD international conference on Management of data, pp. 1247–1250. AcM.
Breiman, L., Friedman, J., Stone, C. J., & Olshen, R. A. (1984). Classification and regression trees.
Wadsworth.
Brodley, C. E., & Utgoff, P. E. (1995). Multivariate decision trees. Machine Learning, 19(1), 45–77.
Cheng, W., Kasneci, G., Graepel, T., Stern, D. H., & Herbrich, R. (2011). Automated feature generation from structured knowledge. In Proceedings of the 20th ACM international conference
on Information and knowledge management, pp. 1395–1404. ACM.
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273–297.
Davidov, D., Gabrilovich, E., & Markovitch, S. (2004). Parameterized generation of labeled datasets
for text categorization based on a hierarchical directory. In Proceedings of the 27th annual
international ACM SIGIR conference on Research and development in information retrieval,
pp. 250–257. ACM.
Fix, E., & Hodges Jr, J. L. (1951). Discriminatory analysis-nonparametric discrimination: consistency properties. Tech. rep., DTIC Document.
Friedman, M. (1937). The use of ranks to avoid the assumption of normality implicit in the analysis
of variance. Journal of the american statistical association, 32(200), 675–701.
Gabrilovich, E., & Markovitch, S. (2009). Wikipedia-based semantic interpretation for natural language processing. Journal of Artificial Intelligence Research, 34, 443–498.
Hersh, W. R., Buckley, C., Leone, T. J., & Hickam, D. H. (1994). OHSUMED: an interactive
retrieval evaluation and new large test collection for research. In Proceedings of the 17th
annual international ACM SIGIR conference on Research and development in information
retrieval, pp. 192–201. Springer.
Hirsh, H., & Japkowicz, N. (1994). Bootstrapping training-data representations for inductive learning: A case study in molecular biology. In Proceedings of the 12th National Conference on
Artificial Intelligence, pp. 639–644.
Hoffart, J., Suchanek, F. M., Berberich, K., & Weikum, G. (2013). YAGO2: a spatially and temporally enhanced knowledge base from Wikipedia. Artificial Intelligence, 194, 28–61.
Hoffart, J., Yosef, M. A., Bordino, I., Fürstenau, H., Pinkal, M., Spaniol, M., Taneva, B., Thater,
S., & Weikum, G. (2011). Robust disambiguation of named entities in text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 782–792.
Association for Computational Linguistics.
Joachims, T. (1998). Text categorization with support vector machines: Learning with many relevant
features. In European conference on machine learning, pp. 137–142. Springer.
21
Katz, G., Shin, E. C. R., & Song, D. (2016). Explorekit: Automatic feature generation and selection.
In Data Mining (ICDM), 2016 IEEE 16th International Conference on, pp. 979–984. IEEE.
Kramer, S., & Frank, E. (2000). Bottom-up propositionalization.. In ILP Work-in-progress reports.
LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324.
Levi, K., & Weiss, Y. (2004). Learning object detection from a small number of examples: the
importance of good features. In Computer Vision and Pattern Recognition, 2004. CVPR 2004.
Proceedings of the 2004 IEEE Computer Society Conference on, Vol. 2, pp. II–II. IEEE.
Markovitch, S., & Rosenstein, D. (2002). Feature generation using general constructor functions.
Machine Learning, 49(1), 59–98.
McNamara, D. S., & Kintsch, W. (1996). Learning from texts: Effects of prior knowledge and text
coherence. Discourse processes, 22(3), 247–288.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. In Advances in neural information
processing systems, pp. 3111–3119.
Milne, D., & Witten, I. H. (2013). An open-source toolkit for mining wikipedia. Artificial Intelligence, 194, 222–239.
Moschitti, A., & Basili, R. (2004). Complex linguistic features for text classification: A comprehensive study. In European Conference on Information Retrieval, pp. 181–196. Springer.
Muggleton, S. (1991). Inductive logic programming. New generation computing, 8(4), 295–318.
Paulheim, H., & Fümkranz, J. (2012). Unsupervised generation of data mining features from linked
open data. In Proceedings of the 2nd international conference on web intelligence, mining
and semantics, p. 31. ACM.
Pelikánová, Z. (2014). Google knowledge graph..
Popescul, A., & Ungar, L. H. (2007). Feature generation and selection in multi-relational statistical
learning. STATISTICAL RELATIONAL LEARNING, 453.
Quinlan, J. R. (1986). Induction of decision trees. Machine learning, 1(1), 81–106.
Quinlan, J. R. (1990). Learning logical definitions from relations. Machine learning, 5(3), 239–266.
Ragavan, H., Rendell, L., Shaw, M., & Tessmer, A. (1993). Complex concept acquisition through
directed search and feature caching. In Proceedings of the Thirteenth International Joint
Conference on Artificial Intelligence (IJCAI-1993), pp. 946–951.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations by
error-propagation. In Parallel Distributed Processing: Explorations in the Microstructure of
Cognition. Volume 1, Vol. 1, pp. 318–362. MIT Press, Cambridge, MA.
Sutton, R. S., & Matheus, C. J. (1991). Learning polynomial functions by feature construction.. In
ML, pp. 208–212.
Terziev, Y. (2011). Feature generation using ontologies during induction of decision trees on linked
data..
22
Vrandečić, D., & Krötzsch, M. (2014). Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10), 78–85.
23
| 2cs.AI
|
1
Coupling Data Transmission
Dmitri Truhachev and Christian Schlegel
Abstract
arXiv:1209.5785v3 [cs.IT] 24 Jun 2017
We consider a signaling format where information is modulated via a superposition of independent data streams.
Each data stream is formed by error-correction encoding, constellation mapping, replication and permutation of
symbols, and application of signature sequences. The relations between data bits and the modulation symbols
transmitted over the channel can be represented in the form of a sparse graph. In case when the modulated data streams
are transmitted with time offsets the receiver observes spatial coupling of the individual graph representations into
a graph chain enabling efficient demodulation/decoding. We prove that a two-stage demodulation/decoding method,
in which iterative demodulation based on symbol estimation and interference cancellation is followed by parallel
error correction decoding, achieves capacity on the additive white Gaussian noise (AWGN) channel asymptotically.
We compare the performance of the two-stage receiver to a receiver which utilizes hard feedback between the errorcorrection encoders and the iterative demodulator.
I. I NTRODUCTION
Recently, the technique of spatial graph coupling applied to iterative processing on graphs attracted significant
interest in many areas of communications. The method was first introduced to construct low-density parity-check
convolutional codes (LDPCCCs) [1] that exhibit the so-called threshold saturation behavior [2][3], which occurs
when the limit (threshold) of suboptimal iterative decoding of LDPCCCs asymptotically achieves the optimum
maximum a posteriori probability (MAP) decoding threshold [3] of LDPC block codes with the same structure [19].
The idea of constructing graph structures from connected identical copies of a single graph has since been applied
to compressed sensing [20][36], image recognition, quantum coding [24], and other fields.
A number of applications of spatial graph coupling to communications over multi-user channels have since
been developed. Yedla et. al. studied threshold saturation for a two-user Gaussian multiple-access channel (MAC)
[7] while Kudekar and Kasai studied it for binary erasure MACs [8]. In terms of multi-user signaling formats,
spatially-coupled partitioned-spreading code-division multiple-access (PS-CDMA) was studied in [21], [25]. It was
proven that in the noiseless regime the coupled system can achieve arbitrary multi-user loads with the two-stage
demodulation/decoding method contrary to uncoupled PS-CDMA [11] or dense CDMA in general [28]. Sparse
CDMA, extensively studied by Guo and Wang [9] in the uncoupled regime, was extended via spatial coupling
in [23], [34] where significant threshold improvements over uncoupled CDMA were derived.
This work was partially supported by NSERC Industrial Chair Grant 428809 and TELUS Corporation Canada and presented in part in ISIT
2013 [22].
June 27, 2017
DRAFT
IEEE Transactions on Information Theory, under revision
In this work we focus on a generalized multi-user communication signaling format which lends itself to efficient
spatial coupling and iterative demodulation. Each user’s transmitter chain includes common elements such as
error-correction encoding, higher-order constellation mapping, symbol repetition and permutation, and signature
sequences. The complexity and rate of the system can be flexibly adjusted depending on the communication
scenario. The systems can be operated and studied in both coupled and uncoupled regimes. For specific parameter
settings some classes of coupled generalized modulation [10], coupled PS-CDMA [21], or coupled version of IDMA
introduced by Ping et. at. [12], [13], [16] can be derived as particular cases. The iterative demodulation is a sequence
of symbol estimation and interference cancellation iterations and can be seen as a message-passing process on the
system graph.
In this paper we investigate spectral efficiency achievable by iterative demodulation/decoding in uncoupled and
coupled regimes and quantifying the gap to the channel capacity. In particular, we focus on the two-stage demodulation/decoding schedule where there is no feedback between the error-correction decoders and the demodulator.
The two-stage schedule allows for simplified system design and improved decoding latency compared to either
sequential peeling schedules or turbo demodulation/decoding schedules based on the exchange of soft information
between the demodulator and the error-correction decoders. The main contributions of our work can be summarized
as follows.
A. Main Contributions
•
We rigorously derive a coupled recursion characterizing the evolution of noise-and-interference power throughout demodulation iterations for the coupled and uncoupled systems. We show that for any given number of
demodulation iterations there exists a system graph for which the iterative demodulation operates as message
passing on a set of trees. We then quantify the achievable data rate as a function of the signal-to-noise ratio
(SNR), constellation symbol alphabet, and the finite system parameters: the number of interfering data streams,
repetition factor, and symbol constellation.
•
For the case of a binary symbols we derive an upper bound on the gap between the achievable rate and the
channel capacity. The gap is a function of the SNR and the finite set of system parameters. We show that the
asymptotic gap it is inversely proportional to the capacity itself and vanishes as the system’s SNR increases.
•
We study an alternative demodulation/decoding schedule were hard decision feedback from the error-correction
decoders to the demodulator is allowed, which is also capacity-achieving asymptotically. We derive a bound
on the gap to capacity and compare the achievable rate with that of the two-stage schedule. In this case we
prove that the gap to capacity vanishes for any fixed SNR as the system load increases.
B. Related Work
A related multi-user communications problem is the problem of maximizing the number of users simultaneously
communicating over the MAC channel or achieving high multi-user efficiency [41], which measures the impact of
residual post/detection multi-user interference on a user’s performance. A number of papers have been dedicated to
2
IEEE Transactions on Information Theory, under revision
the study of asymptotics of joint and individual multi-user detection for classic CDMA. Tse and Verdu investigated
the asymptotic performance of joint MAP decoding of random CDMA, and proved that the multiuser efficiency
approaches one as the SNR goes to infinity [42]. Later Tanaka [28], Müller and Gerstacker [43], and Guo
and Verdu [44] studied the performance of joint and individual multi-user detection of CDMA via the nonrigorous replica method, and derived the limits for the rates of coded random CDMA with joint and separate
demodulation/decoding in terms of multi-user efficiency. In a paper by Takeuchi, Tanaka, and Kawabata [35],
developed in parallel to ours [22], coupled sparse CDMA was studied in the asymptotic regime where the number
of users, spreading gain, and the system size were taken to infinity, while the multi-user load was kept constant.
Based on continuous approximation of the respective coupled recursion for the case of binary symbol alphabet
the authors showed saturation of the iterative detection threshold of sparse CDMA to the MAP threshold of dense
CDMA via spatial graph coupling.
The focus of our paper is on rigorously deriving and demonstrating the rates achievable by coupled and uncoupled
sparse multi-user systems. We estimate the resulting gap to capacity explicitly as a function of SNR rather than
multi-user efficiency. We study a system with general symbol alphabets and derive the achievable rate for finite
parameter settings. Our approach allows us to study the system’s performance for various explicitly defined decoding
schedules, such as the demodulation/decoding schedule with hard feedback, for different numbers of iterations, and
receiver signal-processing options. Our framework leads to graphical interpretations of the achievable rates and the
gap to capacity in a form similar to the are theorem for threshold saturation of LDPCCC in BEC.
It’s worth mentioning that the capacity of the multiple-access channel can be achieved using peeling decoding,
a fact that stems directly from the chain rule of the mutual information. Various types of peeling or sequential
cancellation decoders have been extensively studied for a variety of systems [5], [14], [15], [17]. The target of
our work is the iterative demodulation and the low-latency two-stage demodulation/decoding schedule and, in
particularly, on the role of spatial graph coupling in approaching channel capacity.
Interleaved-division modulation (IDM) method used in point-to-point transmission proposed by Hoeher and
Wo [18], has common elements with the multi-user format we consider such as superposition of the data streams
and bit interleaving. The latter is also a feature of bit-interleaved coded modulations in general [6]. Contrary to
the reception techniques for the separation of symbol-synchronous data streams studied by Hoeher and Wu, we
investigate a cancellation based low-complexity iterative interference cancellation approach which is independent
of data stream synchronism and power fluctuations and it specially suited for a coupled system.
Another related point-to-point communications format are the sparse superposition codes studied by Baron and
Joseph [50], and later shown to achieve capacity in the uncoupled unequal power regime [51]. Sparse superposition
codes are based on superimposing data streams based on Gaussian matrices, a technique closely related to compressed sensing reconstruction. A coupled version of sparse superposition codes was recently studied in [49] where
it was shown that given the coupled recursion describing the state evolution of the approximate message-passing
(AMP) decoder, the codes achieve channel capacity asymptotically. The AMP decoder for coupled superposition
3
IEEE Transactions on Information Theory, under revision
code does not lend itself to a rigorous derivation of the coupled recursions, however.
C. Structure of the Paper
System model, transmission format and receiver processing are described in Section II. Section III presents the
performance analysis, derivation of the main coupled recursions and formulation of the main results. Numerical
results are given in Section IV-A. Finally, Section V concludes the paper.
II. S YSTEM M ODEL
We consider a communication scenario in which a number of terminals communicate to a single receiver. The
signal observed at the receiver is formed by a superposition of L independently modulated data streams that originate
at a single or multiple terminals and then superimpose at the receiver.
The processes of generating data stream l, l = 1, 2, · · · , L is depicted in Fig. 1. First, a binary information
sequence ul = u1,l , u2,l , u3,l , · · · , uK,l is encoded for error correction and mapped to a sequence of constellation
symbols denoted by v l = v1,l , v2,l , v3,l , · · · , vN,l , where vn,l ∈ Al . By Al we denote a set of signal constellation
symbols (which can be a PAM constellation, for example). At the next step each symbol of the sequence v l is
replicated M1 times. After the replication the resulting symbols are permuted using a permutation matrix Pl of
size N M1 × N M1 .
0
Finally each symbol vi,l
, i = 1, 2, · · · , N M1 , l = 1, 2, · · · , L of the permuted sequence is multiplied by a
signature sequence si,l of length M2 . We consider pseudo-random energy normalized binary signature sequences,
however, Gaussian or other sequences can alternatively be used. The symbols of the signature sequences satisfy
1 if (n1 , l1 , m1 ) = (n2 , l2 , m2 )
Esn1 ,l1 ,m1 sn2 ,l2 ,m2 =
(1)
0 otherwise .
The resulting sequence ṽ l = ṽ1,l , ṽ2,l , · · · , ṽM N,l is multiplied by an amplitude al and transmitted over the channel.
The overall symbol repetition factor is given by M = M1 M2 .
Therefore, we can see that generation of a data stream is done in two steps. The first step encompasses individual
error-correction encoding and constellation mapping. The second step is the preparation of the data for multi-user
processing at the receiver and includes replication, permutation, and application of the signature sequences.
A. Uncoupled System
The uncoupled system corresponds to a communication scenario in which all data streams are transmitted over
the channel simultaneously. Assuming real-valued Gaussian noise added in channel the received signal is given by
y=
L
X
al ṽ l + n ,
l=1
where n is a vector of the iid Gaussian noise samples with zero mean and variance σ 2 . The uncoupled system can
be represented by the dependency graph shown on the left side of Fig. 2. The top part of the figure demonstrates
4
IEEE Transactions on Information Theory, under revision
al
ul
Error Correction Encoder
and Constellation Mapper
vl
Symbol
Replication
ṽ l
Signature
Sequences
Permutation
Fig. 1. Generation of the lth data stream.
three individual data stream graphs. The resulting graph, observed at the receiver, is shown at the bottom. Variable
nodes representing symbols vj,l are depicted as circles, while channel nodes representing received values yt are
depicted as hexagons.
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
individual data stream
graphs
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
time
superposition
graph
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
coupled system
uncoupled (block)
system
Fig. 2. Graph representation of the uncoupled and spatially coupled systems.
B. Coupled System
A spatially graph coupled system for the presented modulation format arises naturally when the modulated data
streams are transmitted over the channel with time offsets. This is illustrated in Fig. 2 b). The graphs representing
transmitted data streams couple through the channel nodes which correspond to the times at which the data
stream symbols are transmitted over the common channel. Contrary to the uncoupled case, where the data streams
superimpose in groups of size L, and each group is represented by an independent graph at the receiver, the coupled
5
IEEE Transactions on Information Theory, under revision
system is represented by a single connected graph, also called the coupled graph chain. This draws parallels to
coupled graph chains appearing in other applications such as SC-LDPC codes [1]. We will prove that iterative
demodulation and decoding on this graph chain will give us a substantially higher achievable rate.
The hexagonal channel nodes represent received symbols yt , t = 1, 2, 3, · · · . Symbols ṽj,l transmitted at the same
time t share the same channel node. For the sake of representation we assume that the transmission is “symbolsynchronous”. It is important to note, however, that the receiver processing which is based on symbol estimation
and interference cancellation does not depend on this assumption nor does the iterative demodulation analysis.
We denote the time offset of data stream l by τl , l = 1, 2, · · · , L̂ where L̂ is the total number of streams
transmitted. Then the components of the received signal vector y are given by
yt =
L̂
X
al ṽt−τl ,l + nt .
t = 1, 2, 3 · · · .
l=1
In order to study the performance of the coupled system we consider a regularized delay structure defined
as follows. We introduce W , the coupling parameter, and assume that each transmitted data stream consists of
2W + 1 subsections of length Nw = M N/(2W + 1). In addition the data streams are transmitted in groups of
size LW = L/(2W + 1) streams. At time t = 1 the first group of data streams is transmitted, i.e., the first LW
time delays are given by τ1 = 1, τ2 = · · · = τLW = 0. At time NW + 1 the second group of LW data streams is
transmitted with time delays given by τLW +1 = NW , τLW +2 = · · · = τ2LW = 0 and so on, at every time kNW + 1,
k = 0, 1, 2, · · · exactly LW data streams are added. henceforth we will understand the term “coupled system” to
mean a system with regularized coupling.
While the degrees of nodes in the coupled graph chain are predetermined, the positions of the edges are not. We
can define a graph ensemble to describe a coupled system using random permutation matrices Pl for the construction
of the individual data streams. This representation is given in Appendix A along with a matrix representation of
the coupled chain. In addition, signature sequences are also chosen randomly.
The regularized delay structure is assumed for analysis purposes and to demonstrate the capacity-achieving
properties of the coupled system. It is not required for receiver signal processing. Fig. 2 b) depicts six data streams
transmitted with delays τ1 = 1, τ2 = 5, τ3 = 7, τ4 = 11, τ5 = 12, τ6 = 13. In general, the system performance,
depends on the distribution of the delays. A Poisson delay process produces results close to those with regularized
coupling. The study of delay distributions and their impact on performance is, however, beyond the scope of this
paper.
C. System Load and Rate
We define the load α of the uncoupled system as the ratio of the number of transmitted data streams L to the
repetition factor M , α = L/M . If we assume equal power data streams and the same modulation index B in terms
of the number of bits mapped to a constellation symbol for each data stream we can normalize the transmit symbol
p
amplitudes to be al = 1/L. Hence the total systems SNR measured as a ratio of the total transmit power over
noise power equals 1/σ 2 . The total transmit data rate equals RBL/M = αBR information bits per channel use
6
IEEE Transactions on Information Theory, under revision
J (t)
vj,l
vj1 ,l1 vj2 ,l2 vj3 ,l3
st1
st2
st3
vjL ,lL
stM
st
T (j, l)
Fig. 3. The relation between the variable and channel nodes and the sets of indices involved.
where R = K/N is the rate of the error-correction codes common to all data streams. The signal-to-noise ratio per
information bit equals Eb /N0 = 1/(2Rσ 2 ).
In the coupled case the system load in terms of the number of data streams per repetition factor M for the first N
time instances t = 1, 2, · · · , N is smaller than α. Assuming infinite number of data streams in the system L̂ = ∞
system load equals α for t > N and the same expression for the rate and SNR as for the uncoupled case. The
initial reduced load leads to the initial rate loss typical for all coupled systems, that vanishes as L̂ = ∞.
D. Iterative Demodulation Process
The received signal vector y = (y1 , y2 , y3 , · · · ) contains M = M1 M2 replicas of each transmitted symbol vj,l ,
j = 1, 2, · · · , l = 1, 2, · · · , L. Let T (j, l) = {t1 , t2 , · · · , tM } denote the set of indices t such that the received
vector component yt , contains vj,l . Let’s denote the corresponding symbols of the signature sequences multiplied
(t)
by vj,l during the modulation process by sj,l where t ∈ T (j, l). Moreover, by J (t) we denote a set of all index
pairs (j, l) such that vj,l is included in yt (see Fig. 3). The cardinality of the set T (j, l) is always equal to M .
The cardinality of J (t) equals L for uncoupled system and varies depending on t for the coupled system (see
Section II-C). For each bit vj,l we use a set of received signals {yt }t∈T (j,l) to form a vector y j,l . Since each yt ,
t ∈ T (j, l) contains vj,l we have
y j,l = hvj,l + ξ j,l
(t )
(t )
(t
(2)
)
where h = al (sj,l1 , sj,l2 , · · · , sj,lM ). The vector ξ j,l = (ξj,l,t1 , ξj,l,t2 , · · · , ξj,l,tM ) is the noise-and-interference
vector with respect to the signal vj,l . The components of this noise-and-interference vector are given by
X
(t)
ξj,l,t =
al0 sj 0 ,l0 vj 0 ,l0 + nt , t = t1 , t2 , · · · , tM .
(3)
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
Let us denote the covariance matrix of the interference vector ξ j,l by Rj,l . We now perform minimum meansquared error (MMSE) filtering on y j,l to form an SNR-optimal linear estimate of vj,l , given by
zj,l = wTj,l y j,l = wTj,l hvj,l + wTj,l ξ j,l ,
7
(4)
IEEE Transactions on Information Theory, under revision
where
−1 ∗ −1
wTj,l = (I + h∗ R−1
h Rj,l
j,l h)
minimizes ||wj,l y j,l − vj,l ||2 .
Since we know that vj,l belongs to the symbol alphabet Al we can form a conditional expectation estimate
v̂j,l = E(vj,l |zj,l )
(5)
For the case of equiprobable binary symbols vj,l ∈ {1, −1} (2-PAM modulation) v̂j,l takes the form v̂j,l =
tanh(zj,l γj,l ), where γj,l is the SNR of the estimate zj,l . We will compute this SNR in the next section.
Once the estimates v̂j,l are generated for all data symbols vj,l , j = 1, 2, · · · , l = 1, 2, · · · , L, the next iteration
(1)
starts with an interference cancellation step performed by computing y j,l with components
X
(1)
(t)
yj,l,t = yt −
al0 sj 0 ,l0 v̂j 0 ,l0 , t ∈ T (j, l) .
(6)
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
Therefore,
(1)
(1)
y j,l = hvj,l + ξ j,l ,
(7)
(1)
where the components of ξ j,l equal
(1)
X
ξj,l,t =
(t)
al0 sj 0 ,l0 (vj 0 ,l − v̂j 0 ,l0 ) + nt ,
t ∈ T (j, l) .
(8)
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
(1)
We then proceed with calculating new estimates v̂j,l repeating the same procedure that lead to (4) and (5), and the
estimation and interference cancellation steps are repeated for a number of iterations. To avoid reusing the same
(i)
information throughout the iterations, we compute M extrinsic vectors y j,l,t , t ∈ T (j, l) for each bit vj,l at iteration
(i)
i. The components of vector y j,l,t equal
X
(i)
(τ ) (i−1)
yj,l,τ = yτ −
al0 sj 0 ,l0 v̂j 0 ,l0 ,τ = al vj,l +
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
X
(τ )
(i−1)
al0 sj 0 ,l0 (vj 0 ,l0 − v̂j 0 ,l0 ,τ ) + nτ ,
(9)
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
(i)
(i)
where τ ∈ T (j, l), τ 6= t. The vectors y j,l,t are used to form M estimates zj,l,t according to (4) and corresponding
(i)
v̂j,l,t estimates according to (5), one for each replica of vj,l . This rules correspond to message passing on a system’s
(i)
graph. Each variable node (j, l) passes M estimates v̂j,l,t to the connected channel nodes t ∈ T (j, l) while each
(i)
channel node passes zj,l,t where (j 0 , l0 ) ∈ J (t) to the connected variable nodes.
E. Demodulation and Decoding Schedules
The demodulation and error-correction decoding at the receiver can be scheduled in a number of ways. We
consider two possibilities which are of relatively low complexity. One alternative schedule of significantly higher
complexity is the “turbo-schedule” with soft feedback between the decoders and the iterative demodulator.
Two-Stage Schedule
8
IEEE Transactions on Information Theory, under revision
At the first stage I iterative demodulation iterations are performed as described above. The second stage comprises
simultaneous decoding of the forward error correction codes used to encode the information into the data streams.
Hard Decoding Feedback and Cancellation Schedule
Iterative demodulation iteration are performed until one or more data stream’s SNR raises above the errorcorrection code threshold. Once it happens the code is decoded and hard feedback is given to the demodulator, the
result is subtracted and the demodulation process continues.
III. P ERFORMANCE A NALYSIS
At every iteration the demodulator attempts to reduce the amount of inter-stream interference. The central part
of the performance analysis is dedicated to tracking the evolution of noise-and-interference power throughout
the demodulation iterations. We start with an assessment of the performance of the uncoupled system for which
the evolution of the noise-and-interference power is not dependent on the time index t. After that we derive a
characteristic equation that determines when the iterative demodulation process for the uncoupled system converges
to a nearly interference-free state.
We then proceed with deriving the noise-and-interference power evolution for the coupled system and making a
connection to the characteristic equation of the uncoupled system and its fixed points. Based on this connection we
obtain the expressions for the achievable rates and the gap-to capacity. Contrary to some related work we make our
derivation for the finite parameter values M, L, B and then derive the asymptotically achievable rates as a particular
case.
A. Uncoupled System
We focus on the case of equal power data streams where the amplitudes al = a, l = 1, 2, · · · , L. Without loss
p
of generality we normalize data stream powers so that a = 1/L and vary the noise power σ 2 . In this case
the total system SNR per receive symbol equals 1/σ 2 (see Section II-C). We then proceed with considering the
noise-and-interference power evolution equation (9).
Let us denote the noise-and-interference power at iteration i by
2
r
X
1 (t)
def
(i−1)
xi = E
sj 0 ,l0 (vj 0 ,l0 − v̂j 0 ,l0 ,τ ) + nτ =
L
0 0
(j 0 ,l0 )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
(j ,l )∈J (t)
s.t. (j 0 ,l0 )6=(j,l)
=
X
1
(i−1)
E(vj 0 ,l0 − v̂j 0 ,l0 ,τ )2 + σ 2
L
1
(L − 1)µi−1 + σ 2
L
(10)
(11)
where the mean-square error (MSE) of the symbols at iteration i − 1 are given by µi−1 , i.e.
def
(i−1)
µi−1 = E(vj 0 ,l0 − v̂j 0 ,l0 ,τ )2 = E(vj 0 ,l0 − E(vj,l |zj,l ))2
(12)
The expectation is taken over the respective channel node index t, variable node index (j 0 , l0 ), and the system
realizations. When breaking the variance of the noise-and-interference power in (9) into the sum of variances of
9
IEEE Transactions on Information Theory, under revision
(i−1)
the components vj 0 ,l0 − v̂j 0 ,l0 ,τ as in (10) we assume that these are independent. This can be guaranteed by making
sure that the system graph has large girth, the property not required for practical implementation but is important
for the rigor of the analysis.
Coming back to the graph representation of the system if we connect the variable node vj,l to the channel nodes
yt where t ∈ T (j, l) and then connect each channel node yt to vj 0 ,l0 such that vj,l ∈ J (t) we obtain a sub-graph
describing one demodulation iteration for symbol vj,l . If we expand it further by connecting the nodes vj 0 ,l0 to their
channel nodes yτ ∈ T (j, l) and so on we can obtain a sub-graph describing i demodulation iterations for symbol
vj,l . If we the girth of the entire system’s graph is at least 4i + 1 any demodulation sub-graph for any symbols vj,l
will not contain repeated nodes and is, therefore, a tree. This description and conditions closely resemble these for
LDPC block [33] and convolutional codes [39]. The following lemma elaborates on this connection and proves that
a system graph with arbitrary large girth can be constructed for both uncoupled and coupled systems if the data
stream length is chosen large enough.
Lemma 1. For any number of iterations i, and parameters L, M, B, W both coupled and uncoupled systems can
be constructed such that the corresponding graph has girth larger than 4i.
Proof. See Appendix A.
The next step is to express the MSE µi−1 in terms of the noise and interference power xi−1 at iteration i − 1. The
Central Limit Theorem implies that the noise and interference vector ξ j,l converges to a Gaussian random vector
with independent zero-mean components and covariance matrix Rj,l = diag(σt21 , σt22 , · · · , σt2M ) where σt2 denotes
the variance of ξj,l,t , t ∈ T (j, l) as L increases. Typical L values range around 50, 100 and higher indicating
that Gaussian approximation is much more accurate compared to that performed for LDPC codes with less denser
graphs. The Lindeberg condition is satisfied for both coupled and uncoupled equal power systems. The resulting
SNR of the signal zj,l in (5) equals
γj,l = h∗ R−1
j,l h =
1
L
X
t∈T (j,l)
1
.
σt2
(13)
The MSE expression (12) can be computed as MSE of symbol estimation in additive Gaussian noise and is a
function of the symbol constellation and the SNR which we denote by
µi−1 = gmse (γj,l )
The MSE for the case of binary symbol alphabet takes the form
h
√ 2i
g2 (γ) = E (1 − tanh (γ + ξ γ)) ,
(14)
ξ ∼ N (0, 1) .
Series expansion and a number of bounds and approximation for it have been derived in [48].
10
(15)
IEEE Transactions on Information Theory, under revision
In case of the uncoupled system all noise and interference components in (9) have the same power and, therefore,
σt2k = xi−1 for all k forming the diagonal covariance matrix Rj,l =
1
xi−1 I M
where I M is the M × M identity
matrix. Hence (10), (15), (13) lead to the characteristic equation of the uncoupled system
L−1
M −1
xi =
gmse
+ σ2 .
L
Lxi−1
(16)
Using a variable exchage we get the form of the characteristic equation which will be central to our analysis
1
0
+ σ 02
(17)
xi = gmse
α0 x0i−1
where
x0i = xi
L
,
L−1
σ 02 = σ 2
L
,
L−1
α0 =
L−1
.
M −1
(18)
This representation allows us to carry out a performance analysis of the system for finite parameter values of L
and M where the data stream size N is going to infinity for the purpose of allowing any number of i demodulation
iterations to be performed as message passing on respective trees.
Equation (17) is initialized with
x00 = 1 + σ 02
which constitutes the noise-and-interference power before the demodulation iterations start.
Depending on the value of σ 02 (17) can have a different number of fixed points (see Section III-D) but it always
converges to the largest fixed point below the initial state 1 + σ 02 .
B. Coupled System
We consider the regularized delay structure introduced in Section II-B where each data stream is divided into
2W + 1 subsections of length Nw . Since a data stream can only be transmitted a at the beginning of a subsection
we can assume that the receive sequence consists of subsections of length Nw as well and they are indexed by the
index t = 1, 2, 3, · · · . For the case of coupled system the noise-and-interference power and MSE values are not just
dependent on the demodulation iteration i but also on the time index t of the respective receive subsection.
In order to asses noise-and-interference power for subsection t we look again at equation (10). All variable
nodes connected to a channel node that belongs to receive subsection t are within subsections indexed by τ1 =
t − W, t − W + 1, · · · , t + W . In addition, in average (L − 1)/(2W + 1) variable nodes (j 0 , l0 ) connected to the a
single check node belong to each of the 2W + 1 subsections Hence
def
xti =
X
(j 0 ,l0 )∈J (t0 )
L−1
1
(i−1)
E(vj 0 ,l0 − v̂j 0 ,l0 ,τ )2 + σ 2 =
L
L(2W + 1)
t+W
X
1
µτi−1
+ σ2 .
(19)
τ1 =t−W
s.t. (j 0 ,l0 )6=(j,l)
1
In turn, to compute µτi−1
we notice that equation (12) implies computation of the MSE for symbols of a data
stream transmitted at time τ − W in terms of subsections. The M − 1 replicas of each data symbol in this data
11
IEEE Transactions on Information Theory, under revision
2
), τ2 = τ1 − W, · · · , τ2 + W . Again, in average each SNR is experienced by
stream experience SNRs 1/(Lxτi−1
(M − 1)/(2W + 1) symbols and we obtain
1
µτi−1
M −1
2W + 1
= gmse
τ1X
+W
τ2 =τ1 −W
1
Lxτ2
!
.
(20)
Combining (19) and (20) together applying a variable exchange (18) we obtain the characteristic equation of the
coupled system
xti
W
X
1
gmse
=
2W + 1
τ1 =−W
W
X
1
1
1
t+τ1 +τ2
0
α 2W + 1
x
τ =−W i−1
!
+ σ 02 ,
(21)
2
We assume that transmission starts at time t = 1. At every time instant t, a new set consisting of Lw data streams
is transmitted. Starting from t = 2W + 1 also a set of Lw is finished at each section t. As a result, the initial
conditions for recursion (21) can be formulated as
xt0 = 0 t ≤ 0 ,
xt0 =
(22)
t
gmse (0) + σ 02 ;
2W + 1
xt0 = 1 + σ 02
t ∈ [1, 2W + 1]
t > 2W + 1 .
(23)
(24)
C. Demodulation/Decoding Schedules
Two-Stage Schedule
For the case of uncoupled system the residual SINR per symbol after I demodulation iterations equals
1
α0 x0i .
Assuming that each data stream is encoded with the same error correction decoder with SNR threshold θ for the
case
1
≥θ
α0 x0i
we obtain asymptotically error-free performance where the limit is taken in therm of the code length N . For coupled
system the SINR for subsection t after I iterations equals
γIt =
W
X
1
1
t+j
0
α (2W + 1)
x
j=−W I
(see the argument of gmse (·) function in (21). Hence the subsequent error correction decoding performed at the
second stage of the two-stage receive process is successful for the block transmitted at time t iff
γIt ≥ θ.
(25)
Hard Decoding Feedback and Cancellation Schedule
For hard decision decoding and cancellation schedule the receiver performs the decoding of any data stream for
which SINR exceeds the error-correction code threshold θ and cancels the impact of that stream to the received
12
IEEE Transactions on Information Theory, under revision
signal. That means that any data streams with SINR above θ will no longer contribute to the noise-and-interference
power (17), (21). Let us define the MSE function which included the impact of error correction
0,
if γ ≥ θ
def
gmse,ecc (γ) =
gmse (γ), otherwise.
As soon as the SINR reaches the error correction code threshold the function is set to 0 accounting for the fact that
the respective data stream is eliminated from the cancellation process. The characteristic equations for the coupled
and uncoupled system are modified accordingly and we obtain
1
x0i = gmse,ecc
+ σ 02
α0 x0i−1
and
xti
W
X
1
=
gmse,ecc
2W + 1
τ1 =−W
W
X
1
1
1
t+τ1 +τ2
0
α 2W + 1
x
τ =−W i−1
(26)
!
+ σ 02 ,
(27)
2
The system converges iff x0i and xti for every t converge to σ 2 after a number of iterations indicating the residual
of the received signal contain just noise while all data streams have been decoded.
D. Fixed Points of the Characteristic Equation
The largest fixed point of the characteristic equation of the uncoupled system (17) determines how much the
noise and interference power can be reduced by the iterative demodulation process. Since the MSE function is
differentiable and strictly decreasing [48], [45] the values x0i are strictly decreasing with iterations i. Depending of
the modulation symbol alphabet, σ 02 , and α0 (17) may different number of roots.
We can define a threshold load value αs0 (σ 02 ) such that for α0 ∈ [0, αs0 (σ 02 )) (17) has only one root, two roots
for α0 = αs0 (σ 02 ), and there roots for α0 > αs0 (σ 02 ). Fig. 4 shows αs0 as a function of the SNR 1/σ 02 (circles) for
the case of 2-PAM (blue), 4-PAM (magenta) and 8-PAM (black) symbol alphabets.
Consider now the 2-PAM case and denote the three roots of the system by x(1) < x(2) < x(3) . Lemma 2 states
upper and lower bounds on the roots x(1) and x(3) which will be used in the proof of the main result stated in the
next section. In case (17) has single root we denote it by xs and it satisfies Lemma 2 (b).
Lemma 2. Consider σ 02 ≤ 1. Then the following statements are satisfied:
(a) for α0 ∈ [0, C(σ 02 )]
0
σ 02 ≤ x(1) ≤ (1 + e−1/σ )σ 02 ≤ 2σ 02 ,
(28)
(b) for α0 ∈ [4, C(σ 02 )]
(1 + σ 02 )2 −
2
3
≤ x(3) ≤ (1 + σ 02 )2 ,
0
α
(29)
(c) for α0 ∈ [4, C(σ 02 )]
1
1
1
(1 + σ 02 ) ≤ 1 + σ 02 − 0
−
≤ x(3) ≤ 1 + σ 02
2
α (1 + σ 02 ) α02 (1 + σ 02 )3
13
(30)
IEEE Transactions on Information Theory, under revision
6
5
4
3
2
1
0
0
5
10
15
SNR [dB]
20
25
30
35
Fig. 4. Rates achievable by coupled and uncoupled systems in comparison to channel capacity.
where
C(σ 02 ) =
1
1
log2 1 + 02 .
2
σ
(31)
is the capacity of the AWGN channel with SNR 1/σ 02 .
Proof. See Appendix B.
In contrast to the uncoupled system (17), the fixed points x∗ of the coupled system (21) are vectors. Again we
notice that due to the fact that the MSE function is decreasing the components xti of the coupled recursion vectors
are decreasing as well for each t = 1, 2, 3, · · · , and x∗t = limi→∞ xti , t = 1, 2, 3, · · · .
We show (see Appendix C) that the entries of the fixed points x∗ , are concentrated around the fixed points
x(1) and x(3) of the uncoupled characteristic equation (17) depending on the parameters α0 , σ 02 . Since the smallest
fixed point x(1) is close to the value of the noise power σ 2 (Lemma 2 (b)) we will also show that the system’s
performance approaches channel capacity. For fixed σ 02 let ᾱ denote the maximum load for which the coupled
system converges to x(1) , i.e., for any α ≤ ᾱ the solution of the coupled system (21) satisfies limi→∞ xti ≤ x(1) for
any t > 0 and limt→∞ limi→∞ xti ≤ x(1) for sufficiently large W . The critical load ᾱ is given by blue, magenta,
and black squares curves in Fig. 4 for 2-PAM, 4-PAM and 8-PAM symbol alphabets respectively.
In addition Fig. 4 demonstrates the load maximizing the rate achievable by the uncoupled system for each 1/σ 02
which is given by the blue stars (2-PAM), magenta starts (4-PAM) and black stars (8-PAM) curves. While we
introduce the achievable rate expressions in the next section we notice that the load maximizing the rate achievable
by the uncoupled system almost coincides with αs0 which is the highest load such that the system has a single root.
14
IEEE Transactions on Information Theory, under revision
The optimal loads for the uncoupled system coincide with the critical load ᾱ of the coupled system for low SNRs
but then diverge for high SNRs. The divergence points are different for different signal constellations. At high SNRs
the optimal loads of the uncoupled system saturate while the load ᾱ of the coupled system keeps increasing. This
behavior is consistent with the behavior of the individually optimal and jointly optimal MAP decoding thresholds
studied in [43], [44] (using replica method).
E. Achievable Rate and Capacity
For loads α0 < αs0 the uncoupled system’s noise-and-interference power converges to the single root of the
characteristic equation (17) x(1) and the corresponding SINR to (α0 x(1) )−1 . In order to determine the maximum
sum rate achieved by the entire system we assume that we use codes optimal with respect to the SINR to (α0 x(1) )−1
and the symbol alphabet of interest, i.e. codes achieving capacity of the AWGN channel with SNR (α0 x(1) )−1 and
the symbol alphabet of interest as an input. The total communication rate (sum-rate) achievable by the system for
α0 < αs0 is, therefore,
L
Ru (α, σ ) =
CA
M
2
1
α 0 xs
= αCA
1
α0 xs
,
(32)
where CA (γ) denotes the capacity of the AWGN channel with input alphabet A and SNR γ and xs is the single
root of (17). The factor α =
L
M
accounts for the number of data stream in the system and the repetition factor. Here
we recall that system (21) operates with total signal power 1 and noise power σ 2 and the corresponding capacity
of the (real-valued) AWGN channel for these parameters equals C(σ 2 ) (while σ 02 = σ 2 L/(L − 1)). For the case
when the characteristic equation has three roots, i.e., for α0 > αs0
1
2
Ru (α, σ ) = αCA
,
α0 x(3)
Then
2
Rcoup (α, σ ) ≥
(33)
αCA
1
α0 x(1)
, if α ≤ ᾱ, W ≥ W̄
αCA
1
α0 x(3)
, otherwise
(34)
for some W̄ (see Appendix C). We are now ready to state the main result of the paper.
Theorem 1. There exists a W̄ > 0 such that for any W > W̄
C(σ 2 ) − R(ᾱ, σ 2 ) ≤ Gasimptotic (σ 2 ) + Gfinite (L, M, σ 2 )
(35)
where
Gasimptotic (σ 2 ) =
4
C(σ 2 )
Gfinite (L, M, α, σ 2 ) =
1
log2
2
σ2 + 1
σ2 + 1 −
1
L
+
C(σ 2 ) − 1
M −1
(36)
for C(σ 2 ) ≥ 5, and, therefore,
lim
L.M →∞
Gfinite (L, M, α, σ 2 ) = 0 ,
lim Gasimptotic (σ 2 ) = 0 .
σ 2 →0
15
(37)
(38)
IEEE Transactions on Information Theory, under revision
5
Achievable rate [bits/dimension]
4
3
2
AWGN channel capacity
2-PAM uncoupled achievable rate
2-PAM coupled achievable rate
1
4-PAM uncoupled achievable rate
4-PAM coupled achievable rate
8-PAM uncoupled achievable rate
2-PAM coupled achievable rate
0
5
10
15
20
25
30
35
SNR [dB]
Fig. 5. Rates achievable by coupled and uncoupled systems in comparison to channel capacity.
The expression in Theorem 11 demonstrates the asymptotic behavior of the achievable rate and shows that the
gap between the achievable rate and the AWGN channel capacity tends to zero as the SNR 1/σ 2 increases.
The proof of the asymptotic capacity-achieving property involves potential functions [47], [46] which can be used
to characterize the conditions for convergence of the coupled system. While the proof in Appendix C uses [47]
type potential function here we use [46] type potential function to visualize the capacity approaching behavior
of the coupled systems for the two demodulation/decoding schedules discussed in Section II-E. We start with the
two-stage schedule. We consider the asymptotic scenario in this case where L, M → ∞; α = L/M is constant. A
variable exchange in the characteristic equation (17) leads to2
1
= αgmse (u) + ασ 2 .
u
(39)
Graphically the fixed points of (39) are demonstrated in Fig. 6 that shows MSE vs SNR for a 2-PAM system
with load α = 2.5 and σ 2 = 0.0254. The left hand side of the (39) is shown by the blue solid curve while the right
hand side of (39) is depicted by the red curve. The three black stars depict the three fixed points (1/(αx(3) ), αx(3) ),
(1/(αx(2) ), αx(2) ), and (1/(αx(1) ), αx(1) ) of (39). According to Lemma 2 the fixed points are contained between
(1/(α(1 + σ 2 )), α(1 + σ 2 )) and (1/(ασ 2 ), ασ 2 ). The two areas labeled P 1 and P 2 between the curves are equal
(this point is one of the key benefits of the potential function approach taken in [46]) since the parameter values α
and σ 2 are chosen such that the potential function is exactly equal to 0.
1A
slightly better bound on the asymptotic gap can be found in the previous version of this paper. It requires a much lengthier derivation.
2 This
equation appears in [46] where the system we study is discussed in an example. Here we make further steps relating it to the achievable
rate and capacity.
16
IEEE Transactions on Information Theory, under revision
3
2.5
C1
2
MSE
1.5
R1
P1
1
P2
C3
C2
0.5
M
0
-0.5
-2
0
2
4
6
8
10
12
14
SNR
Fig. 6. Graphic representation of the MSE vs SNR, two-stage demodulation/decoding.
The area colored in light red (divided by 2 ln 2) is equal to the total rate achieved by the coupled system
Z 1/(αx(1) )
1
2
Rcoup (α, σ ) =
αgmse (u)du ,
(40)
2
ln
2
0
while the area colored in blue (divided 2 ln 2) equals AWGN channel capacity
Z 1/(ασ2 )
1
1
du .
C(σ 2 ) =
2 ln 2 1/(α(1+σ2 )) u
(41)
We notice that the area R1 belongs to the achievable rate only which the areas C 1 , C 2 , C 3 (a tiny area between
1/(αx(1) ) and 1/(ασ 2 )) belongs exclusively to the AWGN capacity calculation. Both achievable rate and capacity
share the main area M . Similar situation would happen in case of no-binary symbol alphabet and, as well see later,
in case of other decoding schedules where usage of error-correction in the iterative demodulation loop impacts the
MSE curve.
For the pictorial representation we can see that the difference between the achievable rate and capacity can be
extracted from the areas R1 , C 1 , C 2 , C 3 where R1 and C 2 are the largest and play the main role. We can derive
17
IEEE Transactions on Information Theory, under revision
a simple lower bound on the gap-to-capacity
1
1
C(σ 2 ) − Rach (α, σ 2 ) =
(C 1 + C 2 + C 3 − R1 ) ≥
(C 2 − R1 )
2
ln
2
2
ln
2
1
1
1
1
1
1
1
2
2
2
−
=
ασ
−
αgmse (0) + σ + αgmse
+σ
2 ln 2
ασ 2
2 ln 2 α(1 + σ 2 )
4 ln 2 α(1 + σ 2 )
α(1 + σ 2 )
1
1 1
1
1
1
1
σ2
−
+
g
+
=
mse
2 ln 2 1 + σ 2
2 ln 2 1 + σ 2 2 2
α(1 + σ 2 )
α
1 1
1
1
1
σ2
=
− gmse
−
.
(42)
2 ln 2 1 + σ 2 2 2
α(1 + σ 2 )
α
For the case of 2-PAM signals we can generalize it further using the upper bound (1 + s)−1 ≥ g2 (x) derived in [48]
and obtain
1
C(σ ) − Rach (α, σ ) ≥
2 ln 2(1 + σ 2 )
2
2
1 1 α(1 + σ 2 )
σ2
−
−
2 2 1 + α(1 + σ 2 )
α
1
=
2 ln 2(1 + σ 2 )
1
σ2
−
2(1 + α(1 + σ 2 ))
α
(43)
As σ 2 → 0 the right hand side is approximately equal to
1
2α
≈
1
2C(σ 2 )
which is consistent with the behavior
showcased by Theorem 1. The upper bound on the gap derived in Theorem 1 involves estimation of all areas C 1 ,
C 2 , C 3 and relates estimation of the largest root x(3) of (17).
Consider now the hard decoding feedback and cancellation schedule. Once during the demodulation process a
data stream reaches SNR equal or exceeding the threshold θ of the error correction code the data is passed to the
error correction decoder. The decoder is then capable to perform error-free decoding and the effect of the data
stream would be perfectly cancelled once the decoder gives its feedback to the demodulator.
Fixed points of the modified characteristic equation
1
= αgmse,ecc (u) + ασ 2 .
u
(44)
as well as the areas between the curves characterizing the achievable rate and capacity are given in Fig. 7. The
achievable rate expression is now given by
Rcoup,ecc (α, σ 2 ) =
Z
γ
0
1
αgmse (u)du
2 ln 2
(45)
while the AWGN channel capacity is still computed as in (41). The gap to capacity in terms of the area is now
given by
1
1
(C 1 + C 2 − R1 ) ≥
(C 2 − R1 )
2 ln 2
2 ln 2
1
1 1
1
σ2
=
−
g
−
mse
2 ln 2(1 + σ 2 ) 2 2
α(1 + σ 2 )
α
C(σ 2 ) − Rcoup,ecc (α, σ 2 ) =
(46)
which is the same lower bound as in (42). At the same time we can derive an upper bound
1
C(σ 2 ) − Rcoup,ecc (α, σ 2 ) =
(C 1 + C 2 − R1 )
(47)
2 ln 2
1 1
1
1
1
1 1
1
σ2
≤
−
α(1 + σ 2 − xs ) +
− gmse
−
(48)
2
2
2
2 ln 2 2α xs
1+σ
2 ln 2(1 + σ ) 2 2
α(1 + σ )
α
18
IEEE Transactions on Information Theory, under revision
C1
10
8
P1
MSE
6
4
P2
R1
C2
2
M
0
-2
-2
0
2
4
6
8
10
12
SNR
Fig. 7. Graphic representation of the MSE vs SNR, one time hard decoding feedback and cancellation schedule.
where xs is the single root of the characteristic equation (17). If we focus on the case of binary symbols we can
now use Lemma 2 (c) and the lover bound 1 − y ≤ g2 (y) [48] and based on (48) further obtain
1
1 (1 + σ 2 − xs )2
1
1 1
1
σ2
2
2
C(σ ) − Rcoup,ecc (α, σ ) =
+
−
1−
−
2 ln 2 2 xs (1 + σ 2 )
1 + σ2 2 2
α(1 + σ 2 )
α
2
1
1
1 1 α(1+σ2 ) + α2 (1+σ2 )3
1 − 2σ 2 (1 + σ 2 )
≤
+
1
2
2
2 ln 2 2
2α(1 + σ 2 )2
2 (1 + σ )
3
1
1 − 2σ 2 (1 + σ 2 )
1
+
≤
.
≤
2 ln 2 α2 (1 + σ 2 )4
2α(1 + σ 2 )2
2 ln 2α
(49)
For the case of 2-PAM constellation the capacity-achieving property for hard decoding feedback and cancellation
schedule has been proven in [25]. Here we look at the problem in more detail and prove that in case of one time
feedback and cancellation schedule convergence to channel capacity happens for any fixed σ 2 when we let α → ∞.
For the two-stage decoding the system is only capacity achieving asymptotically as σ 2 → 0. We can formulate this
result as a theorem.
Theorem 2. For any σ 2 ≤ 0.1 (necessary for existence of at lease one root) and any α > 4 there exists a W̄ > 0
19
IEEE Transactions on Information Theory, under revision
such that for any W > W̄ such that
C(σ 2 ) − Rcoup,ecc (α, σ 2 ) ≤
1
2 ln 2α
(50)
and, therefore, for any σ 2 ≤ 0.1
lim C(σ 2 ) − Rcoup,ecc (α, σ 2 ) = 0 .
α→0
IV. N UMERICAL R ESULTS
A. Numerical Results
In this section we present numerical results on the spectral efficiency achievable by the coupling data transmission.
Fig. 8 plots achievable transmission rates (sum rates) in bits per second per Hertz as a function of the total SNR
1/σ 2 in dB. The total system power is normalized to 1. The red curve corresponds to the capacity of the real-valued
AWGN channel with total signal power equal to 1. The pink curves correspond to capacities of the AWGN channels
with constrained inputs. The lower curve is for BPSK modulation, and the other curves are for 4,8, and 16 PAM
(from bottom to top).
The black curve with circles plots simulated sum rate of the coupled data transmission. Block length M N = 10000
has been selected for all points. The three simulated point correspond to loads α = 2, 3, 4 with parameter M = 250
and L chosen accordingly. The coupling window W = 12 was considered. The permutation matrices for each
data stream have been chosen randomly with no attempt to eliminate cycles. The simulation was performed for
the demodulator. The SNR at the output of the demodulator is measured and the resulting sum-rate is computed
assuming component codes achieving capacity of the BIAWGN channel for this SNR. We notice that simulated
performance curve is close to the curve computed numerically (blue). Also for each α = 2, 3, 4 the simulated
performance is closer to the channel capacity than the performance of the corresponding 2,4,8, or 16-PAM. Finally,
we notice that the performance curve does not saturate at the modulation load of α = 2.07 as it happens for the
uncoupled system. The slight divergence for the higher load is likely caused by the presence of short cycles in the
system’s graph which becomes denser as the system load increases.
V. C ONCLUSION
We consider modulation of information in a form a superposition of independent equal power and equal rate
data streams. Each stream is formed by repetition and permutation of data and the streams are added up with an
offset initiating the effect of “stream coupling”. We prove that the proposed system used with iterative demodulation
followed by external error control decoding achieves the capacity of the AWGN channel and Gaussian multipleaccess channel asymptotically and asses the gap to capacity as a function of the system’s parameters. For the hard
feedback and cancellation schedule we show that the gap to capacity can vanish at finite SNR.
20
IEEE Transactions on Information Theory, under revision
5
AWGNmCapacity
BPSKmCapacity
4−PAMmCapacity
8−PAMmCapacity
16−PAMmCapacity
SpatiallymCoupledmComputed
SpatiallymCoupledmSNR−Simulated
4.5
4
Rate [bits/dimension]
3.5
3
2.5
2
1.5
1
0.5
0
0
5
10
15
20
25
SNR [dB]
Fig. 8. Computed and simulated rates achievable by the coupled transmission.
R EFERENCES
[1] A. Jiménez Felström and K. Sh. Zigangirov, “Time-varying periodic convolutional codes with low-density parity-check
matrices,” IEEE Trans. Inf. Theory, vol. 45, no. 6, pp. 2181–2191, Sept. 1999.
[2] M. Lentmaier, A. Sridharan, D. J. Costello, Jr., and K. Sh. Zigangirov, “Iterative decoding threshold analysis for LDPC
convolutional codes,” IEEE Trans. Inf. Theory, vol. 56, no. 10, pp. 5274–5289, Oct. 2010.
[3] S. Kudekar, T. J. Richardson, and R. L. Urbanke, “Threshold saturation via spatial coupling: why convolutional LDPC
ensembles perform so well over the BEC,” IEEE Trans. Inf. Theory, vol. 57, no. 2, pp. 803–834, Feb. 2011.
[4] C. Schlegel and D. Truhachev,“Multiple Access Demodulation in the Lifted Signal Graph with Spatial Coupling,” IEEE
Transactions on Information Theory, accepted for publication, 2012, arXiv:1107.4797.
[5] H. Imai and S. Hirakawa, “A new multilevel coding method using error correcting codes,” IEEE Transactions on Information
Theory, vol. 23, pp. 371–377, May 1977.
[6] G. Caire, G. Taricco, and E. Biglieri, “Bit-interleaved coded modulation,” IEEE Trans. Inform. Theory, vol. 44, pp. 927–946,
May 1998.
[7] A. Yedla, P. S. Nguyen, H. D. Pfister, and K. R. Narayanan, “Universal codes for the Gaussian MAC via spatial coupling,”
in Proc. Annual Allerton Conf. on Commun., Control, and Comp., Monticello, IL, Sept. 2011
[8] S. Kudekar and K. Kasai“Spatially coupled codes over the multiple access channel,” IEEE International Symposium on
Information Theory Proceedings (ISIT), pp. 2816–2820, St Petersburg, Russia, 2011.
[9] D. Guo and C. C. Wang, “Multiuser Detection of Sparsely Spread CDMA,” in IEEE Journal on Selected Areas in
Communications, vol. 26, no. 3, pp. 421-431, April 2008.
[10] C. Schlegel, M. Burnashev, and D. Truhachev, “Generalized superposition modulation and iterative demodulation: A
capacity investigation,” Hindawi Journal of Electr. and Comp. Eng., vol. 2010, Sep. 2010.
21
IEEE Transactions on Information Theory, under revision
[11] D. Truhachev, C. Schlegel, and L. Krzymien, “A two-stage capacity-achieving demodulation/decoding method for random
matrix channels,” IEEE Tran. on Inform. Theory, vol. 55, no. 1, pp. 136–146, Jan. 2009.
[12] L. Ping, L. Liu, K. Wu, and W. Leung, “Interleave division multiple-access,” IEEE Trans. Wireless Commun., vol. 5, no. 4,
pp. 938–947, April 2006.
[13] L. Liu, J. Tong, and L. Ping, “Analysis and Optimization of CDMA Systems with Chip-Level Interleavers,” IEEE Journal
on Selected Areas in Communications, pp. 141-150, Jan. 2006.
[14] L. Duan, B. Rimoldi, R. Urbanke, “Approaching the AWGN channel capacity without active shaping,” Proc. IEEE Int.
Symp. Inform. Theory, p. 374, June–July 1997.
[15] X. Ma, L. Ping, “Coded modulation using superimposed binary codes,” IEEE Trans. Inf. Theory, vol. 50, no. 12, pp.
3331–3343, Dec. 2004.
[16] P. A. Hoeher, H. Schoeneich, J. C. Fricke, “Multi-layer interleave-division multiple access: theory and practice,” European
Transactions on Telecommunications, vol. 19, no. 5, pp. 523–536, Aug. 2008.
[17] H. S. Cronie, “Signal shaping for bit-interleaved coded modulation on the AWGN channel”, IEEE Trans. Commun., vol.
58, no. 12, pp. 3428–3435, Dec. 2010.
[18] P. A. Hoeher and T. Wo, “Superposition modulation: myths and facts,” IEEE Commun. Mag., vol. 49, no. 12, pp. 110–116,
Dec. 2011.
[19] S. Kudekar and T. Richardson and R. Urbanke, “Spatially Coupled Ensembles Universally Achieve Capacity under Belief
Propagation,” submitted to IEEE Trans. Inf. Theory, 2012, arXiv:1201.2999.
[20] S. Kudekar and H. D. Pfister, “The effect of spatial coupling on compressive sensing,” in Proc. Allerton Conf. on
Communications, Control, and Computing, Monticello, IL, Sept. 2010.
[21] C. Schlegel and D. Truhachev, “Multiple Access Demodulation in the Lifted Signal Graph with Spatial Coupling,” in Proc.
IEEE Int. Symp. on Inf. Theory, St. Petersburg, Russia, Aug. 2011.
[22] D. Truhachev, “Universal Multiple Access via Spatially Coupling Data Transmission”, IEEE International Symposium on
Information Theory 2013, Istanbul, Turkey, July 2013.
[23] K. Takeuchi, T. Tanaka, and T. Kawabata, “Improvement of BP-based CDMA multiuser detection by spatial coupling,” in
Proc. IEEE Int. Symp. on Inf. Theory, St. Petersburg, Russia, Aug. 2011.
[24] M. Hagiwara, K. Kasai, H. Imai, and K. Sakaniwa, “Spatially coupled quasi-cyclic quantum LDPC codes,” in Proc. IEEE
Int. Symp. on Inf. Theory, St. Petersburg, Russia, Aug. 2011.
[25] D. Truhachev, “Achieving AWGN Multiple Access Channel Capacity with Spatial Graph Coupling,” IEEE Communications
Letters, vol. 16, no. 5, pp. 585–588, May 2012.
[26] D. Guo, S. Shamai, and S. Verdu, “Mutual Information and Minimum Mean-Square Error in Gaussian Channels,” IEEE
Trans. Inf. Theory, vol. 51, no. 4, pp. 1261–1282, April 2005.
[27] S. Kumar, A. Young, N. Macris, and H. Pfister, “A Proof of Threshold Saturation for Irregular LDPC Codes on BMS
Channels,” in Proc. Allerton Conf. on Communication, Control, and Computing, Allerton, Illinois, USA, Sept. 2012.
[28] T. Tanaka, “A statistical mechanics approach to large-system analysis of CDMA multiuser detectors,” IEEE Trans. Inf.
Theory, vol. 48, no. 11, pp. 2888–2910, Nov. 2002.
[29] M. S. Alencar, “A comparison of bounds on the capacity of a binary channel,” in Proc. Global Telecommun. Conf., Nov.
1996, vol. 2, pp. 1273–1275.
[30] A. Yedla, Y.-Y. Jian, P. S Nguyen and H. D. Pfister, “A Simple Proof of Threshold Saturation for Coupled Scalar Recursions,”
in 7th International Symposium on Turbo Codes and Iterative Information Processing, Göthenburg, Sweden, August 2012.
[31] D. Divsalar, H. Jin, and R. J. McEliece, “Coding theorems for “turbo-like” codes,” in Proc. 36th Allerton Conf. on
Communication, Control, and Computing, pp. 201–210, Allerton, Illinois, USA, Sept. 1998.
22
IEEE Transactions on Information Theory, under revision
[32] M. Burnashev, C. Schlegel, W. Krzymien, and Z. Shi, “Characteristics Analysis of Successive Interference Cancellation
Methods”, Problemy Peredachi Informatsii, vol. 40, no. 4, pp. 297–317, Dec. 2004.
[33] R. Gallager, Information Theory and Reliable Communications, Wiley & Sons, New York, 1968.
[34] K. Takeuchi, T. Tanaka, and T. Kawabata, “A Phenomenological Study on Threshold Improvement via Spatial Coupling,”
IEICE Transactions, vol. 95-A, no. 5, pp. 974–977, 2012.
[35] K. Takeuchi, T. Tanaka and T. Kawabata, “Performance Improvement of Iterative Multiuser Detection for Large Sparsely
Spread CDMA Systems by Spatial Coupling,” in IEEE Transactions on Information Theory, vol. 61, no. 4, pp. 1768-1794,
April 2015.
[36] D. L. Donoho, A. Javanmard, and A. Montanari, “Information-Theoretically Optimal Compressed Sensing via Spatial
Coupling and Approximate Message Passing,” IEEE Trans. Inf. Theory, Dec. 2013.
[37] R. G. G ALLAGER , “Low Density Parity Check Codes”, M.I.T. Press, 1963.
[38] M. Lentmaier, D. Truhachev, and K. S. Zigangirov, “On the theory of low density convolutional codes II,” Problems of
Information Transmission, vol. 37, pp. 288–306, Oct.-Dec. 2001.
[39] D. Truhachev, M. Lentmaier, and K. S. Zigangirov, “Mathematical Analysis of Iterative Decoding of LDPC Convolutional
Codes,” Proc. IEEE Int. Symp. on Inf. Theory, p. 191, Washington, USA, June 24–29, 2001.
[40] G. A. Margulis, “Explicit Constructions of Graphs without Short Cycles and Low-Density Codes,” Combinatorica, vol. 2,
no. 1, pp. 71–78., 1982.
[41] S. Verdú, “Multiuser Detection.” Cambridge, U.K.: Cambridge Univ. Press, 1998.
[42] D. N. C. Tse and S. Verdú, “Optimum Asymptotic Multiuser Efficiency of Randomly Spread CDMA,” IEEE Trans. Inf.
Theory, vol. 46, no. 7, pp. 2718–2722, Nov. 2000.
[43] R. R. Müller and W. H. Gerstacker, “On the Capacity Loss Due to Separation of Detection and Decoding,” IEEE Trans.
Inf. Theory, vol. 50, no. 8, pp. 1769–1778, Aug. 2004.
[44] D. Guo and S. Verdú, “Randomly Spread CDMA: Asymptotics Via Statistical Physics,” IEEE Trans. Inf. Theory, vol. 51,
no. 8, pp. 1983–2010, June. 2005.
[45] D. Guo, Y. Wu, S. Shamai, and S. Verdú, “Estimation in Gaussian Noise: Properties of the Minimum Mean-Square Error,”
IEEE Trans. Inf. Theory, vol. 57, no. 4, pp. 2371–2385, Apr. 2011.
[46] S. Kudekar, T. J. Richardson and R. L. Urbanke, “Wave-Like Solutions of General 1-D Spatially Coupled Systems,” in
IEEE Transactions on Information Theory, vol. 61, no. 8, pp. 4117–4157, Aug. 2015.
[47] A. Yedla, Y. Y. Jian, P. S. Nguyen and H. D. Pfister, “A Simple Proof of Maxwell Saturation for Coupled Scalar Recursions,”
in IEEE Transactions on Information Theory, vol. 60, no. 11, pp. 6943-6965, Nov. 2014.
[48] M. Burnashev, C. Schlegel, W. Krzymien, and Z. Shi, “Characteristics Analysis of Successive Interference Cancellation
Methods”, Problemy Peredachi Informatsii, vol. 40, no. 4, pp. 297–317, Dec. 2004.
[49] J. Barbier, M. Dia, N. Macris, “Proof of Threshold Saturation for Spatially Coupled Sparse Superposition Codes,” Proc.
IEEE Int. Symp. on Inf. Theory, Barcelona, Spain, July 2016.
[50] A. Barron and A. Joseph, “Toward fast reliable communication at rates near capacity with Gaussian noise,” in Information
Theory Proceedings (ISIT), 2010 IEEE International Symposium on, June 2010, pp. 315319.
[51] C. Rush, A. Greig, and R. Venkataramanan, “Capacity-achieving Sparse Regression Codes via Approximate Message
Passing Decoding,” in Information Theory Proceedings (ISIT), 2010 IEEE International Symposium on, Hong Kong, July
2015.
23
IEEE Transactions on Information Theory, under revision
A PPENDIX
A. Proof of Lemma 1
We start by observing the fact that each modulated data stream ṽ l can be represented as a product of the data
sequence v l and a binary matrix H which represents repetition and permutation operations,
ṽ l = H l v l ,
i.e., H l = P l R, where P l is an M N × M N binary permutation matrix, and M N × N matrix R is a binary
repetition matrix, repeating each bit of v l M times,
1 1 1 ··· 1
1 1
R=
T
1
···
1
..
.
1
1
1
···
(51)
1
Without loss of generality we consider M1 = M , M2 = 1 here. Further, when we consider coupling of the data
streams in the channel as described in the end of Section II-B we can represent the resulting modulated sequence
s = Hv = CP R̃v,
where v = (v 1 , v 2 , · · · ) is the vector of all data and C is the coupling matrix. The coupling matrix C (see Fig. 9)
consists of blocks shown by the dashed rectangles where each block consists of (2W + 1) × (2W + 1) identity
matrices I of size Nw × Nw . Each consequent block is shifted down by the size of one identity matrix. The
I I
I I
I I
C=
I
I
I
I
I
I I
I I
I I
I
I
I
I I
I I
I I
I
I
I
I
I
I
I
I
I
I
Fig. 9. Coupling Matrix.
permutation matrix P is a block-diagonal matrix which consists of permutation matrices of individual data streams
P = diag (P 1 , P 2 , P 3 , P 4 , · · · ) and the repetition matrix R̃ is constructed as given in (51) but is infinitely long.
Here we leave the signature sequence out of the consideration since they do not affect the structure of the system’s
graph.
24
IEEE Transactions on Information Theory, under revision
The resulting matrix H is a band-diagonal binary matrix with M ones in each row and L ones in each column.
Such a matrix can be a syndrome former (infinite parity check matrix) of an LDPC convolutional code. An example
of such matrix for L = 6 and M = 3 is given in Fig. 10. This particular matrix is constructed using P consisting
of smaller permutation matrix sub-blocks.
Fig. 10. Multiple convolutional permutor matrix.
It is interesting to note that if the addition of our binary modulated data streams was modulo 2 the resulting
sequence would be a codeword of an (L, M ) spatially coupled LDPC code. Such code is graphically represented
by an infinitely-long bipartite Tanner graph [1][38]. The same bipartite graph describes our modulation system (see
Fig. 3 b)). Several researchers showed that it is possible to construct such bipartite graphs of block LDPC codes
without any cycles of length 2I for any fixed integer I [37](Appendix B), [40] if the length of the code is chosen
large enough. Finally, it has been shown that periodic LDPC constitutional codes with graphs of any given girth I
can be constructed from block LDPC code graphs of large girths by simple unwrapping procedure [39][38]. This
directly implies that modulation graphs of any given girth I can be constructed if M N is chosen to be large enough.
Once graph with no short cycles is constructed we note that the demodulation operations, described in Section 2,
actually correspond to message passing on a modulation graph. The operation of symbol estimation (5) (see Fig. 3)
and the operation of interference cancellation (6) both can be regarded as message exchange, always excluding
one node. As a result, on a graph with no cycles of length up to 2I demodulation with I iterations performed to
estimate each bit vj,l is actually operating on a tree. Therefore, the analysis described above assuming no reuse of
information is precise. We note here that construction of modulation graphs with large girth is done for the purpose
25
IEEE Transactions on Information Theory, under revision
of analysis while in practice it is typically enough to expurgate cycles of length four and six.
B. Proof of Lemma 2
Let us define
def
h(x) = g2
1
α0 x
+ σ 02 ≤ σ 02 ,
(52)
where the inequality follows from g2 (y) > 0, for y ∈ (0, ∞). Similarly the fact that g2 (·) is non-negative implies
the lower bound in Lemma (28) (a)
σ 02 ≤ g2
1
α0 x(1)
+ σ 02 = x(1) .
We use the upper bound [48]
√
g2 (y) ≤ πQ( y) = π
Z
∞
√
y
z2
1
√ e− 2 dz
2π
and the condition α0 ≤ C(σ 02 ) of Lemma (28) to upper bound
!
1
1
1 p
02
02
02
02
h(2σ ) ≤ πQ p
C(σ )2σ exp −
+ σ ≤ π√
+ σ 02
C(σ 02 )σ 2
2π
C(σ 02 )2σ 02
1
1 π 1
02
02
= exp −
+ ln + ln(2σ C(σ )) + σ 02
C(σ 02 )σ 02
2 2
2
1 π 1
1
02 1
02
+ ln + ln 2σ
= exp − 02 1
ln(1 + 1/σ )
+ σ 02
2
2 ln 2
σ 2 ln 2 ln(1 + 1/σ 02 ) 2 2
1
≤ exp − 0 + 2 ln(σ 0 ) + σ 02 < 2σ 02 .
σ
(53)
(54)
(55)
(56)
(57)
(58)
The first inequality in (58) is valid for σ 02 ≤ 1 and is due to the fact that the term −1/σ 02 dominates the exponent in
(57) as σ 02 → 0. Bounds (52) and (55)–(58) imply the existence of a root σ 02 ≤ x(1) ≤ 2σ 02 for any α0 ∈ [0, C(σ 02 )]
when σ 02 ≤ 1. The root x(1) then satisfies the first inequality in (58) which is the upper bound given by (28).
Lemma 2 (a) is proved.
The upper bound on x(3) in (29) is straightforward since g2 (y) ≤ 1 for y ∈ [0, ∞] and, therefore,
1
x(3) = g2
+ σ 02 ≤ 1 + σ 02 .
α0 x(3)
To prove the lower bound in (29) we use the lower bound [48]
g2 (y) ≥ 1 − y,
y ∈ [0, ∞] .
(59)
Inequality (59) implies that xr , the largest root of the equation
x=1−
1
+ σ 02 ,
α0 x
(60)
is a lower bound on x(3) . Hence
s
s
"
!#2
!
2
02 2
02
1
+
σ
1
+
σ
4
4
4
2
(3)
x
≥ (xr ) =
≥
+2 1− 0
1+ 1− 0
2− 0
2
α (1 + σ 02 )2
4
α (1 + σ 02 )2
α (1 + σ 02 )2
2
2
2
1 + σ 02
1 + σ 02
1
4
3
≥
− 0+
1− 0
= 1 + σ 02 − 0 .
(61)
2
α
2
α (1 + σ 02 )2
α
26
IEEE Transactions on Information Theory, under revision
Since the value inside the square root in (61) needs to be non-negative the bound can be used for
4
≤ α0 ,
(1 + σ 02 )2
i.e. for α0 ≥ 4 in particular. Lemma 2 (b) is proved.
To prove the lower bound in (30) we simply refine the bound on xr via a Taylor series expansion of
q
1−
4
α0 (1+σ 02 )2 .
Lemma 2 is proved.
C. Proof of Theorem 1
Proof. We start with deriving a convergence condition for the coupled system (21) using the method of potential
functions [47]. The use of potential function for coupled systems was suggested in [34], [35] where the authors
show that these functions relates to the Bethe free energy of continuous dynamical system describing a coupled
model of a code-division multiple-access (CDMA) system and in [30],[46], where the technique was designed for
more general types of coupled recursions. The potential function has also been used in [36] to study a coupled
dynamical system describing reconstruction for compressed sensing.
The main part of the proof is dedicated to the derivation of the limiting load α∗ for which the coupled system
converges to the nearly interference-free state. To do so we use a relation between the mutual information and the
MSE [26] to derive an analytically tractable lower bound on the minimum of the potential function. Finally, we use
bounds on the roots x(1) and x(3) of the convergence equation, provided by Lemma 2, to compute the achievable
communication rate R(α∗ , σ 2 ) and prove that it is within a small gap from the channel capacity C(σ 2 ), and derive
the asymptotic behavior of this gap.
We focus on the case of 2-PAM symbols and for a fixed pair of parameters α0 , σ 02 such that α0 ∈ [4, C(σ 02 )]
consider the recursion
xi = g2
1
0
α (xi−1 + x(1) )
+ σ 02 − x(1)
(62)
equivalent to (17) in a sense that the fixed points of (62) are equal to these of (17) reduced by x(1) = x(1) (α0 , σ 02 )
which is the smallest fixed point of (17) and are given by 0, x(2) − x(1) , and x(3) − x(1) . The coupled recursion
corresponding to (62) is given by
W
X
1
t
xi =
g2
2W + 1
τ1 =−W
W
X
1
1
1
t+τ1 +τ2
α0 2W + 1
x
+ x(1)
τ =−W i−1
!
+ σ 02 − x(1) ,
t > 0, i > 0
(63)
2
and is, in turn, equivalent to (21) from the convergence perspective. The resulting fixed point sequences for (21)
and (63) differ by x(1) . The recursion (62) can be stated in the form
xi = f (g(xi−1 ), α0 )
(64)
where
def
1
α0
1
−x
x(1)
1
def 1
g(x) = (1) − (1)
.
x
x +x
f (x, α) = g2
27
+ σ 02 − x(1) ,
(65)
(66)
IEEE Transactions on Information Theory, under revision
The potential function of the uncoupled system (64) is given [47] by
σ 02 x
x + x(1)
−
− α0
U (x, α ) = ln
x(1)
x(1) (x + x(1) )
0
Z
1
α0 x(1)
g2 (y)dy .
(67)
1
α0 (x+x(1) )
and based on the fact that f, g are increasing and differentiable satisfying admissibility conditions we can utilize
Theorem 1 [47] which states that the fixed points of (62) satisfy
max x∗t ≤ max arg min U (x, α0 ) .
x
t≥0
for sufficiently large W . Hence, if for α0 , σ 02 the potential function U (x, σ 2 ) = 0 for x = 0 and U (x, σ 02 ) > 0 for
x > 0 the components of the fixed point x∗ of the original coupled system (21) do not exceed x(1) . Let us define
α∗ (σ 02 ) = sup α : min 2 U (x, α0 ) ≥ 0 .
(68)
x∈[0,1+σ ]
0
∗
For any α < α (63) converges to 0 and therefore the original coupled system (21) converges to a value not
exceeding x(1) . We will focus of estimating α∗ and the respective load Rcoup (α∗ , σ 02 ) which we now know is
achievable.
The function U (x, α0 ) is plotted in Fig. 11 for σ 02 = 0.0129, 0.003347, and 0.000855 and the corresponding
α∗ = 3, 4, 5. Note that α∗ is a decreasing function of σ 02 .
3
2.5
2
U(x,α)
α=5
1.5
α=4
1
α=3
0.5
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
Fig. 11. Plot of U (x, α0 ) for α0 = 3 and σ 02 = 0.0129 (blue curve), α0 = 4 and σ 02 = 0.003347 (magenta curve), and α0 = 5 and
σ 2 = 0.000855 (black curve).
28
IEEE Transactions on Information Theory, under revision
To find the minimum of U (x, α0 ) defined by (67) for a given α0 and σ 02 we compute its partial derivative with
respect to x and equate it to 0
σ 02
∂U (x, α0 )
1
−
− g2
=
(1)
∂x
x+x
(x + x(1) )2
1
α0 (x + x(1) )
1
=0.
(x + x(1) )2
(69)
which is equivalent to
x+x
(1)
= g2
1
0
α (x + x(1) )
+ σ 02 .
(70)
Since we consider α0 ∈ (4, C(σ 02 )) and 4 > αs0 (70) has three roots and, therefore, U (x, α0 ) has two local minima
achieved at x = 0 and xm = x(3) − x(1) . Moreover, for α0 = α∗ we have U (xm , α∗ ) = 0. Computing U (xm , α0 )
leads to
U (xm , α0 ) = U (x(3) − x(1) , α0 ) = ln
(3)
(3)
(1)
x −x
x
−
σ 02 − α
x(1)
x(1) x(3)
1
α0 x(1)
Z
def
g2 (y)dy = u(α0 , σ 02 ) .
(71)
1
α0 x(3)
We will now focus on the function u(α0 , σ 02 ), compute lower bounds for it and fund out when these are positive
in order to find a lower bound on α∗ (σ 02 ) and the achievable rate R(α, σ 2 ).
The relationship between MSE (i.e. function g2 (·)) and the mutual information for BIAWGN channel, derived
in [26] implies
1
α0 x(1)
Z
g2 (y)dy = 2 ln 2 CBIAWGN
1
α0 x(1)
− CBIAWGN
1
α0 x(3)
.
(72)
1
α0 x(3)
Applying (72) to (71) leads to
1 02
1 02
x(3)
1
1
0
0
−
σ
+
σ
−
2
ln
2α
C
+
2
ln
2α
C
BIAWGN
BIAWGN
x(1)
x(1)
x(3)
α0 x(1)
α0 x(3)
1
+ ξ(α0 , σ 02 ) ,
= 2 ln 2 C(σ 02 ) − α0 CBIAWGN
α0 x(1)
≥ 2 ln 2 C(σ 02 ) − α0 + ξ(α0 , σ 02 ) .
u(α0 , σ 02 ) = ln
(73)
(74)
(75)
where
def
ξ(α0 , σ 02 ) = ln
x(3)
1
1
− (1) σ 02 + (3) σ 02 + 2 ln 2α0 CBIAWGN
x(1)
x
x
1
α0 x(3)
− 2 ln 2
1
1 + σ 02
log2
2
σ 02
.
We can see that in case the potential function u(α0 , σ 02 ) ≥ 0 it consist of the gap between a term α0 CBIAWGN
(76)
1
,
α0 x(3)
which resembles the achievable rate Rcoup (α, σ 2 ) to the channel capacity C(σ 02 ) (74) (evaluated the point σ 02 close
to σ 2 ) and an additional function ξ(α0 , σ 02 ) (76).
In order to lower bound ξ(α0 , σ 02 ) we use the bounds on the capacity of the BIAWGN channel [33][29]
γ − γ2
≤ CBIAWGN (γ) ≤ 1 .
2 ln 2
(77)
We apply the lower bound in (77) is valid for γ < 1 to the last term in (73). Lemma 2 (b) implies that the condition
1/(α0 x(3) ) < 1 is satisfied for α0 > 4. We also apply the upper bound in (77) to the fourth term in (76) and obtain
1
1
1
2 ln 2αCBIAWGN
≥ (3) −
(78)
2 .
0
α0 x(3)
x
α x(3)
29
IEEE Transactions on Information Theory, under revision
Expanding the first term in (76) leads to
1
1
1
x(3)
1 + σ 02
x(3)
σ 02
+
+
.
ln (1) = 2 ln 2
log2
ln
ln
2
σ 02
2 ln 2 1 + σ 02
2 ln 2 x(1)
x
(79)
We use Lemma 2 (a) and (b) to lower bound the second and third terms in (76)
−
1
x
σ 02 +
(1)
1
x
σ 02 ≥ − 1 +
(3)
σ 02
1 + σ 02
(80)
Combining (78), (79), (80) with (76) leads to
ξ(α0 , σ 02 ) ≥ ln
σ 02
1
1
1
x(3)
+
ln
−
+ (3) −
2 .
02
02
(1)
1+σ
1+σ
x
x
α0 x(3)
(81)
Applying Lemma 2 to (81) we lower bound the first, second, fourth and fifth terms and obtain
2
1 + σ 02 − α30
1
1
1
σ 02
1
0
02
+
− 0
ξ(α , σ ) ≥ ln
+ ln
−
1
2
2
1 + σ 02
α (1 + σ 02 )2 −
σ 02 (1 + e− σ ) 1 + σ 02
(1 + σ 02 )
1
1
3
1
1
≥ ln 1 − 0 − ln(1 + e− σ0 ) − 0
2
α
α 1 − α30
def 3
α0
Let us define κ =
3
α0
(82)
∈ [0, 0.75] (since we consider α0 > 4) and perform a series expansion for the left side of
(82)
∞
∞
1
1
1
1 X κi
1X i
3κ
9
κ − ln(1 + e− σ0 ) ≥ −
ξ(α , σ ) ≥ −
−
− ln(1 + e− σ0 ) = − 0 − ln(1 + e− σ0 )
2 i=1 i
3 i=1
2
2α
0
02
(83)
where we use a simple linear bound for the above series. The bound is valid for α0 ≥ 4.
Coming back to inequality (75) and utilizing the lower bound (83) on ξ(α0 , σ 02 ) we obtain
1
9
u(α0 , σ 02 ) ≥ 2 ln 2 C(σ 02 ) − α0 − 0 − ln(1 + e− σ0 )
2α
(84)
for α0 > 4. If we substitute a candidate solution
α0 = α∗∗ = C(σ 02 ) −
for C(σ 2 ) ≥ 5 into (84) and obtain
∗∗
2
u(α , σ ) ≥ 2 ln 2 C(σ 02 ) − C(σ 02 ) +
=
8 ln 2
9
−
C(σ 02 ) 2 C(σ 02 ) −
3.9
C(σ 02 )
3.9
C(σ 02 )
−
(85)
9
2 C(σ 02 ) −
1
3.9
C(σ 02 )
− ln(1 + e− σ0 )
1
3.9
C(σ 02 )
− ln(1 + e− σ0 ) > 0
(86)
(87)
C(σ 02 ) ≥ 5 (leading to α∗∗ > 4). Hence for α∗∗ (85) the potential function is positive. We can now bound the gap
to capacity based on parameters α0 , σ 02
1
1
3.9
02
∗∗
02
C(σ ) − α CBIAWGN
≤ C(σ ) 1 − CBIAWGN
+
∗∗
(1)
∗∗
(1)
C(σ 02 )
α x
α x
Z ∞
Z ∞
3.9
3.9
√
≤ 2 ln 2C(σ 02 )
g2 (γ)dγ +
≤ 2 ln 2C(σ 02 )
πQ( γ)dγ +
02
1
1
C(σ )
C(σ 02 )
α0 x(1)
α∗∗ x(1)
1
3.9
1
3.9
4
02
≤ 2 ln 2C(σ 02 )2π exp − (1) +
≤
2
ln
2C(σ
)2π
exp
−
+
≤
02
02
02
C(σ )
4σ
C(σ )
C(σ 2 )
2x
30
(88)
(89)
(90)
IEEE Transactions on Information Theory, under revision
where we used upper bound on the MSE function from [48], the MSE-capacity relationship [44], and then a upper
√
bound Q( γ) ≤ − exp(γ/2).
We then bound the gap to the actual capacity
C(σ 2 ) − C(σ 02 ) =
1
log2
2
σ2 + 1
2
σ +1−
1
L
and achievable rate
α CBIAWGN
1
α∗∗ x(1)
≤ α∗∗ − α =
L−1
L
α−1
C(σ 2 ) − 1
−
≤
≤
M −1 M
M −1
M −1
∗∗
∗∗
2
∗∗
− Rcoup (α , σ ) ≤ α CBIAWGN
The Theorem is proved.
31
1
α∗∗ x(1)
− αCBIAWGN
1
α∗∗ x(1)
(91)
(92)
| 7cs.IT
|
arXiv:1503.00910v2 [math.AC] 26 Aug 2015
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
Abstract. In this paper we introduce an algorithm for computing the Stanley
depth of a finitely generated multigraded module M over the polynomial ring
K[X1 , . . . , Xn ]. As an application, we give an example of a module whose Stanley depth is strictly greater than the depth of its syzygy module. In particular,
we obtain complete answers for two open questions raised by Herzog in [Her13].
Moreover, we show that the question whether M has Stanley depth at least r can
be reduced to the question whether a certain combinatorially defined polytope P
contains a Zn -lattice point.
1. Introduction
Let K be a field. Let R = K[X1 , . . . , Xn ] be the standard Zn -graded polynomial ring, and let M = ⊕Ma be a finitely generated Zn -graded R-module (also
called multigraded in the sequel). The Stanley depth of M, denoted sdepth M, is a
combinatorial invariant of M related to a conjecture of Stanley from 1982 [Sta82,
Conjecture 5.1], which states that, in the case when K is infinite, the inequality
depth M ≤ sdepth M holds; this is nowadays called the Stanley conjecture. We refer
the reader to [PSFTY09] for a short introduction to the subject and to [Her13] for
a comprehensive survey.
After the initial submission of this paper, a counterexample to Stanley’s conjecture was given by Duval, Goeckner, Klivans, and Martin in [DGKM15]. We would
like to mention that the counterexample may be checked directly by computational
methods.
The Stanley depth is an interesting invariant which naturally arises in various
combinatorial and computational contexts, which remains rather elusive so far,
cf. [SW91, Mur02, BI10, BIS15]. Our goal is to answer the following natural question, which was raised by Herzog:
Question 1.1. [Her13, Question 1.65] Does there exist an algorithm to compute
the Stanley depth of finitely generated multigraded R-modules?
2010 Mathematics Subject Classification. Primary: 05A18; 05E40; Secondary: 16W50.
Key words and phrases. Graded modules; Hilbert depth; Stanley depth; Stanley decomposition.
The first author was partially supported by the project PN-II-RU-TE-2012-3-0161, granted by
the Romanian National Authority for Scientific Research, CNCS – UEFISCDI. The second author
was partially supported by the German Research Council DFG-GRK 1916. The third author
was partially supported by the Spanish Government—Ministerio de Economı́a y Competitividad
(MINECO), grant MTM2012-36917-C03-03.
1
2
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
In the particular cases of R-modules which are either monomial ideals I ⊂ R, or
quotients thereof, this question has been answered by Herzog, Vladoiu, and Zheng
[HVZ09]. The majority of the published articles concerning Stanley depth are related to this result. A key remark is that, in the cases studied by Herzog, Vladoiu
and Zheng, the Hilbert series already determines the module structure. So, the
Stanley depth may be computed directly from the Hilbert series of M. This leads to
another interesting combinatorial invariant, called the Hilbert depth of M, which was
introduced by Bruns, Krattenthaler, and Uliczka in [BKU10]. In fact, the method of
[HVZ09] extends directly to an algorithm for computing the Hilbert depth of finitely
generated multigraded R-modules, introduced by the first and third author [IMF14]
and the first author together with Zarojanu [IZ14]. However, until now, little is
known about the computation of the Stanley depth in general.
For computing either the Stanley or the Hilbert depth one has to consider certain
combinatorial decompositions. They are called Stanley decompositions in the first
case, respectively Hilbert decompositions in the second case.
An interesting fact is that—beside the interest raised among algebraists and combinatorialists by the conjecture of Stanley—Stanley decompositions have a separate
life in applied mathematics. This goes back to Sturmfels and White [SW91], where
it is shown how Stanley decompositions can be used to describe finitely generated
graded algebras, e.g. rings of invariants under some group action. More recently, this
found applications in the normal form theory for systems of differential equations
with nilpotent linear part (see Murdock [Mur02], Murdock and Sanders [MS07],
Sanders [San07]).
It is also worth mentioning that, in the particular case of a normal affine monoid,
suited Stanley (or Hilbert) decompositions have already been used with success in
order to design arguable the fastest available algorithms for computing Hilbert series
(see [BI10] and [BIS15]). Further, these algorithms have been used for computing
the Hilbert series (and subsequently the associated probability generating functions)
corresponding to three well studied (but difficult to compute) voting situations with
four candidates arising from the field of social choice: the Condorcet paradox, the
Condorcet efficiency of plurality voting and Plurality versus Plurality Runoff (see
[Sch13] and [BIS15] for details).
We remark that every Stanley decomposition is inducing a Hilbert decomposition,
but the converse is not true. In fact, in many particular cases studied until now the
converse also holds (for example in [HVZ09] and the related results). More generally,
it makes sense to ask:
Question 1.2. Which Hilbert decompositions are induced by Stanley decompositions?
A precise answer to Question 1.2 implies an answer to Question 1.1. This is
the main contribution of the present article: In Theorem 3.4 we give an effective
criterion to decide whether a given Hilbert decomposition is induced by a Stanley
decomposition. This leads directly to an algorithm for the computation of the
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
3
Stanley depth of a finitely generated multigraded R-module, which we present in
Section 4.
Further, as an application of our main result, we are able to construct a counterexample which gives a negative answer to the following open question, also raised
by Herzog (see Subsection 5.1):
Question 1.3. [Her13, Question 1.63] Let M be a finitely generated multigraded
R-module with syzygy module Zk for k = 1, 2, . . .. Is it true that sdepth Zk+1 ≥
sdepth Zk ?
Moreover, we define and study the structure of the set of all g-determined Stanley
decompositions of a finitely generated multigraded R-module M. In Section 5.2 we
show that this set naturally corresponds to the set of solutions of a certain system of
linear Diophantine inequalities. In other words, the question whether M has Stanley
depth at least r can be reduced to the question whether a certain combinatorially
defined polytope P contains a Zn -lattice point. This polytope P turns out to be an
intersection of a certain affine subspace with the positive orthant and finitely many
polymatroids.
Finally, we would like to point out that we have not been able to either prove or
disprove [Ape03, Conjecture 2] and [Her13, Conjecture 1.64], despite several computational and theoretical attempts. Further research on these open conjectures is
certainly desirable in view of the recent important advance made by Duval, Goeckner, Klivans and Martin in [DGKM15].
The article is organized as follows. In Section 2, we fix the notation, recall the
definitions and the necessary previous results. In Section 3, we formulate and prove
Theorem 3.4, which is the answer to Question 1.2. In Section 4, we deduce an
algorithm for the computation of the Stanley depth. This fully responds to Herzog’s
Question 1.1. In Section 5 we answer Question 1.3 and we present several interesting
applications of the main result.
2. Prerequisites
In this section we recall the basics about both Stanley and Hilbert decompositions.
We refer the reader to [Her13] for a more comprehensive treatment.
Let K be a field, R = K[X1 , . . . , Xn ] be the polynomial ring with the fine Zn grading, and let M be a finitely generated Zn -graded R-module. Throughout the
paper, we denote the cardinality of a set S by |S| and we set [n] := {1, . . . , n}.
Moreover, n-tuples in Zn will be denoted by boldface letters as a, b, . . ., while ai
will denote the i-th component of a ∈ Zn . Further, for a ∈ Nn , we set supp(a) :=
{i ∈ [n] : ai 6= 0} and Xa := X1a1 · · · Xnan .
Definition 2.1.
(1) A Stanley decomposition of M is a finite family (Ri , mi )i∈I ,
in which all mi ∈ M are homogeneous and Ri are subalgebras of R generated
by a subset of the indeterminates of R, such that Ri ∩ Ann mi = 0 for each
4
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
i ∈ I, and
M=
M
mi Ri
(1)
i∈I
as a multigraded K-vector space.
(2) A Hilbert decomposition of M is a finite family (Ri , si )i∈I , where si ∈ Zn and
the Ri are again subalgebras of R generated by a subset of the indeterminates
of R for each i ∈ I, such that
M
Ri (−si )
(2)
M∼
=
i∈I
as a multigraded K-vector space.
Note that every Stanley decomposition (Ri , mi )i∈I of M gives rise to the Hilbert
decomposition (Ri , deg mi )i∈I . In the sequel, we will say that a Hilbert decomposition is induced by a Stanley decomposition if it arises in this way. Moreover, observe
that, in general, the R-module structure of a Stanley decomposition is different from
that of M, and that Hilbert decompositions depend only on the Hilbert series of M,
i.e. they do not take the R-module structure of M into account.
Definition 2.2. The depth of a Hilbert (resp. Stanley) decomposition is the minimal
dimension of the subalgebras Ri in the decomposition. Equivalently, it is the depth
of the right-hand side of (2) (resp. (1)), considered as R-module. The multigraded
Hilbert depth (resp. the Stanley depth) of M is then the maximal depth of a Hilbert
(resp. Stanley) decomposition of M. We write hdepth M and sdepth M for the
multigraded Hilbert resp. Stanley depth.
We denote by the componentwise order on Zn and we set
[a, b] := {c ∈ Zn : a c b}
with a, b ∈ Zn .
For the computation of the Hilbert resp. Stanley depth, one may restrict the
attention to a certain finite class of decompositions. Let us briefly recall the details.
The module M is said to be positively g-determined for g ∈ Nn if Ma = 0 for
a∈
/ Nn and the multiplication map ·Xk : Ma −→ Ma+ek is an isomorphism whenever
ak ≥ gk , see Miller [Mil00]. A characterization of positively g-determined modules
is given by the following:
Proposition 2.3. [Mil00, Proposition 2.5] The module M is positively g-determined
R
R
if and only if the multigraded Betti numbers of M satisfy β0,a
(M) = β1,a
(M) = 0
unless a ∈ [0, g].
In particular, if M has no components with negative degrees, then it is always
g-determined for a sufficiently large g ∈ Nn . For our purpose, the importance of
M being positively g-determined is that it allows us to restrict the search space for
possible Hilbert or Stanley decompositions, as we explain in the following.
For a given Hilbert decomposition (Ri , si )i∈I of M and a multidegree a ∈ Nn , let
C(a) := {i ∈ I : (Ri (−si ))a 6= 0}
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
5
be the set of indices of those Hilbert spaces that contribute to degree a.
Proposition 2.4. Let M be a positively g-determined module. The following statements are equivalent for a Hilbert decomposition (K[Zi ], si )i∈I of M:
(1) si g for all i ∈ I.
(2) {j : (si )j = gj } ⊆ Zi for all i ∈ I.
(3) C(a) = C(a ∧ g) for all a ∈ Nn . (Here, a ∧ g denotes the componentwise
minimum.)
L
(4)
i K[Zi ](−si ) is g-determined as R-module.
Proof. (1)=⇒(2) Let i ∈ I and j ∈ [n] such that gj = (si )j . By assumption (1),
it holds that (si′ )j ≤ gj = (si )j for every i′ ∈ I. Hence, i′ ∈ C(si + ej ) implies
that Xj ∈ Zi′ and K[Zi′ ](−si′ )si 6= 0. In particular, C(si + ej ) ⊆ C(si ). But M is
g-determined, so
|C(si + ej )| = dimK Msi +ej = dimK Msi = |C(si )|
Thus i ∈ C(si ) = C(si + ej ) and the claim follows.
(2) =⇒ (3) We first show that C(a ∧ g) ⊆ C(a) for a ∈ Nn . Let i ∈ C(a ∧ g). It
suffices to prove that for each j ∈ [n] with (si )j < aj , it holds that j ∈ Zi . Note that
(si )j ≤ (a ∧ g)j for every j. Moreover, if (si )j < (a ∧ g)j then i ∈ C(a ∧ g) implies
that j ∈ Zi . On the other hand, if (si )j = (a ∧ g)j and (si )j < aj , then (si )j = gj
and thus j ∈ Zi by assumption. It follows that C(a ∧ g) ⊆ C(a).
Further, M being g-determined implies as above that |C(a)| = |C(a ∧ g)| and thus
C(a) = C(a ∧ g).
(3) =⇒ (1) For each i ∈ I, it holds that i ∈ C(si ) = C(si ∧ g). Therefore
si si ∧ g g.
(1) and (2) ⇐⇒ (4) This follows easily by considering the Betti numbers of
M
K[Zi ](−si ).
i
Note that the conditions are not equivalent if M is not g-determined. Moreover,
the existence of a Hilbert decomposition satisfying these conditions for some g ∈ Nn
does not imply that M is g-determined. Motivated by the preceding proposition we
introduce the following:
Definition 2.5.
(1) A Hilbert decomposition D of M is called g-determined if
M is positively g-determined and D satisfies the equivalent conditions of
Proposition 2.4.
(2) A Stanley decomposition (Ri , mi )i∈I of M is called g-determined if the underlying Hilbert decomposition (Ri , deg mi )i∈I is g-determined.
Every Hilbert decomposition of M is g-determined for a sufficiently large g ∈ Nn .
On the other hand, for a fixed g ∈ Nn there are only finitely many g-determined
Hilbert decompositions. By the following result, it is essentially sufficient to consider
g-determined Hilbert (Stanley) decompositions if M is g-determined:
6
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
Proposition 2.6. Let M be positively g-determined.
(1) There exists a g-determined Hilbert decomposition of M whose depth equals
the Hilbert depth of M.
(2) Similarly, there exists a g-determined Stanley decomposition of M whose
depth equals the Stanley depth of M.
Proof. This is immediate from Corollary 3.4 and Corollary 4.7 of [IMF14], since the
decompositions used there are g-determined.
3. Which Hilbert decompositions are induced by Stanley
decompositions?
In this section we characterize those Hilbert decompositions which are induced
by Stanley decompositions. Throughout the section, we fix a finitely generated Zn graded R-module M and a Hilbert decomposition D = (Ri , si )i∈I of M. Without
loss of generality, we shall assume that both M and D are (positively) g-determined
for some g ∈ Nn . As above, we set
C(a) := {i ∈ I : (Ri (−si ))a 6= 0}
for each multidegree a ∈ Nn . Then, Proposition 4.4 of [IMF14] may be reformulated
as follows:
Proposition 3.1. [IMF14, Proposition 4.4] The given Hilbert decomposition of M is
induced by a Stanley decomposition if and only if there exist homogeneous elements
(mi )i∈I ⊂ M with deg mi = si such that the following holds:
For all i ∈ I we have that Ri ∩ Ann mi = 0, and for all a ∈ Nn , a g, the set
{Xa−si mi : i ∈ C(a)} ⊂ Ma
(3)
is K-linearly independent.
The difficulty for applying this result is that one has to choose the right elements
mi ∈ Msi in order to determine whether a given Hilbert decomposition is induced
by a Stanley decomposition. In the sequel we present a method for circumventing
this problem. The idea is to consider (for all i ∈ I) “generic” elements m̃i ∈ Msi and
to test (for all a ∈ [0, g]) the linear independence of the sets (3) via computations
of determinants. We make this precise in the following manner.
Construction 3.2. For the given Hilbert decomposition (Ri , si )i∈I of M, we construct a collection of matrices (Aa )a∈[0,g] as follows. First, for each a ∈ [0, g], we
choose a basis
P {ba,1, . . . , ba,la } for the K-vector space Ma . Then, for each i ∈ I, we
set m̃i := j Yi,j bsi ,j with indeterminate coefficients Yi,1 , . . . , Yi,lsi .
The matrix Aa has one row for each of the basis vectors of Ma and one column
for each i ∈ C(a). For every such i, expand Xa−si m̃i in the chosen basis of Ma and
write the coefficients into Aa . More explicitly, if
X
Xa−si bsi ,j =
cj,k ba,k
k
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
7
with cj,k ∈ K, then
Xa−si m̃i =
X X
k
We set Aa = (
P
j
j
cj,k Yi,j ba,k .
cj,k Yi,j )i,k . For the ease of reference, we also set
I˜ := {(i, j) : i ∈ I, 1 ≤ j ≤ lsi },
˜
so that the entries of Aa live in the polynomial ring K[Yi,j : (i, j) ∈ I].
Note that the entries of Aa are linear polynomials in the Yi,j . Moreover, the
matrices Aa are square matrices, because the number of rows equals dim Ma , while
the number of columns equals the cardinality of C(a). But this is also dim Ma , as
we started with a Hilbert decomposition.
Example 3.3. We give a simple example to illustrate the construction. Let R =
K[X1 , X2 ] and M = (X1 , X2 ) ⊕ (X1 X2 ) ⊂ R2 . The module M is positively gdetermined for g = (1, 1). Let e1 , e2 be the generators of R2 . We choose as vector
space bases X1 e1 , X2 e1 , X1 X2 e1 and X1 X2 e2 for the corresponding components of
M. Consider the Hilbert decomposition
M∼
= R(−1, 0) ⊕ R(0, −1).
We have m̃1 = Y1,1X1 e1 and m̃2 = Y2,1 X2 e1 . The matrices Aa constructed above are
in this case
Y1,1 Y2,1
A(1,1) =
.
A(0,1) = Y2,1
A(1,0) = Y1,1
0
0
Next theorem is the main result of this paper.
Theorem 3.4. With the notation introduced in Construction 3.2, the following
holds:
(a) Assume |K| = ∞. Then the given Hilbert decomposition of M is induced by
a Stanley decomposition if and only if the determinant of Aa is not the zero
polynomial for all a ∈ [0, g].
Q
(b) Assume |K| = q < ∞. Let P := a∈[0,g] det Aa . Let further P̃ be the polynomial obtained from P as follows: From every exponent of every monomial in
P , subtract q − 1 until the remainder is less than q. Then the given Hilbert
decomposition of M is induced by a Stanley decomposition if and only if
P̃ 6= 0.
Proof. We use the characterization of Proposition 3.1. First, note that the assumption Ri ∩ Ann mi = 0 in Proposition 3.1 is not really needed: If the sets
{Xa−si mi : i ∈ C(a)}
are K-linearly independent for all a ∈ [0, g], then the fact that Ri ∩Ann mi = 0 for all
i follows automatically. To see this, assume for the contrary that Rj ∩ Ann mj 6= 0
for some j. Then there exists a multidegree d ∈ Nn such that Xd mj = 0 and
Xd ∈ Rj . But as M is g-determined, this implies that there exists d′ d ∈ Nn
8
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
′
such that Xd mj = 0 and d′ + sj g (remember that the multiplication map
·Xk : Md+sj −ek −→ Md+sj is an isomorphism if (d + sj )k > gk ). Then the set
′
{Xd mi : i ∈ C(d′ + sj )} contains the zero vector and therefore cannot be linearly
independent.
P
Next, consider a choice of elements mi = j yi,j bsi ,j with (yi,j )(i,j)∈I˜ ⊂ K. We
now observe that for a fixed a ∈ Nn , the set {Xa−si mi : i ∈ C(a)} is K-linearly
independent if and only if det Aa ((yi,j )(i,j)∈I˜) 6= 0. Hence the elements mi build a
Q
Stanley decomposition if and only if a∈[0,g] det Aa ((yi,j )(i,j)∈I˜) 6= 0.
Q If the field is infinite, then it is possible to choose such yi,j if and only if P :=
a∈[0,g] det Aa is not the zero polynomial. This is clearly equivalent to each of the
factors det Aa being nonzero.
˜
If K is finite, then P has a non-zero value over K|I| if and only if it is not contained
q
˜ This set of generators is already a (universal)
in the ideal (Yi,j
− Yi,j : (i, j) ∈ I).
Gröbner basis, hence P is contained in the ideal if and only if its remainder modulo
this Gröbner basis is zero, see Cox, Little, O’Shea [CLO07, p. 82, Corollary 2].
Clearly, P̃ is the remainder of P with respect to this Gröbner basis, so the claim
follows.
Note that this theorem gives an effectively computable criterion to decide whether
a Hilbert decomposition is induced by a Stanley decomposition.
Remark 3.5. Let us add some remarks.
(1) We can say a little more about the structure of det Aa . Endow the polynomial
˜ with a N|I| -grading by setting deg Yi,j := ei . It follows
ring K[Yi,j : (i, j) ∈ I]
from the definition that the entries of a column of Aa corresponding to mi
are homogeneous of degree ei . Hence det Aa is a homogeneous polynomial
(with respect to this grading) and its degree is a 0/1-vector. In particular,
all monomials in det Aa are squarefree.
(2) Consider the case that dimK Ma ≤ 1 for all a ∈ Zn . Then, by the above
remark, the single entry of Aa is either zero or of the form cYi1 for some i ∈ I
and c ∈ K \ {0}. Hence the Hilbert decomposition is induced by a Stanley
decomposition if and only if none of the Aa is the zero matrix. So, in this
case our Theorem 3.4 specializes to [BKU10, Proposition 2.8]. In particular,
the assumption that K is infinite can be removed from [Sta82, Conjecture
5.1] in the case that M is an R-module with dimK Ma ≤ 1 for all a ∈ Zn .
While this seems to be known, we could not find a precise reference for it.
In general, the case distinction on the cardinality of the field cannot be removed.
In fact, if K is finite, then the condition that det Aa 6= 0 for all a is not sufficient.
On the positive side, we know that the determinants of the Aa are polynomials
with squarefree monomials. Hence, if they are nonzero, then they do not vanish
identically even over a finite field. On the other hand, it might not be possible to
find values for the Yi,j such that all determinants are nonzero simultaneously. The
following example shows this phenomenon.
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
9
e5
e4
e3
e1 , e2
Figure 1. The Hilbert decomposition of Example 3.6.
Example 3.6. Let R = K[X1 , X2 ] endowed with the standard Z2 -grading. Consider
the module M with generators e1 , . . . , e5 in degrees (3, 0), (3, 0), (2, 1), (1, 2), (0, 3)
and relations
X2 e1 = X1 e3 ,
X22 e2 = X12 e4 ,
X23 e1 + X23 e2 = X13 e5 .
A Hilbert decomposition of M is given by
R1 = K[X1 , X2 ],
R2 = R3 = R4 = K[X2 ],
R5 = K[X1 , X2 ],
R6 = R7 = R8 = K[X1 ],
and
s1 = s2 = (3, 0), s3 = (2, 1), s4 = (1, 2), s5 = (0, 3), s6 = (2, 2), s7 = (2, 3), s8 = (1, 3),
see Figure 1. In each degree there is a matrix Aa . Let us compute the matrices in
the degrees (3, 1), (3, 2) and (3, 3). For this we set m̃1 = Y11 e1 + Y12 e2 , m̃3 = Y3 e3 ,
m̃4 = Y4 e4 and m̃5 = Y5 e5 . Moreover, we choose X2k e1 , X2k e2 as basis for M(3,k) .
With these conventions, we have the following matrices:
Y11 Y3
Y11 0
Y11 Y5
A(3,1) =
A(3,2) =
A(3,3) =
(4)
Y12 0
Y12 Y4
Y12 Y5
Their determinants are Y12 Y3 , Y11 Y4 , and (Y11 + Y12 )Y5 . Hence over the finite field F2
with two elements, it is not possible to choose values y11 , y12 for Y11 , Y12 , such that
all three determinants are nonzero. Thus the Hilbert decomposition given above is
induced by a Stanley decomposition over F4 , say, but not over F2 .
For later use, we note the following consequence of Theorem 3.4:
Corollary 3.7. Assume that K is infinite and let (Ri , si )i∈I be a Hilbert decomposition of M. Then (Ri , si )i∈I is induced by a Stanley decomposition if and only if
for each a ∈ Nn , a g, there exists a linearly independent subset (mi )i∈C(a) of Ma ,
such that mi ∈ Xa−si Msi for i ∈ C(a).
10
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
Proof. The condition is clearly equivalent to the non-vanishing of the determinants
of Aa for a ∈ [0, g].
4. An algorithm for computing the Stanley depth of a module
In this section we describe how Theorem 3.4 can be used to effectively compute
the Stanley depth of a given (finitely generated Zn -graded) module. We assume (as
in Section 3) that M is a fixed finitely generated Zn -graded R-module and we fix
g ∈ Nn such that M is positively g-determined.
By Proposition 2.6, one only needs to consider g-determined Stanley decompositions. Hence the Stanley depth of M can be expressed as
D is a g-determined Hilbert decomposition of M
.
sdepth M = max depth D :
which is induced by a Stanley decomposition.
A key remark is that there are only finitely many g-determined Hilbert decompositions of M for a fixed g. To actually compute the Stanley depth using this
formula, one needs to
(1) iterate over all g-determined Hilbert decompositions D of M; and
(2) decide whether D is induced by a Stanley decomposition of M.
An algorithm for the first task was presented in [IZ14, Algorithm 1]. In this
section we shall follow this approach and we modify [IZ14, Algorithm 1], so that it
may be used for computing the Stanley depth. We would like to remark at this point
that an alternative approach for this first task is to use a description of the set of gdetermined Hilbert decompositions as the set of lattice points in a certain polytope.
We give a precise description of this polytope later, in Proposition 5.4. So, in fact
one may use standard software to enumerate these points, for example SCIP [Ach09]
or Normaliz [BI10, BIS15]. This idea for enumerating Hilbert decompositions was
originally suggested by W. Bruns and described in Katthän [Kat15, Section 7.2.1].
For the second task, we suggest to apply Theorem 3.4. In order to make this
effective, one has to choose bases for the components Ma of M. One possibility is to
choose standard monomials with respect to some Gröbner bases, cf. Eisenbud [Eis95,
Theorem 15.3]. The computation of the matrices Aa and their determinant can then
be done using standard algorithms from constructive module theory. We refer to
Chapter 15 of [Eis95] or Chapter 10.4 of Becker and Weispfenning [BW93]. A
possible alternative for the second task is provided by Theorem 5.5.
Remark 4.1. For the case distinction of Theorem 3.4, one has to decide whether
the field is finite or not. We describe one way to avoid this. With the notation
introduced in Construction 3.2, let
Y
˜
P = P (D) :=
det Aa ∈ K[Yi,j : (i, j) ∈ I].
a∈[0,g]
If the field is finite, one has to reduce P to P̃ as described in Theorem 3.4, while
in the infinite case one can directly use P . But even in the finite case, P equals P̃
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
11
if the largest exponent in P does not exceed the cardinality of K. Note that this
is trivially true if K is infinite, so we can base the case distinction on the question
whether the largest exponent in P exceeds the cardinality of K.
4.1. Enumerating g-determined Hilbert decompositions via Hilbert partitions. In the following, we present a modified version of [IZ14, Algorithm 1] for
the computation of the Stanley depth, see Algorithm 4.4 below. Hence we obtain
an algorithm for the computation of the Stanley depth of M.
As the algorithm in [IZ14] is formulated in terms of Hilbert partitions, we recall
the necessary definitions from [IMF14].
Let the polynomial
X
HM (t)g :=
(dimK Ma )ta
0ag
n
be the truncated Z -graded Hilbert series of M. For a, b ∈ Zn such that a b, we
set
X
Q[a, b](t) :=
tc
acb
and call it the polynomial induced by the interval [a, b].
Definition 4.2 ([IMF14]). We define a Hilbert partition of the polynomial HM (t)g
to be a finite sum
X
P : HM (t)g =
Q[ai , bi ](t)
i∈I
of polynomials induced by the intervals [ai , bi ].
Note that there are only finitely many Hilbert partitions of HM (t)g . On one
hand, every Hilbert partition P induces a g-determined Hilbert decomposition D(P)
by the following construction.
P
Construction 4.3 ([IMF14]). Let P : i∈I Q[ai , bi ](t) be a Hilbert partition of
HM (t)g . For 0 a b g we set
G[a, b] := {c ∈ [a, b] : cj = aj for all j with bj = gj }.
Further, for b g let Zb := {j ∈ [n] : bj = gj }, ρ(b) = |Zb | and let K[Zb ] :=
K[Xj : j ∈ Zb ]. Then we define
r
M
M
∼
i
K[Zb ](−c) .
D(P) : M =
i=1
c∈G[ai ,bi ]
On the other hand, by Proposition 2.4, each g-determined Hilbert decomposition
(K[Zi ], si )i∈I is induced by the Hilbert partition
X
P : HM (t)g =
Q[si , bi ](t),
i∈I
where
(bi )j =
(
(si )j
gj
if j ∈
/ Zi ,
if j ∈ Zi .
12
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
Hence the g-determined Hilbert decompositions are exactly those Hilbert decompositions which are induced by a Hilbert partition.
Our modified version [IZ14, Algorithm 1] for the computation of the Stanley depth
is presented in Algorithm 4.4.
Algorithm 4.4: Function that checks if sdepth ≥ s recursively
Data: g ∈ Nn , s ∈ N, an R-module M, a polynomial
H(t) = HM (t)g ∈ N[t1 , ..., tn ], a Container P and q ∈ N ∪ {∞}
Result: true if sdepth M ≥ s
Boolean CheckStanleyDepth(g, s, M, P, P, q);
begin
if H ∈
/ N[t1 , ..., tn ] then
return false;
1
2
3
4
Container E =FindElementsToCover(g, s, H);
if size(E ) = 0 then
Polynomial P (Y ):=ComputeDeterminantsProduct(g, M, P);
P=Reduce(P, q);
if P 6= 0 then
return true;
return false;
else
for i=begin(E ) to i=end(E ) do
Container C[i]:=FindPossibleCovers(g, s, H, E[i]);
if size(C[i])= 0 then
return false;
5
for j=begin(C[i]) to j=end(C[i]) do
Polynomial H̃(t) = H(t) − Q[E[i], C[i][j]](t);
Container P̃:=AddInterval(P, Q[E[i], C[i][j]](t));
if CheckStanleyDepth(g, s, M, H̃, P̃)=true then
return true;
return false;
The differences from [IZ14, Algorithm 1] appear at lines 1–4, 5, and in the usage
of the extra parameters M, P, and q. The container P is used for storing the
intervals in the Hilbert partitions that have been computed, and it can be initialized
empty. The R-module structure of M is needed for computing the matrices Aa (for
all a ∈ [0, g]). Moreover, q is the cardinality of the field, which is needed for the
reduction. Assuming that the reader is familiar with [IZ14, Algorithm 1], we describe
below the new key steps of the algorithm:
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
13
• line 1. If E is empty, then we have computed a complete Hilbert partition
in P (since there are no elements in E to cover). Then we have to check
using Theorem 3.4 whether the Hilbert decomposition D(P) is induced by
a Stanley partition.
• line 2. The function ComputeDeterminantsProduct computes P (D(P))
as in Remark 4.1. Since P depends on the R-module structure of M, we have
to pass it as a parameter.
• line 3. Here we compute the reduction of P with respect to the cardinality
q of the field. We point out that we can skip this step if K is infinite.
• line 4. We apply Theorem 3.4, so we check whether P 6= 0. If the answer
is positive, then we are done. We have reached a good leaf of the searching
tree.
• lines 5. The child P̃ is generated here and further investigated in the recursive
call.
5. Applications and Examples
In this section, we present several applications of Theorem 3.4. To simplify the
discussion we assume throughout this section that |K| = ∞.
5.1. Stanley depth of syzygies. In this subsection, we present an example of an
R-module M, such that sdepth M > sdepth Syz1R (M). This answers Question 63
in [Her13] to the negative. Let us describe the idea of the construction. It was
observed in [IZ14] that there are modules M such that sdepth M < sdepth M ⊕ R.
But it always holds that Syz1R (M ⊕R) = Syz1R (M). Hence, we will look for a module
whose Stanley depth increases sufficiently under adding copies of the ring, to obtain
sdepth Syz1R (M) = sdepth Syz1R (M ⊕ Ra ) < sdepth M ⊕ Ra .
In fact, it is already sufficient to choose M = m, the maximal ideal in some
polynomial ring. It follows from [BKU10, Proposition 3.6] that
n−2
sdepth Syz1R (m) ≤ n − ⌈
⌉,
3
where n is the number of variables. As sdepth m = ⌈ n2 ⌉, we see that in order to use
this upper bound, we need that the Stanley depth of m increases at least by two after
adding any number of copies of the ring. The smallest n where this is possible is six.
Indeed, an easy computation following Popescu [Pop15] shows that the Z-graded
Hilbert depth of m6 ⊕ R9 equals 5, while sdepth Syz1R (m6 ⊕ R9 ) = sdepth Syz1R (m6 ) ≤
⌉ = 4 (see Uliczka [Uli10] for details about the Z-graded Hilbert depth). So
6 − ⌈ 6−2
3
M = m6 ⊕ R9 is our candidate for a counterexample.
We need to compute a Hilbert decomposition D of M with depth D = 5. Unfortunately, this module is already too large for the CoCoA implementation of the
Algorithm in [IZ14]. By Proposition 2.4, it is enough to search for a g-Hilbert decomposition, where g = (1, 1, 1, 1, 1, 1). These decompositions are described by a system
of linear Diophantine inequalities (see Section 5.2 for details) and we can solve the
system with the software SCIP [Ach09]. This yields the Hilbert decomposition of
14
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
M, which is summarized in Table 1. There, an entry such as 2 × [001111, 101111] is
to be interpreted as two copies of the vector space
K[X1 , X3 , X4 , X5 , X6 ](0, 0, −1, −1, −1, −1)
in the Hilbert decomposition.
4 × [000000, 111110] 2 × [000000, 111101] 3 × [000000, 111011]
[000001, 111101]
[000001, 111011]
[000001, 110111]
[000001, 101111]
[000001, 011111]
[000010, 110111]
[000010, 101111]
[000010, 011111]
[000100, 111101]
[000100, 110111]
[000100, 101111]
[000100, 011111]
[001000, 101111]
[010000, 011111]
[100000, 110111]
[000111, 110111]
[001011, 111011]
[001101, 101111]
[001110, 111110]
[010011, 011111]
[010101, 011111]
[010110, 111110]
[011001, 111101]
[011010, 011111]
[011100, 111101]
[100011, 111011]
[100101, 110111]
[100110, 101111]
[101001, 111101]
[101010, 111110]
[101100, 111101]
[110001, 110111]
[110010, 111110]
[110100, 110111]
[111000, 111011] 2 × [001111, 101111]
[101011, 111011]
[110011, 111011]
[111100, 111110]
3 × [011111, 011111] 2 × [101111, 101111] 2 × [110111, 110111]
[111011, 111011] 2 × [111101, 111101]
[111110, 111110]
10 × [111111, 111111]
Table 1. A Hilbert decomposition D of M with depth D = 5.
In particular, the Hilbert depth of M equals 5. It remains to show that this
Hilbert decomposition is induced by a Stanley decomposition of M. Then we can
conclude that
sdepth M = 5 > 4 ≥ sdepth Syz1R (M).
For this we prove the following general result:
Proposition 5.1. Let m ⊂ R be the maximal monomial ideal. Assume that K is
infinite. Then for all α, β ∈ N it holds that
hdepth m⊕α ⊕ R⊕β = sdepth m⊕α ⊕ R⊕β .
In fact, every Hilbert decomposition of this module is induced by a Stanley decomposition.
Proof. Let M := m⊕α ⊕ R⊕β . Further, let e1 , . . . , eα , f1 , . . . , fβ be the natural set of
generators of R⊕α ⊕ R⊕β and consider M as a submodule of this module.
In every nonzero multidegree a ∈ Nn , the elements Xa e1 , . . . , Xa eα , Xa f1 , . . . , Xa fβ
form a vector space basis of Ma . Moreover, a vector space basis of M0 is given by
f1 , . . . , fβ .
Now consider a Hilbert decomposition (Ri , si )i∈I of M. We distinguish two kinds
of summands in this decomposition. First, there are those i where si = 0. Here we
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
15
P
set mi := j Zij fj and we call these generators of the first type. As we start from
a Hilbert decomposition, it is clear that there are exactly
of
Pdim M0 = β generators
P
the first type. Further, for i with si 6= 0 we set mi := j Yij Xsi ej + j Zij Xsi fj .
We call these the generators of the second type.
Next we consider the corresponding matrices as in Theorem 3.4. In the multidegree 0, it is easy to see that A0 is a generic (square) matrix in the variables Zij , and
thus its determinant is non-zero. So consider a multidegree a 6= 0. Both types of
generators can contribute to Ma , so the matrix Aa has the following shape:
0
Y∗∗
Z∗∗
Z∗∗
u
α
β
Here u stands for the number of generators of the first type contributing to the
multidegree a. Note that every entry on the antidiagonal of Aa is non-zero. Indeed,
because the sum of the indices of the matrix entries is α+β +1, while for every entry
of the zero-block this sum is at most α + u ≤ α + β. Hence the antidiagonal gives
a non-zero monomial in the Leibniz expansion of the determinant, and as all nonzero entries of the matrix are different variables, therefore cancelation cannot occur.
Thus the determinant is non-zero and the claim follows from Theorem 3.4.
Remark 5.2.
(1) Proposition 5.1 does not hold as stated for arbitrary ideals.
Consider the case R = K[X1 , X2 ] and M = (X1 X2 ) ⊕ R. Then K ⊕
X1 K[X1 , X2 ] ⊕ X2 K[X1 , X2 ] is a Hilbert decomposition of M that is not
induced by a Stanley decomposition.
(2) The result also does not hold if one adds shifted copies of the ring. Consider R = K[X1 , X2 ] and M = (X1 , X2 ) ⊕ R(−1, −1). Then X1 K[X1 , X2 ] ⊕
X2 K[X1 , X2 ] is a Hilbert decomposition of M which is not induced by a
Stanley decomposition. In fact, by adding shifted copies of the ring, one can
always obtain a Hilbert decomposition of Hilbert depth n for an arbitrary
graded module M. For this, consider a finite free resolution of M,
0 → Fp → Fp−1 → · · · → F0 → M → 0.
Then the sum of the Hilbert series of the even modules equals the Hilbert
series of M plus the sum of the Hilbert series of the odd modules, so the
former is a Hilbert decomposition of the latter.
Based on several examples, we conjecture the following strengthening of Proposition 5.1:
16
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
Conjecture 5.3. For every number of variables and any α, β ∈ N, the Z-graded
Hilbert depth [Uli10] and the Stanley depth of m⊕α ⊕ Rβ coincide.
5.2. The set of g-determined Stanley decompositions. In this section, we
show that the set of all g-determined Stanley decompositions can be described by
a (large) system of linear Diophantine inequalities, or, equivalently, by the set of
Zn -lattice points inside a polytope P.
Consider a finitely generated Nn -graded R-module M which is g-determined for
some g ∈ Nn . Let
Ω := {(K[Z], a) : a ∈ Nn , a g, Z ⊆ [n], {j : gj = aj } ⊆ Z}
be the set of all possible building blocks for a g-determined Hilbert decomposition
of M (according to Proposition 2.4).
We write NΩ for the free commutative monoid with generators {eω : ω ∈ Ω}. A
g-determined
P Hilbert decomposition (Ri , si )i∈I of M can then be identified with the
element i∈I e(Ri ,si ) ∈ NΩ . For a vector u ∈ NΩ , we write u(a, Z) for the component
of u corresponding to Z ⊆ [n] and a ∈ [0, g]. Note that for (K[Z], b) ∈ Ω and a b,
it holds that (K[Z](−b))a 6= 0 if and only if supp(a − b) ⊆ Z. Now, g-determined
Hilbert decompositions may be characterized easily.
Proposition 5.4. A vector u ∈ NΩ corresponds to a g-determined Hilbert decomposition of M if and only if it satisfies the following equalities:
X
X
u(b, Z) = dimK Ma
for a ∈ [0, g].
(5)
b∈[0,a]
Z⊆[n]
supp(a−b)⊆Z
So, the set of g-determined Hilbert decompositions corresponds naturally to the
set of Zn -lattice points in the polytope H of non-negative solutions to (5). The set of
g-determined Stanley decompositions is a subset of this. By the following result, this
subset may be defined by linear inequalities as well, i.e. the g-determined Hilbert
decomposition of M which are induced by g-determined Stanley decompositions
correspond to the Zn -lattice points in a certain polytope P. This is the main result
of this subsection.
Theorem 5.5. A vector u ∈ NΩ corresponds to a g-determined Hilbert decomposition of M which is induced by a g-determined Stanley decomposition, if and only if
it satisfies both (5) and in addition the following inequalities:
X
X
X
Xa−b Mb
for a ∈ [0, g], J ⊆ [0, a].
(6)
u(b, Z) ≤ dimK
b∈J
Z⊆[n]
supp(a−b)⊆Z
b∈J
Here, the sum on the right-hand side is a sum of vector spaces.
Remark 5.6. The system of inequalities (6) is rather large, so it does not seem to
be feasible for the actual computation of the Stanley depth. However, the theorem
shows that the set of all Stanley decomposition has a nice structure. Note that the
integer solutions of (6) for a fixed a ∈ [0, g] form a discrete polymatroid, cf. Herzog
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
17
and Hibi [HH02]. So the set of g-determined Stanley decompositions may also be
seen as an intersection of discrete polymatroids with the polytope H.
The proof uses Rado’s theorem, which we recall for the reader’s convenience.
Recall that a transversal of a set system A1 , . . . , Ar is a collection of pairwise different
elements a1 ∈ A1 , a2 ∈ A2 , . . . , ar ∈ Ar .
Theorem 5.7 (Rado’s theorem, VIII.2.3 [Aig76]). Let M be a matroid on a ground
set B with rank function r and let A : A1 , . . . , Ar ⊆ B be a collection of subsets of
B. Then A has an independent transversal if and only if
!
[
|I| ≤ r
Ai
i∈I
for every subset I ⊂ [r].
We use the following variant of Rado’s theorem.
Corollary 5.8. Let V be a vector space and V : V1 , . . . , Vs a collection of linear
subspaces of V . For u ∈ Ns , the following are equivalent:
(1) There exists an independent transversal of V, i.e. a linearly independent
family of vectors v1 ∈ V1 , v2 ∈ V2 , . . . , vs ∈ Vs .
(2) For each subset I ⊆ {1, . . . , s}, the following inequality holds:
X
|I| ≤ dimK
Vi .
i∈I
Here, the sum on the right-hand side is a sum of vector spaces.
Proof. The inequality is clearly necessary, so we only need to show
S the sufficiency.
Let Ai be a basis for Vi , 1 ≤ i ≤ s. Consider the union M := i Ai as a matroid.
By Rado’s theorem 5.7, A1 , . . . , As has an independent transversal if and only if
!
[
X
|I| ≤ dimK span
Ai = dimK
Vi
i∈I
i∈I
for every subset I ⊂ [s]. Hence the inequality in our claim is sufficient.
Proof of Theorem 5.5. Assume that u ∈ NΩ is indeed a g-determined Hilbert decomposition of the module M. By Corollary 3.7, the Hilbert decomposition u corresponds to a Stanley decomposition if and only if for each a ∈ [0, g], there are
linearly independent elements (m(b, Z, i))(b,Z,i)∈Λ , such that m(b, Z, i) ∈ Xa−b Mb
for all (b, Z, i) ∈ Λ, where
Λ := Λ(a, u) := {(b, Z, i) : b ∈ [0, a], Z ⊂ [n], supp(a − b) ⊂ Z, 1 ≤ i ≤ u(b, Z)}.
So in particular, the inequality in our claim is necessary.
We apply the preceding Corollary 5.8 to the vector space Ma and the collection
(Xa−bMb )(b,Z,i)∈Λ of subspaces. For a subset I ⊂ Λ, consider
I¯ := {(b, Z, i) ∈ Λ : (b, Z ′ , i′ ) ∈ I for some Z ′ , i′ }.
18
BOGDAN ICHIM, LUKAS KATTHÄN, AND JULIO JOSÉ MOYANO-FERNÁNDEZ
It clearly holds that
X
Xa−b Mb =
(b,Z,i)∈I
X
Xa−b Mb ,
(b,Z,i)∈I¯
¯ and these are in bijection with
hence it suffices to consider subsets of the form I,
subsets J ⊆ [0, a]. Hence our inequalities are also sufficient.
Acknowledgements
The authors are greatly indebted to Jürgen Herzog for bringing Question 1.3 to
our attention.
References
[Ach09]
T. Achterberg. Scip: Solving constraint integer programs. Mathematical Programming
Computation, 1(1):1–41, 2009.
[Aig76]
M. Aigner. Kombinatorik; II. Matroide und Transversaltheorie. Springer, 1976.
[Ape03]
J. Apel. On a conjecture of R.P. Stanley; Part I–monomial ideals. Journal of Algebraic
Combinatorics, 17(1):39–56, 2003.
[BI10]
W. Bruns and B. Ichim. Normaliz: Algorithms for affine monoids and rational cones.
Journal of Algebra, 324(5):1098 – 1113, 2010.
[BIS15]
W. Bruns, B. Ichim, and C. Söger. The power of pyramid decomposition in Normaliz.
To appear in Journal of Symbolic Computation, 2015.
[BKU10]
W. Bruns, Chr. Krattenthaler, and J. Uliczka. Stanley decompositions and Hilbert
depth in the Koszul complex. Journal of Commutative Algebra, 2:327–357, 2010.
[BW93]
Th. Becker and V. Weispfenning. Gröbner bases. Springer, 1993.
[CLO07]
D. Cox, J. Little, and D. O’Shea. Ideals, varieties, and algorithms: an introduction to
computational algebraic geometry and commutative algebra. Springer, 2007.
[DGKM15] A. M. Duval, B. Goeckner, C. J. Klivans, and J. L. Martin. A non-partitionable
Cohen-Macaulay simplicial complex. Preprint, arXiv:1504.04279, 2015.
[Eis95]
D. Eisenbud. Commutative algebra with a view toward algebraic geometry. Springer,
1995.
[Her13]
J. Herzog. A survey on Stanley depth. In Monomial Ideals, Computations and Applications, pages 3–45. Springer, 2013.
[HH02]
J. Herzog and T. Hibi. Discrete polymatroids. Journal of Algebraic Combinatorics,
16(3):239–268, 2002.
[HVZ09]
J. Herzog, M. Vladoiu, and X. Zheng. How to compute the Stanley depth of a monomial
ideal. Journal of Algebra, 322(9):3151 – 3169, 2009.
[IMF14]
B. Ichim and J. J. Moyano-Fernández. How to compute the multigraded Hilbert depth
of a module. Mathematische Nachrichten, 287(11–12):1274 – 1287, 2014.
[IZ14]
B. Ichim and A. Zarojanu. An algorithm for computing the multigraded Hilbert depth
of a module. Experimental Mathematics, 23:322–331, 2014.
[Kat15]
L. Katthän. Stanley depth and simplicial spanning trees. Journal of Algebraic Combinatorics, 42(2):507–536, 2015.
[Mil00]
E. Miller. The Alexander duality functors and local duality with monomial support.
Journal of Algebra, 231:180–234, 2000.
[MS07]
J. Murdock and J. Sanders. A new transvectant algorithm for nilpotent normal forms.
Journal of Differential Equations, 238(1):234–256, 2007.
[Mur02]
J. Murdock. On the structure of nilpotent normal form modules. Journal of Differential
Equations, 180(1):198–237, 2002.
HOW TO COMPUTE THE STANLEY DEPTH OF A MODULE
19
[Pop15]
A. Popescu. An algorithm to compute the Hilbert depth. Journal of Symbolic Computation, 66:1–7, 2015.
[PSFTY09] M.R. Pournaki, S.A. Seyed Fakhari, M. Tousi, and S. Yassemi. What is Stanley Depth?
Notices of the AMS, 56(9):1106–1108, 2009.
[San07]
J. Sanders. Stanley decomposition of the joint covariants of three quadratics. Regular
and Chaotic Dynamics, 12(6):732–735, 2007.
[Sch13]
A. Schürmann. Exploiting polyhedral symmetries in social choice. Social Choice and
Welfare, 40(4):1097–1110, 2013.
[Sta82]
R. Stanley. Linear Diophantine equations and local cohomology. Invent. Math., 68:175–
193, 1982.
[SW91]
B. Sturmfels and N. White. Computing combinatorial decompositions of rings. Combinatorica, 11(3):275–293, 1991.
[Uli10]
J. Uliczka. Remarks on Hilbert series of graded modules over polynomial rings.
Manuscripta Math., 132:159–168, 2010.
Simion Stoilow Institute of Mathematics of the Romanian Academy, Research
Unit 5, C.P. 1-764, 014700 Bucharest, Romania
E-mail address: [email protected]
Universität Osnabrück, FB Mathematik/Informatik, 49069 Osnabrück, Germany
E-mail address: [email protected]
Universitat Jaume I, Campus de Riu Sec, Departamento de Matemáticas & Institut
Universitari de Matemàtiques i Aplicacions de Castelló, 12071 Castellón de la
Plana, Spain
E-mail address: [email protected]
| 0math.AC
|
arXiv:1709.06126v2 [cs.CV] 31 Oct 2017
How intelligent are convolutional neural
networks?
Zhennan Yan
Xiang Sean Zhou
Siemens Medical Solutions USA, Inc.
65 Valley Stream Parkway, Malvern, PA 19355, USA
[email protected]
November 2, 2017
Abstract
Motivated by the Gestalt pattern theory in psychology and visual perception,
and the Winograd Challenge for language understanding, we design synthetic experiments to investigate a deep learning algorithm’s ability to infer simple (at least
for human) semantic visual concepts, such as symmetry, counting, and uniformity,
etc., from examples. A visual concept is represented by randomly generated, positive as well as negative, example images. We then test the ability and speed of
algorithms (and humans) to learn the concept from these images. The training
and testing are performed progressively in multiple rounds, with each subsequent
round deliberately designed to be more complex and confusing than the previous
round(s), especially if the true meaning of the concept was not grasped by the
learner. However, if the semantic concept was understood, all the deliberate tests
would become trivially easy. Our experiments show that humans can often infer
a semantic concept quickly after looking at only a very small number of examples (this is often referred to as an “aha moment”: a moment of sudden realization, inspiration, insight, recognition, or comprehension), and performs perfectly
during all testing rounds (except for careless mistakes), even on samples that lie
completely outside of the training distribution. On the contrary, deep convolutional neural networks or DCNNs could approximate some concepts statistically,
but only after seeing many (×104 ) more examples. And it will still make obvious mistakes, especially during deliberate testing rounds or on samples outside
the training distributions. This signals a lack of true “understanding”, or in other
words, a failure to reach the right “formula” for the semantics. We did find that
some concepts are easier for DCNN than others. For example, simple “counting”
(i.e., 1 to 3 objects) is more learnable than “symmetry”, while “uniformity” (e.g.,
in terms of shape variation) or “conformance” (e.g., in terms of group behavior)
are much more difficult for DCNN to learn. To conclude, we propose an “Aha
Challenge” for visual perception, calling for focused and quantitative research on
Gestalt-style machine intelligence using limited training examples.
1
1
Introduction
Recently, deep learning methods [1, 2] have improved state-of-the-art in many domains, such as speech recognition, visual object detection and recognition, machine
translation, etc. Deep Convolutional Neural Network (DCNN) is one of the most successful deep learning architectures. It has been widely adopted by the research communities, since Krizhevsky et al. [3] used a DCNN to almost halve the error rate
in the ImageNet competition in 2012. Since then, DCNN’s have achieved great successes in various computer vision tasks, approaching or even surpassing human-level
performance in some tasks [4, 5].
The advantages of DCNN include the hierarchy of self-learned features and its
ability to learn complex non-linear functions via direct end-to-end training [1]. The hierarchy of features, representing low-level image details as well as high-level abstract
properties, can be learned from the raw data without any hand-crafted manipulation. At
the same time, the complex function, which bridges the gap between raw data and the
learning task, is learned from the examples by the back-propagation algorithm. Usually, there are millions of weights in a DCNN model. Since the complex model could
capture data variance very well, it could absorb a large amount of training samples to
avoid over-fitting. Comparing with conventional machine learning algorithms, deep
learning has a clear advantage in terms of easy implementation, seemingly unlimited
learning capacity, and unprecedented performance, which makes it very attractive to
both research community and industry in this era of big data. However, due to the
end-to-end training of the “black-box” layers, the general reasoning ability of DCNN
is still not fully explored or understood, and the appropriate size of data set to train a
DCNN is still mostly empirical.
The first CNNs only contain convolutional and pooling layers, which are directly
inspired by the biological neural cells and the hierarchical architecture in animal visual
systems [6, 7]. After years of evolution, CNNs have become deeper and deeper. More
nonlinear and complex layers and structures are utilized in DCNNs, such as ReLU and
normalization layers [3], dropout layer [8], residual connection [9], inception modules
[10, 11] and so on. Complexity brought opacity—although there are ways to visualize
what patterns might have been learned by intermediate layers [12, 13], thus shedding
some lights on the inner working of a CNN, overall we still lack a thorough understanding of the learning process. Some studies have shown that carefully designed
adversarial noises could mislead the learned model [14, 15], casting doubt on the generalization capability of these new models— Did they really learn? what did they
learn?
In this paper, we use simple and well-defined concepts, such as “symmetry”, “counting”, and “uniformity”, and synthetic and clean examples to test and compare the “intelligence” of algorithms and humans. The relationship between the number of training
data and the classification accuracy is used to evaluate the learning speed. Unseen test
data, especially those drawn from outside the training distribution, are used to evaluate
the degree of learning at the semantic level. Implicit “end-to-end” learning of such
concepts is a stepping stone to the scaling up of artificial intelligence for many realworld tasks such as diagnostic imaging, where symmetry (e.g., of the brain), counting
(e.g., of vertebrae), and uniformity (of tissue texture or anatomical structures) are key
2
features for the detection of certain diseases.
2
Related work
There have been some attempts to analyze the complexity and learning capacity of artificial neural networks over the last decades [16, 17]. Recently, Basu et. al. [18] derived
upper bounds on the VC dimension of CNN for texture classification tasks. Szegedy et.
al. [14] reported counter-intuitive properties of neural networks, and found adversarial
examples with hardly perceptible perturbation that could mislead the algorithm. Since
then, many more successful attacks at deep learning were reported [19, 20, 21].
Deep networks’ vulnerability to adversarial examples led to active research for a
defense mechanism. Goodfellow et. al. [22] proposed adversarial training, and Hinton
et. al. [23] proposed to distill the knowledge by model compression. Goodfellow et.
al. [24] also proposed a zero-sum game framework for estimating generative models
via an adversarial process, namely Generative Adversarial Nets (GAN). A GAN balances an adversarial data generator and a discriminator during training. And the trained
discriminator is more tolerate to adversarial examples as a result.
While GAN related research reveals and repairs a “statistical” weakness of CNNs,
our study focuses on “semantic” level limitation of CNNs. In other words, we ask
“how well can a CNN discover patterns or concepts from data” — a capability that
is a hallmark of natural intelligence. For example, can a CNN learn the concept of
“Symmetry” (we will start with the simplest “bilateral symmetry”) from examples?
how fast (i.e., with how many training samples) can it learn it? and how general does it
understand the concept? and ultimately, can the same network architecture be trained
to learn another concept, e.g., “counting”?
Few-shot learning [25, 26, 27], one-shot learning [28] or even zero-shot learning
[29, 30] are trying to adopt a classifier to accommodate new classes not seen in training, given only a few examples, one example, or no example at all, respectively. The
goal is to transfer learned knowledge and make the model generalizable to new classes
or tasks. However, the new patterns tested in these papers are mostly analogous or homologous to the learned patterns. They have not tested the kind of semantic Gestalt visual concepts, which are more diverse and more challenging for machines, but mostly
trivially easy for humans. Nevertheless, zero-shot learning capability of DCNN was
observed occasionally in some rounds within our experiments. For example, see the
near 100% performance on the “Deliberate test 1” rows under Setting 2 and Setting
3 in Table 3. There were work on heuristic program to solve visual analogy IQ test
[31], and explicit modeling of higher level visual concepts based on low level textons
[32, 33]. We approach these topics from a different angle, using a classification problem to implicitly embed the concepts, and focus on the testing of the limit of end-to-end
learning capacity of machines.
A similar question was posed in the language understanding domain by Winograd
[34], where a collection of questions can be easily understood, thus answered by a
reasonably intelligent human, but not easily at all by an algorithm. The concepts selected in this study are also easily discoverable by a reasonably intelligent human, and
our experiments confirm this quantitatively. One aspect that is unique to human visual
3
perception process is that there is often an ”Aha!” moment, after which the concept
is fully understood, and error rate drops to zero or near zero immediately. This is not
observed in algorithms. We believe the reason behind this cognitive gap is the same as
that underlines the Winograd Schema Challenge, one that is related to the accumulative human experience (both environmental and societal) or some may call it “general
intelligence”.
There have also been a long history of research in the classic computer vision domain called Gestalt visual perception theory [35, 36]. Our study draws upon some
insights gained from this research field. However, with a few simple concepts, we are
only scratching the surface. As future work, there are many interesting concepts to
explore, such as similarity or uniformity, continuity or conformance, and proximity or
grouping, etc.
3
Study design
The inception v3 networks developed by Szegedy et al. [11] obtained high classification
accuracy with relatively less computational cost. We use this network in this study as
a representative method of DCNNs.
Three types of visual recognition/classification tasks are studied in the following
sections. The first one is based on the concept of “symmetry” (section 4), the second
one is based on “counting” (section 5), and the last one is based on “grouping or conformance behavior” (section 6). First two tasks include several sub-tasks, which may
require additional learning of concepts such as “uniformity” or “grouping”. All tasks
are designed as binary classification problems.
We create synthetic data sets for these image recognition problems. All images are
generated in size of 200 × 200.
To handle the binary tasks and these synthetic images, we adapt the inception v3
model by changing the input size, and replacing the original softmax output layer by
one hidden fully connected layer of 1024 nodes (with relu activation) plus one new
softmax layer of 2 nodes. The weights of the network (except the new layers) are
initialized from the pre-trained model on ImageNet database.
4
Symmetry
In this section, we investigate the learnability of the concept “symmetry” by CNNs as
well as by humans. We focus on the simple form of bilateral symmetry, but test both
global (in section 4.1) and local (in section 4.2) symmetry. Global symmetry means
that all the positive example images have bilateral symmetry, while local symmetry
means that all positive images contains only symmetric shapes. Therefore, the local
symmetry test is also a test of a “uniformity” concept.
To bring the study closer to real-world use cases, and to make it a bit more interesting, we conduct another test exploiting the symmetry in human faces (section 4.3).
4
Figure 1: Example of synthetic samples in A1 . Left: 9 symmetric samples; right: 9
asymmetric samples.
4.1
Global symmetry
Some examples of global symmetry are shown in Fig. 1. In this case, images are created
by randomly sample control points and connecting them to form polygons. The interpolation between control points can be either a straight line or a bezier curve, which is
determined randomly. From these random shapes, symmetric images are generated by
creating polygons from symmetrical control points or mirroring asymmetric shapes.
To program a computer algorithm explicitly to detect such bilateral symmetry is
trivial: just fold the image along the mid-line and then check the differences of overlapping pixels. However, for an algorithm to learn this seemingly simple operation
implicitly, from examples only, is not trivial at all — especially if the benchmark is
human performance. Humans can often grasp the concept quickly, after seeing only a
few dozen examples.
We conduct four rounds of training and testing, designed to evaluate model generalization from different angles and at different levels. The first round is a statistical test
within the same sample distribution as the training. This round is a “smoke test” to ensure that the network is working. The second and third rounds of testing are deliberate
and adversarial manipulation of the training set in order to test for “understanding”, or
the lack thereof, of the real concept. The last round of test uses samples outside and far
away from the training distribution, checking again for concept-level comprehension
from another angle.
In the first round, we generate examples randomly similar to Fig. 1. Denoting A1
as the training set, B1 as the validation set, and C1 as the testing set, we can train model
using A1 and B1 , then test on C1 . To understand the generalization ability of the model
on the concept level, we also generate new “deliberate” test samples D1 based on A1 ,
with small but just enough modifications such that the class labels are all reversed.
D1 should be easy for a learner that has learned the concept, but very confusing for a
learner that merely memorized (i.e., overfit to) A1 , because every sample in D1 has a
similar-looking counterpart in A1 .
5
Figure 2: Example of deliberate samples from D1 (A1 ). Left: 9 symmetric samples,
which are generated from asymmetric ones in Fig. 1; right: 9 asymmetric samples,
which are generated from symmetric ones in Fig. 1.
Denote D(·) as the operation to create deliberate test from existing samples. Various linear or non-linear functions can be applied in D(·). We use different operations
in each round: D1 , D2 , and D3 . Specifically, D1 (·) on symmetric samples is removing
some random part of the foreground in either left or right half to achieve asymmetry,
while D1 (·) on asymmetric samples is mirroring one side of the image on the other side
to achieve symmetry, and then, with a chance of 50%, symmetrically erasing some part
of the image (this last step is to avoid bias by adding similar erasion pattern on both
classes). Some of the samples in set D1 (A1 ) are shown in Fig. 2.
In the second round, we construct a new training set A2 = A1 ∪ D1 (A1 ), validation
set B2 = B1 ∪ D1 (B1 ), and train a new model using A2 and B2 . Then, we test it on
D2 (A2 ). The D2 (·) operation is based on scaling of, or adding small shape object(s)
randomly (with a 50% chance) to, the symmetric samples in A2 . More specifically, to
create deliberate symmetric samples, we either scale the whole image, or add a pair of
identical shapes at random but symmetric positions (in order to maintain symmetry).
To create deliberate asymmetric samples, we either scale one side (left or right) of the
image, or add a shape on one side. For scaling, we increase or decrease the size by
30% ∼ 50%. For the added shape, we evenly sample triangle, square, or ball, with the
same size of 25 pixels and random intensity ∈ [128, 255). If the added small shape is
located inside a foreground shape in the image, the intensity of the additional shape is
set to 0 (making a hole). Random samples from D2 (A2 ) are shown in Fig. 3.
In the third round, again we generate new training set A3 = A2 ∪ D2 (A2 ), and
validation set B3 = B2 ∪ D2 (B2 ). New model is trained using A3 and B3 , and tested
on D3 (A3 ). Samples in D3 (A3 ) are created by adding more small shape objects into
the symmetric images in A3 . The strategy of creating new symmetric samples is similar to D2 (·), but more objects/shapes are added. To create asymmetric samples, we
randomly chose among three different approaches: (1) two random objects are added
on the left and right sides, at asymmetric locations; (2) two random objects of different
shape are added at symmetric positions; (3) two random objects of the same shape, but
6
Figure 3: Example of 2nd round deliberate samples from D2 (A2 ). Left: 9 symmetric
samples; right: 9 asymmetric samples.
Figure 4: Example of 3rd round deliberate samples from D3 (A3 ). Left: 9 symmetric
samples; right: 9 asymmetric samples.
of different sizes, are added at symmetric positions. Examples from D3 (A3 ) are shown
in Fig. 4.
In the last round, we generate brand new testing samples A4 to evaluate the previous three models. These new samples are generated by placing shape objects (triangle,
square and ball) at symmetric or asymmetric locations for the two classes. These samples lie completely outside of the previous training distributions, therefore, can serve as
a good challenge for those models that have failed to learn the concept at the semantic
level. Some samples from A4 are shown in Fig. 5.
4.2
Local Symmetry
In this sub-task, each image contains multiple objects. A positive image contains symmetric objects only, while a negative image contains at least one asymmetric object.
7
Figure 5: Example of 4th round samples from A4 . Left: 9 symmetric samples; right: 9
asymmetric samples.
Figure 6: Example of synthetic training samples for local symmetry task. Left: 6
samples only containing symmetric patterns; right: 6 samples containing at least one
asymmetric pattern.
8
Figure 7: Example of deliberate testing samples for local symmetry task. Left: 6
samples only containing symmetric patterns; right: 6 samples containing at least one
asymmetric pattern.
Some examples are shown in Fig. 6. In these image samples, objects include equilateral triangle, square, ball and connected-component polygon, which is generated in the
similar way in A1 . Asymmetric objects are those asymmetric polygons. The sizes of
objects are in range of [30, 40]. No rotation is applied to objects to assure the bilateral symmetry at the object level. After training, we create new deliberate samples to
test the generalization. Two sets of deliberate samples are generated. The first testing set consists of new objects of the same size range. The symmetric objects include
hexagram, 4-leaf flower (F4), and 2-leaf flower (F2). The asymmetric objects are created by using similar operation of D2 (·), which is scaling of, or adding object to, the
symmetric polygon to make it asymmetric. Some examples are shown in Fig. 7. The
second testing set is constructed by using the training objects of larger sizes, in range
of [40, 45].
4.3
Normal/Tampered human face
The most obvious symmetry we see every day, and are very sensitive to, is probably
in the human face, so we design an experiment to distinguish normal and tampered
human faces. Although human faces are not completely symmetrical, higher degree of
facial symmetry has been shown to be correlated with beauty, attractiveness [37], and
personality [38]. Recently, deep learning has shown great power in face identification
and verification [4, 39]. In this sub-task, we aim to train and test DCNN’s awareness
to facial symmetry.
We collected frontal face images from two public databases. Yale cropped face
database[40, 41] and AT&T face database[42].The tampered faces are created by fusing
two half faces together (each half came from a different subject). Some examples are
shown in Fig. 8.
To fuse two faces, we first use DLIB implementation of real-time face pose estimation algorithm [43] to detect 68 facial landmarks. These landmarks are defined in
[44]. Secondly, we rotate each face so that the mid-line passing through the nose is
vertical. Thirdly, we normalize the two faces to be fused to the same height, keeping
the original width. Then, we align two faces together based on the face center (defined
9
Figure 8: Example of real human faces (top 2 rows) vs. tempered faces (bottom 2 rows)
in training set.
as mass center of eyes, nose and mouth). Finally, we merge two faces into one by using
a sigmoid horizontal blending filter which is defined as:
ω=
1
(1 + exp( x−W/2
))
σ
,
I(x, y) = ω ∗ I1 (x, y) + (1 − ω) ∗ I2 (x, y)
(1)
(2)
in which, (x,y) is the coordinates of image pixel; W is the image width; I, I1 and I2
are target image and two source images respectively; σ = 4 in this study.
5
Counting
The second major task in this study is counting. We design two sub-tasks related to
counting in this section. The first one is simply counting objects. The second and also
more difficult one is counting object types.
Six basic shapes are used to generate shape objects: equilateral triangle, square,
ball, hexagram, 4-leaf flower (F4), and 2-leaf flower (F2). All shapes in an image can
have different sizes, positions, intensities, and orientations in some cases. Fig. 9 shows
a sample image. The size of each shape is measured by the longer edge of its bounding
box.
10
Figure 9: Example of synthetic objects.
5.1
Counting objects
To formulate counting problem as binary classification task, we design data set so that
the two classes of image have different number of objects in each image sample.
Not to make the problem too trivial, a task of counting 3 objects vs. other-than-3
objects is designed. Each positive image contains 3 objects, while each negative image
contains a different number (1, 2, 4 or 5) of objects. The experiments are conducted in
three settings: in the first setting, training images only contain balls as objects. In the
second setting, training objects include triangle, square and ball, but each image only
contain one type of object. The third setting is similar to the second one, but each image
may contain mixed types of object. The objects could be located in various positions
with different intensities and orientation. The size is in range of [20, 30]. Examples of
training images are shown in Fig. 10.
Then, two deliberate testings are conducted: (1) using new objects (hexagram, F4,
F2) of size ∈ [20, 30] to test shape sensitivity; (2) using the training objects of different
size ∈ [30, 40] to test scale sensitivity.
5.2
Counting types (shape uniformity/diversity)
In this task, we design another binary classification problem: one type of shape vs. two
types of shapes. More specifically, each positive image contains multiple objects of the
same type, while each negative image contains two types of objects. The number of
objects in each image varies. Example are shown in Fig. 11. This task is harder than
counting objects, since it requires local shape discrimination and global reasoning at
the same time.
Similar to the last section, we conduct two deliberate testings to test the generalization of the trained model. In training, we only use triangle and square to create image
samples. In deliberate testing of new shapes, we use ball, hexagram and F4.
11
Setting 1
Setting 2
Setting 3
Figure 10: Example of training samples for object counting task. For each of three
settings: left — 6 positive examples containing 3 objects each; right — 6 negative
examples containing other-than-3 objects each.
12
Figure 11: Examples for task of courting types. Left: mixture of same type of shape;
right: mixture of two different types of shape.
Figure 12: Examples for task of common fate or grouping. Left: all objects are pointing
at the same target point; right: objects are not pointing at the same target.
6
Common Fate / Synchrony
In this section, another Gestalt-style experiment is designed to test “grouping or conformance behavior”, where multiple objects (e.g., pointy triangles) in a positive image
are all behaving in a consistent manner, facing a single target (e.g., a dot). Whereas
a negative image would contain objects that do not conform in the same way. Some
image examples are shown in Fig. 12.
Multiple rounds of training and testing are conducted, with the first round using
randomly oriented triangles as negative images. In subsequent rounds, we test on some
deliberately constructed testing samples. These could be negative samples with only a
few (1 or 2) outlier objects that are not targeting the focus point; or samples with fewer
or more objects; or combinations of the variations; plus size variations of the triangles.
7
Experimental results
Since our synthetic data sets are all gray images while the network requires 3-channel
input, the images are converted to 3-channel by duplicating three times. In all experiments, training takes 70 epochs, and model selection is based on minimum error on the
13
Model
M1
M1
M1
M2
M2
M3
M3
M4
Test Data
A1
C1
D1 (A1 )
A2
D2 (A2 )
A3
D3 (A3 )
D3 (A3 )
Accuracy (%)
99.98
99.9
81.31
99.84
90.85
99.73
98.39
99.63
Precision (%)
99.95
99.8
72.79
99.7
84.55
99.48
96.9
99.27
Recall (%)
100
100
100
99.99
99.98
99.99
99.98
100
Table 1: Results of global symmetry in accuracy, precision and recall.
validation set. Training batch size is 40 for all synthetic data sets. We report error rate
(ER), accuracy (ACC), recall and precision, all in percentage.
For data augmentation, we apply randomly (1) rotation in range of 5 degrees; (2)
horizontal and vertical shift in range of 2% of image width and height; (3) flip horizontally and/or vertically.
7.1
Results of global symmetry
We label symmetry as class 0, and asymmetry as class 1. In the first round, we generate
A1 , B1 and C1 . Each set has 8000 examples, 4000 in each class. Denote M1 as the
first trained model, all errors are misclassified asymmetric samples, so the recall of
symmetric image is 100%. In the second round, model M2 is trained on the enlarged
datasets A2 and B2 . Each of the data set has 16000 examples, 8000 in each class.
Similarly, in the third round training, M3 is trained on A3 and B3 , each contains 32000
examples. We report all training and deliberate testing results in table 1.
The ER’s in the three rounds of training and testing are plot in Fig. 13. There are
two main observations from this plot. One is that as the number of training samples increases (by incorporating more deliberate samples), the testing errors become smaller.
This shows the importance of large and representative training data. The other one
is that the generalization of the trained model is not automatically achieved when the
training set is relatively small, since the deliberate testing results are much worse than
the training error. For example, the ER of D1 (A1 ) is about 190 times worse than that of
A1 . However, this is not the case in human testing, which we will report subsequently.
To better understand the behavior of the trained models, we create a new set (denoted by A4 ) with 4000 images in each class. These samples are generated by using
simple shapes (triangle, square and ball) only. Some examples are shown in Fig. 5. 1st
model ER: 6.33%, precision 88.77% and recall 100%. 2nd model ER: 0.83%, precision
98.38% and recall 100%. 3rd model ER: 0.74%, precision 98.57%, recall 99.98%. We
can see that the models trained in the three rounds have improved ability of symmetry
recognition as more deliberate samples are included in training. It is interesting to see
that almost all the errors come from the misclassified asymmetric images. This makes
sense because the asymmetric patterns have much larger variance than symmetric ones.
14
Figure 13: Training/testing errors of the first 3 rounds.
Figure 14: Example of final testing samples in C4 . Left: 9 symmetric samples; right:
9 asymmetric samples.
15
Figure 15: Example of error cases from 3rd deliberate test set using the 4th round
model. All errors come from misclassified asymmetric samples.
Finally, we conduct the 4th round of training by combining training sets A4 and A3 ,
and test the final model M4 on 8000 new testing samples (denoted by C4 ). These final
testing samples are created like A4 by using new types of shape (hexagram, F4 and F2).
Some examples are shown in Fig. 14. The final model achieves 0% ER on C4 . We also
test the final model on D3 (A3 ). The ER decreases from 1.61% to 0.37%.By looking at
the error cases produced by the final model (samples are shown in Fig. 15), we can see
that the model only fail in some fine-scale details in some asymmetric images.
These final numbers, although very impressive, indicate the lack of semantic-level
learning of this relatively simple concept. And the lack of sensitivity to fine-scale
details is consistent with the anecdotal errors reported in the literature.
Human test
We developed a simple game to test human performance in the same task, which
consists of maximum three rounds of training and four rounds of testing. The three sets
of training samples come from A1 , A2 and A3 . The four testing sets come from C1 ,
D1 (A1 ), D2 (A2 ) and D3 (A3 ).
To maximize consistency of the tests across different human subjects, a written instruction (see Appendix) is used and verbal communication was kept to a minimum.
Each subject is firstly shown 12 pairs of positive and negative examples from two
classes. A subject can request to see more examples (in increments of 3) from ei16
Test Data
Statistical test
Deliberate test 1
Deliberate test 2
Accuracy (%)
98.39
56.57
97.47
Precision (%)
96.9
54.07
95.28
Recall (%)
99.98
87.4
99.9
Table 2: Results of local symmetry in accuracy, precision and recall.
ther class. As soon as the subject believes that he/she has learned the classification
rule, we test them on 20 random examples. If succeed with no errors, a next round of
testing, each with 20 random examples, will be conducted. In case the subject make
mistakes in any round of testing, a new training session using corresponding training
set will be given.
We tested on 15 subjects individually in two groups. The first group of 7 subjects
were tested on biased training samples, in which each symmetric image has only one
connected component as the foreground object. The second group of 8 subjects were
tested on unbiased training samples, in which symmetric samples have more variance
(can have separated symmetric objects in an image). In the first group, 5 subjects succeeded to pass all four tests by learning from 24, 24, 30, 39, 48 examples, respectively.
Interestingly, 2 subjects in the first group failed to learn the rule. These two subjects
came up with different hypotheses for discriminative object properties, such as angle,
roundness, connectivity, “looks like living things?” and so on. In the second group, all
subjects succeeded to pass all tests by learning from 24, 24, 30, 33, 36, 48, 64, 300
examples, respectively. Among 13 passing subjects, four learned from only 12 pairs
of examples in the first training round; the rest required more examples from 1st or
2nd training round (made a few errors and quickly corrected them after seeing more
deliberate training samples).
We found that the most tricky testing examples are a couple of the asymmetric
images, in which the small and subtle asymmetry was missed by human eyes, when
testing was conducted in a fast pace.
7.2
Results of local symmetry
We create 8000 training samples and 8000 validation samples to train the model. The
local symmetric and asymmetric patterns are created in the similar way as creating A1
in global symmetry task. Symmetric patterns also include small objects, like, triangle,
square and ball. All foreground objects have size range ∈ [30, 40]. After training, 8000
statistical testing samples are created in the same manor so that they have same data
distribution as the training set.
Deliberate test 1: This test set is created by using new types of objects. Some
examples are shown in Fig. 7. Deliberate test 2: This test set is created by using the
same types of objects as in training, but larger size. The new sizes range in [40, 45].
All results are reported in table 2. From the deliberate testing results, we can see
that the learned model for local symmetry recognition is sensitive to unseen objects,
but not sensitive to different object sizes.
17
Figure 16: All error cases from testing set. Top samples are misclassified as tampered,
while bottom samples are misclassified as real.
7.3
Results of normal vs. tampered human face
We firstly use Yale-cropped and AT&T data sets for training and statistical testing.
Train 1574 samples: 462 real (positive) and 1112 tampered (negative). Test 1450 samples: 338 real and 1112 tampered. Due to limited samples, we use the training set as
validation set in training process. The subjects in testing set are in the same population
of the training set, but the faces have varied illumination and facial expressions. For the
tampered face class, two faces I1 and I2 are swapped to create testing samples. Test
accuracy: 98.76%, precision 98.48%, recall 96.15%. Error cases are shown in Fig. 16.
When we tested the best symmetry model M4 from section 4.1 trained on A4 and
A3 on this face problem (only test on well aligned and cropped Yale set: 380 real
faces and 1406 tampered faces), the accuracy is 81.92%, and precision 64.47%, recall
33.42%. Most errors come from misclassified real faces. Random examples of error
cases are shown in Fig. 17. It shows that the symmetry model cannot identify real faces
well, which should be an easy task for human.
Human test
Fig. 8 was shown to two children, a 12-year-old and a 14-year-old, with the question
“what is the difference between the two groups of images?” Within 20 and 5 seconds
respectively, both realized that the second group of faces are all combinations from
different people. We did not yet conduct the human test on a larger scale, which is part
of the future work to establish more human performance benchmarks.
7.4
Results of counting objects
As described in section 5.1, three settings of training examples are generated (all have
the same number of training samples, 2000 positive and 2000 negative) to learn three
18
Figure 17: Example of error cases by pre-trained symmetry model. Top samples are
misclassified as tampered, while bottom samples are misclassified as real.
19
Test Data
Setting 1
Statistical test
Deliberate test 1
Deliberate test 2
Setting 2
Statistical test
Deliberate test 1
Deliberate test 2
Setting 3
Statistical test
Deliberate test 1
Deliberate test 2
Accuracy (%)
Precision (%)
Recall (%)
99.98
95.63
62.5
100
98.98
79.83
99.95
92.2
33.45
99.98
99.95
88.55
100
99.9
97.53
99.95
100
79.1
99.98
99.95
85.12
99.95
99.9
96.43
100
100
72.95
Table 3: Results of object counting in accuracy, precision and recall.
models. Different testing sets are generated to evaluate these models. For each setting,
new statistical testing samples, which have the same data distribution as the training set,
are generated and tested. Then two deliberate testing sets are generated. Deliberate
testing 1: 4000 testing samples are generated by using completely different objects
(hexagram, F4, and F2). Deliberate testing 2: 4000 testing samples are generated by
using the same objects as training (triangle, square, ball), but the sizes are in range of
[30, 40], which is about 50% bigger than training shapes ([20, 30]). All testing results
are reported in table 3. For the three training settings, the ER’s of deliberate testing 2
become around 1800, 600, 700 times worse than that of statistical testings, respectively.
The model trained in setting 2 outperforms that of other two settings in deliberate
testing sets. The results show that training examples with more types of object are
useful in this task. However, mixed types of object in one image may be distracting.
With only limited number of training examples, such variance could mislead or delay
the training convergence to some extent.
Human test
Our hypothesis is that humans can quickly discover the counting rule after given a
few of the Setting 1 examples as shown in Fig. 10. Once their brains are primed with
this concept, Setting 2 and 3 testing will become easy, with zero training, or at most a
couple more training samples for confirmation, and the subsequent testing performance
will be 100%, with invariance to size, shape, or location etc. A quick human test of a
few subjects confirmed this hypothesis. A larger scale study is left as future work.
7.5
Results of counting types
In training, we use triangle and square to generate all images (4000 training and 4000
validation samples). The default size range is [20, 30]. Deliberate testing 1: We include 3 more object types into testing, i.e. ball, hexagram and F4. In this testing set,
each sample in single-type class may contain one of the three new shapes. Each sample
20
Test Data
Statistical test
Deliberate test 1
Deliberate test 2 [30,40]
Deliberate test 2 [40,50]
Additional test 1
Additional test 2
Accuracy (%)
99.98
73.67
86.6
56.7
83.1
72.75
Precision (%)
99.95
76.68
97.66
78.39
79.53
69.36
Recall (%)
100
68.05
75.0
18.5
89.15
81.5
Table 4: Results of type counting in accuracy, precision and recall.
in the two-type class includes any 2 combination of the 5 shapes. Deliberate testing 2:
We also test on new samples with larger objects. Two sets are created with size range
[30, 40] and [40, 50], respectively.
Additional test 1: Similar to global symmetry task, we can add new samples to
train new model, which is expected to have less over-fitting and better generalization.
We add F4 into training shapes to train a new model and test on the same testing set
as in deliberate testing 1. Additional test 2: Following deliberate testing 2, we train a
new model using samples with shapes of small and large sizes, specifically sizes range
in [20, 25] ∪ [40, 45], and test on size [30,35] to see whether there is improvement of
generalization.
Results are reported in table 4, which show that the learned model is sensitive to
new object shapes and scales. Even after additional training with new training samples,
the models still do not get the logic of type counting.
7.6
Results of common fate
We trained 5 models in sequential rounds, each time adding some new variations of
image samples into the training set. The first dataset contains the samples (4000 training, 4000 validation) similar to Fig. 12, but each image has n ∈ [10, 17] objects and
one target point. In negative samples, all objects are facing random directions. In the
second set, 4000 new samples are added in both training and validation, where all objects except one or two outliers in the negative images are facing the target (outliers
are facing at least 60 degrees away from the target). In the third set, new samples are
added again, where all images contain n ∈ [30, 34] objects, and negative images each
have only one outlier. In the fourth set, new samples similar to the last round are added,
but each with fewer objects (n ∈ [5, 7]). Some negative image examples are shown in
Fig. 18. In the fifth set, the number of training/validation samples are doubled. In each
training round, the best model was picked based on accuracy on the validation set.
In this task, we observed the same behavior as in the other tasks, where deliberate
tests yielded clear accuracy degradation, and training with deliberate samples would
bring the accuracy back up. Below we report a different set of accuracy numbers, all
based on the same hold-out test set, which is different from all the training sets.
The hold-out testing images are more sparsely populated than all the training images, with each image containing only two objects and each negative sample has one
21
Figure 18: Examples of additional negative images in the 3rd and 4th rounds. Left:
negative images with more objects and 1 outlier in 3rd round; right: negative images
with less objects and 1 outlier in 4th round.
Figure 19: Learning curve (green) of Inception-V3 in common fate task based on the
hold-out testing set. The blue point indicates the performance on the altered test set
with larger triangles.
22
Figure 20: Error cases on the hold-out test set by the 5th model.
outlier. This hold-out set was used to evaluate the performance of all trained models,
in order to comparably observe the learning progression. The learning curve (green)
is plotted in Fig. 19. The learning curve shows a promising accuracy (99.58%) for the
final model trained and validated with 64000 images. However, the model still makes
errors. Some of the error cases are shown in Fig. 20.
Finally, we altered the hold-out test set, to make all triangles bigger (by doubling
their size). The testing accuracy dropped to around 74% (the blue point in Fig. 19).
Interestingly, all errors are false negatives.
Human test
Seven human subjects participated in the classification task (using the first set). All
subjects used less than 20 training samples to reach the Aha-moment of grasping the
semantic concept.
One of the subjects was tested for unsupervised learning as a first step. Four training
images, two positive and two negative, are shown to the subject with the question
“can you separate these into two groups?” The human subject succeeded in both the
clustering task as well as the subsequent classification task (“can you put these images
into the two groups you just identified?), with 100% accuracy.
8
8.1
Discussions
The intelligence gap
It seems that in at least two aspects machine learning has not closed the “intelligence
gap”: one is that it needs many more training samples than humans do; and the other is
that the “aha moment” seems to be still uniquely human, and machines have not show
any sign of mastering this trick.
The comparison of learning curves by human and deep learning in symmetry task
23
Figure 21: Learning curve of human vs Inception-V3 in global symmetry task.
is shown in Fig. 21. The human accuracy is calculated as the percentage of head count
who reach the “aha moment”, at sample size 24, 36, 48, 64 and 300. We can see that
human can learn the rule from examples quickly and come to the “aha moment” after
seeing very few examples. On the contrary, DCNN can only statistically approximate
the bilateral symmetry after a large number (40,000) of training examples, and still
cannot reach 100% accuracy on unseen data.
Artificial Neural Networks have indeed shown more sophisticated “intelligence”
than counting. One example was demonstrated by Hoshen and Peleg [45]: they were
able to get a simple Neural Network to learn the mathematical operation of addition
from pictures of numbers as inputs. Considering that human children typically learn
counting first before they are taught how to add numbers, this achievement is quite
impressive. Nevertheless, DCNN’s rather rudimentary visual counting ability serves as
a reminder that there is still some way to go to achieve a basic but semantic level of
artificial visual intelligence.
8.2
Study implications
Imagine that you have access to anonymized brain scans from 10,000 patients. 2000
of them eventually died of a certain disease, and the rest lived long and were free of
that disease. Now you want to find out whether the brain scans could have predicted
the disease. You could try an “end-to-end” learning using DCNN by feeding the brain
scans as training examples. Will it work? Or will you need to understand more about
the disease and its (potentially very complex) manifestations in the brain scan?
Let’s assume that the disease manifestation in the brain is actually based on differ-
24
ent configurations of some “hot spots” (i.e., high intensity areas), with an asymmetric
configuration indicating disease, while symmetry predicts no-disease. Or, the number
of hot spots within a certain region of the brain is indicative of the disease: having,
say, three or less hot spots means healthy, and more means disease, regardless of shape
or size of each hot spot. Or, it is the diversity/heterogeneity of those hot spots that
mattered. Or, it is a combination of these rules for different parts of the brain that are
working together to determine the patient’s fate. But you do not know any of these
“rules” a priori.
Based on the experiments in this paper, you have reason to stop and think, and to
doubt that a direct application of DCNN will always solve the problem. And you know
roughly how many more examples you may need if you do suspect a particular rule is
at work. Or, at least you know how to design synthetic experiments to shed some lights
on this problem, e.g., to look for a performance limit, in case you have failed a first
round of “end-to-end” learning.
8.3
Is visual intelligence learned or hard-wired?
One might argue that symmetry is not a fair test case, because humans may be hardwired to be sensitive to it. Indeed, symmetry is very common in our world, both in
natural anatomy and in man-made objects. And being sensitive to symmetry probably
carries a strong evolutionary advantage. But then again, one can also argue that most,
if not all, of human visual intelligence may be somewhat hard-wired. If this is true,
then does it mean that an “end-to-end” learning-from-examples paradigm may not be
sufficient to achieve full visual intelligence? And that proper understanding, followed
by direct hard-coding (i.e., semantic modeling), of such wirings may be at least an
indispensable stepping-stone?
8.4
Why synthetic testing?
Another question could be, why do we need synthetic examples, while we have already
large annotated data sets like ImageNet? We believe popular learning targets such as
cats, dogs, human faces, or more generally, animals on earth, are not the best targets
for testing a learning machine rigorously. Animal species today are very sparsely distributed on the evolutionary continuum due to multiple prehistorical mass extinctions—
there is no other animals with zebra stripes, or any other animal with a long nose like
an elephant. When a learning machine sees a striped table cloth and calls it a “zebra”,
or calls a child wearing a long-nosed mask an “elephant”, is it wrong?—After all, the
table cloth could be a cutout from a massive zebra picture, and “elephant” would be
the exact right name to call a child during her Halloween adventure. Therefore, when
a CNN is fooled by some noise-like adversarial overlays, could it be that it is seeing a
“zebra stripes”- or “elephant-nose”-type of unique feature for that object? What really
sets humans apart is that humans see the full picture: the table and the child together
with all the other distractions. In order to see the full picture, then, CNNs have to learn
all the objects and concepts at the same time. This is not yet possible today.
In this paper, we step back and investigate at a much smaller scale — we focus
on simple and well-defined concepts, and clean and synthetic examples to test and
25
compare the “intelligence” of algorithms and humans, in a quantitative way.
8.5
Error analysis
For symmetry tasks, DCNN made more mistakes in classifying asymmetric patterns as
symmetric ones. This behavior is somewhat like that of a human. The reason may lie,
on the one hand, in the fact that symmetry is a precisely defined concept in which even
a one-pixel-difference would make it asymmetric; and on the other hand, in the limited
ability for fine-scale discrimination, of humans as well as of machines. Humans often
make careless mistakes or suffer from sensory defects (e.g.,short-sightedness); while
today’s DCNNs are not sensitive to fine-grained variations in the images, unless redesigned specifically for a particular task [46].
In the local symmetry study and facial study, the DCNN models learned very well
on seen shapes/distributions. However, it demonstrated limited generalization in terms
of new objects (second row in Table 2), even though most of the errors should be very
easy for human to avoid (e.g. error cases in Fig. 16).
The counting experiments show that DCNN is too sensitive to object scale in the
object-counting and common fate task; and failed to generalize in scale and object type
in the object-type-counting task. With limited amount of training samples, the DCNN
model cannot learn the counting rules precisely. It tends to learn a small kernel space
that could just cover the training sample distribution. For example, in the last scale testing experiment, the learned model can learn to fit the small and large scales, yet when
tested at medium scale, the error rate spikes up. Similarly, objects with a new unseen
shape seem to be able to easily confuse the learned models. In the common fate task,
we see clearly that scale-invariance is not achievable without exhaustive enumeration
or explicit modeling/coding.
Humans have the prior knowledge (or tendency) that each connected component is
counted once, regardless of its shape or size variations. Our trained DCNN model does
not have this explicit knowledge or tendency. So, it tends to, interestingly, count twice
or more for objects larger than what it saw in the training set. But can we really judge
that this behavior is wrong? or merely “uninformed”? I.e., uninformed of our world in
which scale-invariance is a common law due to both the natural growth phenomenon
and our perspective visual system. Imagine an alien world where things grow in density
instead of size, and a sensing system based on parallel waveforms or touch only, then
scale-invariance would be a very alien concept. This brings the question: how much
of our visual intelligence is specific to this world that we live in, and to this bodily
construct that we possess? and not universal at all?
In the end, a machine may have to live among us to learn the whole experience,
before acquiring Gestalt-style-intelligence in visual perception.
8.6
Sample size and general intelligence
Our experiments show that some “intelligence gaps” can be asymptotically closed with
sufficiently large training data, albeit at very different convergence rate. One can hope
that refinement of the network architecture, trained with ever larger data sets, will bring
26
about an eventual breakthrough. However, others may believe that “general intelligence”, the kind of intelligence that understands relationships and complex rules, and
that sees the full picture, can be reached only by tackling the Gestalt-type “aha moment” problem (plus the Winograd Schema Challenge [34]) first. Our object-counting
experiments support the latter view. For one thing, a DCNN can learn that an object
of many different sizes is the same thing as long as examples of each size is present in
the training set. But it could never learn or infer by itself the rule that “size does not
matter!”—this rule has to be hard-coded.
Even if a data-driven deep learning algorithm can learn complex rules in a given
domain, it may still have a difficult time combining rules or prior knowledge from
multiple different perspectives (e.g., “size does not matter”, “rotation does not matter”,
but “shape matters”). With an increasing number of potentially hidden rules in effect,
the required sample size for “end-to-end” training will grow exponentially.
8.7
Future work: Aha Challenge, a call for participation
The datasets used in this work and scripts to generate the synthetic images are available
online at https://github.com/zhennany/synthetic.
As future work, we would like to propose an “Aha Challenge” to the research community. The idea is to push together focused and quantitative research on algorithmic
vs. natural visual intelligence, toward Gestalt-style pattern recognition based on a relatively small number of training images.
The goal of this challenge is two-fold: to achieve a deeper, and quantitative, understanding of the nature of human visual intelligence; and to improve higher- or semanticlevel “intelligence” in visual learning algorithms.
Participation can be in three forms: submission of new algorithms; submission of
human study results; and proposal for new types of tests.
An algorithm submission should be tested on multiple visual learning tasks, each
designed with deliberate adversarial testing that a human can easily pass (once an “aha
moment” has been reached).
A human subject study should be conducted as comparable as possible to the algorithm training process, with minimal extra information or hints provided, besides
showing the training images to the human subjects.
This challenge is analogous to the ”Winograd Schema Challenge” [34] in the language understanding domain. As demonstrated by the Winograd Challenge, if the
machine does not have all the knowledge of this human world, it cannot correctly infer
all the meanings of, or relationships among, words or visual objects.
We invite submissions of either new types of tests, new algorithms, or human study
results, to establish “intelligence” baselines, and to build machines that one day will be
able to exclaim “Aha! I see!”.
APPENDIX: Written instruction for the human test
Below is the written instruction used in the human test for the symmetry study.
27
You are participating in a game of “visual classification”. The goal
for you is to learn to classify visual patterns into two classes, after seeing
examples of both classes.
Here is how the game will proceed:
1. You will be shown examples of visual patterns from two classes,
class 0 on the left of the screen and class 1 on the right;
2. As soon as you believe that you have learned the classification rule,
you can stop the training;
3. You will be shown 20 random test examples. Please label each as
either class 0 or class 1;
4. If you succeed with no errors, 3 additional rounds of testing, each
with 20 patterns, will be conducted to confirm your learning;
5. In case you make mistakes in any round of testing, a new training
session will be given to strengthen your learning. Again you can stop the
training once you believe you have learned the classifier;
6. The game stops after a maximum of 4 rounds.
Please keep this confidential so that we can invite more participants.
References
[1] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444.
[2] J. Schmidhuber, Deep learning in neural networks: An overview, Neural networks 61
(2015) 85–117.
[3] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional
neural networks, in: Advances in neural information processing systems, 2012, pp. 1097–
1105.
[4] Y. Taigman, M. Yang, M. Ranzato, L. Wolf, Deepface: Closing the gap to human-level
performance in face verification, in: Proceedings of the IEEE conference on computer
vision and pattern recognition, 2014, pp. 1701–1708.
[5] K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: Surpassing human-level performance on imagenet classification, in: Proceedings of the IEEE international conference
on computer vision, 2015, pp. 1026–1034.
[6] D. H. Hubel, T. N. Wiesel, Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex, The Journal of physiology 160 (1) (1962) 106–154.
[7] D. J. Felleman, D. C. Van Essen, Distributed hierarchical processing in the primate cerebral
cortex., Cerebral cortex (New York, NY: 1991) 1 (1) (1991) 1–47.
[8] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: a simple way to prevent neural networks from overfitting., Journal of machine learning research
15 (1) (2014) 1929–1958.
[9] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp.
770–778.
[10] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke,
A. Rabinovich, Going deeper with convolutions, in: Proceedings of the IEEE conference
on computer vision and pattern recognition, 2015, pp. 1–9.
28
[11] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2016, pp. 2818–2826.
[12] M. D. Zeiler, R. Fergus, Visualizing and understanding convolutional networks, in: European conference on computer vision, Springer, 2014, pp. 818–833.
[13] A. Mahendran, A. Vedaldi, Understanding deep image representations by inverting them,
in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2015,
pp. 5188–5196.
[14] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, R. Fergus, Intriguing properties of neural networks, arXiv preprint arXiv:1312.6199.
[15] I. Evtimov, K. Eykholt, E. Fernandes, T. Kohno, B. Li, A. Prakash, A. Rahmati,
D. Song, Robust physical-world attacks on machine learning models, arXiv preprint
arXiv:1707.08945.
[16] C. L. Giles, T. Maxwell, Learning, invariance, and generalization in high-order neural networks, Applied optics 26 (23) (1987) 4972–4978.
[17] C. M. Bishop, Neural networks for pattern recognition, Oxford university press, 1995.
[18] S. Basu, M. Karki, S. Mukhopadhyay, S. Ganguly, R. Nemani, R. DiBiano, S. Gayaka, A
theoretical analysis of deep neural networks for texture classification, in: Neural Networks
(IJCNN), 2016 International Joint Conference on, IEEE, 2016, pp. 992–999.
[19] N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik, A. Swami, Practical blackbox attacks against deep learning systems using adversarial examples, arXiv preprint
arXiv:1602.02697.
[20] V. Behzadan, A. Munir, Vulnerability of deep reinforcement learning to policy induction
attacks, arXiv preprint arXiv:1701.04143.
[21] S. Huang, N. Papernot, I. Goodfellow, Y. Duan, P. Abbeel, Adversarial attacks on neural
network policies, arXiv preprint arXiv:1702.02284.
[22] I. J. Goodfellow, J. Shlens, C. Szegedy, Explaining and harnessing adversarial examples,
arXiv preprint arXiv:1412.6572.
[23] G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv preprint
arXiv:1503.02531.
[24] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville,
Y. Bengio, Generative adversarial nets, in: Z. Ghahramani, M. Welling, C. Cortes, N. D.
Lawrence, K. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems
27, Curran Associates, Inc., 2014, pp. 2672–2680.
[25] J. Snell, K. Swersky, R. S. Zemel, Prototypical networks for few-shot learning, arXiv
preprint arXiv:1703.05175.
[26] S. Ravi, H. Larochelle, Optimization as a model for few-shot learning, in: International
Conference on Learning Representations, 2017.
[27] B. M. Lake, R. Salakhutdinov, J. B. Tenenbaum, Human-level concept learning through
probabilistic program induction, Science 350 (6266) (2015) 1332–1338.
[28] L. Fei-Fei, R. Fergus, P. Perona, One-shot learning of object categories, IEEE transactions
on pattern analysis and machine intelligence 28 (4) (2006) 594–611.
[29] M. Palatucci, D. Pomerleau, G. E. Hinton, T. M. Mitchell, Zero-shot learning with semantic
output codes, in: Advances in neural information processing systems, 2009, pp. 1410–
1418.
29
[30] R. Socher, M. Ganjoo, C. D. Manning, A. Ng, Zero-shot learning through cross-modal
transfer, in: Advances in neural information processing systems, 2013, pp. 935–943.
[31] T. G. Evans, A heuristic program to solve geometric-analogy problems, in: Proceedings of
the April 21-23, 1964, spring joint computer conference, ACM, 1964, pp. 327–338.
[32] S.-C. Zhu, C.-E. Guo, Y. Wang, Z. Xu, What are textons?, International Journal of Computer Vision 62 (1) (2005) 121–143.
[33] T.-F. Wu, G.-S. Xia, S.-C. Zhu, Compositional boosting for computing hierarchical image
structures, in: Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on, IEEE, 2007, pp. 1–8.
[34] H. J. Levesque, E. Davis, L. Morgenstern, The winograd schema challenge., in: AAAI
Spring Symposium: Logical Formalizations of Commonsense Reasoning, Vol. 46, 2011,
p. 47.
[35] J. Wagemans, J. H. Elder, M. Kubovy, S. E. Palmer, M. A. Peterson, M. Singh, R. von der
Heydt, A century of gestalt psychology in visual perception: I. perceptual grouping and
figure–ground organization., Psychological bulletin 138 (6) (2012) 1172.
[36] F. Jäkel, M. Singh, F. A. Wichmann, M. H. Herzog, An overview of quantitative approaches
in gestalt perception, Vision research 126 (2016) 3–8.
[37] K. Grammer, R. Thornhill, Human (homo sapiens) facial attractiveness and sexual selection: the role of symmetry and averageness., Journal of comparative psychology 108 (3)
(1994) 233.
[38] B. Fink, N. Neave, J. T. Manning, K. Grammer, Facial symmetry and the bigfivepersonality factors, Personality and individual differences 39 (3) (2005) 523–529.
[39] Y. Sun, Y. Chen, X. Wang, X. Tang, Deep learning face representation by joint
identification-verification, in: Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence,
K. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 27, Curran
Associates, Inc., 2014, pp. 1988–1996.
[40] A. Georghiades, P. Belhumeur, D. Kriegman, From few to many: Illumination cone models
for face recognition under variable lighting and pose, IEEE Trans. Pattern Anal. Mach.
Intelligence 23 (6) (2001) 643–660.
[41] K. Lee, J. Ho, D. Kriegman, Acquiring linear subspaces for face recognition under variable
lighting, IEEE Trans. Pattern Anal. Mach. Intelligence 27 (5) (2005) 684–698.
[42] F. S. Samaria, A. C. Harter, Parameterisation of a stochastic model for human face identification, in: Applications of Computer Vision, 1994., Proceedings of the Second IEEE
Workshop on, IEEE, 1994, pp. 138–142.
[43] V. Kazemi, J. Sullivan, One millisecond face alignment with an ensemble of regression
trees, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1867–1874.
[44] C. Sagonas, G. Tzimiropoulos, S. Zafeiriou, M. Pantic, 300 faces in-the-wild challenge:
The first facial landmark localization challenge, in: Proceedings of the IEEE International
Conference on Computer Vision Workshops, 2013, pp. 397–403.
[45] Y. Hoshen, S. Peleg, Visual learning of arithmetic operation., in: AAAI, 2016, pp. 3733–
3739.
[46] T.-Y. Lin, A. RoyChowdhury, S. Maji, Bilinear cnn models for fine-grained visual recognition, in: Proceedings of the IEEE International Conference on Computer Vision, 2015, pp.
1449–1457.
30
| 7cs.IT
|
arXiv:1711.05485v2 [math.AC] 16 Nov 2017
Prüfer intersection of valuation domains of a field of rational
functions
G. Peruginelli∗
November 17, 2017
Abstract
Let V be a rank one valuation domain with quotient field K. We characterize the subsets S
of V for which the ring of integer-valued polynomials Int(S, V ) = {f ∈ K[X] | f (S) ⊆ V } is a
Prüfer domain. The characterization is obtained by means of the notion of pseudo-monotone
sequence and pseudo-limit in the sense of Chabert, which generalize the classical notions of
pseudo-sequence and pseudo-limit by Ostrowski and Kaplansky, respectively. We show that
Int(S, V ) is Prüfer if and only if no element of the algebraic closure K of K is a pseudo-limit of
a pseudo-monotone sequence contained in S, with respect to some extension of V to K. This
result expands a recent result by Loper and Werner.
Keywords: Prüfer domain, pseudo-convergent sequence, pseudo-limit, residually transcendental extension, integer-valued polynomial
MSC Primary 13F05, Secondary 13F20, 13A18
1
Introduction
An integral domain D is Prüfer if DM is a valuation domain for each maximal ideal M of D. A
Prüfer domain D enjoys an abundance of properties (see for example [11]), among which there is
the fact that D is integrally closed. By a celebrated result of Krull, every integrally closed domain
with quotient field K can be represented as an intersection of valuation domains of K. Conversely,
it is of extreme importance to establish when a given family of valuation domains of a given field
K intersects in a Prüfer domain with quotient field K. This problem has also connections to real
algebraic geometry, since the real holomorphy rings of a formally real function field is well-known to
be a Prüfer domain (see for example [10, §2.1]). Different authors have investigated this problem:
for example, Gilmer and Roquette gave explicit construction of Prüfer domains constructed as
intersection of valuation domains, or, which is the same thing, as the integral closure of some
subring (see [12] and [24], respectively). Recently, Olberding gave a geometric criterion on a subset
Z of T
the Zariski-Riemann space of all the valuation domains of a field in order for the holomorphy
ring V ∈Z V to be a Prüfer domain; this criterion is given in terms of projective morphisms of Z,
considered as a locally ringed space, into the projective line (see [19]). In [20] Olberding gave a
∗ Dipartimento di Matematica ”Tullio Levi-Civita”, Università di Padova, Via Trieste 63, 35121 Padova, Italy.
E-mail: [email protected].
1
sufficient condition on a family of rank one valuation domains which satisfies certain assumptions
so that the intersection of the elements of the family is a Prüfer domain.
In this paper we focus our attention to the relevant class of polynomial rings called integervalued polynomials. Classically, given an integral domain D with quotient field K and a subset S
of D, the ring of integer-valued polynomials over S is defined as:
Int(S, D) = {f ∈ K[X] | f (S) ⊆ D}.
For S = D, we set Int(D, D) = Int(D). We refer to [6] for a detailed treatment of this kind of rings.
If D is Noetherian, Chabert and McQuillan independently gave sufficient and necessary conditions
on D so that Int(D) is Prüfer (see [6, Theorem VI.1.7]). Later on, Loper generalized their result to
a general domain D (see [15]). The problem of establishing when Int(S, D) is a Prüfer domain for
a general subset S of D is considerably more difficult, see [16] for a recent survey on this problem.
Since a necessary condition for Int(S, D) to be Prüfer is that D is Prüfer, it is reasonable to work
locally. Henceforth, we consider D to be equal to a valuation domain V .
The ring Int(S, V ) can be represented in the following way as an intersection of a family of
valuation domains of the field of rational functions K(X) and the polynomial ring K[X] (which
likewise can be represented as an intersection of valuation domains lying over the trivial valuation
domain K):
\
Int(S, V ) = K[X] ∩
Ws
s∈S
where, for each s ∈ S, Ws is the valuation domain of those rational functions which are integervalued at s, i.e.: Ws = {ϕ ∈ K(X) | ϕ(s) ∈ V }. In the language of Roquette [24], a rational
function ϕ ∈ K(X) is holomorphic at Ws (or, equivalently, ϕ has no pole at Ws ) if and only if ϕ is
integer-valued at s. Clearly, Ws lies over V , and, in the case V has rank one, Ws has rank two. The
topology on the subspace of the Riemann-Zariski space of K(X) formed by the valuation domains
Ws , s ∈ S, has been extensively studied in [23], when V has rank one: in particular, {Ws | s ∈ S}
as a subspace of the Zariski-Riemann space of all the valuation domains of K(X) is homeomorphic
to S, considered as a subset of V , endowed with the V -adic topology.
For a general valuation domain V , we have the following well-known result (which of course can
be deduced by the above result of Loper in [15]):
Theorem 1.1. [6, Lemma VI.1.4, Proposition VI.1.5] Let V be a valuation domain. Then Int(V )
is a Prüfer domain if and only if V is a DVR with finite residue field.
The first result on when Int(S, V ) is Prüfer dates back to McQuillan: he showed that if S is a
finite set then Int(S, V ) is Prüfer (more generally, he showed that for a finite subset S of an integral
domain D, Int(S, D) is Prüfer if and only if D is Prüfer, see [18]). Later on, Cahen, Chabert and
Loper turned their attention to infinite subsets S of a valuation domain V , and gave the following
sufficient condition (here, precompact means that the topological closure of S in the completion of
V is compact).
Theorem 1.2. [7, Theorem 4.1] Let V be a valuation domain and S a subset of V . If S is a
precompact subset of V then Int(S, V ) is a Prüfer domain.
Whether the precompact condition on S is also a necessary condition or not was a natural
question posed in [7]. Recently, Park proved that if S is an additive subgroup of V , then Int(S, V )
is a Prüfer domain if and only if S is precompact ([22, Theorem 2.7]). If V is a rank one discrete
2
valuation domain we have a complete characterization: in fact, it is sufficient and necessary that
S is precompact in order for Int(S, V ) to be Prüfer ([7, Corollary 4.3]). Unfortunately, for a nondiscrete rank one valuation domain V the precompact condition turned out to be not necessary, as
Loper and Werner showed soon after Park by considering subsets S of V whose elements comprise
a pseudo-convergent sequence in the sense of Ostrowski (for all the definitions related to this notion
see §2.1 below). It is worth recalling that the first time this notion has been successfully used
in the realm of integer-valued polynomials is in two articles of Chabert (see [8, 9]). Loper and
Werner made a thorough study of the rings of polynomials which are integer-valued over a pseudoconvergent sequence E = {sn }n∈N of a rank one valuation domain V , obtaining in particular the
following characterization of when Int(E, V ) is Prüfer.
Theorem 1.3. [17, Theorem 5.2] Let V be a rank one valuation domain and E = {sn }n∈N a
pseudo-convergent sequence in V . Then Int(E, V ) is a Prüfer domain if and only if either E is of
transcendental type or the breadth ideal of E is the zero ideal.
In particular, if E is a pseudo-convergent sequence with non-zero breadth ideal and of transcendental type, then E is not precompact and Int(E, V ) is a Prüfer domain ([17, Example 5.12]).
In this paper, we give a sufficient and necessary condition on a general subset S of a rank one
valuation domain V so that Int(S, V ) is Prüfer, generalizing the above result by Loper and Werner.
Throughout the paper, we assume that V is a rank one valuation domain with maximal ideal M
and quotient field K. We denote by v the associated valuation and by Γv the value group. In
particular, Γv is an ordered subgroup of the reals, so that Γv ⊆ R. Our approach proceeds as
follows. We employ a criterion for an integrally closed domain D to be Prüfer (which can be found
for example in the book of Zariski and Samuel [25]): it is sufficient and necessary that, for each
valuation overring V of D with center a prime ideal P on D, the extension of the residue field of
V over the quotient field of D/P is not transcendental. In our setting, a valuation overring W of
Int(S, V ) which do not satisfy the previous property is a residually transcendental extension of V
(i.e.: W lies over V and the residue field of W is a transcendental extension of the residue field of
V ). These valuation domains of the field of rational functions have been completely described by
Alexandru and Popescu. Putting together these facts, we show that the lack of the Prüfer property
for Int(S, V ) occurs precisely when S contains a pseudo-monotone sequence in the sense of Chabert
which admits a pseudo-limit in the algebraic closure of K (with respect to a suitable extension of
V ). These notions generalize the notions of pseudo-convergent sequence and pseudo-limit in the
sense of Ostrowski and Kaplansky, respectively.
Here is a summary of this paper. In §2.1 we introduce the notion of pseudo-monotone sequence
and pseudo-limit given by Chabert. In §2.2 we recall a result of Chabert about the fact that the
polynomial closure of a subset S of V , defined as the largest subset of V over which all the polynomials of Int(S, V ) are integer-valued, is a topological closure. In §3 we recall the aforementioned
criterion for an integrally closed domain to be Prüfer and some known facts about residually transcendental extensions of a valuation domain, which are crucial for our discussion. Finally, in §4,
we give our main result which classifies the subsets S of a rank one valuation domain V for which
Int(S, V ) is Prüfer (see Theorem 4.17). This result is accomplished by describing when an element
α ∈ K is a pseudo-limit of a pseudo-monotone sequence contained in S: this happens when a closed
ball B(α, γ) = {x ∈ K | v(x − α) ≥ γ} is contained in the polynomial closure of S. From this
point of view, the assumption of Theorem 1.2 is equivalent to the fact that S does not contain
any pseudo-monotone sequence, which is a sufficent but not necessary condition for Int(S, V ) to be
Prüfer, as the above example of Loper and Werner shows.
3
2
Preliminaries
2.1
Pseudo-monotone sequences
We introduce the following notion, which is given by Chabert in [8]. It contains the classical
definition of pseudo-convergent sequence of a valuation domain by Ostrowski in [21] and exploited
by Kaplansky in [13] to describe immediate extensions of a valued field.
Definition 2.1. Let E = {sn }n∈N be a sequence in K. We say that E is a pseudo-monotone
sequence (with respect to the valuation v) if the sequence {v(sn+1 − sn )}n∈N is monotone, that is,
one of the following conditions holds:
i) v(sn+1 − sn ) < v(sn+2 − sn+1 ), ∀n ∈ N.
ii) v(sn − sm ) = γ ∈ Γv , for all n 6= m ∈ N.
iii) v(sn+1 − sn ) > v(sn+2 − sn+1 ), ∀n ∈ N.
More precisely, we say that E is pseudo-convergent, pseudo-stationary or pseudo-divergent in each of
the three different cases, respectively. Case i) is precisely the original definition given by Ostrowski
in [21, §11, p. 368]. Let α ∈ K. We say that α is a pseudo-limit of E in each of the three different
cases above if:
i) v(α − sn ) < v(α − sn+1 ), ∀n ∈ N, or, equivalently, v(α − sn ) = v(sn+1 − sn ), ∀n ∈ N.
ii) v(α − sn ) = γ, for all n ∈ N.
iii) v(α − sn ) > v(α − sn+1 ), ∀n ∈ N, or, equivalently, v(α − sn+1 ) = v(sn+1 − sn ), ∀n ∈ N.
We remark that case i) is the definition of pseudo-limit as given by Kaplansky in [13]. Given a
subset S of K and an element α in K, we say that α is a pseudo-limit of S if α is a pseudo-limit
of a pseudo-monotone sequence of elements of S.
The following limit in R ∪ {∞} is called the breadth of a pseudo-monotone sequence E ([21, p.
368]):
δ = lim v(sn+1 − sn )
n→∞
Note that since {v(sn+1 − sn )}n∈N is either increasing, decreasing or stationary, the above limit
is well-defined and δ may not be in Γv . In the latter case, V is necessarily not discrete. Note that,
if E = {sn }n∈N ⊂ V and α is a pseudo-limit of E, then it is easy to see that the breadth δ is greater
than or equal to 0 and α ∈ V . We now give some remarks and further definitions for each of the
three cases above.
2.1.1
Pseudo-convergent sequences
Definition 2.2. Let E = {sn }n∈N be a pseudo-convergent sequence in V . The following ideal of
V:
Br(E) = {b ∈ V | v(b) > v(sn+1 − sn ), ∀n ∈ N}
is called the breadth ideal of E.
We say that E is of transcendental type if v(f (sn )) eventually stabilizes for every f ∈ K[X]. If
for some f ∈ K[X] the sequence v(f (sn )) is eventually increasing then we say that E is of algebraic
type.
4
Clearly, the breadth ideal is the zero ideal if and only if δ = +∞. If V is a discrete rank one
valuation domain (DVR), then the breadth ideal is necessarily equal to the zero ideal. In general,
this last condition holds exactly when E is a classical Cauchy sequence and then the definition
of pseudo-limit boils down to the classical notion of limit (which in this case is unique). In the
following, to avoid confusion, a pseudo-convergent sequence is supposed to have non-zero breadth
ideal, and similary, an element α ∈ K is a pseudo-limit of a sequence E if E is a pseudo-convergent
sequence in this strict sense. Moreover, in this case if α ∈ K is a pseudo-limit for E, then {α}+Br(E)
is the set of all the pseudo-limits for E ([13, Lemma 3]).
The following easy lemma gives a link between the breadth and the breadth ideal for a pseudoconvergent sequence (the inf is considered in R).
Lemma 2.3. Let E = {sn }n∈N ⊂ V be a pseudo-convergent sequence with non-zero breadth ideal.
Let
δ ′ = inf{v(b) | b ∈ Br(E)}
Then δ ′ = δ, the breadth of E. Moreover, δ ∈ Γv ⇔ Br(E) is a principal ideal.
Proof. Since v(sn+1 − sn ) < v(b), for all b ∈ Br(E), we have δ ≤ δ ′ . Suppose that δ < δ ′ , then
since Γv is dense in R there exists γ ∈ Γv such that δ < γ < δ ′ . Let γ = v(a), for some a ∈ K. If
a ∈ Br(E) then γ ≥ δ ′ which is not possible. If a ∈
/ Br(E), then there exists N ∈ N such that for
all n ≥ N we have v(sn+1 − sn ) ≥ γ, which also is not possible. Hence δ = δ ′ . The last claim is
straightforward.
In particular, the set of all the pseudo-limits of a pseudoconvergent sequence with non-zero
breadth ideal and with a pseudo-limit α ∈ K is equal to the ball B(α, δ) = {x ∈ K | v(x − α) ≥ δ}.
2.1.2
Pseudo-stationary sequences
Let E = {sn }n∈N ⊆ V be a pseudo-stationary sequence. Note that, in this case the breadth δ of E
is by definition in Γv , so that δ = v(d), for some d ∈ K. Moreover, the residue field V /M is infinite.
In fact, if s′n = sdn , for each n ∈ N, then E ′ = {s′n }n∈N ⊂ V ∗ is a pseudo-stationary sequence with
breadth 0, so that there are infinitely many residue classes modulo the maximal ideal M .
Suppose now that α ∈ K is pseudo-limit of a pseudo-stationary sequence E. In [8, Remark
4.7] it is remarked that any element of B̊(α, γ) = {β ∈ K | v(α − β) > γ} is a pseudo-limit of E.
However, if β ∈ K is such that v(α − β) = γ, then v(sn − β) ≥ γ for every n ∈ N. Since for all
n 6= m we have γ = v(sn − sm ) = v(sn − β + β − m), for at most one n′ ∈ N we may have the strict
inequality v(sn′ − β) > γ. Hence, up to removing one element from E, any element of B(α, γ) is a
pseudo-limit of E. In this broader sense, any element of E itself is a pseudo-limit of E.
2.1.3
Pseudo-divergent sequences
If V is discrete, then there are no pseudo-divergent sequences contained in V . Moreover, if α ∈ K
is a pseudo-limit of a pseudo-divergent sequence E = {sn }n∈N ⊂ K with breadth γ, then the set of
all the pseudo-limits in K of E is equal to the open ball B̊(α, γ) = {x ∈ K | v(x − α) > δ} (see [8,
Remark 4.7]).
Remark 2.4. We have seen that if V admits a pseudo-monotone sequence E = {sn }n∈N with
breadth γ ∈ R, then V is either non-discrete or the residue field V /M is infinite. If E is pseudostationary, then V /M is necessarily infinite and if E is pseudo-divergent or pseudo-convergent with
5
non-zero breadth ideal then V is necessarily non-discrete. In particular, the only pseudo-monotone
sequences in a DVR are the pseudo-stationary sequences.
2.2
Polynomial closure
Definition 2.5. Let S be a subset of K. The polynomial closure of S is the largest subset of K
over which the polynomials of Int(S, V ) are integer-valued, namely:
S = {s ∈ K | ∀f ∈ Int(S, V ), f (s) ∈ V }
Equivalently, the polynomial closure of S is the largest subset of K such that Int(S, V ) = Int(S, V ).
A subset S of K for which the equality S = S is called polynomially closed.
The main result of Chabert in [8] is the following theorem, which will be essential in §4 for the
proof of our main result.
Theorem 2.6. [8, Theorem 5.3] Let V be a valuation domain of rank one. Then the polynomial
closure is a topological closure, that is, there exists a topology on K for which the closed sets are
exactly the polynomially closed sets. A basis for the closed sets for this topology is given by the finite
unions of closed balls B(a, γ) = {x ∈ K | v(x − a) ≥ γ}, for a ∈ K and γ ∈ Γv .
Definition 2.7. The topology on the valued field K which has the polynomially closed subsets as
closed sets is called polynomial topology.
Example 2.8. Given α ∈ K and γ ∈ R, in [8, Proposition 3.2] it is proved that the polynomial
closure of the open ball B̊(α, γ) = {x ∈ K | v(x − α) > γ} is equal to:
B(α, γ), γ = inf{λ ∈ Γv | λ > γ}, if either v is discrete or γ ∈
/ Γv
B̊(α, γ) =
B(α, γ), otherwise
Remark 2.9. In particular, given α ∈ K, the subsets of the form:
{x ∈ K | v(x − si ) < γi , i = 1, . . . , r}
where si ∈ K and γi ∈ Γv are such that v(α − si ) < γi , for i = 1, . . . , r, form a fundamental system
of open neighborhoods of α for the polynomial topology.
3
Residually transcendental extensions
The following criterion, as appears for example in [25, Theorem 10, chapt. VI, §5] or [11, Theorem
19.15], in particular establishes when an integrally closed domain D is Prüfer: D must not admit
valuation overrings of a certain specified kind. Recall that a valuation overring V of an integral
domain D is a valuation domain V contained between D and its quotient field K. The center of a
valuation overring V of D is the intersection of the maximal ideal of V with D.
Theorem 3.1. Let D be an integrally closed domain and P a prime ideal of D. Then DP is a
valuation domain if and only if there is no valuation overring V of D centered in P such that the
residue field of V is transcendental over the quotient field of D/P .
6
By means of this Theorem, we are going to show that an integrally closed domain Int(S, V ),
S ⊆ V , is not Prüfer exactly when it admits a valuation overring which is an extension of V to
K(X) and whose residue field extension is transcendental.
Definition 3.2. A valuation domain W of the field of rational functions K(X) is a residually
transcendental extension of V = W ∩ K (or simply residually transcendental extension if V is
understood) if the residue field of W is a transcendental extension of the residue field of V .
The residually transcendental extensions of V to K(X) have been completely described by
Alexander and Popescu ([1]). In order to describe these valuation domains, we need to introduce
the following class of valuations on K(X).
Definition 3.3. Let α ∈ K and δ an element of a value group Γ which contains Γv . For f ∈ K[X]
such that f (X) = a0 + a1 (X − α) + . . . + an (X − α)n , we set:
vα,δ (f ) = inf{v(ai ) + iδ | i = 0, . . . , n}
The function vα,δ naturally extends to a valuation on K(X) ([5, Chapt. VI, §. 10, Lemme 1]). We
denote by Vα,δ the valuation domain associated to vα,δ , i.e.: Vα,δ = {ϕ ∈ K(X) | vα,δ (ϕ) ≥ 0}.
Clearly, Vα,δ lies over V . We let also Mα,δ = {ϕ ∈ K(X) | vα,δ (ϕ) > 0} be the maximal ideal of
Vα,δ .
Remark 3.4. Note that, if γ ∈ Γv , γ ≥ 0 and d ∈ V is any element such that v(d) = γ, then it is
easy to see that:
X −α
X −α
X −α
, Mα,γ ∩ K[X] = M
Vα,γ = V
, Vα,γ ∩ K[X] = V
d
d
d
M [ X−α ]
d
P
In general, if γ ∈ R, then Vα,γ ∩ K[X] = {f (X) = k≥0 ak (X − α)k ∈ K[X] | v(ak ) + kγ ≥ 0, ∀k}.
As in [9, §4], in this case we set V [(X − α)/γ] to be Vα,γ ∩ K[X].
In [2, p. 580] the authors say that vα,δ is residually transcendental if and only if δ has finite
order over Γv . For the sake of the reader we give a self-contained proof here.
Lemma 3.5. Let α ∈ K and δ an element of a value group Γ which contains Γv . Then vα,δ is
residually transcendental if and only if δ has finite order over Γv , i.e., there exists n ∈ N such that
nδ ∈ Γv .
Proof. Suppose there exists n ≥ 1 such that nδ = γ = v(c) ∈ Γv , for some c ∈ K. Clearly, the vα,δ n
is zero. We claim that over the residue field of V the polynomial
adic valuation of f (X) = (X−α)
c
f (X) is transcendental. In fact, suppose there exist ad−1 , . . . , a0 ∈ V such that
d
f + ad−1 f
d−1
+ . . . + a1 f + a0 = 0
that is
g = f d + ad−1 f d−1 + a1 f + a0 ∈ Mα,δ
However, if we set ad = 1, we have:
vα,δ (g) = inf{v(ai ) − inδ + niδ | i = 0, . . . , d} = inf{v(ai ) | i = 0, . . . , d} = 0
7
which is a contradiction.
Pd
∗
Conversely, suppose that nδ ∈
/ Γv , for each n ≥ 1. Let g ∈ Vα,δ
, say g(X) = i≥0 ai (X − α)i ;
then we have
vα,δ (g) = inf{v(ai ) + iδ | i = 0, . . . , d} = 0 ⇔ v(a0 ) = 0 & v(ai ) + iδ > 0, ∀i = 1, . . . , d
because on the assumption on δ. Then g(X) is congruent to g(0) = a0 modulo M so that over the
residue field g(X) is algebraic.
Remark 3.6. Suppose we are in the case where δ ∈ Γv (so that, in particular, vα,δ is residually
n
transcendental). If we consider the following expansion f (X) = b0 + b1 X−α
+ . . . + bn ( X−α
d
d ) ,
where d ∈ K is such that v(d) = δ, then vα,δ (f ) = inf{v(bi ) | i = 0, . . . , n}. In particular, by [5,
Chapt. VI, §10, Prop. 2], vα,δ is the unique valuation on K(X) = K( X−α
d ) for which the image of
X−α
X−α
in
the
residue
field
is
transcendental
over
V
/M
(note
that
has
valuation zero).
d
d
Let K be a fixed algebraic closure of K 1 and Γv = Γv ⊗Z Q. The following theorem characterizes
the residually transcendental extensions of V to K(X) (see also [14, Theorem 3.11] for an alternative
and more recent approach). The theorem holds for any valuation domain (i.e., no matter of its
dimension). For the sake of the reader we give a sketch of the proof.
Theorem 3.7. [1, Proposition 2 & Théorème 11] Let W be a residually transcendental extension
of V to K(X). Then there exist α ∈ K, γ ∈ Γv and a valuation W of K lying over V such that
W = W α,γ ∩ K(X).
Proof. Let W be an extension of W to K(X). It is clear that W is residually transcendental over
W ∩ K. Thus, without loss of generality we may assume that K is algebraically closed. Now, by
[1, Proposition 2], W is the valuation domain associated to a valuation w on K(X) which on a
polynomial f ∈ K[X] is defined as:
X
ai (aX − b)i
w(f ) = inf {v(ai )}, if f (X) =
i
i
for some a, b ∈ K, a 6= 0. Now, if we write f (X) = i bi (X − α)i where bi = ai ai and α = b/a, we
get that v(ai ) = v(bi ) − iv(a), so finally w = vα,γ , where γ = −v(a) (see also Remark 3.6).
P
Definition 3.8. Given α ∈ K, γ ∈ Γv and a valuation domain W of K lying over V , we denote
W
W
the valuation domain W α,γ ∩ K(X). If W is understood we denote Vα,γ
by Vα,γ .
by Vα,γ
W
Remark 3.9. Let (α, γ) ∈ K × Γv be fixed. The valuation domain Vα,γ
depends on the chosen
′
extension W of V to K. For example, let w, w be the two valuations of Q(i) which extend the
5-adic valuation v5 on Q: since 5 = (2 − i)(2 + i), we let w = w2−i be the (2 − i)-adic valuation on
Q(i) and w′ = w2+i be the (2 + i)-adic valuation on Q(i). Let f (X) = −X + 2 = −(X − i) + 2 − i.
Then
wi,1 (f (X)) = inf{v5 (−1) + 1, w2−i (2 − i)} = 1
′
wi,1
(f (X)) = inf{v5 (−1) + 1, w2+i (2 − i)} = 0
1 K is not to be confused with the polynomial closure of K, which is K itself. Since both symbols are now equally
customary, we decide to change neither of them.
8
′
In particular, f ∈ Q[X] is a unit in Wi,1
and is in the maximal ideal of Wi,1 . Therefore the
contractions of these valuation domains to Q(X) cannot be the same.
Therefore, whenever we write Vα,γ without any reference to an extension of V to K, we are
implicitly assuming that such an extension has been fixed in advance. Note that there is no
ambiguity in writing Vα,γ whenever (α, γ) ∈ K × Γv .
Note also that we may assume that there exists a finite field extension F of K and a valuation
domain W of F lying over V such that α is in F and γ is in Γw , the value group of W .
Lemma 3.10. Suppose that Γv is not discrete. Let α ∈ K, γ ∈ R and {γn }n∈N be a strictly
monotone sequence (either increasing or decreasing) of elements of Γv converging to γ. Then, for
each f ∈ K[X], we have
lim vα,γn (f ) = vα,γ (f )
n→∞
Proof. Let f ∈ K[X], f (X) = a0 +a1 (X −α)+. . .+ad (X −α)d . Suppose that vα,γ1 (f ) = v(ai )+iγ1 ,
for some i ∈ {0, . . . , d}, so that, in particular, v(ai )+iγ1 ≤ v(aj )+jγ1 , for all j ∈ {0, . . . , d}. Clearly,
v(ai ) + iγ2 ≤ v(aj ) + jγ2 , ∀j ∈ {0, . . . , d}
so that, in particular, vα,γ2 (f ) = v(ai ) + iγ2 . If we continue in this way, we have
vα,γn (f ) = v(ai ) + iγn ≤ v(aj ) + jγn , ∀n ∈ N
Now, since {γn }n∈N is strictly monotone converging to γ, we have:
lim vα,γn (f ) = v(ai ) + iγ ≤ v(aj ) + jγ, ∀j ∈ {0, . . . , d}
n→∞
The last inequality implies that vα,γ (f ) = v(ai )+iγ = limn→∞ vα,γn (f ), as we wanted to prove.
The next result shows that the family of rings Vα,γ ∩K[X], α ∈ K, γ ∈ R, has a natural ordering;
it generalizes [4, Proposition 1.1].
Proposition 3.11. Let γ1 , γ2 ∈ R and α1 , α2 ∈ K. Then the following conditions are equivalent:
i) Vα1 ,γ1 ∩ K[X] ⊆ Vα2 ,γ2 ∩ K[X]
ii) vα1 ,γ1 (f ) ≤ vα2 ,γ2 (f ), ∀f ∈ K[X]
iii) γ1 ≤ γ2 and v(α1 − α2 ) ≥ γ1
In particular, the following conditions are equivalent:
1) Vα1 ,γ1 ∩ K[X] = Vα2 ,γ2 ∩ K[X]
2) vα1 ,γ1 (f ) = vα2 ,γ2 (f ), ∀f ∈ K[X]
3) γ1 = γ2 and v(α1 − α2 ) ≥ γ1
4) Vα1 ,γ1 = Vα2 ,γ2
9
Proof. We prove first the proposition in the case γ1 ∈ Γv .
1
1
∈ Vα∗1 ,γ1 ∩ K[X] ⇒ X−α
∈ Vα2 ,γ2 .
i) ⇒ iii). Let d ∈ K be such that v(d) = γ1 . Then X−α
d
d
Thus,
X − α1
X − α2
α2 − α1
vα2 ,γ2
= vα2 ,γ2
= inf{γ2 − γ1 , v(α2 − α1 ) − γ1 } ≥ 0
(3.12)
+
d
d
d
so that γ1 ≤ γ2 and v(α2 − α1 ) ≥ γ1 .
1
iii) ⇒ ii). Let d ∈ K be such that v(d) = γ1 . Since v(α1 − α2 ) ≥ γ1 we have V [ X−α
d ] =
X−α2
α2 −α1
X−α2
X−α1
X−α2
V [ d + d ] = V [ d ] and similarly M [ d ] = M [ d ], so that vα1 ,γ1 = vα2 ,γ1 by
Remark 3.4. Now, since γ1 ≤ γ2 , for f (X) = a0 + a1 (X − α2 ) + . . . , ad (X − α2 )d ∈ K[X] we have:
vα1 ,γ1 (f ) = vα2 ,γ1 (f ) = inf {v(ai ) + iγ1 } ≤ inf {v(ai ) + iγ2 } = vα2 ,γ2 (f )
i
i
ii) ⇒ i). Let f ∈ Vα1 ,γ1 ∩ K[X]. Then vα2 ,γ2 (f ) ≥ vα1 ,γ1 (f ) ≥ 0 which implies that f ∈
Vα2 ,γ2 ∩ K[X].
We prove now the last equivalences. The equivalences of 1),2) and 3) follow at once by the
previous equivalences. Clearly, 4) implies 1) and 2) implies 4), since every rational function can be
written as the ratio of two polynomials and vα1 ,γ1 (f /g) = vα1 ,γ1 (f ) − vα1 ,γ1 (g).
We prove now the proposition in full generality, thus, γ1 ∈ R \ Γv .
i) ⇒ iii). Suppose that Γv is not discrete. Let c ∈ K be such that v(c) < γ1 . Then fc (X) =
X−α1
is in Vα1 ,γ1 because vα1 ,γ1 (fc ) = γ1 − v(c) > 0. Therefore, fc ∈ Vα2 ,γ2 by assumption. As
c
in (3.12) we get that vα2 ,γ2 (f ) = inf{γ2 − v(c), v(α1 − α2 ) − v(c)} ≥ 0, so that γ2 > v(c) and
v(α1 − α2 ) > v(c). Since v(c) is an arbitrary chosen element smaller than γ1 , we get γ2 ≥ γ1 and
v(α1 − α2 ) > γ1 . If Γv is discrete, then we choose {γn }n∈N in Γv ⊗Z Q and we consider everything
over the algebraic closure of K (fixing an extension of V to K), proceed as above and then contract
down to K.
iii) ⇒ ii). If Γv is not discrete, let {γn }n∈N be a strictly decreasing sequence in Γv converging to
γ1 . Since v(α1 −α2 ) > γ1 , there exists N ∈ N such that for all n ≥ N , we have v(α1 −α2 ) ≥ γn > γ1 ,
so that by the first part of the proposition, we have vα1 ,γn = vα2 ,γn . Hence, by Lemma 3.10 we have
vα1 ,γ1 = vα2 ,γ1 . If follows now immediately that for each f ∈ K[X] we have vα1 ,γ1 (f ) = vα2 ,γ1 (f ) ≤
vα2 ,γ2 (f ). We pass to Γv ⊗ Q if Γv is discrete.
ii) ⇒ i). This case is proved in the same way as in the case γ1 ∈ Γv .
The equivalences of 1),2),3) and 4) are proved as above.
The following lemma is well-known but we give a different proof based on the previous criterion
Theorem 3.1.
Lemma 3.13. Let (α, γ) ∈ K × Γv . Then Vα,γ ∩ K[X] is not a Prüfer domain.
Proof. By Remark 3.4, Vα,γ ∩ K[X] = V [ X−α
d ], where d ∈ K is such that v(d) = γ. We consider
X−α
)
of
V
[
]
(note
that this maximal ideal strictly contains the center
the maximal ideal (M, X−α
d
d
′
of Vα,γ in Vα,γ ∩ K[X], which is equal to M [ X−α
d ], see Remark 3.4). We claim that for each γ ∈ Γv
X−α
such that γ < γ ′ , (M, d ) is the center of Vα,γ ′ in Vα,γ ∩K[X], thus we have the following equality:
(M,
X −α
) = Mα,γ ′ ∩ (Vα,γ ∩ K[X])
d
10
is in
Since the left-hand side of the last equality is a maximal ideal, it is sufficient to show that X−α
d
X−α
′
′
′
=
c
·
,
where
c,
d
∈
K
are
such
that
v(d
)
=
γ
and
Mα,γ ′ . This follows from the fact that X−α
′
d
d
X−α
v(c) = γ ′ − γ > 0. Now, the residue field of V [ X−α
]
at
(M,
)
is
isomorphic
to
V
/M
.
Moreover,
d
d
X−α
Vα,γ ′ is a residually transcendental extension of V , so that the localization of V [ X−α
d ] at (M, d )
2
is not a valuation domain by Theorem 3.1 .
Lemma 3.14. Let W be a valuation domain of K(X) such that W ∩ K[X] is not Prüfer. Then
W ∩ K[X] is not Prüfer.
In particular, if (α, γ) ∈ K × Γv , then Vα,γ ∩ K[X] is not a Prüfer domain.
Proof. If R = W ∩ K[X] is Prüfer, then its integral closure R in K(X) is Prüfer (K(X) ⊆ K(X)
is an algebraic extension). But it is immediate to see that R ⊆ W ∩ K[X]: in fact, R ⊂ K[X] ⇒
R ⊂ K[X] and R ⊂ V = W ∩ K(X) ⇒ R ⊂ V ⊆ W . In particular, W ∩ K[X] would be Prüfer, a
contradiction.
For the last claim, we know that W α,γ ∩ K[X] is not a Prüfer domain. Then by the previous
claim, V α,γ ∩ K[X] = Vα,γ ∩ K[X] is not a Prüfer domain.
The following easy result shows that if Vα,γ , (α, γ) ∈ K × Γv , is a valuation overring of Int(S, V ),
then α ∈ V and γ ≥ 0.
Lemma 3.15. Let α ∈ K and γ ∈ Γv . We have V [X] ⊂ Vα,γ if and only if α ∈ V and γ ≥ 0.
Proof. The ’if’ direction is clear. For the ’only if’ part, recall that Vα,γ ∩ K[X] = V [ X−α
d ], where
X−α
X−α
d ∈ K is such that v(d) = γ. Then since X ∈ V [ d ] and we have X = d · d + α, which
necessarily implies that d, α ∈ V .
Theorem 3.16. Let R ⊆ K[X] be an integrally closed domain with quotient field K(X) and such
that D = R ∩ K is a Prüfer domain with quotient field K. Then R is Prüfer if and only if there is
no valuation overring V of D and (α, γ) ∈ K × Γv such that R ⊂ Vα,γ .
The theorem is false without the assumption R ⊆ K[X]. For example, R = Vα,γ , (α, γ) ∈ K ×Γv ,
is a Prüfer domain.
Proof. Note that R ⊂ Vα,γ ⇔ R ⊂ K[X] ∩ Vα,γ . By Lemma 3.14, K[X] ∩ Vα,γ is not Prüfer. Hence,
the ’only if’ part follows immediately, since an overring of a Prüfer domain is Prüfer.
Conversely, if R is not Prüfer, then by Theorem 3.1 there exists a valuation overring W of R
with maximal ideal MW such that R/(MW ∩ R) ⊂ W/MW is a transcendental extension. Note that
W ∩ K = V is a valuation overring of D, and since the latter ring is Prüfer, DP = V , where P is
the center of V in D. We claim that W is a residually transcendental extension of V , so that by
2 It
can be noted that this could have also been proved directly.
11
Theorem 3.7 we have W = Vα,γ , for some (α, γ) ∈ K × Γv . Indeed, we have the following diagram:
R/(MW
O
W/MW
O
♣7
♣
♣
♣
♣
♣♣♣
♣♣♣
∩ R)
V /MV
7
♣♣♣
♣
♣
♣♣♣
♣♣♣
D/P
By Theorem 3.1 D/P ⊂ V /MV is not a transcendental extension (because D is assumed to be
Prüfer), therefore the extension V /MV ⊂ W/MW is transcendental and thus W is a residually
transcendental extension of V . The proof is now complete by Theorem 3.7.
4
Pseudo-monotone sequences and polynomial closure
For the next lemma, see also [9, Lemma 4.1, Theorem 4.3, Proposition 4.5].
Lemma 4.1. Let α ∈ K and γ ∈ R. We have
V [(X − α)/γ] ⊆ Int(B(α, γ), V )
(4.2)
In particular, if S ⊆ V is such that Int(S, V ) ⊂ Vα,γ , then B(α, γ) ⊆ S. The other containment of
(4.2) holds if and only if either V is not discrete or V /M is infinite. In particular, if one of these
last conditions holds, then B(α, γ) ⊆ S implies Int(S, V ) ⊂ Vα,γ .
Proof. It is straighforward to verify the above containment. Moreover, we have
Int(S, V ) = Int(S, V ) ⊆ Vα,γ ∩ K[X] = V [(X − α)/γ]
(the last equality follows by Remark 3.4), so the second claim follows. The third claim follows by
[9, Proposition 4.5]. Finally, the last claim is straightforward.
We remark that the last statement is false if V is a DVR with finite residue field: in fact, in that
case Int(V ) is Prüfer by Theorem 1.1 so by Theorem 3.16 Int(V ) 6⊂ V0,0 (note that V = B(0, 0)).
The following important result by Chabert shows the connection between pseudo-monotone
sequences and polynomial closure.
Proposition 4.3. [8, Prop. 4.8] Let S ⊆ V be a subset. Let {sn }n∈N be a pseudo-monotone
sequence in S with breadth γ ∈ R and pseudo-limit α ∈ V . Then B(α, γ) ⊆ S, or, equivalently
Int(S, V ) ⊂ Vα,γ .
Note that by Remark 2.4 either V is not discrete or its residue field is infinite, so the last
equivalence of the above Proposition follows by Lemma 4.1.
The aim of this section is to show that Proposition 4.3 can be reversed, in the sense that if
B(α, γ) is the largest ball centered in α ∈ K which is contained in the polynomial closure of S,
then there exists a pseudo-monotone sequence E of S with pseudo-limit α and breadth γ.
12
Remark 4.4. Let E = {sn }n∈N ⊂ V be a pseudo-monotone sequence in V with breadth γ and
pseudo-limit α ∈ V . Suppose that γ = v(d) ∈ Γv , for some d ∈ V . By Proposition 4.3 and Remark
3.4, we have
X −α
Int(E, V ) ⊆ V
(4.5)
d
If E is either pseudo-monotone or pseudo-divergent, then v(sn − α) ≥ v(d), so the containment in
(4.5) is an equality. If E is pseudo-stationary, by §2.1.2 any element of E is a pseudo-limit so we
have the equality:
X − sn
, ∀n ∈ N
Int(E, V ) = V
d
If E is pseudo-convergent, then v(sn − α) < γ for all n ∈ N, so we only have the strict containment
in (4.5).
If E is pseudo-convergent or pseudo-divergent, then γ may not be in Γv and may also not be
torsion over Γv (in which case Vα,γ is not residually transcendental over V , by Lemma 3.5). However
if this condition holds, there exists γ ′ ∈ Γv , γ ′ > γ so that by Proposition 3.11, Int(E, V ) ⊆
Vα,γ ∩ K[X] ⊂ Vα,γ ′ ∩ K[X] ⊂ Vα,γ ′ , and Vα,γ ′ is residually transcendental over V . Therefore, by
Theorem 3.16, Int(E, V ) is not a Prüfer domain.
Let α ∈ K and γ ∈ Γv . For the next lemma, we set
∂B(α, γ) = {x ∈ K | v(x − α) = γ}
Lemma 4.6. Let α ∈ V and γ ∈ Γv . If either V is not discrete or V /M is infinite then
∂B(α, γ) = B(α, γ)
If V is a DVR with finite residue field, then ∂B(α, γ) is polynomially closed.
Proof. Let d ∈ V be such that v(d) = γ. Note that under the isomorphism X 7→ X−α
d , ∂B(α, γ) =
{x ∈ K | v(x − α) = γ} is isomorphic to V ∗ and similarly B(α, γ) is isomorphic to V . Therefore,
∂B(α, γ) = B(α, γ) if and only if V ∗ = V . We prove now the last equality under the current
hypothesis.
If V /M is infinite, then there exists a sequence E = {sn }n∈N ⊂ V and an element s ∈ V ∗ such
that v(sn − sm ) = v(sn − s) = 0, ∀n 6= m. Thus, E is pseudo-stationary with pseudo-limit s and
by Proposition 4.3 we may conclude.
S
/ M , ∀i = 1, . . . , n. Since the polynomial
If V /M is finite, let V ∗ = i=1,...,n (ai + M ), where ai ∈
S
closure in this context is a topological closure by Theorem 2.6, we have V ∗ = i=1,...,n (ai + M ) =
S
i=1,...,n (ai + M ) by [6, Proposition IV.1.5, p. 75]. The polynomial closure of M is equal either to
V if V is not discrete or to M itself if V is discrete (see Example 2.8). The proof is complete.
For a subset S of V such that Int(S, V ) is not Prüfer, we know by Theorem 3.16 that there
exist α ∈ K and γ ∈ Γv = Γv ⊗Z Q such that Int(S, V ) ⊂ Vα,γ . The next result shows that we can
reduce to the case of α ∈ V and γ ∈ Γv .
Proposition 4.7. Let S be a subset of an integrally closed domain D with quotient field K. Let F
be an algebraic extension of K and DF the integral closure of D in F . Then the integral closure of
Int(S, D) in F (X) is the ring Int(S, DF ).
13
Proof. It is well-known that Int(S, DF ) ⊂ F [X] is integrally closed (see [6, Proposition IV.4.1]), so
we have just to show that every element of Int(S, DF ) is integral over Int(S, D). Up to enlarging the
field F , we may assume that F is normal over K (e.g., the algebraic closure of K). We are going to
show that Int(S, D) ⊂ Int(S, DF ) is an integral ring extension under this further assumption. Let
then f ∈ Int(S, DF ), since f ∈ F [X], we know that f satisfies a monic equation over the polynomial
ring K[X]:
f n + gn−1 f n−1 + . . . + g1 f + g0 = 0, gi ∈ K[X], i = 0, . . . , n − 1.
We claim that gi ∈ Int(S, D), for i = 0, . . . , n − 1. Let Φ(T ) = T n + gn−1 T n−1 + . . . + g0 ∈ K[X][T ].
The roots of Φ(T ) are exactly the conjugates of f under the action of the Galois group Gal(F/K),
which acts on the coefficients of the polynomial f . If σ ∈ Gal(F/K), then σ(f ) ∈ F [X], and, more
precisely, σ(f ) ∈ Int(S, DF ). In fact, for each s ∈ S ⊂ K, since σ leaves each element of K invariant,
we have σ(f )(s) = σ(f (s)) which still is an element of DF (which likewise is left invariant under the
action of Gal(F/K)). Now, since each coefficient gi (X) of Φ(T ) lies in K[X] and is an elementary
symmetric function on the elements σ(f ), σ ∈ Gal(F/K), we have that gi (s) ∈ DF ∩ K = D, for
each s ∈ S, thus gi ∈ Int(S, D), as claimed.
Proposition 4.8. Let (α, γ) ∈ F × Γw , where F is a finite field extension of K and W is a
valuation domain of F lying over V . If S is a subset of V such that Int(S, V ) ⊂ Wα,γ ∩ K(X), then
Int(S, W ) ⊂ Wα,γ .
Note that the polynomials in Int(S, V ) have coefficients in K, the quotient field of V , while the
polynomials in Int(S, W ) have coefficients in F , the quotient field of W .
Proof. Let S be a subset of V such that Int(S, V ) ⊂ Wα,γ ∩ K(X) = Vα,γ . The integral closure VF
of V in F is equal to an intersection of finitely many rank 1 valuation domains W = W1 , . . . , Wn
of F lying over V . In particular,
\
Int(S, Wi )
Int(S, VF ) =
i=1,...,n
Let MW be the maximal ideal of W . If T = VF \(MW ∩VF ), then T −1 VF = W and since localization
commutes with finite intersections we have:
\
T −1 Int(S, Wi )
T −1 Int(S, VF ) =
i=1,...,n
Let f ∈ K[X], say f (X) = g(X)
d , for some g ∈ VF [X] and d ∈ VF . Let γi = wi (d), for each i ≥ 2.
By the approximation theorem for independent valuations ([5, Corollaire 1, Chapt. VI, §7]), there
exists t ∈ VF such that w(t) = 0 and wi (t) = γi . In particular, t ∈ T . Then t · f ∈ Int(S, Wi ), so
that T −1 Int(S, Wi ) = K[X] for all i ≥ 2. Clearly, T −1 Int(S, W ) = Int(S, W ). Hence,
T −1 Int(S, VF ) = Int(S, W )
(4.9)
Since Int(S, VF ) is the integral closure of Int(S, V ) in F (X) by Proposition 4.7 and Vα,γ = Wα,γ ∩
K(X) is an overring of Int(S, V ), it follows that Wα,γ is an overring of Int(S, VF ).
Let now f ∈ Int(S, W ). By (4.9) there exists d ∈ T such that d · f ∈ Int(S, VF ) ⊂ Wα,γ . Since
d ∈ W ∗ is also a unit in Wα,γ , it follows that f ∈ Wα,γ . This shows that Int(S, W ) ⊂ Wα,γ . Since
S ⊆ V ⊂ W , by Lemma 3.15, α ∈ W and γ ≥ 0.
14
For a subset S of V , α ∈ K and γ ∈ R, we set
Sα,<γ ={s ∈ S | v(s − α) < γ}
Sα,>γ ={s ∈ S | v(s − α) > γ}
Sα,γ ={s ∈ S | v(s − α) = γ}
Note that if Sα,γ is not empty then γ ∈ Γv .
Proposition 4.10. Let S ⊆ V , α ∈ V and γ ∈ Γv . Then B(α, γ) ⊆ Sα,γ if and only if there
exists a pseudo-monotone sequence E = {sn }n∈N in Sα,γ with breadth γ such that E is either
pseudo-stationary and with pseudo-limit α or pseudo-divergent and with a pseudo-limit in Sα,γ .
In particular, if V is a DVR, then E can only be pseudo-stationary.
Proof. The ’if’ part follows from Proposition 4.3.
Conversely, suppose that B(α, γ) ⊆ Sα,γ . Note that this assumption necessarily implies that
either V is non-discrete or V /M is infinite. In fact, Sα,γ ⊆ ∂B(α, γ) and ∂B(α, γ) is polynomially
closed if V is a DVR with finite residue field (Lemma 4.6), so in this case Sα,γ could not contain
B(α, γ).
Suppose that there is no pseudo-divergent sequence E = {sn }n∈N in Sα,γ with breadth γ and
with pseudo-limit s ∈ Sα,γ . This is equivalent to the following: for each s ∈ Sα,γ , let γs = inf{v(s −
s′ ) | s′ ∈ Sα,γ , v(s′ −s) > γ}. Then γs > γ, for each s ∈ Sα,γ (note that, a priori, for s 6= s′ ∈ Sα,γ we
have v(s−s′ ) ≥ γ). We construct now a pseudo-stationary sequence in Sα,γ with pseudo-limit α and
breadth γ. Let s1 ∈ Sα,γ . Then α belongs to the open set U1 = {x ∈ K | v(x − s1 ) < γs1 } and since
α ∈ Sα,γ by assumption, there exists s2 ∈ Sα,γ ∩ {x ∈ K | v(x − s1 ) < γs1 }. By definition of γs1 , we
must have v(s1 − s2 ) = γ. Now, we consider the open set U2 = {x ∈ K | v(x − si ) < γsi , i = 1, 2}.
Since α ∈ U2 there exists s3 ∈ U2 ∩ Sα,γ , so that v(s3 − si ) < γsi , for i = 1, 2, which implies
that v(s3 − si ) = γ, for i = 1, 2. If we continue in this way, we get a pseudo-stationary sequence
E = {sn }n∈N ⊆ Sα,γ with pseudo-limit α and breadth γ, as wanted.
The last claim follows immediately from §2.1.3.
In the next two results, we consider an integral domain R ⊂ K[X] with quotient field K(X)
such that R ⊂ Vα,γ ′ , for some (α, γ ′ ) ∈ V × Γv (in particular, R is not Prüfer by Theorem 3.16).
If we keep α fixed, we know by Proposition 3.11 that the set of rings {K[X] ∩ Vα,γ ′ | γ ′ ∈ Γv }, is
linearly ordered. It makes then sense to consider the infimum in R of the set {γ ′ ∈ Γv | R ⊂ Vα,γ ′ }.
Lemma 4.11. Let R ⊂ K[X] be an integral domain with quotient field K(X). Suppose R ⊂ Vα,γ ′
for some α ∈ V and γ ′ ∈ Γv and let γ = inf{γ ′ ∈ Γv | R ⊂ Vα,γ ′ } ∈ R. Then R ⊂ Vα,γ .
In particular, γ is a minimum if and only if γ ∈ Γv . Note that, if V is nondiscrete, it may well
be that γ ∈ R \ Γv .
Proof. Let f ∈ R, with f (X) = a0 + a1 (X − α) + . . . + ad (X − α)d . Then
f ∈ Vα,γ ′ ⇔ inf{v(ai ) + iγ ′ | i = 0, . . . , d} ≥ 0 ⇔ a0 ∈ V, γ ′ ≥ −
v(ai )
, i = 1, . . . , d
i
Since γ is the infimum of the γ ′ with the above property in particular we have
a0 ∈ V, γ ≥ −
v(ai )
, i = 1, . . . , d
i
15
that is, vα,γ (f ) = inf{v(ai ) + iγ | i = 0, . . . , d} ≥ 0 ⇔ f ∈ Vα,γ .
By Lemma 4.1, the next theorem shows that if B(α, γ) is the largest ball centered in α contained
in the polynomial closure S of S, then there exists a pseudo-monotone sequence in S with breadth
γ and pseudo-limit in V , which is equal either to α or to α + t, where t ∈ V has valuation γ. This
result is a converse to Proposition 4.3.
Theorem 4.12. Let S ⊆ V be a subset such that Int(S, V ) ⊂ Vα,γ ′ , for some α ∈ V and γ ′ ∈ Γv .
Let γ = inf{γ ′ ∈ Γv | Int(S, V ) ⊂ Vα,γ ′ } ∈ R. Then there exists a pseudo-monotone sequence
E = {sn }n∈N ⊆ S with breadth γ such that one of the following conditions holds:
1) E ⊆ Sα,<γ is pseudo-convergent with pseudo-limit α.
2) E ⊆ Sα,>γ is pseudo-divergent with pseudo-limit α.
3) E ⊆ Sα,γ is pseudo-divergent with a pseudo-limit s ∈ Sα,γ .
4) E ⊆ Sα,γ is pseudo-stationary with pseudo-limit α.
Moreover, condition 1) holds if and only if sup{v(s − α) | s ∈ Sα,<γ } = γ, condition 2) holds if and
only if inf{v(s − α) | s ∈ Sα,>γ } = γ and conditions 3) or 4) hold if and only if B(α, γ) ⊆ Sα,γ . In
these last two cases, γ is a minimum⇔ γ ∈ Γv .
In particular, if V is discrete, γ is a minimum and only case 4) holds.
Remark 4.13. Note that by Lemma 4.11 we have Int(S, V ) ⊂ Vα,γ either in the case γ ∈ Γv ⇔ γ
is a minimum or γ ∈ R \ Γv . Note also that in case 3), where s ∈ Sα,γ is a pseudo-limit of a
pseudo-divergent sequence E = {sn }n∈N ⊆ Sα,γ , we have Vα,γ = Vs,γ , by Proposition 3.11.
Proof. Since by Theorem 3.16 Int(S, V ) is not a Prüfer domain, by Theorem 1.1 either V is not
discrete or its residue field is infinite, otherwise Int(V ) would be Prüfer and in particular its overring
Int(S, V ) would be Prüfer, too.
We consider the following real numbers:
γ1 = sup{v(s − α) | s ∈ Sα,<γ }
γ2 = inf{v(s − α) | s ∈ Sα,>γ }
Clearly, we have γ1 ≤ γ ≤ γ2 . Since Sα,>γ ⊆ B(α, γ2 ) and every closed ball is polynomially closed
(Theorem 2.6), we have:
Sα,>γ ⊆ B(α, γ2 )
(4.14)
If γ1 = γ then there exists a pseudo-convergent sequence {sn }n∈N ⊆ Sα,<γ with pseudo-limit α and
breadth γ, that is:
v(sn − α) < v(sn+1 − α) ր γ
Similrly, if γ2 = γ then there exists a pseudo-divergent sequence {sn }n∈N ⊆ Sα,>γ with α as
pseudo-limit and breadth γ:
v(sn − α) > v(sn+1 − α) ց γ
Hence, if either γ1 = γ or γ2 = γ we are done.
Suppose from now on that
γ1 < γ < γ2
16
We are going to show that under these conditions γ is a minimum (or, equivalently, γ ∈ Γv ), by
means of the fact that Sα,γ is non-empty. We claim first that {v(s − α) | s ∈ Sα,<γ )} is finite; in
fact, if that were not true, there would exist a sequence {sn }n∈N ⊆ Sα,<γ either pseudo-convergent
or pseudo-divergent with breadth γ ′ < γ, so that Int(S, V ) ⊂ Vα,γ ′ by Proposition 4.3, contrary to
the assumption on γ. Therefore γ1 is a maximum and:
{v(s − α) | s ∈ Sα,<γ } = {γ1 , . . . , γr }, γr < . . . < γ1 < γ
For each i = 1, . . . , r we set:
Sα,γi = {s ∈ Sα,<γ | v(s − α) = γi }
so that
[
Sα,<γ =
Sα,γi
i=1,...,r
For each i = 1, . . . , r, there is no pseudo-stationary sequence {sn }n∈N ⊂ Sα,γi with breadth γi and
pseudo-limit α, otherwise by Proposition 4.3 Int(S, V ) ⊂ Vα,γi , in contradiction with the definition
of γ. Hence, for each i ∈ {1, . . . , r}, there exist finitely many si,j ∈ Sα,γi , j ∈ Ii , such that the
following holds:
∀s ∈ Sα,γi , ∃j ∈ Ii such that v(s − si,j ) > γi
For each j ∈ Ii , we set Sα,γi ,j = {s ∈ Sα,γi | v(s − si,j ) > γi } and γi,j = inf{v(s − si,j ) | s ∈ Sα,γi ,j }.
If γi,j = γi then there exists a pseudo-divergent sequence with pseudo-limit si,j and breadth γi so
that by Proposition 4.3 we would have B(si,j , γi ) = B(α, γi ) ⊆ S (recall that v(α − si,j ) = γi ),
which is equivalent to Int(S, V ) ⊂ Vα,γi by Lemma 4.1, contrary to the assumption on γ. Thus,
γi,j > γi for all j ∈ Ii (and for all i ∈ 1, . . . , r). Finally, we have showed that
[
Sα,γi ⊆
B(si,j , γi,j )
j∈Ii
so that
Sα,<γ =
[
i=1,...,r
Sα,γi ⊆
[
[
B(si,j , γi,j )
(4.15)
i=1,...,r j∈Ii
Let γ ′ ∈ Γv be such that γ ≤ γ ′ < γ2 and Int(S, V ) ⊂ Vα,γ ′ ; note that if γ is not a min, then in
particular Γv is non-discrete and we may choose γ ′ ∈ Γv such that γ < γ ′ < γ2 ; if γ is a min, then
we choose γ ′ = γ. Since the polynomial closure is a topological closure (Theorem 2.6), by Lemma
4.1 we have
B(α, γ ′ ) ⊆ S = Sα,<γ ∪ Sα,γ ∪ Sα,>γ
We claim that ∂B(α, γ ′ ) ⊆ Sα,γ . In fact, let β ∈ V be such that v(β − α) = γ ′ . If β ∈ Sα,<γ
then by (4.15) we have v(β − si,j ) ≥ γi,j for some i ∈ {1, . . . , r} and j ∈ Ii , which implies that
v(α − si,j ) = v(α − β + β − si,j ) > γi , a contradiction. Similarly, β is not in Sα,>γ , by(4.14) and
the fact that γ ′ < γ2 . Therefore ∂B(α, γ ′ ) ⊆ Sα,γ , as claimed. In particular, Sα,γ 6= ∅, γ ∈ Γv and
so γ = min{γ ′ ∈ Γv | Int(S, V ) ⊂ Vα,γ ′ }. We may therefore assume that γ ′ = γ. By Lemma 4.6
(recall that either V is not discrete or its residue field is infinite) we have B(α, γ) ⊆ Sα,γ , so that,
by Proposition 4.10, there exists a pseudo-monotone sequence E = {sn }n∈N ⊆ Sα,γ with breadth
γ such that either E is pseudo-stationary and has pseudo-limit α or E is pseudo-divergent and has
pseudo-limit in Sα,γ . The proof is now complete.
17
As we have already said, if Int(S, V ) is not Prüfer then there might be residually transcendental
valuation overrings which are not of the form Vα,γ with (α, γ) ∈ K × Γv . For example, if E =
{sn }n∈N ⊂ V is a pseudo-convergent sequence of algebraic type without pseudo-limits in K, then
Int(E, V ) is not Prüfer by Theorem 1.3; by Theorem 4.12 is not difficult to show that Int(S, V ) 6⊂
Vα,γ for every (α, γ) ∈ K × Γv . However, we may reduce our discussion to this case by means of
Proposition 4.8, as the following proposition shows.
Proposition 4.16. Let S ⊆ V be such that Int(S, V ) ⊂ Vα,γ ′ = Wα,γ ′ ∩ K(X) for some (α, γ ′ ) ∈
F × Γw , where F is a finite extension of K and W is a valuation domain of F lying over V . Let
γ = inf{γ ′ ∈ Γw | Int(S, V ) ⊂ Vα,γ ′ = Wα,γ ′ ∩ K(X)}. Then there exists a pseudo-monotone
sequence E = {sn }n∈N ⊆ S with breadth γ and pseudo-limit which is equal either to α or belongs to
{x ∈ W | w(x − α) = γ}.
If V is a DVR, then E is pseudo-stationary, the breadth γ is in Γv and there exists β ∈ V which
is a pseudo-limit of E, so that, in particular, Vα,γ = Vβ,γ .
Proof. Suppose that Int(S, V ) ⊂ Vα,γ ′ . By Proposition 4.8, Int(S, W ) ⊂ Wα,γ ′ . Note that, by
Lemma 3.15, α ∈ W and γ ′ ≥ 0. By Theorem 4.12, there exists a pseudo-monotone sequence
{sn }n∈N ⊆ S whose breadth is γ and has pseudo-limit which is either α or belongs to {x ∈ W |
w(x − α) = γ}.
In the case V is discrete, by Theorem 4.12 E is necessarily pseudo-stationary with breadth γ
and α is a pseudo-limit of E. Since the sn ’s are elements of K, w(sn − sm ) = v(sn − sm ), so that
γ ∈ Γv . Moreover, by §2.1.2, any element of E = {sn }n∈N can be considered as a pseudo-limit of
E, so the last equality follows from Proposition 3.11.
Finally, the next theorem characterizes the subsets S of V for which Int(S, V ) is Prüfer. Note
that, for a pseudo-monotone sequence E = {sn }n∈N contained in K and an element α ∈ K, we say
that α is a pseudo-limit of E if there exists a valuation w of K(α) which lies above v such that α
is a pseudo-limit of E with respect to w (clearly, E is a pseudo-monotone sequence with respect to
w). With this terminology, the next theorem shows that Int(S, V ) is Prüfer if and only if S does
not admit any pseudo-limit in K.
Theorem 4.17. Let S ⊆ V . Then Int(S, V ) is a Prüfer domain if and only if S does not contain
a pseudo-monotone sequence E = {sn }n∈N which has a pseudo-limit α ∈ K.
Proof. Suppose there exists a pseudo-monotone sequence E = {sn }n∈N ⊆ S with breadth γ ∈ R
and pseudo-limit α ∈ K. If E is pseudo-stationary, then we know that γ ∈ Γv and any element s of
E is a pseudo-limit of E (§2.1.2). Then by Proposition 4.3, Int(S, V ) ⊂ Vs,γ = Vα,γ , so by Theorem
3.16 Int(S, V ) is not Prüfer. Suppose now that E is either pseudo-convergent or pseudo-divergent,
and let F be a finite extension of K which contains α. Let W be a valuation domain of F lying over
V (which is necessarily of rank one) for which α is a pseudo-limit of E (which clearly is a pseudomonotone sequence with respect to the associated valuation w). Clearly α ∈ W . By Proposition
4.3, it follows that Int(S, W ) ⊂ Wα,γ and contracting down to K[X] we get Int(S, V ) ⊂ Vα,γ , so by
Theorem 3.16 and Remark 4.4 Int(S, V ) is not Prüfer.
Conversely, suppose that Int(S, V ) is not Prüfer. By Theorem 3.16 there exists (α, γ ′ ) ∈ K × Γv
such that Int(S, V ) ⊂ Vα,γ ′ . As we have already done (see Remark 3.9), let F be a finite extension of
K and W a valuation domain of F lying over V such that (α, γ ′ ) ∈ F ×Γw and Vα,γ ′ = Wα,γ ′ ∩K(X).
Let γ = inf{γ ′ ∈ Γw | Int(S, V ) ⊂ Vα,γ ′ = Wα,γ ′ ∩ K(X)} ∈ R. By Proposition 4.16, there exists
a pseudo-monotone sequence E = {sn }n∈N ⊆ S with breadth γ and pseudo-limit which is equal
either to α or to α + t, where t ∈ W is such that w(t) = γ.
18
We summarize here some known results about Int(S, V ) being a Prüfer domain along with
some other new, when V is a DVR. As already remarked by Loper and Werner in [17], if V is a
non-discrete rank one valuation domain, there are subsets S of V which are not precompact but
Int(S, V ) is Prüfer (see the introduction).
Corollary 4.18. Let V be a DVR and S ⊆ V . Then the following conditions are equivalent:
i) Int(S, V ) is Prüfer.
ii) there is no pseudo-stationary sequence contained in S.
iii) S is precompact.
iv) there is no (α, γ) ∈ V × Γv such that Int(S, V ) ⊂ Vα,γ .
Proof. It is easy to see that ii) and iii) are equivalent if V is discrete. In fact, S is precompact if
and only if S modulo M n is finite for each n ≥ 1 ([7, Proposition 1.2]). Now, if the latter condition
holds, then there cannot be any pseudo-stationary sequence in V by §2.1.2. Conversely, if S modulo
M n is infinite for some n ≥ 1, then there exists {sm }m∈N ⊆ S such that v(sm − sk ) < n for each
m 6= k. Since the values {v(sm − sk ) | m 6= k} are finite, we can extract a subsequence from
{sm }m∈N which is pseudo-stationary, as claimed.
The fact that i) and iv) are equivalent follows from Proposition 4.16 and Theorem 4.12.
We show then that i) and ii) are equivalent. By Proposition 4.3, if there is a pseudo-stationary
sequence with breadth γ contained in S, then Int(S, V ) ⊂ Vα,γ , so by Lemma 3.13 Int(S, V ) is not
Prüfer. Suppose now that Int(S, V ) is not Prüfer: by Theorem 3.16 there exists (α, γ) ∈ F × Γw ,
where F is a finite extension of K and W a valuation domain of F lying over V with α ∈ W and
γ ∈ Γw such that Int(S, V ) ⊂ Vα,γ . Suppose also that γ ∈ Γw is minimal with this property. Then
by Proposition 4.16 it would follow that γ ∈ Γv and there exists a pseudo-stationary sequence in S
with breadth γ, a contradiction.
Actually, iii) implies ii) also when V is not discrete. More generally, when S is a precompact
subset of V , then there is no pseudo-monotone sequence containe in S, so that by Theorem 4.17
we get again the result of Theorem 1.2.
Acknowledgements
This research has been supported by the grant ”Ing. G. Schirillo” of the Istituto Nazionale di Alta
Matematica and also by the University of Padova.
The author wishes to thank also Jean-Luc Chabert for pointing out some inaccuracies.
References
[1] V. Alexandru, N. Popescu, Sur une classe de prolongements à K(X) d’une valuation sur un
corps K, Rev. Roumaine Math. Pures Appl. 33 (1988), no. 5, 393-400.
[2] V. Alexandru, N. Popescu, Al. Zaharescu, A theorem of characterization of residual transcendental extensions of a valuation. J. Math. Kyoto Univ. 28 (1988), no. 4, 579-592.
19
[3] V. Alexandru, N. Popescu, Al. Zaharescu, Minimal pairs of definition of a residual transcendental extension of a valuation, J. Math. Kyoto Univ. 30 (1990), no. 2, 207-225.
[4] V. Alexandru, N. Popescu, Al. Zaharescu, All valuations of K(X), J. Math. Kyoto Univ. 30
(1990), no. 2, 281296.
[5] N. Bourbaki, Algèbre commutative, Hermann, Paris, 1961.
[6] J.-P. Cahen, J.-L. Chabert, Integer-valued polynomials, Amer. Math. Soc. Surveys and Monographs, 48, Providence, 1997.
[7] J.-P. Cahen, J.-L. Chabert, K. A. Loper, High dimension Prüfer domains of integer-valued
polynomials. Mathematics in the new millennium (Seoul, 2000). J. Korean Math. Soc. 38 (2001),
no. 5, 915-935.
[8] J.-L. Chabert, On the polynomial closure in a valued field, J. Number Theory 130 (2010), 458468.
[9] J.-L. Chabert, Integer-valued polynomial in valued fields with an application to discrete dynamical systems, Commutative algebra and its applications, 103134, Walter de Gruyter, Berlin,
2009.
[10] M. Fontana, J. A. Huckaba, I. J. Papick, Prüfer domains. Monographs and Textbooks in Pure
and Applied Mathematics, 203. Marcel Dekker, Inc., New York, 1997.
[11] R. Gilmer, Multiplicative ideal theory. Queen’s Papers in Pure and Applied Mathematics, 90.
Queen’s University, Kingston, ON, 1992.
[12] R. Gilmer, Two constructions of Prüfer domains. J. Reine Angew. Math. 239/240 (1969),
153-162.
[13] I Kaplansky, Maximal Fields with Valuation, Duke Math. J. 9, 3030-321 (1942).
[14] F.-V. Kuhlmann, Value groups, residue fields, and bad places of rational function fields. Trans.
Amer. Math. Soc. 356 (2004), no. 11, 4559-4600.
[15] K. A. Loper, A classification of all D such that Int(D) is a Prüfer domain. Proc. Amer. Math.
Soc. 126 (1998), no. 3, 657-660.
[16] K. A. Loper, M. Syvuk, Prüfer Domains of Integer-Valued Polynomials, in “Multiplicative
Ideal Theory and Factorization Theory - Commutative and Non-Commutative Perspectives”,
Springer Proceedings in Mathematics and Statistics Volume 170, Springer (2016), p. 219-231.
[17] K. A. Loper, N. Werner, Pseudo-convergent sequences and Prüfer domains of integer-valued
polynomials, J. Commut. Algebra 8 (2016), no. 3, 411-429.
[18] D. L. McQuillan, Rings of integer-valued polynomials determined by finite sets, Proc. Roy. Irish
Acad. Sect. A 85 (1985), no. 2, 177-184.
[19] B. Olberding, On the geometry of Prüfer intersections of valuation rings. Pacific J. Math. 273
(2015), no. 2, 353-368.
20
[20] B. Olberding, A principal ideal theorem for compact sets of rank one valuation rings. J. Algebra
489 (2017), 399-426.
[21] A. Ostrowski, Untersuchungen zur arithmetischen Theorie der Körper, Math. Z. 39 (1935),
269-404.
[22] M. H. Park, Prüfer domains of integer-valued polynomials on a subset, J. Pure Appl. Algebra
219 (2015), no. 7, 2713-2723.
[23] G. Peruginelli, Transcendental extensions of a valuation domain of rank one, Proc. Amer.
Math. Soc. 145 (2017), no. 10, 4211-4226.
[24] P. Roquette, Principal ideal theorems for holomorphy rings in fields. J. Reine Angew. Math.
262/263 (1973), 361-374.
[25] O. Zariski, P. Samuel, Commutative Algebra, vol. II, Springer-Verlag, New York-HeidelbergBerlin, 1975.
21
| 0math.AC
|
High Performance Risk Aggregation:
Addressing the Data Processing Challenge the Hadoop MapReduce Way
∗
Z. Yao, B. Varghese and A. Rau-Chaplin
Faculty of Computer Science, Dalhousie University, Halifax, Canada
arXiv:1311.5686v1 [cs.DC] 22 Nov 2013
{yao, varghese, arc}@cs.dal.ca
ABSTRACT
Monte Carlo simulations employed for the analysis of portfolios of catastrophic risk process large volumes of data. Often
times these simulations are not performed in real-time scenarios as they are slow and consume large data. Such simulations can benefit from a framework that exploits parallelism for addressing the computational challenge and facilitates a distributed file system for addressing the data challenge. To this end, the Apache Hadoop framework is chosen
for the simulation reported in this paper so that the computational challenge can be tackled using the MapReduce
model and the data challenge can be addressed using the
Hadoop Distributed File System. A parallel algorithm for
the analysis of aggregate risk is proposed and implemented
using the MapReduce model in this paper. An evaluation of
the performance of the algorithm indicates that the Hadoop
MapReduce model offers a framework for processing large
data in aggregate risk analysis. A simulation of aggregate
risk employing 100,000 trials with 1000 catastrophic events
per trial on a typical exposure set and contract structure is
performed on multiple worker nodes in less than 6 minutes.
The result indicates the scope and feasibility of MapReduce
for tackling the computational and data challenge in the
analysis of aggregate risk for real-time use.
Keywords
hadoop mapreduce; risk aggregation; risk analysis; data processing; high-performance analytics
1.
INTRODUCTION
In the domain of large-scale computational analysis of
risk, large amounts of data need to be rapidly processed
and millions of simulations need to be quickly performed
(for example, [1, 2, 3]). This can be achieved only if data is
∗Corresponding author. E-mail: [email protected]. Webpage: http://www.blessonv.com
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
efficiently managed and parallelism is exploited within algorithms employed in the simulations. The domain, therefore,
inherently opens avenues to exploit the synergy that can
be achieved by bringing together state-of-the-art techniques
in data processing and management and high-performance
computing. Research on aggregate analysis of risks [5, 6, 7]
using high-performance computing is sparse at best. The
research reported in this paper is motivated towards exploring techniques for employing high-performance computing
not only to speed up the simulation but also to process and
manage data efficiently for the aggregate analysis of risk.
In this context the MapReduce model [4, 8, 9] is used for
achieving high-performance aggregate analysis of risks.
The aggregate analysis of risk is a Monte Carlo simulation performed on a portfolio of risks that an insurer or
reinsurer holds. A portfolio can cover risks related to catastrophic events such as earthquakes, floods or hurricanes, and
may comprise tens of thousands of contracts. The contracts
generally follow an ‘eXcess of Loss’ (XL) [10, 11] structure
providing coverage for single event occurrences or multiple
event occurrences, or a combination of both single and multiple event occurrences. Each trial in the aggregate analysis simulation represents a view of the occurrence of catastrophic events and the order in which they occur within a
contractual year. The trial also provides information on how
the occurrence of an event in a contractual year will interact
with complex treaty terms to produce an aggregated loss. A
pre-simulated Year Event Table (YET) containing between
several thousands and millions of alternative views of a single contractual year is input for the aggregate analysis. The
output of aggregate analysis is a Year Loss Table (YLT).
From a YLT, an insurer or a reinsurer can derive important portfolio risk metrics such as the Probable Maximum
Loss (PML) [12, 13] and the Tail Value-at-Risk (TVaR) [14,
15] which are used for both internal risk management and
reporting to regulators and rating agencies.
In this paper, the analysis of portfolios of catastrophic risk
is proposed and implemented using a MapReduce model on
the Hadoop [16, 17, 18] platform. The algorithm rapidly
consumes large amounts of data in the form of the YET
and Event Loss Tables (ELT). Therefore, the challenges of
organising input data and processing it efficiently, and applying parallelism within the algorithm are considered. The
MapReduce model lends itself well towards solving embarrassingly parallel problems such as the aggregate analysis
of risk, and is hence chosen to implement the algorithm.
The algorithm employs two MapReduce rounds to perform
both the numerical computations as well as to manage and
process data efficiently. The algorithm is implemented on
the Apache Hadoop platform. The Hadoop Distributed File
System (HDFS) and the Distributed Cache (DC) are key
components offered by the Hadoop platform in addressing
the data challenges. The preliminary results obtained from
the experiments of the analysis indicate that the MapReduce model can be used to scale the analysis over multiple
nodes of a cluster; parallelism can be exploited in the analysis for achieving faster numerical computations and data
management.
The remainder of this paper is organised as follows. Section 2 considers the sequential and MapReduce algorithm
for the analysis of aggregate risk. Section 3 presents the implementation of the MapReduce algorithm on the Apache
Hadoop Platform and the preliminary results obtained from
the experimental studies. Section 4 concludes this paper by
considering future work.
Algorithm 1: Sequential Aggregate Risk Analysis
Input : Y ET , ELT pool, P F
Output: Y LT
1
2
3
4
5
6
7
8
9
10
11
12
13
2.
ANALYSIS OF AGGREGATE RISK
The sequential and MapReduce algorithm of the analysis
of aggregate risk is presented in this section. There are three
inputs to the algorithm for the analysis of aggregate risk,
namely the Y ET , P F , and a pool of ELT s. The Y ET
is the Year Event Table which is the representation of a
pre-simulated occurrence of Events E in the form of trials
T . Each Trial captures the sequence of the occurrences of
Events for a year using Time-stamps in the form of Event
Time-stamp pairs. The P F is a portfolio that represents
a group of Programs, P , which in turn represents a set of
Layers, L that covers a set of ELT s using financial terms.
The ELT is the Event Loss Table which represents the losses
that correspond to an event based on an exposure (one event
can appear over different ELTs with different losses).
The intermediary output of the algorithm are the Layer
Loss Table LLT consisting Trial-Loss pairs. The final output
of the algorithm is Y LT , which is the Year Loss Table that
contains the losses covered by a portfolio.
2.1
Sequential Algorithm
Algorithm 1 shows the sequential analysis of aggregate
risk. The algorithm scans through the hierarchy of the portfolio, P F ; firstly through the Programs, P , followed by the
Layers, L, then the Event Loss Tables, ELT s. Line nos. 5-8
shows how the loss associated with an Event in the ELT is
computed. For this, the loss, lE associated with an Event,
E is retrieved, after which contractual financial terms to the
benefit of the Layer are applied to the losses and are summed
0
up as lE
.
In line nos. 9 and 10, two Occurrence Financial Terms,
namely the Occurrence Retention and the Occurrence Limit
0
are applied to the loss, lE
and summed up as lT . The lT
losses correspond to the total loss in one trial. Occurrence
Retention refers to the retention or deductible of the insured
for an individual occurrence loss, where as Occurrence Limit
refers to the limit or coverage the insurer will pay for occurrence losses in excess of the retention. The Occurrence
Financial Terms capture specific contractual properties of
’eXcess of Loss’ treaties as they apply to individual event
occurrences only.
In line nos. 12 and 13, two Aggregate Financial Terms,
namely the Aggregate Retention and the Aggregate Limit
are applied to the loss, lT to produce aggregated loss for
a Trial. Aggregate Retention refers to the retention or de-
14
15
16
17
18
19
20
for each Program, P do
for each Layer, L, in P do
for each Trial, T , in Y ET do
for each Event, E, in T do
for each ELT covered by L do
Lookup E in the ELT and find
corresponding loss, lE
0
0
lE
← lE
+ lE
end
0
Apply Occurrence Financial Terms to lE
0
lT ← lT + lE
end
Apply Aggregate Financial Terms to lT
Populate Trial-Loss pairs in LLT using lT
end
end
Sum losses of Trial-Loss pairs in all LLT
Populate Trial-Loss pairs in P LT
end
Aggregate losses of Trial-Loss pairs in P LT
Populate Y LT
ductible of the insured for an annual cumulative loss, where
as Aggregate Limit refers to the limit or coverage the insurer will pay for annual cumulative losses in excess of the
aggregate retention. The Aggregate Financial terms captures contractual properties as they apply to multiple event
occurrences. The Trial-Loss pairs are then used to populate
Layer Loss Tables LLT s; each Layer is represented using a
Layer Loss Table consisting of Trial-Loss pairs.
In line nos. 16 and 17, the trial losses are aggregated from
the Layer level to the Program level. The losses are represented again as a Trial-Loss pair and are used to populate
Program Loss Tables P LT s; each Program is represented
using a Program Loss Table.
In line nos. 19 and 20, the trial losses are aggregated from
the Program level to the Portfolio level. The Trial-Loss pairs
are populated in the Year Loss Table Y LT which represents
the output of the analysis of aggregate risk. Financial functions or filters are then applied on the aggregate loss values.
2.2
MapReduce Algorithm
Algorithm 2: Parallel Aggregate Risk Analysis
Input : Y ET , ELT pool, P F
Output: Y LT
1
2
3
4
5
6
forall the Programs of P do
forall the Layers L in P do
LLT ← M apReduce1 (L, Y ET )
end
end
Y LT ← M apReduce2 (LLT s)
MapReduce is a programming model developed by Google
for processing large amount of data on large clusters. A map
and a reduce function are adopted in this model to execute
a problem that can be decomposed into sub-problems with
no dependencies; therefore the model is most attractive for
embarrassingly parallel problems. This model is scalable
across large number of computing resources. In addition
to the computations, the fault tolerance of the execution,
for example, handling machine failures are taken care by
the MapReduce model. An open-source software framework
that supports the MapReduce model, Apache Hadoop [16,
17, 18], is used in the research reported in this paper.
The MapReduce model lends itself well towards solving
embarrassingly parallel problems, and therefore, the analysis
of aggregate risk is explored on MapReduce. In the analysis
of aggregate risks, the Programs contained in the Portfolio
are independent of each other, the Layers contained in a
Program are independent of each other and further the Trials in the Year Event Table are independent of each other.
This indicates that the problem of analysing aggregate risks
requires a large number of computations which can be performed as independent parallel problems.
Another reason of choice for the MapReduce model is that
it can handle large data processing for the analysis of aggregate risks. For example, consider a Year Event Table comprising one million simulations, which is approximately 30
GB. So for a Portfolio comprising 2 Programs, each with 10
Layers, the approximate volume of data that needs to be
processed is 600 GB.
Further MapReduce implementations such as Hadoop provide dynamic job scheduling based on the availability of cluster resources and distributed file system fault tolerance.
Algorithm 2 shows the MapReduce analysis of aggregate
risk. The aim of this algorithm is similar to the sequential
algorithm in which the algorithm scans through the Portfolio, P F ; firstly through the Programs, P , and then through
the Layers, L. The first round of MapReduce jobs, denoted
as M apReduce1 are launched for all the Layers. The Map
function (refer Algorithm 3) scans through all the Event
Loss Tables ELT s covered by the Layers L to compute the
0
in parallel for every Event in the ELT. The compulosses lE
tations of loss lT at the Layer level are performed in parallel
by the Reduce function (refer Algorithm 4). The output of
M apReduce1 is a Layer Loss Table LLT .
The second round of MapReduce jobs, denoted as
M apReduce2 are launched for aggregating all the LLT s in
each Program to a Y LT .
Algorithm 3: Map function in M apReduce1 of the
Analysis of Aggregate Risk
Input : < T , E >
0
Output: < T , lE
>
1
2
3
4
5
6
for each ELT covered by L do
Lookup E in the ELT and find corresponding loss,
lE
Apply Financial Terms to lE
0
0
lE
← lE
+ lE
end
0
Emit(T , lE
)
The master node of the cluster solving a problem partitions the input data to intermediate files effectively splitting
the problem into sub-problems. The sub-problems are distributed to the worker nodes by the master node, often referred to as the ‘Map’ step performed by the Mapper. The
Algorithm 4: Reduce Function in M apReduce1 of the
Analysis of Aggregate Risk
Input : T , L0E
Output: < T , lT >
1
2
3
4
5
6
0
for each lE
in L0E do
0
Apply Occurrence Financial Terms to lE
0
lT ← lT + lE
end
Apply Aggregate Financial Terms to lT
Emit(T , lT )
map function executed by the Mapper receives as input a
< key, value > pair to generate a set of < intermediate
key, intermediate value > pairs. The results of the decomposed sub-problems are then combined by the Reducer
referred to as the ‘Reduce’ step. The Reduce function executed by each Reducer merges the < intermediate key, intermediate
value > pairs to generate a final output. The Reduce function receives all the values corresponding to the same intermediate key.
Algorithm 3 and Algorithm 4 show how parallelism is
achieved by using the Map and Reduce functions in a first
round at the Layer level. Algorithm 3 shows the Map function whose inputs are a set of T, E from the Y ET , and the
0
> which corresponds
output is a Trial-Loss pair < T, lE
to an Event. To estimate the loss, it is necessary to scan
through every Event Loss Table ELT covered by a Layer L
(line nos. 1-5). The loss, lE associated with an Event, E in
the ELT is fetched from memory in line no. 2. Contractual
financial terms to the benefit of the Layer are applied to the
0
(line no. 4).
losses (line no. 3) to aggregate the losses as lE
0
>.
The loss for every Event in a Trial is emitted as < T, lE
Algorithm 4 shows the Reduce function in the first MapReduce round. The inputs are the Trial T and the set of losses
0
) corresponding to that Trial, represented as L0E , and the
(lE
output is a Trial-Loss pair < T, lT >. For every loss value
0
in the set of losses L0E , the Occurence Financial Terms,
lE
namely Occurrence Retention and the Occurrence Limit, are
0
(line no. 2) and summed up as lT (line no. 3).
applied to lE
The Aggregate Financial Terms, namely Aggregate Retention and Aggregate Limit are applied to lT (line no. 5).
The aggregated loss for a Trial, lT is emitted as < T, lT >
to populate the Layer Loss Table.
Algorithm 5 and Algorithm 6 show how parallelism is
achieved by using the Map and Reduce functions in a second
round for aggregating all Layer Loss Tables to produce the
Y LT . Algorithm 5 shows the Map function whose inputs are
a set of Layer Loss Tables LLT s, and the output is a TrialLoss pair < T, lT > which corresponds to the Layer-wise loss
for Trial T .
Algorithm 5: Map function in M apReduce2 of the
Analysis of Aggregate Risk
Input : LLT s
Output: < T , lT >
1
2
3
for each T in LLT do
Emit(< T, lT >)
end
Algorithm 6: Reduce function in M apReduce2 of the
Analysis of Aggregate Risk
Input : < T, LT >
Output: < T , lT0 >
1
2
3
4
for each lT in LT do
lT0 ← lT0 + lT
end
Emit(< T, lT0 >)
Algorithm 6 shows the Reduce function whose inputs are
a set of losses corresponding to a Trial in all Layers LT , and
the output is a Trial-Loss pair < T, lT0 > which is an entry
to populate the final output, the Year Loss Table Y LT . The
function sums up trial losses lT across all Layers to produce
a portfolio-wise aggregate loss lT0 .
3.
IMPLEMENTATION AND EXPERIMENTS
ON THE HADOOP PLATFORM
The experimental platform for implementing the MapReduce algorithm is a heterogeneous cluster comprising (a) a
master node which is an IBM blade of two XEON 2.67 GHz
processors comprising six cores, memory of 20 GB per processor and a hard drive of 500 GB with an additional 7 TB
RAID array, and (b) six worker nodes each with an Opteron
Dual Core 2216 2.4 GHz processor comprising four cores,
memory of 4 GB RAM and a hard drive of 150 GB. The
nodes are interconnected via Infiniband.
Apache Hadoop, an open-source software framework is
used for implementing the MapReduce analysis of aggregate risk. Other available frameworks [19, 20] require the
use of additional interfaces, commercial or web-based, for
deploying an application and were therefore not chosen.
The Hadoop framework works in the following way for
a MapReduce round. First of all the data files from the
Hadoop Distributed File System (HDFS) is loaded using
the InputFormat interface. HDFS provides a functionality called Distributed Cache for distributing small data files
which are shared by the nodes of the cluster. The Distributed Cache provides local access to shared data. The
InputFormat interface specifies the input the Mapper and
splits the input data as required by the Mapper. The Mapper
interface receives the partitioned data and emits intermediate key-value pairs. The Partitioner interface receives the
intermediate key-value pairs and controls the partitioning of
these keys for the Reducer interface. Then the Reducer interface receives the partitioned intermediate key-value pairs
and generates the final output of this MapReduce round.
The output is received by the OutputFormat interface and
provides it back to HDFS.
The input data for MapReduce ARA which are the Year
Event Table Y ET , the pool of Event Loss Table ELT and
the Portfolio P F specification are stored on HDFS. The master node executes Algorithm 2 to generate the Year Loss
Table Y LT which is again stored on the HDFS. The two
MapReduce rounds are illustrated in Figure 1.
In the first MapReduce round the InputFormat interface
splits the Y ET based on the number of Mappers specified
for the MapReduce round. The Mappers are configured such
that they also receive the ELT s covered by one Layer which
are contained in the distributed cache. The Mapper applies
Financial Terms to the losses. In this implementation combining the ELT s is considered for achieving fast lookup. A
typical ELT would contain entries in the form of an Event
ID and related loss information. When the ELT s are combined they contain an Event ID and the loss information
related to all the individual ELT s. This reduces the number of lookups for retrieving loss information related to an
Event when the Events in a Trial contained in the Y ET are
scanned through by the Mapper. The Mapper emits a trialEvent Loss pair which is collected by the Partitioner. The
Partitioner delivers the Trial-Event Loss pairs to the Reducers; one Reducer receives all the Trial-Event Loss pairs
related to a specific trial. The Reducer applies the Occurrence Financial and Aggregate Financial Terms to the losses
emitted to it by the Mapper. Then the OutputFormat writes
the output of the first MapReduce round as Layer Loss Tables LLT to the HDFS.
In the second MapReduce round the InputFormat receives
all the LLT s from HDFS. The InputFormat interface splits
the set of LLT s and distributes them to the Mappers. The
Mapper interface emits Layer-wise Trial-Loss pairs. The
Partitioner receives all the Trial-Loss pairs and partitions
them based on the Trial for each Reducer. The Reducer
interface uses the partitioned Trial-Loss pairs and combines
them to Portfolio-wise Trial-Loss pairs. Then the OutputFormat writes the output of the second MapReduce round
as a Year Loss Table Y LT to the HDFS.
3.1
Results
Experiments were performed for one Portfolio comprising
one Program and one Layer and sixteen Event Loss Tables.
The Year Event Table has 100,000 Trials, with each Trial
comprising 1000 Events. The experiments are performed
for up to 12 workers as there are 12 cores available on the
cluster employed for the experiments.
Figure 2 shows two bar graphs for the total time taken
in seconds for the MapReduce rounds when the workers are
varied between 1 and 12; Figure 2a for the first MapReduce
round and Figure 2b for the second MapReduce round. In
the first MapReduce round the best timing performance is
achieved on 12 Mappers and 12 Reducers taking a total of
370 seconds, with 280 seconds for the Mapper and 90 seconds
for the Reducer. Over 85% efficiency is achieved in each case
using multiple worker nodes compared to 1 worker. This
round is most efficient on 3 workers achieving an efficiency of
97% and the performance deteriorates beyond the use of four
workers on the cluster employed. In the second MapReduce
round the best timing performance is achieved again on 12
Mapper and 12 Reducers taking a total of 13.9 seconds, with
7.2 seconds for the Mapper and 6.7 seconds for the Reducer.
Using 2 workers has the best efficiency of 74%; the efficiency
deteriorates beyond this. The second MapReduce round has
performed poorly compared to the first round as there are
large I/O and initialisation overheads on the workers.
Figure 3 shows two bar graphs in which the time for the
first MapReduce round is profiled. For the Mapper the time
taken into account is for (a) applying Financial Terms, (b)
local I/O operations, and (c) data delivery from the HDFS
to the InputFormat, from the InputFormat to the Mapper,
and from the Mapper to the Partitioner. For the Reducer the
time taken into account is for (a) applying Occurrence and
Aggregate Financial Terms, (b) local I/O operations, and (c)
data delivery from the Partitioner to the Reducer, from the
(a) First MapReduce round
(b) Second MapReduce round
Figure 1: MapReduce rounds in the Hadoop implementation
Reducer to the OutputFormat and from the OuputFormat to
HDFS. For both the Mappers and the Reducers it is observed
that over half the total time is taken for local I/O operations.
In the case of the Mapper the mathematical computations
only take 1/4th the total time, and the total time taken for
data delivery from the HDFS to the InputFormat, and from
the InputFormat to the Mapper and from the Mapper to
the Partitioner is only 1/4th the total time. In the case of
the Reducer the mathematical computations take 1/3rd the
total time, whereas the total time taken for data delivery
from the Partitioner to the Reducer, from the Reducer to
the OutputFormat, and from the OutputFormat to HDFS is
nearly 1/6th the total time. This indicates that the local I/O
operations on the cluster employed is expensive though the
performance of Hadoop is exploited for both the numerical
computations and for large data processing and delivery.
Figure 4 shows a bar graph for the time taken in seconds
for the second MapReduce round on 12 workers when the
number of Layers are varied from 1 to 5000. There is a
steady increase in the time taken for data processing and
data delivery by the Mapper and the Reducer. Gradually
the time step decreases resulting in the flattening of the
trend curve.
Figure 5 shows the relative speed up achieved using MapReduce for aggregate risk analysis. There is close to linear
speed up achieved until seven worker nodes are employed,
and beyond seven nodes the gap between linear and relative
speed up increases. This is reflected in the efficiency of the
simulation for different number of workers shown in Figure
6. Over 90% efficiency is achieved up to seven worker nodes.
Beyond seven workers efficiency drops. For all the workers
over 50% of the time is required for local I/O operations, and
around 22% of the time is required for applying financial
terms. Between 15%-22% of the time is required for data
delivery, with a slight increase in time for each additional
worker employed. This is possibly due to the overhead involved in using a centralised RAID data storage, which can
be minimised if distributed file replication techniques are
employed.
The results indicate that there is scope for achieving high
efficiency and speedup for numerical computations and large
data processing and delivery within the Hadoop system.
However, it is apparent that large overheads for local I/O
operations on the workers and for data transfer onto a centralised RAID system are restraining performance. This
large overhead is a resultant of the bottleneck in the connectivity between the worker nodes and the latency in processing data from local drives and the redundant data transfer to
the centralised RAID storage. Therefore, efforts need to be
made towards reducing the I/O overhead and seeking alternative distributed strategies to incorporate data replication
for exploiting the full benefit of the Hadoop MapReduce
model.
(a) First MapReduce round
(a) No. of Mappers vs Time time taken for (a) applying Financial Terms, (b) local I/O operation by each Mapper, and
(c) data delivery
(b) Second MapReduce round
Figure 2: Number of workers vs total time taken in seconds
for the MapReduce rounds in the Hadoop implementation
(b) No. of Reducers vs Time time taken for (a) applying Occurrence and Aggregate Financial Terms, (b) local I/O operation
by each Reducer, and (c) data delivery
Figure 3: Profiled time for the first MapReduce round in the
Hadoop implementation
4.
CONCLUSION
Simulations for the analysis of portfolios of catastrophic
risk need to manage and process large volumes of data in the
form of a Year Event Table and Event Loss Tables. To be
able to employ the simulations in real-time the algorithms
need to rapidly process the data which poses both computational and data management challenges. In this paper,
how the MapReduce model using the Hadoop framework can
meet the requirements of rapidly consuming large volumes
of data for the analysis of portfolios of catastrophic risk to
address the challenges has been presented. An embarrassingly parallel algorithm for aggregate risk analysis is proposed and implemented using the Map and Reduce functions
on the Apache Hadoop framework. The data challenges can
be surmounted by employing the Hadoop Distributed File
System and the Distributed Cache both offered by Apache
Hadoop. A simulation of aggregate risk employing 100,000
trials with 1000 catastrophic events per trial performed on
multiple worker nodes using two MapReduce rounds takes
less than 6 minutes. The experimental results show the feasibility of employing MapReduce for parallel numerical computations and data management of aggregate risk analysis
in real-time.
Future work will be directed towards optimising the implementation for reducing the local I/O overheads to achieve
better speedup. Efforts will be made towards incorporating
additional financial filters, such as secondary uncertainty for
fine-grain analysis of aggregate risk.
5.
REFERENCES
[1] G. Connor, L. R. Goldberg and R. A. Korajczyk,
“Portfolio Risk Analysis,” Princeton University Press,
2010.
[2] A. Melnikov, “Risk Analysis in Finance and
Insurance,” Second Edition, CRC Press, 2011.
[3] A. K. Bahl, O. Baltzer, A. Rau-Chaplin and B.
Varghese, “Parallel Simulations for Analysing
Portfolios of Catastrophic Event Risk,” Workshop of
the International Conference for High Performance
Computing, Networking, Storage and Analysis (SC),
2012.
Figure 4: Number of Layers vs the total time in seconds for
the second MapReduce round
Figure 5: Speedup achieved for Aggregate Risk Analysis
using MapReduce on Apache Hadoop
Figure 6: Efficiency achieved for Aggregate Risk Analysis
using MapReduce on Apache Hadoop
[4] J. Dean and S. Ghemawat, “MapReduce: Simplified
Data Processing on Large Clusters,” Communications
of the ACM, Vol. 51, No. 1, 2008, pp. 107-113.
[5] R. R. Anderson and W. Dong, “Pricing Catastrophe
Reinsurance with Reinstatement Provisions Using a
Catastrophe Model,” Casualty Actuarial Society
Forum, Summer 1998, pp. 303-322.
[6] G. G. Meyers, F. L. Klinker and D. A. Lalonde, “The
Aggregation and Correlation of Reinsurance
Exposure,” Casualty Actuarial Society Forum, Spring
2003, pp. 69-152.
[7] W. Dong, H. Shah and F. Wong, “A Rational
Approach to Pricing of Catastrophe Insurance,”
Journal of Risk and Uncertainty, Vol. 12, 1996, pp.
201-218.
[8] K. -H. Lee, Y. -J. Lee, H. Choi, Y. D. Chung and B.
Moon, “Parallel Data Processing with MapReduce: A
Survey,” SIGMOD Record, Vol. 40, No. 4, 2011, pp.
11-20.
[9] T. Condie, N. Conway, P. Alvaro, J. M. Hellerstein, K.
Elmeleegy and R. Sears, “MapReduce Online,” EECS
Department, University of California, Berkeley, USA,
Oct 2009, Technical Report No. UCB/EECS-2009-136.
[10] D. Cummins, C. Lewis and R. Phillips, “Pricing
Excess-of-Loss Reinsurance Contracts Against
Catastrophic Loss,” The Financing of Catastrophe
Risk, Editors: K. A. Froot, University of Chicago
Press, 1999, pp. 93-148.
[11] Y. -S. Lee, “The Mathematics of Excess of Loss
Coverages and Retrospective Rating - A Graphical
Approach,” Casualty Actuarial Society Forum, 1988,
pp. 49-77.
[12] G. Woo, “Natural Catastrophe Probable Maximum
Loss,” British Actuarial Journal, Vol. 8, 2002.
[13] M. E. Wilkinson, “Estimating Probable Maximum
Loss with Order Statistics,” Casualty Actuarial
Society Forum, 1982, pp. 195-209.
[14] A. A. Gaivoronski and G. Pflug, “Value-at-Risk in
Portfolio Optimisation: Properties and Computational
Approach,” Journal of Risk, Vol. 7, No. 2, 2005, pp.
1-31.
[15] P. Glasserman, P. Heidelberger and P. Shahabuddin,
“Portfolio Value-at-Risk with Heavy-tailed Risk
Factors,” Mathematical Finance, Vol. 12, No. 3, 2002,
pp. 239-269.
[16] T. White, “Hadoop: The Definitive Guide,” 1st
Edition, O’Reilly Media, Inc., 2009.
[17] Apache Hadoop Project: http://hadoop.apache.org/
[Last Accessed: 10 April, 2013]
[18] K. Shvachko, K. Hairong, S. Radia and R. Chansler,
“The Hadoop Distributed File System,” Proceedings of
the 26th IEEE Symposium on Mass Storage Systems
and Technologies, 2010, pp. 1-10.
[19] Amazon Elastic MapReduce (EMR): http://aws.
amazon.com/elasticmapreduce/ [Last accessed: 10
April, 2013]
[20] Google MapReduce: https://developers.google.com/
appengine/docs/python/dataprocessing/overview
[Last accessed: 10 April, 2013]
| 5cs.CE
|
Orthogonal Compaction Using Additional Bends
Michael Jünger1 , Petra Mutzel2 , and Christiane Spisla2
arXiv:1706.06514v1 [cs.DS] 20 Jun 2017
1
University of Cologne, Cologne, Germany
[email protected]
2
TU Dortmund University, Dortmund, Germany
{petra.mutzel,christiane.spisla}@cs.tu-dortmund.de
Abstract. Compacting orthogonal drawings is a challenging task. Usually algorithms try to compute drawings with small area or edge length
while preserving the underlying orthogonal shape. We present a onedimensional compaction algorithm that alters the orthogonal shape of
edges for better geometric results. An experimental evaluation shows
that we were able to reduce the total edge length and the drawing area,
but at the expense of additional bends.
1
Introduction
The compaction problem in orthogonal graph drawing deals with constructing an
area-efficient drawing on the orthogonal grid. Every edge is drawn as a sequence
of horizonal and vertical segments, where the vertices and bends are placed
on grid points. Compaction has been studied in the context of the topologyshape-metrics approach [3]. Here, in a first phase a combinatorial embedding is
computed that determines the topology of the layout with the goal to minimize
the number of crossings. In the second phase, a dimensionless orthogonal shape
of the graph is determined by fixing the angles between adjacent edges and the
bends along the edges. The goal is to minimize the number of bends. In the
third phase, metrics are added to the orthogonal shape. In this context, first the
coordinates of vertices and bends are assigned to grid points so that the given
orthogonal shape is maintained. Finally, the (orthogonal) compaction problem
asks for a drawing minimizing geometric aestetic criteria, such as the area of the
drawing or the total edge length. The shape is not allowed to change.
Since the orthogonal compaction problem is NP-hard [16], in practice heuristics are used that fix the x- (or y-, resp.) coordinates and solve the resulting compaction problem in one dimension. Given an initial drawing, the one-dimensional
compaction problem with the goal of minimizing the height (or width, resp.) of
the layout can be transformed to the longest path problem in a directed acyclic
graph. If in addition the total edge length shall be minimized, the problem can
be solved by computing a minimum cost flow. The topology-shape-metrics approach aims at drawings with a small number of crossings, a small number of
bends, and a small drawing area. These goals are addressed in this order. And
indeed, compared with other drawing methods, the number of crossings and
bends is relatively small [8]. However, the layouts often contain large areas of
(a)
(b)
Fig. 1. (a) A drawing with large areas of white space due to shape restrictions. (b)
Introducing two additional bends to one edge of the drawing leads to a smaller drawing.
white space. It seems that the goal of getting a small drawing area has not been
achieved so far. Consider the drawing in Fig. 1(a) which contains large areas of
white space due to shape restrictions. By introducing two bends on one of the
edges the drawing area can be reduced drastically. This motivates us to study a
compaction problem in which the shape conditions are relaxed.
This brings us back to the origin of the orthogonal compaction problem in
VLSI-design (see, e.g., [15]). In contrast to the compaction problem considered in
graph drawing, even the permutation of wires along the boundary of a component
(and hence, changing the embedding) is allowed.
We suggest a moderate relaxation of the orthogonal compaction problem.
More precisely, we suggest to study the one-dimensional monotone flexible edge
compaction problem with fixed vertex star geometry, henceforth the Fled-Five
compaction problem, which asks for the minimization of the vertical (horizontal, resp.) edge length and allows changing the orthogonal shape of the edge,
but preserves the x-monotonicity (y-monotonicity, resp.) of edge segments and
prohibits changing the directions of the initial edge segments around the vertices. We present a polynomial-time algorithm based on a network flow model
that solves the Fled-Five compaction problem to optimality. Our computational
results show that repeated application of Fled-Five compaction in x- and ydirection is able to reduce the total edge length and the drawing area at the
expense of additional bends.
This paper is organized as follows. We recall the state of the art in Sect. 2 and
some basic definition about orthogonal graph drawing and especially the compaction phase in Sect. 3. We present our new algorithm in Sect. 4 and evaluate
it experimentally in Sect. 5.
2
State-of-the-Art
Patrignani [16] has shown that planar orthogonal compaction is in general NPhard and Bannister et al. [2] gave inapproximability results for the nonplanar
case. But for some special cases there exist polynomial algorithms, e.g. if all faces
are of rectangular shape [7] or if all faces are so-called turn-regular [4] or have a
unique completion [14]. Klau et al. [14] suggested a branch-and-cut approach to
solve an integer linear program based on extending a pair of constraint graphs.
However, in practice, heuristics are used which iteratively fix the x-, and then
the y-coordinates, and solve the resulting one-dimensional compaction problem.
This process is repeated until no further progress is made. One-dimensional
compaction algorithms often use either network flow techniques or a longest
path method in order to assign integer coordinates to the vertices, see e.g. [11]
for an overview. An experimental comparison of planar compaction algorithms
was presented by Klau et al. [13].
Although there has been done some work to improve the quality of a drawing
by changing its shape, e.g. [9], most compaction algorithms take as input an
orthogonal representation and try to produce compact drawings with respect
to that representation. This can lead to an unnecessarily large drawing area
with unused space. Often better results in terms of area and edge length can be
achieved if the orthogonal shape can be altered, as we have seen in Fig. 1. On
the other hand, it might be desirable to not change a given drawing too much
in order to preserve the mental map.
3
Notation and Preliminary Results
In this section we give basic definitions and notations. For more details on orthogonal drawings and graph drawing in general see e.g. [7], [11] or [17].
3.1
Orthogonal Graph Drawing
For the rest of this paper we restrict ourselves to undirected 4-graphs, i.e. graphs
whose vertices have at most four incident edges. A graph G = (V, E) with |V | = n
and |E| = m is called planar if it admits a drawing Γ in the plane without edge
crossings. Such a planar drawing of G induces a (planar) embedding, which is
represented by a circular ordered list of bordering edges for every face. The
unbounded region of a planar drawing is called external face.
An orthogonal representation H is an extension of a planar embedding that
gives combinatorial information about the orthogonal shape of a drawing. For
every edge we provide information about the bends encountered while traversing
the edge and the angle formed at vertices by two consecutive edges. If one of
these angles is 270◦ or 360◦ we associate with v a reflex corner. An orthogonal
representation is called normalized if it has no bends. An orthogonal grid drawing
Γ of G is a drawing in which every edge is drawn as a sequence of horizontal and
vertical edge segments and every vertex and bend has integer coordinates. Such a
drawing induces an orthogonal representation HΓ and a star geometry for every
vertex fixing the directions of the initial line segments of its incident edges. If
an edge first turns to the right and then to the left, or vice versa, we call this a
double bend and the edge segment between those two bends a middle segment.
Every orthogonal representation can be normalized by replacing all bends in HΓ
with dummy vertices of degree two, thus adding vertices to G and Γ .
Since our new approach is based on a network flow model for one-dimensional
compaction, we will give a brief introduction to minimum cost flows here. For
more information about network flows, see [1]. Let N = (VN , EN ) be a directed
graph. Whenever we talk about flows we will call N a network, the members
of VN nodes and the members of EN arcs (in contrast to vertices and edges
in an undirected graph). Every arc a has a lower bound l(a) ∈ IR≥0 , an upper
bound u(a) ∈ IR≥0 ∪ {∞} and a nonnegative cost c(a). A demand b(n) ∈ IR is
associated with every node. We call a function x : A → IR≥0 a flow if x satisfies
the following conditions:
l(a) ≤ x(a) ≤ u(a)
X
X
x(a) −
x(a) = b(k)
a=(k,l)
∀ a ∈ EN
(capacity constraint)
(1)
∀ k ∈ VN
(flow conservation)
(2)
a=(j,k)
P
A minimum cost flow is a flow x with minimum total cost cx = a∈EN x(a)c(a)
under all feasible flows. The minimum cost flows we are interested in can be
computed in O(|VN |3/2 log |VN |) time [6].
3.2
Compaction of Orthogonal Drawings
We focus on the vertical (orthogonal) compaction problem that receives as input
a planar grid drawing Γ of a graph G with an orthogonal representation HΓ ,
and asks for another planar orthogonal drawing Γ 0 of G realizing HΓ so that the
vertical edge length is minimized subject to fixed x-coordinates. In the following
we describe a flow-based method for vertical compaction similar to the coordinate
assignment algorithm in Di Battista et al. [7].
Assume we have an initial grid drawing Γ . In a first step we normalize HΓ
resulting in Γ and H Γ . Then we add vertical visibility edges (so-called dissecting
edges). We insert a vertical edge connecting each reflex corner with the vertex
or edge that is visible in vertical direction, possibly introducing dummy vertices.
e Γ . This way we get rid of all reflex corners and have a
This gives us Γe and H
representation with rectangular faces. Now we are ready to construct the network
e Γ we add a node
Ny for vertical compaction. For each face f in representation H
nf to Ny with demand b(nf ) = 0 and for every vertical edge e with left face fl
and right face fr we insert an arc ae = (nfl , nfr ) with lower bound l(ae ) = 1
and upper bound u(ae ) = ∞. If e is a dummy edge ae gets zero cost, otherwise
a cost of one. See Fig. 2 for an example.
Suppose we have computed a feasible flow. Now a unit of flow corresponds
to one unit of vertical edge length. The x-coordinates remain unchanged. The
capacity constraint assures that every edge gets a minimum length of one and
the flow conservation constraint guarantees that every face is drawn as a proper
rectangle. Finally, the visibility edges and artificial vertices can be removed.
Lemma 1. The above algorithm solves the vertical compaction problem to optimality.
(a)
(b)
Fig. 2. The flow networks Ny for vertical compaction (a) and Nx for horizontal compaction.(b). White vertices are dummy vertices and dashed edges are dummy edges.
Proof. In the vertical compaction problem we have to preserve the vertical visibility properties of Γ in order to avoid overlapping graph elements. This is
achieved by the newly added visibility edges. Because of the one-to-one correspondence of vertical length of an edge segment in the resulting drawing Γ 0 and
the amount of flow carried by a non-zero-cost arc ae ∈ Ny , the result of the
minimum cost flow gives us a minimal vertical length assignment.
t
u
4
The Fled-Five Compaction Approach
In this section we study the following relaxation of the one-dimensional compaction problem called the monotone flexible edge compaction problem with fixed
vertex star geometry (Fled-Five compaction problem). For vertical compaction
it is defined as follows: Given a planar grid drawing Γ of a 4-graph G with an
orthogonal representation HΓ , compute another planar grid drawing of G that
minimizes the vertical edge length subject to fixed x-coordinates in which all horizontal edge segments of Γ are drawn x-monotonically the vertex star geometry
for all vertices as well as the planar embedding is maintained.
In contrast to the classical compaction problem, it is not required here to realize the entire orthogonal representation, but only the local geometric surrounding of each vertex. The fixed x-coordinates and the x-monotonicity prohibits the
lengthening of the total horizontal edge length (see Fig. 3).
We adapt the network flow approach described in Sect. 3.2. But now we are
able to introduce or remove double bends of certain edges in order to improve
the vertical edge length. We illustrate the concept by the following example.
Figure 4(a) shows an optimally compacted drawing with respect to its orthogonal
representation. The blue arcs show the arcs of the vertical compaction network
(compare Fig. 2). For better readability we omitted the network nodes belonging
to the faces. In Fig. 4(b), the same graph is shown, but with a different orthogonal
representation. The new double bend in the middle edge leads to a smaller
drawing. In this drawing it is possible to send flow between the two internal
(a)
(b)
(c)
(d)
Fig. 3. Modifications in Fled-Five. (a) The original edge of length three. (b) Modification allowed in Fled-Five. (c) not allowed because of changed x-coordinate and (d)
not allowed, because x-monotonicity is violated.
(a)
(b)
(c)
Fig. 4. A graph and the arcs of the corresponding flow network (a,b) of the traditional
algorithm and (c) in our algorithm
faces, since they are separated by a vertical edge segment. In the corresponding
flow network there is a new network arc (red) connecting the upper and the lower
face. This leads to the key idea of our approach. Introducing arcs in the vertical
flow network between horizontally separated face nodes enables us to shift flow
between them and therefore exchange vertical edge length at the expense of a
double bend (see Fig. 4(c)). We can also reverse this thought. If there is an
unnecessary double bend we can get rid of it by not sending flow over the arc
corresponding to its middle segment and thus removing the middle segment from
the drawing.
But we have to be careful here. First, we are compacting in vertical direction,
so we cannot change the x-coordinates. If an edge has a double bend, it needs
to have a horizontal expansion of at least two. Thus we will not consider edges
of length one as possible candidates for getting a double bend. Second, adding a
double bend to e introduces two reflex corners, one in both adjacent faces of e.
Two double bends may even cause an edge overlap, see Fig. 5(a). We need to
ensure that the computed flow corresponds to a feasible drawing. Therefore we
will treat each grid point along an edge that could be part of a double bend
as a potential reflex corner, which we will eliminate by dissecting. We will now
describe the algorithm in detail.
Our algorithm works in three phases: dissection, computation of a minimum
cost flow and transforming the flow into a new drawing. First we normalize HΓ
to H Γ . For dissection we split the horizontal edges of Γ by placing an artificial
bend vertex on every inner grid point of an horizontal edge. The bend vertices
may later be transformed to double bends. Then we dissect Γ as described in
2
2
2
(a)
2
1
2
2
1
2
2
2
(b)
Fig. 5. Graph with modified flow network, a minimum cost flow (all unlabeled arcs
carry one unit of flow) and the resulting drawing (a) without additional dissecting
edges and (b) including them. The bold edge in (a) indicates an edge overlap and the
grey vertices result from normalization, the white vertices in (b) are bend vertices.
Sect. 3 and treat every bend vertex as a reflex corner in its adjacent faces (see
5(b)). That means, we may dissect the drawing in vertical stripes of unit length.
Doing so, we solve both of the problems mentioned above. Since we consider only
edges of length at least two, we guarantee that the edge will be long enough for a
double bend and by inserting the visibility edges we keep the vertical separation
e Γ and G.
e
of graph elements intact. This results in the extensions Γe and H
Observation 1. After this dissection method the number of vertices and edges
e is O(A2 ), where A = hΓ · wΓ and hΓ (wΓ ) is the height (width) of Γ .
in G
Next we construct the network from Sect. 3.2. Additionally to the already
introduced arcs (blue arcs in Fig. 5) we add two arcs a↑v and a↓v for every bend
vertex (red curved arcs in Fig. 5). Notice that every such bend vertex v has four
adjacent faces, one at the lower and upper left and right. If it lies on the external
face, two of the adjacent faces may coincide. Arc a↑v goes from the lower left face
of v to its upper right face and a↓v goes from the upper left to the lower right
face. Flow on one of these arcs will lead to a double bend in the edge segment of
G that is split by v. The lower bound of these arcs is zero, the upper bound is
e that is a middle segment
set to infinity and the cost is one. For every edge e ∈ G
in G we decrease the lower bound of its arc ae to zero, since this is a vertical
edge segment, that we may delete (orange arc in Fig. 5).
After computing a minimum cost flow in this network, we interpret the result
e we translate the amount of
in the following way: For every vertical edge e of G
flow on the arc ae of Ny to the length of e. Let a↑v be an arc corresponding
to a bend vertex v carrying k units of flow. Let e be the split horizontal edge
and xv the x-coordinate of v. Then e will start at its left endpoint in horizontal
direction, bend downwards at x-coordinate xv , proceed for k units in y-direction
and then continue to the right to its other endpoint. If we deal with an arc of the
e be an
form a↓v the corresponding edge will do an upward bend at xv . Let e0 ∈ G
edge that corresponds to a middle segment and let ae0 be the corresponding arc.
If ae0 carries no flow we do not assign any vertical length to e0 , hence the double
bend to which e0 belongs collapses. Notice that flow on every non-zero-cost arc
corresponds to vertical edge length. Finally, we remove all dummy edges and
vertices. See Fig. 5 for an example.
Let us assume that the width and height of Γ are bounded by the number of
its vertices and bends. Otherwise there would be a grid line without any vertex
or bend on it. We can delete such grid lines iteratively until we reach our bound.
Theorem 1. Let Γ be a planar grid drawing of a 4-graph G with an orthogonal
representation HΓ . Let n̄ be the number of vertices and bends of Γ . Then the
above described extended network-based compaction algorithm takes O(n̄3 log n̄)
time and solves the vertical Fled-Five compaction problem to optimality.
Proof. Because of the visibility edges we will maintain visibility properties of Γ .
Every vertical edge segment of G that is not a middle segment gets a minimum
length of one due to the lower bound of the corresponding arc in Ny . Every
bend vertex can savely be turned to a double bend if its corresponding arc
carries flow, since we add visibility edges to the top and bottom of it. By this
modification we do not change the vertex star geometry of Γ , because bend
vertices are not part of G. Every middle segment can be removed if there is no
flow on the corresponding arc. Because its edge e ∈ Γ has a horizontal expansion
of at least two, e will still have a proper length and due to the dissection phase
visibility properties are maintained. This modification also does not affect the
vertex star geometry of Γ , since a middle segment is not adjacent to a vertex
of G. So every face will have a rectangular shape, no matter what modification
we apply, and due to flow conservation every face and thus the entire graph is
drawn consistently. Similar to Lemma 3.2 the length of vertical segments of Γ 0
is equal to the cost of the computed minimum cost flow. The horizontal edge
segments maintain their x-monotonicity, because we only add vertical segments
to the drawing. Hence the horizontal edge lengths and the x-coordinates of the
vertices in Γ stay the same. Because the minimum cost flow gives us a minimal
vertical length assignment we have an optimal solution for the vertical Fled-Five
compaction problem.
For the running time we first consider the dissection phase. By Observation 1
we end up with O(n̄2 ) vertices. For inserting dummy edges based on visibility
properties we can use a sweep-line algorithm that runs in O(n̄2 log n̄) time in our
e
case. The construction of the flow network runs in linear time in the size of G.
The network is planar and has O(n̄2 ) nodes. For planar flow networks with n
nodes and O(n) arcs there exists a O(n3/2 log n) time algorithm for computing
a minimum cost flow [6]. Therefore we have a total running time of O(n̄3 log n̄).
t
u
The cubic part of the running time comes from the possibly quadratic number
of bend vertices and dissecting edges. So if we restrict the number of bend vertices
to be linear, we can decrease the running time to O(n̄3/2 log n̄).
Controlling the Number of New Bends. Although the above compaction approach may reduce the total edge length, the number of bends in the resulting
drawing may increase. A possible approach to control the number of new bends
could be to bound the number of specific arcs used by a feasible flow. This is
known as the binary case of the budget-constrained minimum cost flow problem
for which Holzhauser et al. [10] showed strong NP-completeness. In fact, in this
model we have no other control over the number of additional bends than restricting the number of bend vertices. Each such vertex can generate two new
bends.
5
Experimental Evaluation
In our experimental evaluation we compare the new compaction approach FF
with the flow-based method TRAD described in Sect. 3. Both lead to a planar drawing with optimal total edge length for the according one-dimensional
compaction problem.
The algorithm was implemented in C++ using the OGDF library [5]. We
have computed orthogonal drawings in the traditional way using the topologyshape-metrics approach, applying a constructive compaction method in order
to get a feasible planar grid drawing with a normalized orthogonal representation H. We have taken these drawings as input for both approaches FF and
TRAD that repeatedly apply horizontal and vertical compaction steps until no
further improvement can be achieved. Recall that although both approaches are
optimal for the one-dimensional compaction problems, repeated application of
horizontal and vertical compaction steps does in general not lead to drawings
with optimal total edge length for the two-dimensional compaction problems.
Notice also that due to the alternating repetitions in FF the monotonicity of
edge segments may no longer be maintained.
We state the following hypotheses regarding the results of our new approach
FF compared to those of TRAD:
(H1) The total edge length and therefore the area of the drawings will decrease.
We will also examine the change of the maximum edge length. It is hard
to predict the behaviour, because on the one hand adding a double bend
lengthens an edge, but it could also lead to shorter edges in other places.
(H2) The number of bends will rise significantly.
(H3) Although there will be many more dissecting edges, the running time will
not increase drastically.
We tested our algorithm on three test suites. The standard benchmark set
called Rome graphs introduced in [8] consists of about 11,000 real-world and real-
30
800
total
max
area
bends
time per iteration
time
vis. edges
700
increase in %
improvement in %
35
25
20
15
10
5
600
500
400
300
200
100
0
10-50
51-100
101-150
151-200
0
200-314
10-50
51-100
nodes
101-150
151-200
200-314
301-400
401-500
301-400
401-500
nodes
(a)
300
total
max
area
15
250
increase in %
improvement in %
20
10
5
0
-5
bends
time per iteration
time
vis. edges
200
150
100
50
-10
-15
40-100
101-200
201-300
301-400
0
401-500
40-100
101-200
nodes
201-300
nodes
(b)
improvement in %
45
40
300
total
max
area
250
increase in %
50
35
30
25
20
15
10
bends
time per iteration
time
vis. edges
200
150
100
50
5
0
40-100
101-200
201-300
301-400
0
401-500
nodes
40-100
101-200
201-300
nodes
(c)
Fig. 6. Average change of various criteria from TRAD to FF in percent for (a) the
Rome graphs, (b) the quasi trees, and (c) the biconnected graphs
world like graphs with 10 to 100 vertices. The second set of graphs are quasitrees which are known to be hard to compact. They have already been used
in the compaction literature (e.g., [13],[12]). From this, we selected a subset
of 155 graphs with 40 to 450 vertices. The last set consists of 240 randomly
generated biconnected planar 4-graphs with 40 to 500 vertices. All used graphs
were, if necessary, initially turned into planar 4-graphs by planarizing them with
methods of OGDF and replacing vertices with k > 4 outgoing edges with faces
of size k. This results in graph sizes of up to 314 vertices for the Rome test suite
and up to 475 vertices for the quasi-trees before the orthogonalization step.
The input instances are available on https://ls11-www.cs.tu-dortmund.de/
mutzel/compaction. All tests were run on an Intel Xeon E5-2640v3 2.6GHz
CPU with 128 GB RAM.
(H1) Figure 6 shows on the left the average decrease of the total and maximum
edge length as well as of the area in percent. In all three graph classes the
200
TRAD
FF
5000
maximum edge length
total edge length
6000
4000
3000
2000
1000
0
0
100
200
300
400
180
140
120
100
80
60
40
20
0
500
TRAD
FF
160
0
100
200
nodes
300
120
TRAD
FF
total running time in sec.
350
bends
250
200
150
100
50
0
0
100
200
300
nodes
300
400
500
400
500
nodes
400
500
100
TRAD
FF
80
60
40
20
0
0
100
200
300
nodes
Fig. 7. Absolute results of TRAD and FF for biconnected graphs
total edge length as well as the area has improved. We were able to decrease
the total edge length by up to 36.1% and the area by up to 68.7%. In general,
larger graphs allow a larger improvement. It turned out that for the majority
of the instances we were also able to reduce the maximum edge length. Only
for the quasi-trees the longest edges produced by FF are longer on average. But
in general the results are mixed, reaching from a lengthening by 87.5% to a
shortening by 58.1%. Figure 7 shows absolute values for the total edge length
and the maximum edge length measured for the biconnected graphs. (The plots
for the other graph classes can be found in the Appendix.)
(H2) The right side of Fig. 6 displays the average increase of the number of
bends. If an instance had no bends after TRAD, we use the number of bends after
FF as relative increase to take also these instances into account for computing
the average increase of bends. As expected the drawings produced by FF have
many more bends, especially for the Rome graphs. In one of the worst cases the
number of bends went up from zero to 72. Figure 7 shows the absolute values for
the number of bends measured for the biconnected graphs. Figure 8 gives the
relation between the number of invested bends and the improvement in terms
of area and total edge length for the biconneced graphs. The data for the other
two graph classes look similar.
(H3) We measured the total running time and the number of performed compaction iterations. For better comparison we also display the running time per
iteration, because FF tends to do more iterations. The reason for that is that
in one compaction step of FF changes the drawing and therefore the flow network of the next step more significantly, leading to more new possibilities in
compacting. The right side of Fig. 6 displays the average increase of the running
time and the number of visibility edges. In all cases the number of additional
dissecting edges has increased significantly. The running time per iteration has
decrease of tot. edge length in %
decrease of area in %
70
60
50
40
30
20
10
0
0
100
200
300
400
500
600
40
35
30
25
20
15
10
5
0
0
100
200
300
400
increase of bends in %
increase of bends in %
(a)
(b)
500
600
Fig. 8. Relation between additional bends and (a) improvement of area and (b) total
edge length for the biconnected graphs
also increased, but not more than 50% on average. Only for 252 instances of
the Rome graphs and one biconnected graph the running time per iteration has
more than doubled.
The results support the hypotheses. Our new approach is able to improve the
drawing area and the total edge length by introducing additional bends. Figure 9
shows two examples drawn with TRAD and FF, respectively.
(c)
(a)
(d)
(b)
Fig. 9. Biconnected graph with 220 vertices and 304 edges and Rome graph with 55
vertices and 66 edges (after planarization and expanding high degree vertices): (a)
TRAD: total edge length 1459, area 3060, bends 31 (b) FF: total edge length 1107, area
1320, bends 133 (c) TRAD: total edge length 168, area 240, bends 1 (d) FF: total edge
length 140, area 140, bends 13. Edges with additional bends are red.
Acknowledgements. This work was supported by the DFG under the project
Compact Graph Drawing with Port Constraints (DFG MU 1129/10-1). We would
like to thank Gunnar Klau for providing us with most of our test instances. We
also gratefully acknowledge helpful discussions with Martin Gronemann, Sven
Mallach, and Daniel Schmidt.
References
1. Ahuja, R.K., Magnanti, T.L., Orlin, J.B.: Network Flows: Theory, Algorithms, and
Applications. Prentice-Hall, Inc., Upper Saddle River, NJ, USA (1993)
2. Bannister, M.J., Eppstein, D., Simons, J.A.: Inapproximability of orthogonal compaction. J. Graph Algorithms Appl. 16(3), 651–673 (2012), https://doi.org/10.
7155/jgaa.00263
3. Batini, C., Nardelli, E., Tamassia, R.: A layout algorithm for data flow diagrams.
IEEE Transactions on Software Engineering SE-12(4), 538–546 (1986)
4. Bridgeman, S.S., Di Battista, G., Didimo, W., Liotta, G., Tamassia, R., Vismara, L.: Turn-regularity and optimal area drawings of orthogonal representations.
Comput. Geom. 16(1), 53–93 (2000), https://doi.org/10.1016/S0925-7721(99)
00054-1
5. Chimani, M., Gutwenger, C., Jünger, M., Klau, G.W., Klein, K., Mutzel, P.: The
Open Graph Drawing Framework (OGDF). In: Tamassia, R. (ed.) Handbook of
Graph Drawing and Visualization, chap. 17, pp. 543–569. CRC Press (2013)
6. Cornelsen, S., Karrenbauer, A.: Accelerated bend minimization. J. Graph Algorithms Appl. 16(3), 635–650 (2012), http://dx.doi.org/10.7155/jgaa.00265
7. Di Battista, G., Eades, P., Tamassia, R., Tollis, I.G.: Graph Drawing: Algorithms
for the Visualization of Graphs. Prentice-Hall, Upper Saddle River, NJ, USA (1999)
8. Di Battista, G., Garg, A., Liotta, G., Tamassia, R., Tassinari, E., Vargiu, F.: An
experimental comparison of four graph drawing algorithms. Comput. Geom. 7,
303–325 (1997), https://doi.org/10.1016/S0925-7721(96)00005-3
9. Fößmeier, U., Heß, C., Kaufmann, M.: On improving orthogonal drawings: The 4malgorithm. In: Whitesides, S. (ed.) Graph Drawing, 6th International Symposium,
GD’98, Proceedings. LNCS, vol. 1547, pp. 125–137. Springer (1998), https://doi.
org/10.1007/3-540-37623-2$_$10
10. Holzhauser, M., Krumke, S.O., Thielen, C.: Budget-constrained minimum cost
flows. J. Comb. Optim. 31(4), 1720–1745 (2016), https://doi.org/10.1007/
s10878-015-9865-y
11. Kaufmann, M., Wagner, D. (eds.): Drawing Graphs, Methods and Models, LNCS,
vol. 2025. Springer (2001)
12. Klau, G.W.: A combinatorial approach to orthogonal placement problems. Ph.D.
thesis, Saarland University, Saarbrücken, Germany (2002), http://scidok.sulb.
uni-saarland.de/volltexte/2004/196/index.html
13. Klau, G.W., Klein, K., Mutzel, P.: An experimental comparison of orthogonal
compaction algorithms. In: Marks, J. (ed.) Graph Drawing, 8th International
Symposium, GD 2000, Proceedings. LNCS, vol. 1984, pp. 37–51. Springer (2001),
https://doi.org/10.1007/3-540-44541-2$_$5
14. Klau, G.W., Mutzel, P.: Optimal compaction of orthogonal grid drawings. In:
Cornuéjols, G., Burkard, R.E., Woeginger, G.J. (eds.) Integer Programming and
Combinatorial Optimization, 7th International IPCO Conference, 1999, Proceedings. LLNCS, vol. 1610, pp. 304–319. Springer (1999), https://doi.org/10.1007/
3-540-48777-8$_$23
15. Lengauer, T.: Combinatorial Algorithms for Integrated Circuit Layout. John Wiley
& Sons, Inc., New York, NY, USA (1990)
16. Patrignani, M.: On the complexity of orthogonal compaction. In: Algorithms and
Data Structures, 6th International Workshop, WADS ’99, Proceedings. LNCS, vol.
1663, pp. 56–61. Springer (1999), https://doi.org/10.1007/3-540-48447-7$_$7
17. Tamassia, R. (ed.): Handbook on Graph Drawing and Visualization.
Chapman
and
Hall/CRC
(2013),
https://www.crcpress.com/
Handbook-of-Graph-Drawing-and-Visualization/Tamassia/9781584884125
Appendix
Additional Experimental Results
Figures 10 and 11 show the absolute values for the quasi trees and the Rome
graphs. The three plots in Fig. 12 complement the absolute values for the biconnected graphs of Fig. 7. Figures 13 and 14 display the relation between additional
bends and improvement of area and total edge length for the quasi trees and for
the Rome graphs.
2000
TRAD
FF
5000
1600
1400
4000
1200
area
total edge length
6000
TRAD
FF
1800
1000
800
3000
2000
600
400
1000
200
0
0
100
200
300
400
0
500
0
100
200
nodes
200
TRAD
FF
90
160
70
140
60
120
50
40
60
40
10
500
400
500
20
0
100
200
300
400
0
500
0
100
200
nodes
3
total running time in sec.
TRAD
FF
2.5
2
1.5
1
0.5
0
100
200
300
nodes
12
300
400
8
6
4
2
0
500
TRAD
FF
10
0
100
nodes
200
300
nodes
1400
TRAD
FF
1200
visibility edges
time per iteration in seconds
400
80
20
0
500
100
30
0
400
TRAD
FF
180
80
bends
maximum edge length
100
300
nodes
1000
800
600
400
200
0
0
100
200
300
400
500
nodes
Fig. 10. Absolute results of TRAD and FF for quasi-trees
3000
TRAD
FF
TRAD
FF
2500
2000
2000
1500
area
total edge length
2500
1000
1500
1000
500
0
500
0
50
100
150
200
250
0
300
0
50
100
nodes
TRAD
FF
100
200
250
300
200
250
300
200
250
300
TRAD
FF
200
80
60
150
100
40
50
20
0
0
50
100
150
200
250
0
300
0
50
100
nodes
3
total running time in sec.
TRAD
FF
2.5
2
1.5
1
0.5
0
0
50
100
150
150
nodes
14
200
250
10
8
6
4
2
0
300
TRAD
FF
12
0
50
100
nodes
150
nodes
1200
TRAD
FF
1000
visibility edges
time per iteration in seconds
150
nodes
250
bends
maximum edge length
120
800
600
400
200
0
0
50
100
150
200
250
300
nodes
Fig. 11. Absolute results of TRAD and FF for Rome graphs
12000
time per iteration in seconds
14000
TRAD
FF
area
10000
8000
6000
4000
2000
0
0
100
200
300
400
500
30
TRAD
FF
25
20
15
10
5
0
0
100
200
nodes
300
400
500
nodes
4000
visibility edges
3500
TRAD
FF
3000
2500
2000
1500
1000
500
0
0
100
200
300
400
500
nodes
decrease of area in %
50
40
30
20
10
0
-10
0
50
100
150
200
250
300
350
400
decrease of total edge length in %
Fig. 12. Remaining absolute results of TRAD and FF for biconnected graphs
30
25
20
15
10
5
0
0
50
increase of bends in %
100
150
200
250
300
350
400
increase of bends in %
60
decrease of area in %
50
40
30
20
10
0
-10
-20
-30
-40
0
1000
2000
3000
4000
5000
increase of bends in %
6000
7000
decrease of total edge length in %
Fig. 13. Relation between additional bends and improvement of area and total edge
length for the quasi trees
30
25
20
15
10
5
0
0
1000
2000
3000
4000
5000
6000
7000
increase of bends in %
Fig. 14. Relation between additional bends and improvement of area and total edge
length for the Rome graphs
| 8cs.DS
|
Time-Domain Multi-Beam Selection and Its
Performance Improvement for mmWave Systems
Hsiao-Lan Chiang, Wolfgang Rave, and Gerhard Fettweis
arXiv:1803.07046v1 [cs.IT] 19 Mar 2018
Vodafone Chair for Mobile Communications, Technische Universität Dresden, Germany
Email: {hsiao-lan.chiang, rave, gerhard.fettweis}@ifn.et.tu-dresden.de
Abstract—Multi-beam selection is one of the crucial technologies in hybrid beamforming systems for frequency-selective
fading channels. Addressing the problem in the frequency domain
facilitates the procedure of acquiring observations for analog
beam selection. However, it is difficult to improve the quality of
the contaminated observations at low SNR. To this end, this paper
uses an idea that the significant observations are sparse in the time
domain to further enhance the quality of signals as well as the
beam selection performance. By exploiting properties of channel
impulse responses and circular convolutions in the time domain,
we can reduce the size of a Toeplitz matrix in deconvolution
to generate periodic true values of coupling coefficients plus
random noise signals. An arithmetic mean of these signals yields
refined observations with minor noise effects and provides more
accurate sparse multipath delay information. As a result, only
the refined observations associated with the estimated multipath
delay indices have to be taken into account for the analog beam
selection problem.
I. I NTRODUCTION
With the rapid increase of data rates in wireless communications, bandwidth shortage is getting more critical. Accordingly,
there is a growing interest in using millimeter wave (mmWave)
for future wireless communications taking advantage of the
enormous amount of available spectrum [1]. In mmWave
systems, a combination of analog beamforming (operating
in passband) [2], [3] and digital beamforming (operating in
baseband) [4] is one of the low-cost solutions to higher data
rate transmission, and this combination is commonly called
hybrid beamforming [5]-[8]. To implement hybrid beamforming at a transmitter and a receiver simultaneously is certainly
intractable. Therefore, our previous works in [9], [10] focus on
finding the key parameters of the hybrid beamforming gain to
alleviate the problem, and eventually all that matters about the
hybrid beamforming performance is the analog beam selection.
The problem of analog beam selection for frequencyselective fading channels can be stated as a sum-power (or
energy) maximization across all subcarriers [10], [11]. From
Parseval’s theorem, we know that it is equivalent to calculating
the energy of the observations for the analog beam selection
in the delay (or time) domain. Particularly, the observations in
the delay domain can be interpreted as coupling coefficients
of a matrix-valued channel impulse response (CIR) and all
possible analog beam pairs plus noise. Considering an OFDM
system, it is easier to obtain the observations in the frequency
domain. However, these signals seriously suffer from the noise
in the low SNR regime, and it needs more effort to refine them
in the frequency domain than in the delay domain because
the significant observations are not sparse in the frequency
domain. To this end, this paper presents a low-complexity
beam selection method and its performance improvement in
the delay domain.
The delay-domain convolution operation can be constructed
as a matrix multiplication, where one of the inputs (that is, the
training sequence) is converted into a Toeplitz matrix. Then,
left multiplying the received signal vector by the inverse of
the Toeplitz matrix leads to the observations for the analog
beam selection. In OFDM systems, the length LC of a cyclic
prefix (CP) is much less than one OFDM symbol duration with
L samples but is enough to cover the maximum delay spread
[12], which means that at most LC observations in one OFDM
symbol can be used for the beam selection. Unfortunately, the
LC observations are unreliable in the low SNR regime.
In order to improve the quality of the observations for the
beam selection, we generate the training
j k sequence of length
LC with a certain period M = LLC within one OFDM
symbol duration at the transmitter. After deconvolution by a
small-size Toeplitz matrix, we have M periodic signals of
length LC plus random noise signals. An arithmetic mean
of these signals yields the refined observations, where the
effective noise variance is reduced by a factor of M . According
to one of the transmission numerologies in 3GPP 5G New
Radio (NR) [13], M ≈ 14 so that the mean absolute error
(MAE) between the energy estimate and its true value can be
significantly reduced. In addition, if the refined observations
are reliable enough to find the delay indices, eventually only
a few number of signals corresponding to the estimated delay
indices are the significant observations for the analog beam
selection.
The following notations are used throughout this paper. a is
a scalar, a is a column vector, and A is a matrix. an denotes
the nth column vector of A; ai,j denotes the (i, j)th entry of
A. AT and AH denote the transpose and Hermitian transpose
of A respectively. [A]n,: denotes the nth row vector of A. IN
and 0N ×M denote respectively the N ×N identity and N ×M
zero matrices. a[l] ∗ b[l] denotes the linear convolution of a[l]
and b[l].
II. S YSTEM M ODEL
A system having a transmitter with an NT -element uniform
linear antenna array (ULA) communicates NRF data streams
to a receiver with an NR -element ULA as shown in Fig. 1. The
NRF analog beamforming vectors at the transmitter in matrix
)
NRF
:
NR
NT
and FS is the sampling rate,
H[l] =
NRF
P
X
αp δ[l − lp ] · aA (φA,p )aD (φD,p )H
| {z }
p=1
cp [l]
=
$QDORJ
EHDPIRUPLQJ
(3)
cp [l]aA (φA,p )aH
D (φD,p )
p=1
Fig. 1. Both a transmitter and a receiver have NRF analog beamforming
vectors and a baseband (BB) signal processing block including digital beamforming (DBF). This paper focuses on an analog beam selection problem,
which dominates the complexity and performance of hybrid beamforming
systems [10].
F = [f1 , · · · , fNRF ] are selected from a predefined codebook
F = {f̃nf ∈ CNT ×1 , nf = 1, · · · , NF } with the nth
f member
represented as [2]
f̃nf
P
X
aD (φD,p ) = √
1 h j λ2π sin(φT ,nf )∆d
1, e 0
,··· ,
=√
NT
e
2π
jλ
sin(φT ,nf )·(NT −1)∆d
0
where αp ∈ R>0 is the attenuation coefficient for path p and
PP
2
p=1 |αp | = 1. Note that the path loss values influenced by
an environment and geometry are mentioned in the average
received power ρ in (2). cp [l] characterizes the CIR for path p
at sample l and we assume that cp [l] = 0 when l ≥ LC ,
where LC is the CP length. The departure array response
vector aD (φD,p ) is a function of angle of departure (AoD),
φD,p ∼ U(− π2 , π2 ), for path p,
iT
2π
1
sin(φD,p )∆d
j
,··· ,
1, e λ0
NT
e
, (1)
where φT,nf stands for the nth
f candidate of the steering angles
at the transmitter, ∆d = λ20 is the distance between two
neighboring antennas, and λ0 is the wavelength at the carrier
frequency. At the receiver, the NRF analog beamforming
vectors in matrix W = [w1 , · · · , wNRF ] are selected from
the other codebook defined as W = {w̃nw ∈ CNR ×1 , nw =
1, · · · , NW }, where the members can be generated by the
same rule as (1). The analog beamforming matrices F and
W are assumed to be constant within one OFDM symbol
duration owing to hardware constraints.
Via a coupling of two analog beamforming matrices and
a multipath matrix-valued CIR H[l] ∈ CNR ×NT , where l =
0, · · · , L − 1 denotes the sample in one OFDM symbol, the lth
sampled received signal vector r[l] ∈ CNRF ×1 can be written
as
√
(2)
r[l] = ρ · WH H[l] ∗ Fs[l] + WH n[l],
where ρ stands for the average received power containing the
transmit power, transmit antenna gain, receive antenna gain,
and path loss, s[l] ∈ CNRF ×1 is the transmitted signal vector,
and n[l] ∈ CNR ×1 is an NR -dimensional independent and
identically distributed (i.i.d.) complex Gaussian random vector,
n[l] ∼ CN (0NR ×1 , σn2 INR ).
mmWave channel models have been widely studied recently
[14], [15]. Based on the references, a simplified mmWave
CIR matrix H[l] can be expressed as the sum of P outer
products of the array response vectors associated with the
normalized-quantized delay lp = bτp FS c ∈ N0 (the set of
natural numbers contains zero), where τp ∈ R≥0 (the set of
positive real numbers contains zero) is the delay for path p
T
2π
j λ sin(φD,p )(NT −1)∆d
0
, (4)
and the arrival array response vector aA (φA,p ), where φA,p ∼
U(− π2 , π2 ), has a similar form as (4).
III. T IME -D OMAIN A NALOG B EAM S ELECTION
A. Observations for analog beam selection
In order to acquire the observations for the analog beam
selection, we simply assume that all the beam pairs selected
from F and W are trained by a known training sequence.
Hypothetically there is no data transmission and reception
before the transmitter and receiver select the preferable analog
beam pairs. Hence, one can use a training sequence of length
L in one OFDM symbol, {s[0], · · · , s[L − 1]}, to train one
beam pair. The lth sampled scalar of the received signals by
using the beam pair (f̃nf , w̃nw ) can therefore be expressed as
rnw ,nf [l] =
=
√
√
ρ · w̃nHw H[l] ∗ f̃nf s[l] + w̃nHw n[l]
| {z }
znw ,nf [l]
ρ·
w̃nHw H[l]
(5)
∗ f̃nf s[l] + znw ,nf [l],
where nf = 1, · · · , NF , nw = 1, · · · , NW , the combined
noise znw ,nf [l] ∼ CN (0, σn2 ) still has a Gaussian distribution
with mean zero and variance σn2 due to the equal-magnitude
elements of w̃nw .
To implement deconvolution of the received signal and
get the observations for the beam selection, we intend to
decouple the angle- and delay-domain components in rnw ,nf [l]
by replacing the channel matrix H[l] with (3). Consequently,
rnw ,nf [l] can be further written as follows:
(f̂nrf , ŵnrf ) =
rnw ,nf [l]
P
=
√ X H
ρ·
w̃nw aA (φA,p )aH
D (φD,p )f̃nf (cp [l] ∗ s[l]) + znw ,nf [l]
{z
}
p=1 |
,ηp,nw ,nf
=
√
ρ·
P
X
arg max
f̃nf ∈ F \F 0 , w̃nw ∈ W\W 0
gnw ,nf ,
(10)
where nrf = 1, · · · , NRF , gnw ,nf is the energy of the
observations
L−1
X
2
(11)
gnw ,nf =
ynw ,nf [l] ,
l=0
ηp,nw ,nf · (cp [l] ∗ s[l]) + znw ,nf [l],
0
p=1
(6)
where |ηp,nw ,nf | = |w̃nHw aA (φA,p )| · |aH
D (φD,p )f̃nf | is the
multiplication of beamforming gains at the transmitter and
receiver.
Then, we collect L samples in a vector and express the
linear convolution as a multiplication by a Toeplitz matrix S
[16]
T
rnw ,nf = rnw ,nf [0], · · · , rnw ,nf [L − 1]
=
√
ρ·S
P
X
ηp,nw ,nf cp + znw ,nf ,
(7)
p=1
where l = 0, · · · , L − 1, theoretically lead to the optimal solution. Let us write down the corresponding objective function
where
s[0]
..
.
S=
s[L − 1]
···
..
.
···
s[1]
.. ∈ CL×L ,
.
s[0]
T
cp = [cp [0], · · · , cp [L − 1]] ∈ CL×1 ,
T
znw ,nf = znw ,nf [0], · · · , znw ,nf [L − 1] ∈ CL×1 .
=
√
√
ρ·
(8b)
ρ·
=
ynNF
[l]
w ,nf
2
(9)
+ ξnw ,nf [l],
where l = 0, · · · , L − 1. One can design the training sequence
so that ξnw ,nf [l] has a complex Gaussian distribution with
mean zero and a variance of σξ2 .
B. Problem statement
The observations {ynw ,nf [l] ∀nw , nf , l} can be interpreted
as coupling coefficients of the channel and the trained beam
pairs. If the coupling coefficients are acquired in the frequency
domain, our previous work in [10] introduces how to use them
to select the analog beam pairs. Simply speaking, the problem
of frequency-domain analog beam selection can be formulated
as finding the beam pairs that maximize the sum of the power
of the observations across all subcarriers. From Parseval’s
theorem, we know that the objective function is equivalent to
the sum of the power across all samples in the delay domain.
As a result, the delay-domain analog beam selection can be
expressed as the following maximization problem:
P
X
√
(13)
ρ w̃nHw H[lp ]f̃nf
2
,
p=1
(8c)
ξnw ,nf [l]
w̃nHw H[l]f̃nf
L−1
X
l=0
P
X
ηp,nw ,nf cp [l] + S−1 l,: znw ,nf
{z
}
|
p=1
gnNF
,
w ,nf
(8a)
The L observations can therefore be obtained by premultiplying rnw ,nf by S−1 , where det(S) 6= 0, given by
ynw ,nf [l] = S−1 l,: rnw ,nf
=
F = {f̂n , n = 1, · · · , nrf − 1} and W 0 = {ŵn , n =
1, · · · , nrf − 1} are the sets including the selected analog
beamforming vectors from iteration 1 to nrf − 1. The energy estimate gnw ,nf is also the objective function used in
frequency-domain analog beam selection problem [10]. However, in the frequency domain, we do not have the information
that ynw ,nf [l], l = LC , · · · , L − 1, only contain noise signals.
In the beam selection problem stated in (10), the sum of the
power of L noise-free observations
√
(12)
ynNF
[l] , ρ w̃nHw H[l]f̃nf ,
w ,nf
where the second equality follows from that H[l] = 0 when
l ∈
/ {lp , p = 1, · · · , P } and lp , p = 1, · · · , P , are different
to each other. Compared with (11), it is clear that in (13)
only P (rather than L) observations associated with the P
delay indices have to be taken into account. Therefore, our
goal is to reduce the error between the energy estimate gnw ,nf
and its true value gnNF
without an additional computational
w ,nf
overhead.
IV. P ERFORMANCE I MPROVEMENT OF A NALOG B EAM
S ELECTION IN T IME D OMAIN
A. Performance metric
From the discussion in the previous subsection, we know
that there is a higher probability to find the optimal solution
when the error between gnw ,nf and gnNF
approximates to
w ,nf
zero. Therefore, we use the MAE between gnw ,nf and gnNF
w ,nf
as a performance metric to quantify the performance of beam
selection, which is stated in Theorem 1. We consider the MAE
rather than the mean squared error (MSE) due to that fact
are energy signals; it is redundant to
that gnw ,nf and gnNF
w ,nf
calculate the squared error between these two values.
Theorem 1. Given matrix-valued CIRs H[l], l = 0, · · · , L−1,
one has the energy estimates
gn
w ,nf
=
L−1
X
l=0
√
ρ w̃nHw H[l]f̃nf + ξnw ,nf [l]
2
∀nw , nf (14)
these averaged (or refined) observations, the energy estimate
in (11) becomes
and the corresponding true values
=
g NF
n ,n
w
P
X
√
f
ρ w̃nHw H[lp ]f̃nf
2
∀nw , nf .
(15)
p=1
Then the MAE between g n ,n and g NF
is upper bounded
nw ,nf
w
f
by
h
i
MAE(g n ,n ) , E g n ,n − g NF
n
,n
w
w
f
f
(16)
w f
≤ E εnw ,nf + E [ν] ,
where
εnw ,nf ∼ N 0, 2σξ2 g NF
n ,n
w
f
(17)
and
ν ∼ Γ(L, σξ2 ).
(18)
Proof: See Appendix A.
gn0 w ,nf =
LX
C −1
2
yn0 w ,nf [lc ] .
(21)
lc =0
Based on the derivation of Theorem 1, the MAE between
the estimate gn0 w ,nf and its true value gnNF
conditioned on
w ,nf
the same channel realizations, H[l], l = 0, · · · , L − 1, is given
by
i
h
MAE g 0n ,n , E g 0n ,n − g NF
n ,n
w
f
h w f i w f
(22)
0
≤ E εnw ,nf + E [ν 0 ] ,
where ε0nw ,nf has a Gaussian distribution
!
!
σξ2
0
NF
εnw ,nf ∼ N 0, 2
g n ,n ,
w
f
M
(23)
and ν 0 follows a gamma distribution
B. Refine observations by averaging random noise signals
In OFDM systems, the CP length (LC ) is designed to cover
the maximum or root-mean-square (RMS) delay spread, which
means that the number of useful observations in one OFDM
symbol is less than or equal to LC . To improve the quality of
the observations, we use a property of circular convolutions
introduced as follows. First, simply modifying the transmitted
training sequence of length L as M = LLC (assume LLC ∈ N+ )
repeated sequence blocks, where the length of each block is
LC . Such periodic training sequence blocks make the linear
convolution in (6) become
cp [l] ∗ s[l] =
=
L−1
X
cp [n]s[l − n]
n=0
LX
C −1
(19)
cp [n]s[l − n]
n=0
= (cp ∗LC s)[l],
where the second equality follows from that cp [l] = 0 when
l ≥ LC , and ∗LC denotes a circular convolution over the cyclic
group of integers modulo LC . Then, following from (9), we
can use a Toeplitz matrix of small size LC × LC (generated
by one training sequence block) to sequentially implement the
deconvolution of M received periodic blocks. An arithmetic
mean of the M outputs of the deconvolution leads to a result
suffering from less noise effect
0
ν ∼Γ
L σξ2
,
M M
!
.
(24)
Compared with (17), (18), the noise effect caused by ε0nw ,nf
and ν 0 can be effectively reduced when M is large. For
example, one of the use cases in 3GPP 5G NR [13] shows
1
1
that the CP ratio M
≈ 14
.
C. Further refine observations by using knowledge of multipath delay
In the previous subsection, we present how to enhance
the quality of the observations. Without any information of
multipath delay, the LC signals in (20), {yn0 w ,nf [lc ], lc =
0, · · · , LC −1}, with respect to a certain beam pair (f̃nf , w̃nw )
are regarded as useful observations. Nevertheless, only P
sparse observations corresponding to the P CIRs are exactly
useful. Fortunately, we can borrow the idea of the analog
beam selection in (10) to find the multipath delay indices
because the signals {yn0 w ,nf [lc ], ∀nw , nf , lc } are represented
in the discrete delay-angle domain, where lc and (nw , nf )
respectively denote the delay- and angle-domain indices. Accordingly, the multipath delay estimation can be stated as the
following problem: given {yn0 w ,nf [lc ], ∀nw , nf , lc }, one can
calculate the sum of the power of NW NF observations across
all steering angles as
f [lc ] =
NW X
NF
X
2
yn0 w ,nf [lc ] ,
(25)
nw =1 nf =1
yn0 w ,nf [lc ] = ynNF
[l ] +
w ,nf c
M
1 X
ξn ,n [(m − 1)LC + lc ]
M m=1 w f
|
{z
}
0
ξn
[lc ]
w ,n
f
= ynNF
[l ] + ξn0 w ,nf [lc ],
w ,nf c
(20)
where lc = 0, · · · , LC − 1, and the variance of ξn0 w ,nf [lc ] ∼
σ2
CN 0, Mξ is effectively reduced by a factor of M . By using
and solve the constrained maximization problem
ˆlp̂ =
arg max
f [lc ],
lc ∈{0,··· ,LC −1}\L
(
f [lc ] ≥ µ,
s.t.
L = {ˆln , n = 1, · · · , p̂ − 1},
(26)
where p̂ = 1, · · · , P̂ denotes the path index whose received
power across all steering angles is greater than or equal to
60
Ref [10]
Refined obs.
Estimated delay
Explicit delay
50
40
30
MAE
a pre-defined threshold µ, and L is the set containing the
selected path indices from iteration 1 to p̂−1. Here we consider
the sum of the power of NW NF observations in the angle
domain; therefore
2 the threshold can be simply assumed to be
σ
µ = NW NF Mξ . Since mmWave channels are sparse in
nature, the multipath delay indices can be estimated by using
this characteristic to further improve the performance [17].
Due to the page limit, we do not provide more discussion.
According to the estimated delay indices, only P̂ refined
observations are used for the analog beam selection problem
(assume P ≤ P̂ < LC but {ˆlp̂ ∀p̂} does not necessarily
include {lp ∀p}) and the corresponding objective function is
given by
P̂
X
2
(27)
gn00w ,nf =
yn0 w ,nf [ˆlp̂ ] .
24 dB
20
10
11 dB
0
−10
−20
−20
−15
−10
−5
0
SNR (dB)
5
10
15
20
p̂=1
Similarly, conditioned on the same channel realizations, H[l],
l = 0, · · · , L − 1, we have the MAE between gn00w ,nf and its
true value gnNF
upper bounded by
w ,nf
h
i
MAE(g 00n ,n ) , E g 00n ,n − g NF
nw ,nf
w
w
f
f
h
i
NF
00NF
≤ g n ,n − g n ,n + E ε00nw ,nf + E [ν 00 ] ,
w
w
f
f
(28)
where
P̂
X
2
NF
ˆlp̂ ]
y
=
[
(29)
g 00NF
n ,n
n ,n
w
w
f
f
p̂=1
≤ g NF
with equality iff {ˆlp̂ ∀p̂} ⊇ {lp ∀p}.
and g 00NF
nw ,nf
nw ,nf
Furthermore, ε00nw ,nf and ν 00 are given as follows:
!
!
σξ2
00
00NF
εnw ,nf ∼ N 0, 2
g n ,n
(30)
w
f
M
and
σξ2
ν ∼ Γ P̂ ,
M
00
ε0nw ,nf
!
Fig. 2. MAE between energy estimates and their true value, where Ref
uses L = 2048 unrefined observations, and others use LC = 128 refined
observations with and without knowledge of multipath delay.
In the codebooks,
32
n
steering angleo candidates are:
−1 (nf −16)
180◦
, nf = 1, · · · , 32 [18].
π · sin
16
As discussed in Section III-B, the true value of the energy yields the optimal solution of the problem in (10).
Let us denote the indices of the optimal beam pairs as
(n̊w,nrf ,n̊f,nrf ) ∀nrf , and then use the MAE as a performance metric to evaluate the performance of the proposed and reference methods with respect to the beam pairs
(n̊w,nrf ,n̊f,nrf ) ∀nrf . In Fig. 2, the curves labeled as Ref,
Refined obs., and Estimated delay are respectively calculated
by the following equations:
NRF
1 X
Ref =
MAE(gn̊w,nrf ,n̊f,nrf ), (32a)
NRF n =1
rf
.
(31)
0
Compared with
and ν , although the variance of ε00nw ,nf
and the shape parameter of ν 00 become smaller, MAE(g 00n ,n )
w
f
is not necessarily less than MAE(g 0n ,n ) if the difference
w
f
and g 00NF
is too large. It depends on the
between g NF
nw ,nf
nw ,nf
performance of multipath delay estimation.
V. N UMERICAL R ESULTS
The system parameters used in the simulations are listed
below:
Number of antennas
NT = NR = 32
Number of RF chains
NRF = 2
Number of samples per OFDM symbol L = 2048
CP length
LC = 128
Codebook size
NF = NW = 32
Number of paths
P = 10
In addition, the effective noise variance is given
by σξ2 = ρ · 10−γ/10 , where γ (dB) is the SNR.
Refined obs. =
1
NRF
N
RF
X
MAE(gn̊0 w,n
rf
,n̊f,nrf ),
(32b)
nrf =1
NRF
1 X
MAE(gn̊00w,n ,n̊f,n ), (32c)
Estimated delay =
rf
rf
NRF n =1
rf
where the energy estimate in (32a) is equivalent to the sum
of the power of observations across all subcarriers, which is
the objective function of the frequency-domain analog beam
selection problem in [10].
From (16) and (22), we can find the upper bounds of
(32a) and (32b), and they are dominated by the gamma
distributed random variables when the values of shape and
scale parameters are large. As a result, (32a) and (32b) can be
approximated by
(32a) ≈ E [ν] ,
0
(32b) ≈ E [ν ] .
(33a)
(33b)
Delay estimation error rate (%)
60
50
40
30
20
10
0
−20
−15
−10
−5
0
SNR (dB)
5
10
15
20
Fig. 3. Estimation error rate of P delay indices in curve Estimated delay in
Fig. 2.
and therefore the difference in MAE between Ref and Refined
obs. is given by
E [ν]
10 log10
= 10 log10 M 2 = 24.08 dB.
0
E [ν ]
In (32c), if we only use P̂ refined observations associated
with P̂ estimated delay indices (the estimation error rate is
shown in Fig. 3), the MAE can be reduced by 3 dB compared
with Refined obs., see curve Estimated delay. Ideally, if the
set containing P̂ estimated delay indices is equal to {lp ∀p},
following from (28), the corresponding MAE is upper bounded
by
MAE(g 00n ,n |given {lp ∀p} )
h w f
i
≤ E ε00nw ,nf |given {lp ∀p} + E ν 00 |given {lp ∀p}
h
i
= E ε0nw ,nf + E ν 00 |given {lp ∀p}
where
00
ν |given {lp ∀p}
σξ2
∼ Γ P,
M
(34)
,
(35)
NRF
1 X
MAE(g 00n̊
|given {lp ∀p} ).
w,nrf ,n̊f,nrf
NRF n =1
rf
(36)
In the low SNR regime, MAE(g 00n ,n |given {lp ∀p} ) is domw
f
inated by the gamma distributed random variable as well.
Hence, the difference in MAE between Refined obs. and
Explicit delay approximates to
!
= 10 log10
The mmWave channel sparsity in the delay domain is widely
acknowledged as a powerful cue for analog beam selection.
Different to the conventional methods addressing the feature in
the frequency domain, this paper presents a new perspective in
the delay domain and shows that the significant observations
used for the analog beam selection are also sparse. To improve
the quality of the observations, we propose a solution that
transmits the periodic training sequence of length equal to a
CP length. An arithmetic mean can accordingly reduce the
noise variance to refine the observations. Then based on the
refined signals represented in the delay-angle domain, the
sparse significant observations can be simply captured by
finding the maximum term in the sum of the power of the
refined signals across angle.
A. Proof of Theorem 1
!
Explicit delay =
E [ν 0 ]
00
E ν |given {lp ∀p}
VI. C ONCLUSION
VII. A PPENDIX
and the simulation results are shown in curve Explicit delay
calculated by
10 log10
available, we try to find P̂ paths whose sum of the received
power across all steering angles are greater than or equal to
σ2
the pre-defined threshold µ = Mξ NW NF . When SNR <
0 dB, the delay estimation error rate of more than 30% leads
to an MAE reduction of approximately 3 dB, compared with
Refined obs. On the other hand, when SNR ≥ 10 dB, the delay
estimation error rate approximates to zero. However, the gap
between Estimated delay and Explicit delay in Fig. 2 is still
quite obvious, which means that P̂ P and therefore not
only the useful observations but also a large number of noise
signals are reserved. The delay estimation approach can be
further enhanced by, for example, modifying the threshold;
nevertheless, it is beyond the scope of this paper.
L
MP
= 11.07 dB.
Given channel matrices H[l], l = 0, · · · , L − 1, the objective function gnw ,nf in (11) becomes (38), where the first
term g NF
is a constant, the second term εnw ,nf has a
nw ,nf
,
normal distribution with mean zero and variance 2σξ2 g NF
nw ,nf
2 NF
εnw ,nf ∼ N (0, 2σξ g n ,n ), and the third term ν has a gamma
w
f
distribution with the shape parameter L and scale parameter
σξ2 , ν ∼ Γ(L, σξ2 ). Therefore, the MAE between g n ,n and
w
f
g NF
(denoted as MAE(g n ,n )) is given by
nw ,nf
w
f
h
i
MAE(g n ,n ) , E g n ,n − g NF
nw ,nf
w
f
h w f
i
NF
= E g n ,n + εnw ,nf + ν − g NF
nw ,nf
w f
= E εnw ,nf + ν
≤ E εnw ,nf + E [ν] .
(37)
When the SNR increases, (36) is not only dominated by the
gamma distributed random variable, so the difference between
Refined obs. and Explicit delay cannot be simply approximated
by (37).
In Fig. 3, it shows the estimation error rate of P delay
indices in curve Estimated delay in Fig. 2. As mentioned in
Section IV-C, since the exact number P = 10 of paths is not
ACKNOWLEDGMENT
The research leading to these results has received funding
from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 671551 (5GXHaul) and the TUD-NEC project “mmWave Antenna Array
gn
w ,nf
=
L−1
X
√
ρ w̃nHw H[l]f̃nf + ξnw ,nf [l]
2
l=0
=
L−1
X
y NF
[l] + ξnw ,nf [l]
n ,n
w
l=0
=
L−1
X
2
2
NF
+
I
y
R y NF
[l]
+
ξ
[l]
+
ξ
nw ,nf [l]
nw ,nf [l]
n ,n
n ,n
w
l=0
=
L−1
X
2
f
w
f
f
2
2 L−1
X
R y NF
[l] + I y NF
[l] +
[l] R ξnw ,nf [l] + 2I y NF
[l] I ξnw ,nf [l]
2R y NF
n ,n
n ,n
n ,n
n ,n
w
l=0
|
w
f
w
f
{z
l=0
}
,g NF
|
+
R ξnw ,nf [l]
2
+ I ξnw ,nf [l]
{z
(38)
f
}
,εnw ,nf
nw ,nf
L−1
X
w
f
2
l=0
|
{z
}
,ν
= g NF
+ εnw ,nf + ν,
n ,n
w
f
Concept Study”, a cooperation project between Technische
Universität Dresden (TUD), Germany, and NEC, Japan.
R EFERENCES
[1] T. Rappaport, R. Heath, R. Daniels, and J. Murdock, Millimeter Wave
Wireless Communications. Prentice Hall, 2014.
[2] J. Liberti and T. Rappaport, Smart antennas for wireless communications: IS-95 and third generation CDMA applications. Prentice Hall,
1999.
[3] A. Hajimiri, H. Hashemi, A. Natarajan, X. Guan, and A. Komijani,
“Integrated phased array systems in silicon,” Proc. IEEE, vol. 93, no. 9,
pp. 1637–1655, Sep. 2005.
[4] B. van Veen and K. M. Buckley, “Beamforming: A versatile approach
to spatial filtering,” IEEE ASSP Magazine, vol. 5, pp. 4–24, Apr. 1988.
[5] X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-phase-shift-based
RF-baseband codesign for MIMO antenna selection,” IEEE Trans. Signal
Process., vol. 53, no. 11, pp. 4091–4103, Nov. 2005.
[6] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
“Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
[7] C. H. Yu, M. P. Chang, and J. C. Guey, “Beam space selection for
high rank millimeter wave communication,” in IEEE Veh. Technol. Conf.
(VTC Spring), Glasgow, UK, May 2015, pp. 1–5.
[8] S. Han, C. l. I, Z. Xu, and C. Rowell, “Large-scale antenna systems
with hybrid analog and digital beamforming for millimeter wave 5G,”
IEEE Commun. Mag., vol. 53, no. 1, pp. 186–194, Jan. 2015.
[9] H. L. Chiang, W. Rave, T. Kadur, and G. Fettweis, “Hybrid beamforming
based on implicit channel state information for millimeter wave links,”
Sep. 2017. [Online]. Available: https://arxiv.org/abs/1709.07273
[10] ——, “Frequency-selective hybrid beamforming based on implicit CSI
for millimeter wave systems,” in IEEE Int. Conf. on Commun. (ICC),
Kansas City, MO, USA, May 2018.
[11] A. Alkhateeb and R. W. Heath, “Frequency selective hybrid precoding
for limited feedback millimeter wave systems,” IEEE Trans. Commun.,
vol. 64, no. 5, pp. 1801–1818, May 2016.
[12] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge University Press, 2005.
[13] 3GPP TS 38.211 V1.0.0, “Physical channels and modulation (Release
15),” Tech. Rep., 2017.
[14] 3GPP TR 38.900 V14.3.1, “Study on channel model for frequency
spectrum above 6 GHz (Release 14),” Tech. Rep., 2017.
[15] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models for
future wireless communication system design,” IEEE Trans. Commun.,
vol. 63, no. 9, pp. 3029–3056, Sep. 2015.
[16] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Prentice Hall, 1997.
[17] G. Gui, Q. Wan, W. Peng, and F. Adachi, “Sparse multipath channel
estimation using compressive sampling matching pursuit algorithm,” in
IEEE VTS Asia Pacific Wireless Commun. Symp. (APWCS), Kaohsiung,
Taiwan, May 2010, pp. 10–14.
[18] H. L. Chiang, T. Kadur, and G. Fettweis, “Analyses of orthogonal and
non-orthogonal steering vectors at millimeter wave systems,” in IEEE
Int. Symp. on A World of Wireless, Mobile and Multimedia Networks
(WoWMoM), Coimbra, Portugal, Jun. 2016, pp. 1–6.
| 7cs.IT
|
Tests for separability in nonparametric covariance operators
of random surfaces
J. A. D. Aston∗, D. Pigoli and S. Tavakoli†
arXiv:1505.02023v4 [stat.ME] 3 Jun 2016
Statistical Laboratory
Department of Pure Mathematics and Mathematical Statistics
University of Cambridge
Abstract
. The assumption of separability of the covariance operator for a random image or hypersurface can be of substantial use in applications, especially in situations where the accurate
estimation of the full covariance structure is unfeasible, either for computational reasons, or
due to a small sample size. However, inferential tools to verify this assumption are somewhat lacking in high-dimensional or functional data analysis settings, where this assumption
is most relevant. We propose here to test separability by focusing on K-dimensional projections of the difference between the covariance operator and a nonparametric separable
approximation. The subspace we project onto is one generated by the eigenfunctions of
the covariance operator estimated under the separability hypothesis, negating the need to
ever estimate the full non-separable covariance. We show that the rescaled difference of
the sample covariance operator with its separable approximation is asymptotically Gaussian. As a by-product of this result, we derive asymptotically pivotal tests under Gaussian
assumptions, and propose bootstrap methods for approximating the distribution of the test
statistics. We probe the finite sample performance through simulations studies, and present
an application to log-spectrogram images from a phonetic linguistics dataset.
Keywords: Acoustic Phonetic Data, Bootstrap, Dimensional Reduction, Functional Data,
Partial Trace, Sparsity.
1
Introduction
Many applications involve hypersurface data, data that is both functional (as in functional
data analysis, see e.g. Ramsay & Silverman 2005, Ferraty & Vieu 2006, Horváth & Kokoszka
2012a, Wang et al. 2015) and multidimensional. Examples abound and include images from
medical devices such as MRI (Lindquist 2008) or PET (Worsley et al. 1996), spectrograms
∗
Research Supported by EPSRC grant EP/K021672/2.
Address for correspondence: Shahin Tavakoli, Statistical laboratory, Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Wilberforce Road, Cambridge, CB3 0WB, United Kingdom. Email:
[email protected]
†
1
derived from audio signals (Rabiner & Schafer 1978, and as in the application we consider
in Section 4) or geolocalized data (see, e.g., Secchi et al. 2015). In these kinds of problem,
the number of available observations (hypersurfaces) is often small relative to the highdimensional nature of the individual observation, and not usually large enough to estimate
a full multivariate covariance function.
It is usually, therefore, necessary to make some simplifying assumptions about the data
or their covariance structure. If the covariance structure is of interest, such as for PCA or
network modeling, for instance, it is usually assumed to have some kind of lower dimensional structure. Traditionally, this translates into a sparsity assumption: one assumes that
most entries of the covariance matrix or function are zero. Though being relevant for a
number of applications (Tibshirani 2014), this traditional definition of sparsity may not be
appropriate in some cases, such as in imaging, as this can give rise to artefacts in the analysis (for example, holes in an image). In such problems, where the data is multidimensional,
a natural assumption that can be made is that the covariance is separable. This assumption
greatly simplifies both the estimation and the computational cost in dealing with multivariate covariance functions, while still allowing for a positive definite covariance to be specified. In the context of space-time data X(s, t), for instance, where s ∈ [−S, S]d , S > 0,
denotes the location in space, and t ∈ [0, T ], T > 0, is the time index, the assumption of
separability translates into
c(s, t, s0 , t0 ) = c1 (s, s0 )c2 (t, t0 ),
s, s0 ∈ [−S, S]d ; t, t0 ∈ [0, T ],
(1.1)
where c, c1 , and c2 , are respectively the full covariance function, the space covariance
function and the time covariance function. In words, this means that the full covariance
function factorises as a product of the spatial covariance function with the time covariance
function.
The separability assumption (see e.g. Gneiting et al. 2007, Genton 2007) simplifies the
covariance structure of the process and makes it far easier to estimate; in some sense, the
separability assumption results in a estimator of the covariance which has less variance,
at the expense of a possible bias. As an illustrative example, consider that we observe a
discretized version of the process through measurements on a two dimensional grid (without
loss of generality, as the same arguments apply for any dimension greater than 2) being a
q × p matrix (of course, the functional data analysis approach taken here does not assume
that the replications of the process are observed on same grid, nor that they are observed
on a grid). Since we are not assuming a parametric form for the covariance, the degrees of
freedom in the full covariance are qp(qp + 1)/2, while the separability assumption reduces
them to q(q + 1)/2 + p(p + 1)/2. This reflects a dramatic reduction in the dimension of the
problem even for moderate value of q, p, and overcomes both computational and estimation
problems due to the relatively small sample sizes available in applications. For example, for
q = p = 10, we have qp(qp + 1)/2 = 5050 degrees of freedom, however, if the separability
holds, then we have only q(q + 1) + p(p + 1) = 110 degrees of freedom. Of course, this
is only one example, and our approach is not restricted to data on a grid, but this illustrates
the computational savings that such assumptions can possess.
Three related computational classes of problem can be identified. In the first case, the
full covariance structure can be computed and stored. In the second one, it is still possible,
although burdensome, to compute the full covariance matrix but it can not be stored, while
the last class includes problems where even computation of the full covariance is infeasible.
The values of q, p that set the boundaries for these classes depend of course on the available
hardware and they are rapidly changing. At the present time however, for widely available
systems, storage is feasible up to q, p ≈ 100 while computation becomes unfeasible when
2
RI
M
ural
ruct
St
full covariance
separable covariance
Hard−drive (1TB)
1e+00
RAM (16GB)
1e−06
Memory (GB)
1e+06
RI
I
f fM
o
fMR
e
full
Slic
10
50
100
500
5000
p
Figure 1: Memory required to store the full covariance and the separable covariance of p × p
matrix data, as a function of p. Several types of data related to Neuroimaging (structural and
functional Magnetic Resonance Imaging) are used as exemplars of data sizes, as they naturally
have multidimensional structure.
q, p get close to 1000 (see Figure 1). On the contrary, a separable covariance structure can be
usually both computed and stored without effort even for these sizes of problem. We would
like to stress however that the constraints coming from the need for statistical accuracy are
usually tighter. The estimation of the full covariance structure even for q, p = 100 presents
about 5 × 107 unknown parameters, when typical sample sizes are in the order of hundreds
at most. If we are able to assume separability, we can rely on far more accurate estimates.
While the separability assumption can be very useful, and is indeed often implicitly
made in many higher dimensional applications when using isotropic smoothing (Worsley
et al. 1996, Lindquist 2008), very little has been done to develop tools to assess its validity
on a case by case basis. In the classical multivariate setup, some tests for the separability
assumption are available. These have been mainly developed in the field of spatial statistics
(see Lu & Zimmerman 2005, Fuentes 2006, and references therein), where the discussion
of separable covariance functions is well-established, or for applications involving repeated
measures (Mitchell et al. 2005). These methods, however, rely on the estimation of the full
multidimensional covariance structure, which can be troublesome. It is sometimes possible to circumvent this problem by considering a parametric model for the full covariance
structure (Simpson 2010, Simpson et al. 2014, Liu et al. 2014). On the contrary, when the
3
covariance is being non-parametrically specified, as will be the case in this paper, estimation
of the full covariance is at best computationally complex with large estimation errors, and
in many cases simply computationally infeasible. Indeed, we highlight that, while the focus
of this paper is on checking the viability of a separable structure for the covariance, this is
done without any parametric assumption on the form of c1 (s, s0 ) and c2 (t, t0 ), thus allowing for the maximum flexibility. This is opposed to assuming a parametric separable form
with only few unknown parameters, which is usually too restrictive in many applications,
something that has led to separability being rightly criticised and viewed with suspicion in
the spatio-temporal statistics literature (Gneiting 2002, Gneiting et al. 2007). Moreover,
the methods we develop here are aimed to applications typical of functional data, where
replicates from the underlying random process are available. This is different from the
spatio-temporal setting, where usually only one realization of the process is observed. See
also Constantinou et al. (2015) for another approach to test for separability in functional
data.
It is important to notice that a separable covariance structure (or equivalently, a separable correlation structure) is not necessarily connected with the original data being separable.
Furthermore, sums or differences of separable hypersurfaces are not necessarily separable.
On the other hand, the error structure may be separable even if the mean is not. Given that
in many applications of functional data analysis, the estimation of the covariance is the first
step in the analysis, we concentrate on covariance separability. Indeed, covariance separability is an extremely useful assumption as it implies separability of the eigenfunctions,
allowing computationally efficient estimation of the eigenfunctions (and principal components). Even if separability is misspecified, separable eigenfunctions can still form a basis
representation for the data, they simply no longer carry optimal efficiency guarantees in
this case (Aston & Kirch 2012), but can often have near-optimality under the appropriate
assumptions (Chen et al. 2015) .
In this paper, we propose a test to verify if the data at hand are in agreement with a
separability assumption. Our test does not require the estimation of the full covariance
structure, but only the estimation of the separable structure (1.1), thus avoiding both the
computational issues and the diminished accuracy involved in the former. To do this, we
rely on a strategy from Functional Data Analysis (Ramsay & Silverman 2002, 2005, Ferraty
& Vieu 2006, Ramsay et al. 2009, Horváth & Kokoszka 2012b), which consists in projecting
the observations onto a carefully chosen low-dimensional subspace. The key fact for the
success of our approach is that, under the null hypothesis, it is possible to determine this
subspace using only the marginal covariance functions. While the optimal choice for the
dimension of this subspace is a non-trivial problem, some insight can be obtained through
our extensive simulation studies (Section 4.1). Ultimately, the proposed test checks the
separability in the chosen subspace, which will often be the focus of following analyses.
The paper proceeds as follows. In Section 2, we examine the ideas behind separability, propose a separable approximation of a covariance operator, and study the asymptotics
of the difference between the sample covariance operator and its separable approximation.
This difference will be the building block of the testing procedures introduced in Section 3,
and whose distribution we propose to approximate by bootstrap techniques. In Section 4,
we investigate by means of simulation studies the finite sample behaviour of our testing procedures and apply our methods to acoustic phonetic data. A conclusion, given in Section 5,
summarizes the main contributions of this paper. Proofs are collected in appendices A, B,
and C, while implementation details, theoretical background and additional figures can be
found in the appendices E, D and F. All the tests introduced in the paper are available as an
R package covsep (Tavakoli 2016), available on the Comprehensive R Archive Network
4
(CRAN).
For notational simplicity, the proposed method will be described for two dimensional
functional data (e.g. random surfaces), hence a four dimensional covariance structure (i.e.
the covariance of a random surface), but the generalization to higher dimensional cases is
straightforward. The methodology is developed in general for data that take values in a
Hilbert space, but the case of square integrable surfaces—being relevant for the case of
acoustic phonetic data—is used throughout the paper as a demonstration. We recall that
the proposed approach is not restricted to data observed on a regular grid, although for
simplicity of exposition we consider here the case where data are observed densely and a
pre-processing smoothing step allows to consider the smooth surfaces as our observations,
as happens for example the case of the acoustic phonetic data described in Section 4. If data
are observed sparsely, the proposed approach can still be applied but there may be the need
to use more appropriate estimators for the marginal covariance functions (see, e.g. Yao et al.
n.d.) and these need to satisfy the properties described in Section 2.
2 Separable Covariances: definitions, estimators and asymptotic results
While the general idea of the factorization of a multi-dimensional covariance structure as the
product of lower dimensional covariances is easy to describe, the development of a testing
procedure asks for a rigorous mathematical definition and the introduction of some technical
results. In this section we propose a definition of separability for covariance operators, show
how it is possible to estimate a separable version of a covariance operator and evaluate the
difference between the empirical covariance operator and its separable version. Moreover,
we derive some asymptotic results for these estimators. To do this, we first set the problem
in the framework of random elements in Hilbert spaces and their covariance operators.
The benefit in doing this is twofold. First, our results become applicable in more general
settings (e.g. multidimensional functional data, data on multidimensional grids, fixed size
rectangular random matrices) and do not depend on a specific choice of smoothness of
the data (which is implicitly assumed when modeling the data as e.g. square integrable
surfaces). They only rely on the Hilbert space structure of the space in which the data lie.
Second, it highlights the importance of the partial trace operator in the estimation of the
separable covariance structure, and how the properties of the partial trace (Appendix C)
play a crucial role in the asymptotic behavior of the proposed test statistics. However, to
ease explanation, we use the case of the Hilbert space of square integrable surfaces (which
shall be used in our linguistic application, see Section 4) as an illustration of our testing
procedure.
2.1
Notation
Let us first introduce some definitions and notation about operators in a Hilbert space (see
e.g. Gohberg et al. 1990, Kadison & Ringrose 1997a, Ringrose 1971). Let H be a real separable Hilbert space (that is, a Hilbert space with a countable orthonormal basis), whose inner
product and norm are denoted by h·, ·i and k·k, respectively. The space of bounded (linear)
operators on H is denoted by S∞ (H), and its norm is |||T |||∞ = supx6=0 kT xk/kxk. The
space of Hilbert–Schmidt operators
on H is denoted by S2 (H), and is a Hilbert space with
P
the inner-product hS, T iS2 = i≥1 hSei , T ei i and induced norm |||·|||2 , where (ei )i≥1 ⊂ H
is an orthonormal basis of H. The space of trace-class operator on H is denoted by S1 (H),
5
P
and consists of all compact operators T with finite trace-norm, i.e. |||T |||1 = n≥1 sn (T ) <
∞, where sn (T ) ≥ 0 denotes the n-th singular
Pvalue of T . For any trace-class operator
T ∈ S1 (H), we define its trace by Tr(T ) =
i≥1 hT ei , ei i, where (ei )i≥1 ⊂ H is an
orthonormal basis, and the sum is independent of the choice of the orthonormal basis.
If H1 , H2 are real separable Hilbert spaces, we denote by H = H1 ⊗ P
H2 their tensor
N
product Hilbert space, which is obtained by the completion of all finite sums i,j=1 ui ⊗ vj ,
ui ∈ H1 , vj ∈ H2 , under the inner-product hu ⊗ v, z ⊗ wi = hu, zi hv, wi , u, z ∈ H1 , z, w ∈
H2 (see e.g. Kadison & Ringrose 1997a). If C1 ∈ S∞ (H1 ), C2 ∈ S∞ (H2 ), we denote by
e C2 the unique linear operator on H1 ⊗ H2 satisfying
C1 ⊗
e C2 (u ⊗ v) = C1 u ⊗ C2 v, for all u ∈ H1 , v ∈ H2 .
C1 ⊗
(2.1)
e C2
It is a bounded operator on H, with C1 ⊗
= |||C1 |||∞ |||C2 |||∞ . Furthermore, if
∞
e
e C2
C1 ∈ S1 (H1 ) and C2 ∈ S1 (H2 ), then C1 ⊗ C2 ∈ S1 (H1 ⊗ H2 ) and C1 ⊗
=
1
|||C1 |||1 |||C2 |||1 . We denote by Tr1 : S1 (H1 ⊗ H2 ) → S1 (H2 ) the partial trace with respect
e B) = Tr(A)B, for all
to H1 . It is the unique bounded linear operator satisfying Tr1 (A ⊗
A ∈ S1 (H1 ), B ∈ S1 (H2 ). Tr2 : S1 (H1 ⊗ H2 ) → S1 (H1 ) is defined symmetrically (see
Appendix C for more details).
If X ∈ H is a random element with E kXk < ∞, then µ = E X ∈ H, the mean of X,
2
is well defined. Furthermore, if E kXk < ∞, then C = E [(X − µ) ⊗2 (X − µ)] defines
the covariance operator of X, where f ⊗2 g is the operator on H defined by (f ⊗2 g)h =
hh, gi f , for f, g, h ∈ H. The covariance operator C is a trace-class hermitian operator on
H, and encodes all the second-order fluctuations of X around its mean.
Using this nomenclature, we are going to deal with random variables belonging to a
tensor product Hilbert space. This framework encompasses the situation where X is a
d
random surface, for example a space-time indexed data, i.e. X
= X(s, t), s ∈ [−S, S] , t ∈
2
d
[0, T ], S, T > 0, by setting H = L [−S, S] × [0, T ], R , for instance (notice however
that additional smoothness assumptions on X would lead to assume that X belongs to
some other Hilbert space). In this case, the covariance operator of the random element
X ∈ L2 [−S, S]d × [0, T ], R satisfies
Z
Z T
Cf (s, t) =
c(s, t, s0 , t0 )f (s0 , t0 )ds0 dt0 , s ∈ [−S, S]d , t ∈ [0, T ],
[−S,S]d
0
f ∈ L2 [−S, S]d × [0, T ], R , where c(s, t, s0 , t0 ) = cov [X(s, t), X(s0 , t0 )] is the covariance function of X. The space of square integrable surfaces,
L2 [−S, S]d × [0, T ], R ,
is a tensor product Hilbert space because it can can be identified with
L2 [−S, S]d , R ⊗ L2 ([0, T ], R).
2.2
Separability
We recall now that we want to define separability so that the covariance function can
be
written as c(s, t, s0 , t0 ) = c1 (s, s0 )c2 (t, t0 ), for some c1 ∈ L2 [−S, S]d × [−S, S]d , R and
c2 ∈ L2 ([0, T ] × [0, T ], R). This can be extended to the covariance operator of a random
elements X ∈ H = H1 ⊗ H2 , where H1 , H2 are arbitrary separable real Hilbert spaces.
We call its covariance operator C separable if
e C2 ,
C = C1 ⊗
6
(2.2)
e C2
where C1 , respectively C2 , are trace-class operators on H1 , respectively on H2 , and C1 ⊗
e C2 =
is defined in (2.1). Notice that though the decomposition (2.2) is not unique, since C1 ⊗
e (α−1 C2 ) for any α 6= 0, this will not cause any problem at a later stage since we
(αC1 ) ⊗
e C2 , which is identifiable.
will ultimately be dealing with the product C1 ⊗
e C2 are known. If X1 , . . . , XN i.i.d.
In practice, neither C nor C1 ⊗
∼ X and (2.2) holds,
b
the sample covariance operator CN is not necessarily separable in finite samples. However,
we can estimate a separable approximation of it by
b1,N ⊗
b2,N ,
e C
C
b1,N = Tr2 (C
bN )/
where C
(2.3) is that
(2.3)
q
q
bN ), C
b2,N = Tr1 (C
bN )/ Tr(C
bN ). The intuition behind
Tr(C
e Tr1 (T ),
Tr(T )T = Tr2 (T ) ⊗
e B, A ∈ S1 (H1 ), B ∈ S1 (H2 ), with
for all T ∈ S1 (H1 ⊗ H2 ) of the form T = A ⊗
Tr(T ) 6= 0.
Let us consider again what this means when X is a random element of L2 [−S, S]d × [0, T ], R —
i.e. the realization of a space-time process—of which we observe N i.i.d. replications
X1 , . . . , XN ∼ X. In this case, Proposition C.2 tells us that if the covariance function
b1,N and C
b2,N are defined by
is continuous, the operators C
Z
b1,N f (s) =
C
b
c1,N (s, s0 )f (s)ds, f ∈ L2 [−S, S]d , R ,
[−S,S]d
T
Z
b2,N g(t) =
C
0
b
c2,N (t, t0 )g(t)dt,
g ∈ L2 ([0, T ], R),
where
b
c1,N (s, s0 ) = qR
c̃1,N (s, s0 )
,
c̃ (s, s)ds
[−S,S]d 1,N
c̃2,N (t, t0 )
b
c2,N (t, t0 ) = qR
,
T
c̃
(t,
t)dt
0 2,N
and
Z T
N Z
1 X T
0
0
Xi (s, t) − X(s, t) Xi (s , t) − X(s , t) dt =
c̃1,N (s, s ) =
cN (s, t, s0 , t)dt,
N i=1 0
0
Z
N Z
1 X
0
0
0
c̃2,N (t, t ) =
Xi (s, t) − X(s, t) Xi (s, t ) − X(s, t ) ds =
cN (s, t, s, t0 )ds,
N i=1 [−S,S]d
[−S,S]d
0
X(s, t) =
N
N
1 X
1 X
Xi (s, t), b
cN (s, t, s0 , t0 ) =
Xi (s, t) − X(s, t) Xi (s0 , t0 ) − X(s0 , t0 ) ,
N i=1
N i=1
for all s, s0 ∈ [−S, S]d , t, t0 ∈ [0, T ]. The assumption of separability here means that the
estimated covariance is written as a product of a purely spatial component and a purely
temporal component, thus making both modeling and estimation easier in many practical
applications.
We stress again that we aim to develop a test statistic that solely relies on the estimation
of the separable components C1 and C2 , and does not require the estimation of the full
e C2 , the
covariance C. We can expect that under the null hypothesis H0 : C = C1 ⊗
bN − C
b1,N ⊗
b2,N between the sample covariance operator and its
e C
difference DN = C
7
separable approximation should take small values. We propose therefore to construct our
test statistic by projecting DN onto the first eigenfunctions of C, since P
these encode the
directions P
along which X has the most variability. If we denote by C1 = i≥1 λi ui ⊗2 ui
and C2 = j≥1 γj vj ⊗2 vj the Mercer decompositions of C1 and C2 , we have
e C2 =
C = C1 ⊗
X
λi γj (ui ⊗ vj )
i,j≥1
O
2
(ui ⊗ vj ),
where we have used results from Appendix D.1. The eigenfunctions of C are therefore of
the form ur ⊗ vs , where ur ∈ H1 is the r-th eigenfunction of C1 and vs ∈ H2 is the s-th
eigenfunction of C2 . We define a test statistic based on the projection
√
(2.4)
TN (r, s) = N hDN (ûr ⊗ v̂s ), ûr ⊗ v̂s i , r, s ≥ 1 fixed,
where we have replaced the eigenfunctions of C1 and C2 by their empirical counterpart,
i.e.
b1,N , respectively C
b2,N , are given by C
b1,N = P
the Mercer decompositions of C
λ̂
û
⊗
ûi ,
i
i
i≥1
P
b
b
respectively C2,N =
γ̂j v̂j ⊗ v̂j . Notice that though the eigenfunctions of C1,N and
j≥1
b2,N are defined up to a multiplicative constant α = ±1, our test statistic is well deC
fined. The key fact for the practical implementation of the method is that TN (r, s) can
be computed without
the need to estimate (and store in memory)
the operator DN , since
√ 1 PN
2
bj − λ̂r γ̂s . In particular, the computaTN (r, s) = N N k=1 Xk − X N , vbi ⊗ u
tion of TN (r, s) does not require an estimation of the full covariance operator C, but only
the estimation of the marginal covariance operators C1 and C2 , and their eigenstructure.
2.3
Asymptotics
The theoretical justification for using a projection of DN to define a test procedure is that,
p
e C2 , we have |||DN |||1 −→
under the null hypothesis H0 : C = C1 ⊗
0 as N → ∞,
i.e. DN convergences in probability
to
zero
with
respect
to
the
trace
norm.
In fact, we will
√
show in Theorem 2.3 that N DN is asymptotically Gaussian under the following regularity
conditions:
Condition 2.1. X is a random element of the real Hilbert space H satisfying
∞
X
h
i1/4
4
E hX, ej i
< ∞,
(2.5)
j=1
for some orthonormal basis (ej )j≥1 of H.
The implications of this condition can be better understood in light of the following
remark.
remark 2.2 (Mas (2006)).
4
1. Condition 2.1 implies that E kXk < ∞.
√
4
2. If E kXk < ∞, then N (CN − C) converges in distribution to a Gaussian random
element of S2 (H) for N → ∞, with respect to the Hilbert–Schmidt
topology. Under
√
Condition 2.1, a stronger form of convergence holds: N (CN − C) converges in
distribution to a random element of S1 (H) for N → ∞, with respect to the tracenorm topology.
8
3. If X is Gaussian and (λj )j≥1 is the sequence
p of eigenvalues of its covariance operaP
tor, a sufficient condition for (2.5) is j≥1 λj < ∞.
Condition 2.1 requires fourth order moments rather than the usual second order moments often assumed in functional data, as in this case we are interested in investigating the
variation of the second moment,
and hence require assumptions on the fourth order strucbN = 1 PN (Xi − X) ⊗2 (Xi − X), where X = N −1 PN Xk . The
ture. Recall that C
j=1
k=1
N
bN )
e Tr1 (C
bN − Tr2 (CbN ) ⊗
following result establishes the asymptotic distribution of DN = C
:
bN )
Tr(C
theorem 2.3. Let H1 , H2 be separable real Hilbert spaces, X1 , . . . , XN ∼ X be i.i.d.
random elements on H1 ⊗ H2 with covariance operator C, and Tr C 6= 0.
If X satisfies Condition 2.1 (with H = H1 ⊗ H2 ), then, under the null hypothesis
e C2 ,
H0 : C = C1 ⊗
C1 ∈ S1 (H1 ), C2 ∈ S1 (H2 ),
we have
√
N
b
b
e
bN − Tr2 (CN ) ⊗ Tr1 (CN )
C
b
Tr(CN )
!
d
−→ Z,
as N → ∞,
(2.6)
where Z is a Gaussian random element of S1 (H1 ⊗ H2 ) with mean zero, whose covariance
structure is given in Lemma A.1.
√
bN − C) to converge in distribution
Condition 2.1 is used here because we need N (C
in the topology of the space S1 (H1 ⊗ H2 ); it could be replaced by any (weaker) condition
ensuring such convergence. The assumption Tr C 6= 0 is equivalent to assuming that X is
not almost surely constant.
e C2 =
Proof of Theorem 2.3. First, notice that C = C1 ⊗
Therefore, using the linearity of the partial trace, we get
!
bN ) ⊗
bN )
√
√
e Tr1 (C
Tr
(
C
2
bN −
bN − C)
= N (C
N C
b
Tr(CN )
e Tr1 (C)
Tr2 (C) ⊗
Tr(C)
under H0 .
√
bN ) ⊗
bN )
e Tr1 (C
e Tr1 (C) Tr2 (C
Tr2 (C) ⊗
+ N
+
bN )
Tr(C)
Tr(C
√
bN − C) C
Tr
N (C
√
bN − C) +
= N (C
bN )
Tr(C
√
bN − C) ⊗
e Tr1 (C)
Tr2
N (C
−
bN )
Tr(C
√
bN ) ⊗
bN − C)
e Tr1
Tr2 (C
N (C
−
.
bN )
Tr(C
√
bN − C), C
bN ,
N (C
=Ψ
where
Ψ(T, S) = T +
e Tr1 (C) Tr2 (S) ⊗
e Tr1 (T )
Tr(T )C Tr2 (T ) ⊗
−
−
;
Tr(S)
Tr(S)
Tr(S)
9
T, S ∈ S1 (H1 ⊗ H2 ).
!
Notice that the function Ψ : S1 (H1 ⊗ H2 ) × S1 (H1 ⊗ H2 ) → S1 (H1 ⊗ H2 ) is continuous at (T, S) ∈ S1 (H1 ⊗ H2 ) × S1√(H1 ⊗ H2 ) in each coordinate, with respect to the trace
bN − C) converges in distribution—under Condinorm, provided Tr(S) 6= 0. Since N (C
tion 2.1—to a Gaussian random element Y ∈ S1 (H1 ⊗ H2 ), with respect to the trace norm
√
bN − C), C
bN converges in distribution to
|||·||| (see Mas 2006, Proposition 5), Ψ
N (C
1
Ψ(Y, C) = Y +
e Tr1 (C) Tr2 (C) ⊗
e Tr1 (Y )
Tr(Y )C
Tr2 (Y ) ⊗
−
−
Tr(C)
Tr(C)
Tr(C)
(2.7)
by the continuous mapping theorem in metric spaces (Billingsley 1999). Ψ(Y, C) is Gaussian because each of the summands of (2.7) are Gaussian. Indeed, the first and second
summands are obviously Gaussian, and the last two summands are Gaussian by Proposition C.3, and Proposition D.2.
We can now give the asymptotic distribution of TN (r, s), defined in (2.4) as the (scaled)
projection of DN in a direction given by the tensor product of the empirical eigenfunctions
ûr and v̂s . The proof of the following result is given in Appendix B.
corollary 2.4. Under the conditions of Theorem 2.3, if I ⊂ {(i, j) : i, j ≥ 1} is a finite set
of indices such that λr γs > 0 for each (r, s) ∈ I, then
d
(TN (r, s))(r,s)∈I −→ N (0, Σ),
as N → ∞.
This means that the vector (TN (r, s))(r,s)∈I is asymptotically multivariate Gaussian, with
asymptotic variance-covariance matrix Σ = Σ(r,s),(r0 ,s0 ) (r,s),(r0 ,s0 )∈I is given by
Σ(r,s),(r0 ,s0 ) = β̃rsr0 s0 +
+
αrs β̃r0 s0 ·· + αr0 s β̃r··s0 + αrs0 β̃r0 ··s + αr0 s0 β̃rs··
Tr(C)
αrs αr0 s0 β̃····
λr λr0 β̃·s·s0
γs γs0 β̃r·r0 ·
+
+
Tr(C)2
Tr(C1 )2
Tr(C2 )2
λr β̃r0 s0 ·s + λr0 β̃rs·s0
γs β̃r0 s0 r· + γs0 β̃rsr0 ·
−
Tr(C1 )
Tr(C2 )
!
γs0 β̃r0 ···
λr0 β̃·s0 ··
αrs
+
−
Tr(C) Tr(C2 )
Tr(C1 )
!
αr0 s0
γs β̃r···
λr β̃·s··
−
+
Tr(C) Tr(C2 ) Tr(C1 )
−
where µ = E [X], αrs = λr γs ,
i
h
2
2
β̃ijkl = E hX − µ, ui ⊗ vj i hX − µ, uk ⊗ vl i ,
and ‘ ·’ denotes summation over the corresponding index, i.e. β̃r·jk =
P
i≥1
β̃rijk .
We note that the asymptotic variance-covariance of (TN (r, s))(r,s)∈I depends on the
second and fourth order moments of X, which is not surprising since it is based on estimators of the covariance of X. Under the additional assumption that X is Gaussian, the
asymptotic variance-covariance of (TN (r, s))(r,s)∈I can be entirely expressed in terms of
the covariance operator C. The proof of the following result is given in Appendix B.
10
corollary 2.5. Assume the conditions of Theorem 2.3 hold, and that X is Gaussian. If
I ⊂ {(i, j) : i, j ≥ 1} is a finite set of indices such that λr γs > 0 for each (r, s) ∈ I, then
d
(TN (r, s))(r,s)∈I −→ N (0, Σ),
as N → ∞.
where
Σ(r,s),(r0 ,s0 ) =
2λr λr0 γs γs0
2
2
0 Tr(C1 ) + |||C1 ||| − (λr + λr 0 ) Tr(C1 )
δ
rr
2
Tr(C)2
2
× δss0 Tr(C2 )2 + |||C2 |||2 − (γs + γs0 ) Tr(C2 ) ,
and δij = 1 if i = j, and zero otherwise. In particular, notice that Σ itself is separable.
It will be seen in the next section that even in the case where we use a bootstrap test,
knowledge of the asymptotic distribution can be very useful to establish a pivotal bootstrap
test, which will be seen to have very good performance in simulation.
3
Separability Tests and Bootstrap Approximations
In this section we use the estimation procedures and the theoretical results presented in
e C2 , against the alternative that C cannot be
Section 2 to develop a test for H0 : C = C1 ⊗
written as a tensor product.
First, it is straightforward to define a testing procedure when X is Gaussian. Indeed, if
we let
!2
N
1 X
2
2
Xk − X, u
br ⊗ vbs − λ̂r γ̂s ,
(3.1)
GN (r, s) = TN (r, s) = N
N
k=1
and
−1
b1,N )2 Tr(C
b2,N )2
σ̂ 2 (r, s) = Tr(C
2λ̂2r γ̂s2
2
b1,N
b1,N )
b1,N )2 + C
− 2λ̂r Tr(C
× Tr(C
2
2
b2,N )2 + C
b2,N
b2,N ) , (3.2)
× Tr(C
− 2γ̂s Tr(C
2
then σ̂ −2 (r, s)GN (r, s) is asymptotically χ21 distributed, and {G2N (r, s) > σ̂ 2 (r, s)χ21 (1 −
α)}, where χ21 (1 − α) is the 1 − α quantile of the χ21 distribution, would be a rejection
region of level approximately α, for α ∈ [0, 1] and N large.
Apart for the distributional assumption for X to be Gaussian, this approach suffers also
the important limitation that it only tests the separability assumption along one eigendirection. It is possible to extend this approach to take into account several eigendirections. For
simplicity, let us consider the case I = {1, . . . , p} × {1, . . . , q}. Denote by TN (I) the
p × q matrix with entries (TN (I))ij = TN (i, j), and let
2
e N (I) = Σ̂−1/2 TN (I)Σ̂−T/2 ,
G
L,I
R,I
(3.3)
where |A|2 denotes the sum of squared entries of a matrix A, A−1/2 denotes the inverse of
(any) square root of the matrix A, A−T/2 = (A−1/2 )T , and the matrices Σ̂L,I , respectively
11
Σ̂R,I , which are estimators of the row, resp. column, asymptotic covariances of TN (I),
e N (I) is asymptotically χ2pq distributed. In the simulaare defined in Appendix E. Then G
tion studies (Section P
4.1), we consider also an approximate version of this Studentized test
2
2
e a (I) =
statistics, G
N
(r,s)∈I TN (r, s)/σ̂ (r, s), which are obtained simply by standardiz2
ing marginally each entry TN (r, s), thus ignoring the dependence between the test statistics
associated with different directions. In order to assess the advantage of Studentization, we
also consider the non-Studentized test statistic
X
GN (I) =
TN2 (r, s).
(r,s)∈I
e N , TN , σ̂ 2 (r, s), Σ̂L,I and Σ̂R,I are described in Appendix E.
The computation details for G
remark 3.1. Notice that the only test whose asymptotic distribution is parameter free is
e N (I), under Gaussian assumptions. It would in principle be possible to construct an
G
analogous test without the Gaussian assumptions (using Corollary 2.4). However, due to
the large number of parameters that would need to be estimated in this case, we expect the
asymptotics to come into force only for very large sample sizes (this is actually the case
under Gaussian assumptions, specially if the set of projections I is large, as can be seen
in Figure 10). For these reasons, we shall investigate bootstrap approximations to the test
statistics.
The choice of the number of eigenfunctions K (the number of elements in I) onto which
one should project is not trivial. The popular choice of including enough eigenfunctions to
explain a fixed percentage of the variability in the dataset may seem inappropriate in this
context, because under the alternative hypothesis there is no guarantee that the separable
eigenfunctions explain that percentage of variation.
For fixed K, notice that the test at least guarantees the separability in the subspace
of the respective K eigenfunctions, which is where the following analysis will be often
focused. On the other hand, since our test statistic looks at an estimator of the non-separable
component
e Tr1 (C)
Tr2 (C) ⊗
,
D=C−
Tr(C)
restricted to the subspace spanned by the eigenfunctions ur ⊗ vs , the test takes small values
(and thus lacks power) when
e (vs ⊗2 vs ) = 0,
hD(ur ⊗ vs ), ur ⊗ vs i = D, (ur ⊗2 ur ) ⊗
S
2
that is when the non-separable component D is orthogonal to
e (vs ⊗2 vs )
(ur ⊗2 ur ) ⊗
with respect to the Hilbert–Schmidt inner product. Thus the proposed test statistic GN (I)
is powerful when D is not orthogonal to the subspace
e (vj ⊗2 vj ), (i, j) ∈ I},
VI = span{(ui ⊗2 ui ) ⊗
and in general the power of the test for finite sample size depends on the properly rescaled
norm of the projection of D onto VI .
In practice, it seems reasonable to use the subset of eigenfunctions that it is possible to
estimate accurately given the available sample sizes. The accuracy of the estimates for the
12
eigendirections can be in turn evaluated with bootstrap methods, see e.g. Hall & HosseiniNasab (2006) for the case of functional data. A good strategy may also be to consider
more than one subset of eigenfunctions and then summarize the response obtained from the
different tests using a Bonferroni correction.
e C2,N ),
As an alternative to these test statistics (based on projections of DN = CN −C1,N ⊗
2
we consider also a test based on the squared Hilbert–Schmidt norm of DN , i.e. |||DN |||2 ,
whose null distribution will be approximated by a bootstrap procedure (this test will be referred to as Hilbert–Schmidt test hereafter). Though it seems that such tests would require
one to store the full sample covariance of the data (which could be infeasible), we describe
in Appendix E a way of circumventing such problem, although the computation of each entry of the full covariance is still needed. Therefore this could be used only for applications
in which the dimension of the discretized covariance matrix is not too large.
In the following, we propose also a bootstrap approach to approximate the distribution
e N (I), G
e a (I) and GN (I), with the aim to improve the finite sample
of the test statistics G
N
properties of the procedure and to relax the distributional assumption on X.
3.1
Parametric Bootstrap
If we assume we know the distribution of X up to its mean µ and its covariance operator
e N (I), G
e a (I), GN (I)
C, i.e. X ∼ F (µ; C), we can approximate the distribution of G
N
2
and |||DN |||2 under the separability hypothesis via a parametric bootstrap procedure. Since
e C2,N , respectively X, is an estimate of C, respectively µ, we simulate B bootstrap
C1,N ⊗
b i.i.d.
e C2,N , for b = 1, . . . , B. For each sample, we
∼ F X, C1,N ⊗
samples X1b , . . . , XN
e N (I), HN = G
e a (I)
compute H b = HN (X b , . . . , X b ), where HN = GN (I), HN = G
N
1
2
N
N
respectively HN = |||DN |||2 , if we wish to use the non-Studentized projection test, the
Studentized projection test, the approximated Studentized version or the Hilbert–Schmidt
test, respectively. A formal description of the algorithm for obtaining the p-value of the test
based on the statistic HN = HN (X1 , . . . , XN ) with the parametric bootstrap can be found
in Appendix E, along with the details for the computation of HN . We highlight that this
procedure does not ask for the estimation of the full covariance structure, but only of its
separable approximation, with the exception of the Hilbert–Schmidt test (and even in this
case, it is possible to avoid the storage of the full covariance).
3.2
Empirical Bootstrap
In many applications it is not possible to assume a distribution for the random element X,
and a non-parametric approach is therefore needed. In this setting, we can use the empirical
e N (I) or |||DN |||2 under
bootstrap to estimate the distribution of the test statistic GN (I), G
2
e C2 . Let HN denote the test statistic whose distribution
the null hypothesis H0 : C = C1 ⊗
is of interest. Based on an i.i.d. sample X1 , . . . , XN ∼ X, we wish to approximate the
∗
distribution of HN with the distribution of some test statistic ∆∗N = ∆N (X1∗ , . . . , XN
),
∗
∗
where X1 , . . . , XN is obtained by drawing with replacement from the set {X1 , . . . , XN }.
∗
Though it is tempting to use ∆∗N = HN (X1∗ , . . . , XN
), this is not an appropriate choice.
Indeed, let us look at the case HN = GN (i, j). Notice that the true covariance of X is
C=
e Tr1 (C)
Tr2 (C) ⊗
+ D,
Tr(C)
13
(3.4)
where D is a possibly non-zero operator, and that
∗
∗
∗
∗
∗
e C2,N
HN
= GN (i, j|X1∗ , . . . , XN
) = N (CN
− C1,N
⊗
)(ûi ⊗ v̂j ), ûi ⊗ v̂j
2
,
∗
∗
∗
∗
∗
∗
where CN
= CN (X1∗ , . . . , XN
), C1,N
= C1,N (X1∗ , . . . , XN
), and C2,N
= C2,N (X1∗ , . . . , XN
).
∗
∗
∗
∗
e C2,N ) ≈ (CN − C1,N ⊗
e C2,N ) ≈ D, the statistic HN would apSince (CN − C1,N ⊗
proximate the distribution of HN under the hypothesis (3.4), which is not what we want.
∗
We therefore propose the following choices of ∆∗N = ∆n (X1∗ , . . . , XN
; X1 , . . . , XN ), depending on the choice of HN :
P
2
1. HN = GN (I), ∆∗N = (i,j)∈I (TN∗ (i, j) − TN (i, j)) .
e N (I), ∆∗ =
2. HN = G
N
Σ̂∗L,I
−1/2
−T/2 2
, where
(T∗N (I) − TN (I)) Σ̂∗R,I
∗
∗
Σ̂∗L,I = Σ̂L,I (X1∗ , . . . , XN
), and Σ̂∗R,I = Σ̂R,I (X1∗ , . . . , XN
). are the row, resp.
column, covariances estimated from the bootstrap sample.
2
∗
2
2
e a (I), ∆∗ = P
3. HN = G
N
N
(i,j)∈I (TN (i, j) − TN (i, j)) /σ̂∗ (i, j), where σ̂∗ (i, j) =
∗
σ̂ 2 (i, j|X1∗ , . . . , XN
).
2
2
∗
∗
∗
4. HN = |||DN |||2 , ∆∗N = |||DN
− DN |||2 , where DN
= DN (X1∗ , . . . , XN
).
The algorithm to approximate the p-value of HN by the empirical bootstrap is described
in detail. The basic idea consists of generating B bootstrap samples, computing ∆∗N for
each bootstrap sample and looking at the proportion of bootstrap samples for which ∆∗N is
larger than the test statistic HN computed from the original sample.
4
Empirical demonstrations of the method
4.1
Simulation studies
We investigated the finite sample behavior of our testing procedures through an intensive
reproducible simulation study (its running time is equivalent to approximately 401 days on
a single CPU computer). We compared the test based on the asymptotic distribution of
e N (I), G
e a (I), and |||DN |||2 , with the p-values
(3.1), as well as the tests based on GN (I), G
N
2
obtained via the parametric bootstrap or the empirical bootstrap.
We generated discretized functional data X1 , . . . , XN ∈ R32×7 under two scenarios.
In the first scenario (Gaussian scenario), the data were generated from a multivariate Gaussian distribution N (0, C). In the second scenario (Non-Gaussian scenario), the data were
generated from a multivariate t distribution with 6 degrees of freedom and non centrality
parameter equal to zero. In the Gaussian scenario, we set C = C(γ) , where
C(γ) (i1 , j1 , i2 , j2 ) = (1 − γ)c1 (i1 , i2 )c2 (j1 , j2 )
(i1 − i2 )2
1
exp −
, (4.1)
+γ
(j1 − j2 )2 + 1
(j1 − j2 )2 + 1
γ ∈ [0, 1]; i1 , i2 = 1, . . . , 32; j1 , j2 = 1, . . . , 7. The covariances c1 and c2 used in the
simulations can be seen in Figure 2. For the Non-Gaussian scenario, we chose a multivariate t distribution with the correlation structure implied by C(γ) , γ ∈ [0, 1]; i1 , i2 =
1, . . . , 32; j1 , j2 = 1, . . . , 7. The parameter γ ∈ [0, 1] controls the departure from the separability of the covariance C(γ) : γ = 0 yields a separable covariance, whereas γ = 1 yields
14
7
a complete non-separable covariance structure (Cressie & Huang 1999). All the simulations
have been performed using the R package covsep (Tavakoli 2016), available on CRAN,
which implements the tests presented in the paper.
5
15
3
5
1.8
1.6
1.4
1.2
1.0
0.8
0.6
1
5
15
25
0.7
0.6
0.5
0.4
0.3
0.2
0.1
25
1 2 3 4 5 6 7
Figure 2: Covariance functions c1 (left) and c2 (right) used in the simulation study.
For each value of γ ∈ {0, 0.01, 0.02, . . . , 0.1} and N ∈ {10, 25, 50, 100}, we performed 1000 replications for each of the above simulations, and estimated the power of the
tests based on the asymptotic distribution of (3.1).
e N (1, 1), GN (1, 1), and |||DN ||| , with
We first also estimated the power of the tests G
2
distributions approximated by a Gaussian parametric bootstrap, and the empirical bootstrap,
with B = 1000. The results are shown in Figure 3. In the Gaussian scenario (Figure 3,
panel (a)), the empirical size of all the proposed tests gets closer to the nominal level (5%)
as N increases (see also Table 2). Nevertheless, the non-Studentized tests GN (1, 1), for
both parametric and empirical bootstrap, seem to have a slower convergence with respect
to the Studentized version, and even for N = 100 the level of these tests appear still higher
than the nominal one (and a CLT-based 95% confidence interval for the true level does not
contain the nominal level in both cases). The empirical bootstrap version of the Hilbert–
Schmidt test also fails to respect the nominal level at N = 100, but its parametric bootstrap
counterpart respects the level, even for N = 25. For N = 25, 50, 100, the most powerful tests (amongst those who respect the nominal level) are the parametric and empirical
e N (1, 1), and they seem to have equal power. The power of the
bootstrap versions of G
Hilbert–Schmidt test based on the parametric bootstrap seems to be competitive only for
N = 100 and γ = 0.1, and is much lower for other values of the parameters. The test based
on the asymptotic distribution does not respect the nominal level for small N but it does
when N increases. Indeed, the convergence to the nominal level seems remarkably fast and
its power is comparable with those of the parametric and empirical bootstrap tests based on
e N (1, 1). Despite being based on an asymptotic result, its performance is quite good also
G
in finite samples, and it is less computationally demanding than the bootstrap tests.
In the non-Gaussian scenario (Figure 3, panel (b)), only the empirical bootstrap version
e N (1, 1) and of the Hilbert–Schmidt test seem to respect the level for N = 10 (see also
of G
Table 2). Amongst these tests, the most powerful one is clearly the empirical bootstrap test
e N (1, 1). Although the Gaussian parametric bootstrap test has higher empirical
based on G
power, it does not have the correct level (as expected) and thus cannot be used in a none N (1, 1)
Gaussian scenario. Notice also that the test based on the asymptotic distribution of G
(under Gaussian assumptions) does not respects the level of the test even for N = 100.
The same holds for the Gaussian bootstrap version of the Hilbert–Schmidt test. Finally,
15
1.0
0.8
N=25
0.6
●
0.4
●
0.2
●
●
●
●
●●
●
●
●
●
●●
●
0.8
0.6
●
●
0.4
●
●
●
●
●
●
●
●
●
●
●●
N=100
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.00
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.2
●
●
●
●
●
●
●
●
N=50
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
0.0
●
●
●
●
Empirical power
●
●
1.00.0
Empirical power
N=10
●
●
●
●
●
●
●
●
●●
0.02
0.04
0.06
0.08
●
0.10 0.00
0.02
0.04
γ
0.06
0.08
0.10
γ
1.0
(a) Gaussian scenario
N=10
0.8
●
0.6
●
●
●
●
●
●
●
●
●
●
●
●
●
1.00.0
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
N=50
●
●
●
0.8
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
N=100
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
0.4
0.6
●
●
●
●
●
●
●
0.2
●
●
●
●
●
0.0
Empirical power
N=25
●
●
0.4
●
●
●
0.2
Empirical power
●
●
0.00
0.02
0.04
0.06
0.08
0.10 0.00
0.02
γ
0.04
0.06
0.08
0.10
γ
(b) Non Gaussian scenario
Figure 3: Empirical power of the testing procedures in the Gaussian scenario (panel (a)) and
non-Gaussian scenario (panel (b)), for N = 16
10, 25, 50, 100 and I = I1 . The results shown
correspond to the test (3.1) based on its asymptotic distribution (····+····), the Gaussian parametric
bootstrap test (solid line with empty circles) and its studentized version (dash-dotted line with
empty circles), the empirical parametric bootstrap test (—4—) and its Studentized version (–
–4– –), the Gaussian parametric Hilbert–Schmidt test (dash-dotted line with filled circles) and
the empirical Hilbert–Schmidt test (dash-dotted line with filled triangles). The horizontal dotted
line indicates the nominal level (5%) of the test. Note that the points have been horizontally
jittered for better visibility.
though the empirical bootstrap version of the Hilbert–Schmidt test respects the level for
N = 10, 25, 50, 100, it has virtually no power for N = 10, 25, 50, and has very low power
for N = 100 (at most 0.3 for γ = 0.1).
As mentioned previously, there is no guarantee that a violation in the separability of
C is mostly reflected in the first separable eigensubspace. Therefore, we consider also a
larger subspace for the test. Figure 9 shows the empirical power for the asymptotic test,
e N (I2 ), as well as
the parametric and empirical bootstrap tests based on the test statistic G
a
e
parametric and bootstrap tests based on the test statistics GN (I), GN (I2 ) where I2 =
{(i, j) : i, j = 1, 2}. In the Gaussian scenario, the asymptotic test is much slower in
e N (1, 1). For
converging to the correct level compared to its univariate version based on G
larger N its power is comparable to that of the parametric and empirical bootstrap based on
e N (I2 ), which in addition respects the nominal level, even for
the Studentized test statistics G
e a (I2 )
N = 10. It is interesting to note that the approximated Studentized bootstrap tests G
N
have a performance which is better than the non Studentized bootstrap tests GN (I2 ) but
e N (I2 ). The Hilbert–Schmidt test is again
far worse than that of the Studentized tests G
outperformed by all the other tests, with the exception of the non-Studentized bootstrap test
when N = 10, 25. The results are similar for the non-Gaussian scenario, apart for the fact
that the asymptotic test does not respect the nominal level (as expected, since it asks for X
to be Gaussian).
To investigate the difference between projecting on one or several eigensubspaces, we
e N (I) for increasing
also compare the power of the empirical bootstrap version of the tests G
projection subspaces, i.e. for I = Il , l = 1, 2, 3, where I1 = {(1, 1)}, I2 = {(i, j) : i, j =
1, 2} and I3 = {(i, j) : i = 1, . . . , 4; j = 1, . . . , 10}. The results are shown in Figure 4
for the Gaussian scenario and Figure 6 for the non-Gaussian scenario. In the Gaussian
e N (I2 ). In this case, projecting onto a
scenario, for N = 10, the most powerful test is G
larger eigensubspace decreases the power of the test dramatically. However, for N ≥ 25
e N (I3 ), albeit only significantly larger than that
the power of the test is the largest for G
e
of GN (I2 ) when γ = 0.01. Our interpretation is that when the sample size is too small,
including too many eigendirection is bound to add only noise that degrades the performance
of the test. However, as long as the separable eigenfunctions are estimated accurately,
projecting in a larger eigenspace improves the performance of test. See also Figure 10 for
the complete simulation results of the projection set I3 .
This prompts us to investigate how the power of the test varies across all projection
subsets
Ir,s = {(i, j) : 1 ≤ i ≤ r, 1 ≤ j ≤ s} ,
e N (I), with distribution approximated
r = 1, . . . , 32, s = 1, . . . , 7,=. The test used is G
by the empirical bootstrap with B = 1000. Figure 5 shows the empirical size and power
of the separability test in the Gaussian scenario for sample size N = 25, and Figure 7,
respectively Figure 8, shows the power for different sample sizes in the Gaussian scenario,
respectively the non-Gaussian scenario.
4.1.1
Discussion of simulation studies
The simulation studies above illustrate how the empirical bootstrap test based on the test
e N (I) usually outperforms its competitors, albeit it is also much more compustatistics G
tationally expensive than the asymptotic test, whose performance are comparable in the
Gaussian scenario for large enough number of observations.
17
1.0
N=10
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
0.8
●
●
●
●●
●
●
●
0.6
●
●
●
0.4
●
●
●
●
●
●
●
●
●
●
●
●
0.2
Empirical power
Empirical
power
N=25
●
●
●
●
●
●
●
1.00.0
●
●
●
●
●
N=50
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
N=100
●
●
●
●
●
0.8
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.6
●
●
0.4
●
0.2
●
0.0
Empirical power
Empirical
power
●
●
●
●
●
●
0.00
●
●
●
0.02
0.04
0.06
0.08
0.10 0.00
γ
0.02
0.04
0.06
0.08
0.10
γ
γ
γ
e N (Il ), for l = 1 (solid line),
Figure 4: Empirical power of the empirical bootstrap version of G
l = 2 (dashed line) and l = 3 (dash-dotted line), in the Gaussian scenario. The horizontal dotted
line indicates the nominal level (5%) of the test. Note that the points have been horizontally
jittered for better visibility.
The choice of the best set of eigendirections to use in the definition of the test statistics
is difficult. It seems that K should be ideally chosen to be increasing with N . This is
reasonable, because larger values of N increase the accuracy of the estimation of the eigenfunctions and therefore we will be able to detect departures from the separability in more
eigendirections, including ones not only associated with the largest eigenvalues. However,
the optimal rate at which K should increase with N is still an open problem, and will certainly depend in a complex way on the eigenstructure of the true underlying covariance
operator C.
This is confirmed by the results reported in Figure 5 and Figures 7 and 8. These indeed
show that taking into account too few eigendirections can result in smaller power, while
including too many of them can also decrease the power.
As an alternative to tests based on projections of DN , the tests based on the squared
2
Hilbert–Schmidt norm of DN , i.e. |||DN |||2 , could potentially detect any departure from
e N (I). But as the simulation study
the separability hypothesis—as opposed to the tests G
illustrates, they might be far less powerful in practice, particularly in situations where the
departure from separability is reflected in only in a few eigendirections. Moreover, this
approach still requires the computation of the full covariance operator (although not its
storage) and is therefore not feasible for all applications.
18
gaussian scenario
s
5
10
15
20
25
1 2 3 4 5 6 7
gamma = 0.005, N = 25
1 2 3 4 5 6 7
s
gamma = 0, N = 25
30
5
10
15
r
0.0
20
25
30
r
0.2
0.4
0.6
0.8
1.0
Figure 5: Empirical size (left) and power (right) of the separability test as functions of the proe N (I), with distribution approximated by the empirical bootstrap
jection set I. The test used is G
with B = 1000. The left plot, respectively the right plot, was simulated from the Gaussian
scenario with γ = 0, respectively γ = 0.005, and N = 25 . Each (r, s) rectangle represents the
level/power of the test based on the projection set I = {(i, j) : 1 ≤ i ≤ r, 1 ≤ j ≤ s}.
4.2
Application to acoustic phonetic data
An interesting case where the proposed methods can be useful are phonetic spectrograms.
These data arise in the analysis of speech records, since relevant features of recorded sounds
can be better explored in a two dimensional time-frequency domain.
In particular, we consider here the dataset of 23 speakers from five different Romance
languages that has been first described in Pigoli et al. (2014). The speakers were recorded
while pronouncing the words corresponding to the numbers from one to ten in their language and the recordings are converted to a sampling rate of 16000 samples per second.
Since not all these words are available for all the speakers, we have a total of 219 speech
records. We focus on the spectrum that speakers produce in each speech recording xL
ik (t),
where L is the language, i = 1, . . . , 10 the pronounced word and k = 1, . . . , nL the speaker,
nL being the number of speakers available for language L. We then use a short-time Fourier
transform to obtain a two dimensional log-spectrogram: we use a Gaussian window function w(·) with a window size of 10 milliseconds and we compute the short-time Fourier
transform as
Z +∞
L
−jωτ
Xik
(ω, t) =
xL
dτ.
ik (τ )w(τ − t)e
−∞
19
The spectrogram is defined as the magnitude of the Fourier transform and the log-spectrogram
(in decibel) is therefore
L
2
SL
ik (ω, t) = 10 log10 (|Xik (ω, t)| ).
The raw log-spectrograms are then smoothed (with the robust spline smoothing method
proposed in Garcia 2010) and aligned in time using an adaptation to 2-D of the procedure in
Tang & Müller (2008). The alignment is needed because a phase distortion can be present in
acoustic signals, due to difference in speech velocity between speakers. Since the different
words of each language have different mean log-spectrograms, the focus of the linguistic
analysis—which is the study cross-linguistics changes—is on the residual log-spectrograms
L
L
Rik
(ω, t) = Sik
(ω, t) − (1/ni )
ni
X
L
Sik
(ω, t).
k=1
Assuming that all the words within the language have the same covariance structure, we
disregard hereafter the information about the pronounced words that generated the residual
log-spectrogram, and use the surface data RjL (ω, t), j = 1, . . . , NL , i.e. the set of observations for the language L including all speakers and words, for the separability test. These
observations are measured on an equispaced grid with 81 points in the frequency direction
and 100 points in the time direction. This translate on a full covariance structure with about
33 × 106 degrees of freedom. Thus, although the discretized covariance matrix is in principle computable, its storage is a problem. More importantly, the accuracy of its estimate
is poor, since we have at most 50 observations within each language. For these reasons, we
would like to investigate if a separable approximation of each covariance is appropriate.
We thus apply the Studentized version of the empirical bootstrap test for separability
to the residual log-spectrograms for each language individually. Here, we take into consideration different choices for set of eigendirections to be used in the definition of the test
e N (I), namely I = I1 = {(1, 1)}, I = I2 = {(r, s) : 1 ≤ r ≤ 2, 1 ≤ s ≤ 3},
statistic G
I = I3 = {(r, s) : 1 ≤ r ≤ 8, 1 ≤ s ≤ 10}. For all cases we use B = 1000 bootstrap
replicates.
The resulting p-values for each language and for each set of indices can be found in
Table 1. Taking into account the multiple testings with a Bonferroni correction, we can
conclude that the separability assumption does not appear to hold. We can also see that
the departure from separability is caught mainly on the first component for the two Spanish
varieties. In conclusion, a separable covariance structure is not a good fit for these languages
and thus, when practitioners use this approximation for computational or modeling reasons,
they should bear in mind that relevant aspects of the covariance structure may be missed in
the analysis.
5
Discussion and conclusions
We presented tests to verify the separability assumption for the covariance operators of
random surfaces (or hypersurfaces) through hypothesis testing. These tests are based on the
difference between the sample covariance operator and its separable approximation—which
we have shown to be asymptotically Gaussian—projected onto subspaces spanned by the
eigenfunctions of the covariance of the data. While the optimal choice for this subspace
is still an open problem and it may depend on the eigenstructure of the full covariance
operator, it is however possible to give some advice on how to choose I in practice:
20
Table 1: P -values for the test for the separability of the covariance operators of the residual
log-spectrograms of the five Romance languages, using the Studentized version of the empirical
bootstrap.
I
I1
I2
I3
French
0.65
0.078
0.001
Italian
< 0.001
0.197
0.002
Portuguese
< 0.001
0.022
0.001
American Spanish
< 0.001
0.36
0.001
Iberian Spanish
< 0.001
0.013
< 0.001
• in many cases, a dimensional reduction based on the separable eigenfunctions is
needed also for the follow up analysis. Then, it is recommended to use the same
subspace for the test procedure as well, so that we are guaranteed at least that the projection of the covariance structure in the subspace that will be used for the analysis is
separable, as shown in Section 3.
• As mentioned in Section 3, it is usually better to focus on the subset of eigenfunctions
that it is possible to estimate accurately with the available data. These can be again
identified with bootstrap methods such as the one described in Hall & Hosseini-Nasab
(2006) or considering the dimension of the sample size. As highlighted by the results
of the simulation studies in Figure 5 and in Figures 7 and 8, the empirical power of the
test starts to decline when eigendirections that cannot be reasonably estimated with
the available sample size are included.
• When in doubt, it is also possible to apply the test to more than one subset of eigenfunctions and then summarize the response using a Bonferroni correction. We follow
this approach in the data application described in Section 4.2.
Though an asymptotic distribution is available in some cases, we also propose to approximate the distribution of our test statistics using either a parametric bootstrap (in case
the distribution of the data is known) or an empirical bootstrap. A simulation study suggests
that the Studentized version of the empirical bootstrap test gives the highest power in nonGaussian settings, and has power comparable to its parametric bootstrap counterpart and
to the asymptotic test in the Gaussian setting. We therefore use the Studentized empirical
bootstrap for the application to linguistic data, since it is not easy to assess the distribution
of the data generating process. The bootstrap test leads to the conclusion that the covariance
structure is indeed not separable.
Our present approach implicitly assumed that the functional observations (e.g. the hypersurfaces) were densely observed. Though this approach is not restricted to data observed
on a grid, it leaves aside the important class of functional data that are sparsely observed
(e.g. Yao et al. n.d.). However, the extension of our methodology to the case of sparsely
observed functional data is also possible, as long as the estimator used for the full covariance is consistent and satisfies a central limit theorem. Indeed, while we have only detailed
the methods for 2-dimensional surfaces, the extension to higher-order multidimensional
functions (such as 3-dimensional volumetric images from applications such as magnetic
resonance imaging) is straightforward.
21
A
The Asymptotic Covariance Structure
lemma A.1. The covariance operator of the random operator
Z, defined in Theorem 2.3, isi
h
N
characterized by the following equality, in which Γ = E (X ⊗ X − C) f (X ⊗ X − C) :
e A2 )Z Tr (B1 ⊗
e B2 )Z =
E Tr (A1 ⊗
(A.1)
O
O
O
Tr [BC]
Tr[B2 C2 ]
g
g
g
e
Tr (A
B)Γ +
Tr (A
IdH )Γ −
Tr A
(B1 ⊗ IdH2 ) Γ
Tr(C)
Tr[C2 ]
(
O
O
Tr[AC]
Tr[BC]
Tr[B1 C1 ]
g
g
e B2 ) Γ +
Tr A
(IdH1 ⊗
Tr (IdH
B)Γ +
Tr[Γ]
−
Tr[C1 ]
Tr[C]
Tr[C]
)
O
O
Tr[B2 C2 ]
Tr[B1 C1 ]
g
g
e IdH2 ))Γ −
e B2 ))Γ
−
Tr (IdH
(B1 ⊗
Tr (IdH
(IdH1 ⊗
Tr[C2 ]
Tr[C1 ]
(
O Tr[BC]
O
Tr[A2 C2 ]
e IdH2 ) g B Γ +
e IdH2 ) g IdH Γ
Tr (A1 ⊗
Tr (A1 ⊗
−
Tr[C2 ]
Tr[C]
O
Tr[B2 C2 ]
g
e IdH2 )
e IdH2 ) Γ
Tr (A1 ⊗
(B1 ⊗
−
Tr[C2 ]
)
O
Tr[B1 C1 ]
g
e B2 ) Γ
e IdH2 )
(IdH1 ⊗
Tr (A1 ⊗
−
Tr[C1 ]
(
O Tr[BC]
O
Tr[A1 C1 ]
e A2 ) g B Γ +
e A2 ) g IdH Γ
−
Tr (IdH1 ⊗
Tr (IdH1 ⊗
Tr[C1 ]
Tr[C]
O
Tr[B2 C2 ]
e A2 ) g (B1 ⊗
e IdH2 ) Γ
Tr (IdH1 ⊗
−
Tr[C2 ]
)
O
Tr[B1 C1 ]
g
e A2 )
e B2 ) Γ ,
−
Tr (IdH1 ⊗
(IdH1 ⊗
Tr[C1 ]
e A2 , B = B 1 ⊗
e B2 , H =
where A1 , B1 ∈ S∞ (H1 ), A2 , B2 ∈ S∞ (H2 ), and A = A1 ⊗
H1 ⊗ H2 , and IdH denotes the identity operator on the Hilbert space H.
Proof. By the linearity of the expectation and the trace, and by the properties of the partial
trace, the computation of (A.1) boils down to the computation of expressions of the form
e A02 )Y Tr (B10 ⊗
e B20 )Y ,
E Tr (A01 ⊗
2
for general A01 , B10 ∈ S∞ (H1 ), A02 , B20 ∈ S∞ (H2 ). Since E |||Y |||1 < ∞, we have
e A02 )Y Tr (B10 ⊗
e B20 )Y ) =
E (Tr (A01 ⊗
O
O
e A02 ) g (B10 ⊗
e B20 ) E Y g Y
= Tr (A01 ⊗
O
e A02 ) g (B10 ⊗
e B20 ) Γ ,
= Tr (A01 ⊗
h
i
N
where Γ = E (X ⊗ X − C) f (X ⊗ X − C) . The computation of (A.1) follows directly.
22
B
Proofs
Proof of Corollary 2.4. To alleviate the notation, we shall assume without loss of generality that µ = E X = 0.hUsing the propertiesiof the tensor product (see Appendix D.1, we
√
e B̂s ) N DN , where Âr = (ûr ⊗2 ûr ), B̂s = (v̂s ⊗2 v̂s ).
get that TN (r, s) = Tr (Âr ⊗
Now notice that though Ar = ur ⊗2 ur and Bs = vs ⊗2 vs are not estimable separately
e -product is identifiable, and is consistently
(since C1 and C2 are not identifiable), their ⊗
e B̂s (in trace norm). Slutsky’s Lemma, Theorem 2.3 and the continuous
estimated by Âr ⊗
mapping theorem imply therefore that (TN (r, s))(r,s)∈I has the same asymptotic distribui
h
√
e Bs ) N DN . This implies that
tion of TeN (r, s)
, where TeN (r, s) = Tr (Ar ⊗
(r,s)∈I
d
e Bs )Z
(TN (r, s))(r,s)∈I −→ Z 0 = Tr (Ar ⊗
,
(r,s)∈I
as N → ∞,
where Z is a mean zero Gaussian random element of S1 (H1 ⊗ H2 ) whose covariance structure is given by Lemma A.1. Z 0 is therefore also Gaussian random element, with mean zero
and covariances
0
0
e Bs )Z Tr Ar0 ⊗
e Bs0 )Z .
Σ(r,s),(r0 ,s0 ) = cov(Z(r,s)
, Z(r
0 ,s0 ) ) = E Tr (Ar ⊗
Using Lemma A.1, we see that the computation of Σ(r,s),(r0 ,s0 ) depends on the terms
e Bs )C = λr γs , Tr [Ar C1 ] = λr , Tr [Bs C2 ] = γs , as well as on the value
Tr (Ar ⊗
of
O
e B10 ) g (A02 ⊗
e B20 ) Γ
Tr (A01 ⊗
0
0
0
0
for general
P A1 , A2 ∈ S∞ (H1 ), B1 , B2 ∈ S∞ (H2 ). Using the Karhunen–Loève expansion
X = i,i0 ≥1 ξii0 ui ⊗ vi0 , where ξii0 = hX, ui ⊗ vi0 i, we get
e (X ⊗2 X − C)
Γ = E (X ⊗2 X − C) ⊗
X
O
e vi0 j 0 g ukl ⊗
e vk 0 l 0
=
βii0 jj 0 kk0 ll0 uij ⊗
i,i0 ,j,j 0 ,k,k0 ,l,l0 ≥1
−
X
O
e vi0 i0 g ujj ⊗
e vj 0 j 0
αii0 αjj 0 uii ⊗
i,i0 ,j,j 0
where we have written uij = ui ⊗2 uj ∈ S1 (H1 ), vij = vi ⊗2 vj ∈ S1 (H2 ), βii0 jj 0 kk0 ll0 =
e vi0 j 0 = (ui ⊗ vi0 ) ⊗2 (uj ⊗ vj 0 ).
E [ξii0 ξjj 0 ξkk0 ξll0 ], αij = λi γj and used the identity uij ⊗
Therefore,
O
0 e
0 g
0 e
0
Tr (A1 ⊗ A2 )
(B1 ⊗ B2 ) Γ =
X
=
βii0 jj 0 kk0 ll0 Tr[A01 uij ] Tr[A02 vi0 j 0 ] Tr[B10 ukl ] Tr[B20 vk0 l0 ]
i,i0 ,j,j 0 ,k,k0 ,l,l0 ≥1
−
X
αii0 αjj 0 Tr[A01 uii ] Tr[B10 ujj ] Tr[A02 vi0 i0 ] Tr[B20 vj 0 j 0 ],
i,i0 ,j,j 0
and the computation of the variance Σ(r,s),(r0 ,s0 ) follows from a straightforward (though
tedious) calculation.
23
Proof of Corollary 2.5. We only need to compute and substitute the values of the fourth
2 2or
der moments terms β̃ijkl in the expression given by Corollary 2.4. Since β̃ijkl = E ξij
ξkl =
2
3αkl
if (i, j) = (k, l), and β̃ijkl = αij αkl if (i, j) 6= (k, l), straightforward calculations
give
2
2
β̃rs·· = 2αrs
+ αrs Tr(C) = β̃r··s ,
β̃···· = Tr(C)2 + 2|||C|||2 ,
2
2
β̃·s·s0 = γs γs0 (Tr(C1 )2 + 2δss0 |||C1 |||2 ),
β̃r·r0 · = λr λr0 (Tr(C2 )2 + 2δrr0 |||C2 |||2 ),
2
β̃rs·s0 = 2δss0 αrs
+ αrs γs0 Tr(C1 ),
2
β̃rsr0 · = 2δrr0 αrs
+ αrs λr0 Tr(C2 ),
2
2
β̃···s = 2γs2 |||C1 |||2 + γs Tr(C1 )2 Tr(C2 ),
β̃r··· = 2λ2r |||C2 |||2 + λr Tr(C1 ) Tr(C2 )2 ,
where δij = 1 if i = j, and zero otherwise. The proof is finished by direct calculations.
C
Partial Traces
Letting S1 (H1 ⊗ H2 ) denote the space of trace-class operators on H1 ⊗ H2 , we define the
partial trace with respect to H1 as the unique linear operator Tr1 : S1 (H1 ⊗ H2 ) → S1 (H2 )
e B) = Tr(A)B for all A ∈ S1 (H1 ), B ∈ S1 (H2 ).
satisfying Tr1 (A ⊗
proposition C.1. The operator Tr1 is well-defined, linear, continuous, and satisfies
|||Tr1 (A)|||1 ≤ |||A|||1 ,
A ∈ S1 (H1 ⊗ H2 )
(C.1)
Furthermore,
e S)T ),
Tr(S Tr1 (T )) = Tr((Id1 ⊗
T ∈ S1 (H1 ⊗ H2 ), S ∈ S∞ (H2 ),
(C.2)
where Id1 is the identity operator on H1 .
Proof. Let us start by proving that the operator Tr1 (·) is well defined. By Lemma D.6, the
space
n
X
e Bi : Ai ∈ S1 (H1 ), Bi ∈ S1 (H2 ), n = 1, 2, . . .}
B0 = {
Ai ⊗
i=1
is a dense subset of S1 (H1P
⊗ H2 ). We therefore only need to show that Tr1 (·) is continuous
n
e Bi . Then, for any S ∈ S∞ (H2 ), we have
on B0 . Let T ∈ B0 , T = i=1 Ai ⊗
Tr(S Tr1 (T )) =
n
X
Tr(Ai ) Tr(SBi ) =
i=1
e S)
= Tr (Id1 ⊗
n
X
e S)(Ai ⊗
e Bi ) =
Tr (Id1 ⊗
i=1
" n
X
#!
e Bi
Ai ⊗
e S)T .
= Tr (Id1 ⊗
i=1
Hence, using the following formula for the trace norm,
|||T |||1 = sup {| Tr(ST )| : |||S|||∞ = 1} ,
we get |||Tr1 (T )|||1 ≤ |||T |||1 for all T ∈ B0 . Thus Tr1 (·) can be extended by continuity to
H1 ⊗ H2 , and (C.1) (of the paper) holds.
Let us now show (C.2) (of the paper). Fix S ∈ S∞ (H2 ), and
define the linear functione Id2 )T and hS (T ) = Tr(S Tr1 (T )),
als gS , hS : S1 (H1 ⊗ H2 ) → R by gS (T ) = Tr (S ⊗
for T ∈ S1 (H1 ⊗ H2 ). By Hölder’s inequality and (C.1) (of the paper), gS and hS are both
continuous. Since they are equal on the dense subset B0 , they are in fact equal everywhere,
and (C.2) (of the paper) follows.
24
We can also define Tr2 : S1 (H1 ⊗ H2 ) → S1 (H1 ) analogously. The following result
gives an explicit formula for the partial traces of integral operators with continuous kernels.
proposition C.2. Let Ds ⊂ Rp , Dt ⊂ Rq be compact subsets, H1 = L2 (Ds , R) , H2 =
L2 (Dt , R), and H = L2 (Ds × Dt , R) = H1 ⊗ H2 . If C ∈ S1 (L2 (Ds × Dt , R))
is a positive definite operator with symmetric continuous kernel c = c(s, t, s0 , t0 ), i.e.
c(s, t, s0 , t0 ) = c(s0 , t0 , s, t) for all s, s0 ∈ Ds , t, t0 ∈ Dt , and
ZZ
Cf (s, t) =
c(s, t, s0 , t0 )f (s0 , t0 )ds0 dt0 ,
f ∈ L2 (Ds × Dt , R),
Ds ×Dt
then Tr1 (C) is the integral operator on L2 (Dt , R) with kernel k(t, t0 ) =
The analogous result also holds for Tr2 (C).
R
Ds
c(s, t, s, t0 )ds.
Proof. Let ε > 0. By Lemma D.7, we know that there exists an integral operator C 0 with
continuous kernel c0 such that |||C − C 0 |||1 ≤ ε/2 and kc − c0 k∞ ≤ ε/2, where C 0 =
PN
e
n=1 An ⊗ Bn , and each An , Bn are finite rank operators, with continuous kernels an ,
respectively bn , and kgk∞ = supx |g(x)|. We have
Z
≤ |||Tr1 (C) − Tr1 (C 0 )|||2
(C.3)
Tr1 (C) −
c(s, ·, s, ·)ds
Ds
2
+
Tr1 (C 0 ) −
Z
c0 (s, ·, s, ·)ds
Ds
Z
2
c0 (s, ·, s, ·)ds −
+
Ds
Z
c(s, ·, s, ·)ds
Ds
.
2
The first term is bounded |||Tr1 (C) − Tr1 (C 0 )|||1 ≤ |||C − C 0 |||1 ≤ ε/2. The second term
is equal to zero since
Tr1 (
N
X
e Bn ) =
An ⊗
n=1
N
X
Tr(An )Bn
n=1
=
N Z
X
n=1
Z
=
Ds
Z
=
An (s, s)dsBn
Ds
N
X
!
e Bn
An ⊗
(s, ·, s, ·)ds.
n=1
c0 (s, ·, s, ·)ds,
Ds
where the second equality comes from the fact that An is a finite rank operator (hence
trace-class) with continuous kernel. The third term of (C.3) is
!1/2
2
0
[c (s, t, s, t ) − c(s, t, s, t )] ds dtdt
Z
ZZ
Dt ×Dt
0
0
0
Ds
≤ |Ds | |Dt | kc0 − ck∞ ≤ ε/2,
where |Ds | =
R
Ds
dx and |Dt | =
R
dy. Therefore,
Z
Tr1 (C) −
c(s, t, s, t0 )ds
Dt
Ds
≤ ε.
2
25
Since
this holds for any ε > 0, Tr1 (C) is equal to the operator with kernel k(t, t0 ) =
R
c(s,
t, s, t0 )ds. The proof of the analogous result for Tr2 (C) is similar.
Ds
The next result states that the partial trace of a Gaussian random trace-class operator is
also Gaussian.
proposition C.3. Let Z ∈ S1 (H1 ⊗ H2 ) be a Gaussian random element. Then Tr1 (Z) ∈
S1 (H2 ) is a Gaussian random element.
Proof. The proof
is finished by noticing that A ∈ S∞ (H2 ), we have Tr(A Tr1 (Z)) =
e A)Z , where the right-hand side is obviously Gaussian.
Tr (Id ⊗
D
Background Results
This section presents some background results that are used in the paper. Some references
for these results are Zhu (2007), Gohberg & Krejn (1971), Gohberg et al. (1990), Kadison
& Ringrose (1997a,b), Ringrose (1971).
D.1
Tensor Products Hilbert Spaces, and Hilbert–Schmidt Operators
Let H1 , H2 be two real separable Hilbert spaces, whose inner products are denoted by h·, ·i1
and h·, ·i2 , respectively. Let H1 ⊗ H2 denote the Hilbert space obtained as the completion
of the space of finite linear combinations of simple tensors u ⊗ v, u ∈ H1 , v ∈ H2 under
the inner product
hu ⊗ v, u0 ⊗ v 0 i = hu, u0 i1 hv, v 0 i2 .
The Hilbert space H1 ⊗ H2 is actually isometrically isomorphic to the space of Hilbert–
Schmidt operators from H2 to H1 , denoted by S2 (H2 , H1 ), which consists of all continuous
linear operators T : H2 → H1 satisfying
X
2
2
|||T |||2 =
kT en k ,
n≥1
where the sum extends over any orthonormal basis (en )n≥1 of H2 . The norm |||·||| is actually
induced by the inner-product inner product
X
hT, SiS2 =
hT en , Sen i1 , T, S ∈ S2 (H2 , H1 ),
n≥1
which is independent of the choice of the basis (the space S2 (H2 , H1 ) is therefore itself
a Hilbert space). The isomorphism between H1 ⊗ H2 and S2 (H2 , H1 ) is given by the
mapping Φ : H1 ⊗ H2 → S2 (H2 , H1 ), defined by Φ(u ⊗ v) = u ⊗2 v for all u ∈ H1 , v ∈
H2 , where u ⊗2 v(v 0 ) = hv 0 , vi u for u ∈ H1 , v, v 0 ∈ H2 . We therefore identify these two
spaces, and might write u ⊗ v instead of u ⊗2 v hereafter.
Notice that
is itself a Hilbert space, if A, B ∈ H, the
N since H = S2 (H1 ) = S2 (H1 , H1 )N
operator A 2 B ∈ S2 (H, H) is defined by (A 2 B) (C) = hC, BiS2 A, for A, B, C ∈
H. Here are some properties of the tensor product · ⊗2 ·:
proposition D.1. Let H be a real separable Hilbert space. For any u, v, f, g ∈ H, A, B ∈
S2 (H)
1. · ⊗2 · is linear on the left, and conjugate-linear on the right,
26
2. hu ⊗2 v, f ⊗2 giS2 = hu, f i hg, vi = h(u ⊗2 v)g, f i,
3. hA, u ⊗2 viS2 = hAv, ui = v ⊗2 u, A† S2 ,
4. Tr(u ⊗2 v) = hu, vi,
5. |||u ⊗2 v|||1 = |||u ⊗2 v|||2 = kukkvk,
6. (u ⊗2 v)(f ⊗2 g) = hf, vi u ⊗2 g,
7. (u ⊗2 v)† = v ⊗2 u,
N
N
8. (A 2 B)† = B 2 A.
Proof. The proof follows from the definition and the properties of the inner product, and is
therefore omitted.
e B) ∈ S∞ (H1 ⊗ H2 ) is
Recall that for A ∈ S∞ (H1 ), B ∈ S∞ (H2 ), the operator (A ⊗
defined by the linear extension of
e B)(u ⊗ v) = Au ⊗ Bv, u ∈ H1 , v ∈ H2
(A ⊗
e B)† = A† ⊗
e B†, A ⊗
e B
e B)(C ⊗
e D) =
Furthermore, we have (A ⊗
= |||A|||∞ |||B|||∞ , and (A ⊗
∞
e CD) for A, C ∈ S∞ (H1 ), B, D ∈ S∞ (H2 ). For u, v ∈ H1 , f, g ∈ H2 , (u ⊗2 v) ⊗
e (f ⊗2 g) =
(AC ⊗
N
e B ∈ S1 (H1 ⊗ H2 ),
(u ⊗ f ) 2 (v ⊗ g) If A ∈ S1 (H1 ), B ∈ S1 (H2 ), then A ⊗
e B
A⊗
1
≤ |||A|||1 |||B|||1 ,
e B) = Tr(A) Tr(B).
and Tr(A ⊗
In the case H1 = L2 [−S, S]d , R , H2 = L2 ([0, T ], R), with S, T > 0, if A ∈
S2 (H1 ), B ∈ S2 (H2 ) are Hilbert–Schmidt operators (hence also integral operators, with
e B ∈ S2 (H1 ⊗ H2 ) = S2 (L2 [−S, S]d × [0, T ], R )
kernels a(s, s), b(t, t), respectively), the operator A ⊗
is also an integral operator with kernel k(s, t, s0 , t0 ) = a(s, s0 )b(t, t0 ), that is,
Z
e B)u(s, t) =
(A ⊗
D.2
0
T
Z
k(s, t, s0 , t0 )u(s0 , t0 )dsdt.
[−S,S]d
Random Elements in Banach Spaces
We understand random elements of a separable Banach space (B, k·k) in the Bochner sense
(e.g. Ryan 2002). A random element X ∈ B satisfying E kXk < ∞ has a mean E X ∈ B,
which satisfies S( E X) = E(SX) for all bounded linear operator S : B → B 0 , where B 0
is another Banach space.
D.3
Random Trace-class Operators
If X ∈ S1 (H) is a random element satisfying E |||X|||1 < ∞, i.e. a random trace-class
operator, then E Tr(AX) = Tr(A E X) for any A ∈ S∞ (H). Furthermore, if X 0 ∈
S1 (H) is another random element such that E(|||X|||1 |||X 0 |||1 ) < ∞, then
O
g 0 0
0 0
E (Tr [AX] Tr[A X ]) = E Tr AX
AX
O
O
g 0
g 0
= E Tr (A
A )(X
X)
27
O
O
g 0
g 0
= Tr (A
A) E X
X
,
for any A, A0 ∈ S∞ (H).
2
The second-order structure of a random element X ∈ S1 (H) satisfying E |||X|||1 < ∞
is encoded by the covariance functional Γ : S∞ (H) × S∞ (H) → R, which is defined by
Γ(A, B) = cov (Tr[AX], Tr[BY ]) .
Since
Γ(A, B) = E (Tr[A(X − µ)] Tr[B(X − µ)])
O
O
g
g
= Tr (A
B) E (X − µ)
(X − µ) ,
the second-order structure is also encoded by the generalized covariance operator
O
g
Γ = E (X − µ)
(X − µ) ∈ S1 (H ⊗ H).
proposition D.2. Let H1 , H2 be real separable Hilbert spaces. Let Y ∈ S1 (H1 ) be a
2
Gaussian random element such that E |||Y |||1 < ∞. Then, for any T ∈ S1 (H2 ) fixed,
e T is a Gaussian random element of S1 (H1 ⊗ H2 ).
Y ⊗
e T )) is Gaussian. This
Proof. We need to show that for all S ∈ S∞ (H1 ⊗ H2 ), Tr(S(Y ⊗
e T )) is Gaussian for all n ≥ 1, where (Sn )
can be reduced to showing that Tr(Sn (Y ⊗
is a sequence of operators in S∞ (H1 ⊗ H2 ) that converges weakly to S. Indeed, letting
Dn = Sn − S, we have
h
i
e T )) − Tr(S(Y ⊗
e T ) 2 = E Tr(Dn (Y ⊗
e T ))2
E Tr(Sn (Y ⊗
e Dn )(Y ⊗
e T ⊗
e Y ⊗
e T)
= E Tr (Dn ⊗
e Dn ) E(Y ⊗
e T ⊗
e Y ⊗
e T) ,
= Tr (Dn ⊗
where the last equality is valid since
e Dn )(Y ⊗
e T ⊗
e Y ⊗
e T)
E (Dn ⊗
2
1
2
2
≤ |||Dn |||∞ |||T |||1 E |||Y |||1 < ∞.
e Dn converges weakly to zero, and since
Lemma D.4 tells us that Dn ⊗
2
2
e T ⊗
e Y ⊗
e T)
E(Y ⊗
≤ |||T |||1 E |||Y |||1 < ∞,
1
h
i
e T )) − Tr(S(Y ⊗
e T ) 2 → 0 as n → ∞.
Lemma D.5 tells us that E Tr(Sn (Y ⊗
Therefore, since the space of Gaussian random variables is close under the L2 (Ω, P) norm,
e T )) is Gaussian if Tr(Sn (Y ⊗
e T )) is Gaussian for all n ≥ 1. Lemma D.3 tells
Tr(S(Y ⊗
Pn
e Bi , where Ai ∈ S∞ (H1 ) and Bi ∈ S∞ (H2 ). In
us that we can choose Sn = i=1 Ai ⊗
this case,
e T )) =
Tr(Sn (Y ⊗
n
X
n
X
e Bi )(Y ⊗
e T) =
Tr (Ai ⊗
Tr(Ai Y ) Tr(Bi T ),
i=1
i=1
which is Gaussian since Tr(Ai Y ) is Gaussian for each i ≥ 1.
28
D.4
Technical results
Recall that (Tn )n≥1 ⊂ S∞ (H) is said to converge weakly to T ∈ S∞ (H) if for all u, v ∈
H, hTn u, vi → hT u, vi as n → ∞.
lemma D.3. Let H1 , H2 be real separable Hilbert spaces. For any S ∈ S∞ (HP
1 ⊗ H2 ),
n
e Bi
there exists a sequences of operators (Sn )n≥1 ⊂ S∞ (H1 ⊗ H2 ) of the form Sn = i=1 Ai ⊗
with Ai ∈ S∞ (H1 ), Bi ∈ S∞ (H2 ), such that Sn converges weakly to S.
Proof. For S ∈ S∞ (H1 ⊗ H2 ), define
N
X
SN =
hSen ⊗ fm , en0 ⊗ fm0 i (en0 ⊗ fm0 ) ⊗2 (en ⊗ fm ),
n,n0 ,m,m0 =1
N
X
=
e (fm0 ⊗ fm ),
hSen ⊗ fm , en0 ⊗ fm0 i (en0 ⊗ en ) ⊗
n,n0 ,m,m0 =1
where (en )n≥1 is an orthonormal basis of H1 , and (fm )m≥1 is an orthonormal basis of H2 .
First, notice that we have the following equality:
hSN ei ⊗ fj , ek ⊗ fl i =
N
X
N
X
hSen ⊗ fm , en0 ⊗ fm0 i hen ⊗ fm , ei ⊗ fj i hen0 ⊗ fm0 , ek ⊗ fl i
n,m=1 n0 ,m0 =1
= hSei ⊗ fj , ek ⊗ fl i 1{i≤N } 1{j≤N } 1{k≤N } 1{l≤N } .
P
Therefore, for general g, h ∈ H1 ⊗ H2 , g = i,j≥1 αij ei ⊗ fj and
h=
X
(D.1)
βkl ek ⊗ fl ,
k,l≥1
we have
hSN h, gi =
X X
αij βkl hSN (ei ⊗ fj ), ek ⊗ fl i
i,j≥1 k,l≥1
=
X X
αij βkl hS(ei ⊗ fj ), ek ⊗ fl i 1{i≤N } 1{j≤N } 1{k≤N } 1{l≤N }
i,j≥1 k,l≥1
(Using (D.1))
*
=
S
N
X
i,j=1
αij ei ⊗ fj ,
N
X
+
βkl ek ⊗ fl
.
k,l=1
Therefore, by continuity of the inner product and the continuity of S, we have limN →∞ hSN h, gi =
hSh, gi.
lemma D.4. Let (Sn )n≥1 ⊂ S∞ (H) be a sequence of operators converging weakly to
e Sn converges weakly to S ⊗
e S ∈ S∞ (H ⊗ H).
S ∈ S∞ (H). Then, Sn ⊗
Proof. For u, v, z, w ∈ H, we have
e Sn (u ⊗ v), z ⊗ w = lim hSn u, zi hSn v, wi
lim Sn ⊗
n→∞
n→∞
= hSu, zi hSv, wi
29
e S (u ⊗ v), z ⊗ w .
S ⊗
P
Now for general g, h ∈ H ⊗ H, let us write g = i,j≥1 αij ei ⊗ ej , h = k,l≥1 βkl ek ⊗ el ,
PN
N
where (ei )i≥1 is an orthonormal basis of H, and let g N =
=
i,j=1 αij ei ⊗ ej , h
PN
k,l=1 βkl ek ⊗ el for N ≥ 1. Also, let
=
P
K = max{sup |||Sn |||∞ , |||S|||∞ },
n≥1
and notice that K < ∞ by the uniform boundedness principle (e.g. Rudin 1991). We have
e Sn )g, h − (S ⊗
e S)g, h
(Sn ⊗
e Sn )g, h | + | (Sn ⊗
e Sn )g N , hN |
≤ | (Sn ⊗
e Sn )g N , hN − (S ⊗
e S)g N , hN
+ (Sn ⊗
e S)g N , hN − (S ⊗
e S)g, h
+ (S ⊗
≤ K 2 kgk h − hN + K 2 g − g N
+
N
X
hN
αij βkl h(Sn − S)ei , ek i hSn ej , el i
i,j,k,l=1
+ hSei , ek i h(Sn − S)ej , el i
+ K 2 kgk h − hN + K 2 g N − g khk
≤ 4K 2 max {kgk, khk} max h − hN , g − g N
+ 2KN 4 (kgk + khk) max {|h(Sn − S)ei , ek i|} .
1≤i,k≤N
Now, for any ε > 0, choose N > 1 such that max
Since N is fixed, we can find an n0 ≥ 0 such that
max |h(Sn − S)ei , ek i| ≤
1≤i,k≤N
h − hN , g − g N
ε
,
6KN 4 (kgk + khk)
≤ ε 6K 2 max {kgk, khk}
for all n ≥ n0 .
e Sn )g, hi − h(S ⊗
e S)g, hi| ≤ ε, therefore Sn ⊗
e Sn
Then, for all n ≥ n0 , we have |h(Sn ⊗
e S.
converges weakly to S ⊗
lemma D.5. Let (Sn )n≥1 ⊂ S∞ (H) be a sequence of operators converging weakly to
S ∈ S∞ (H). Then, for all T ∈ S1 (H), we have
Tr(Sn T ) → Tr(ST ),
n → ∞.
P
Proof. Let T = l≥1 λl ul ⊗2 vl be the singular value decomposition of T . Without loss
of generality, (vl )l≥1 is an orthonormal basis of H. We have
lim Tr(Sn T ) = lim
n→∞
n→∞
= lim
n→∞
=
X
l≥1
30
X
hvl , Sn T vl i
l≥1
X
λl hvl , Sn ul i
l≥1
λl lim hvl , Sn ul i
n→∞
−1
.
=
X
λl hvl , Sul i
l≥1
= Tr(ST ),
where the third equality is justified by the dominated convergence theorem since
supn≥1 {|||Sn |||∞ } |||T |||1 < ∞ by the uniform boundedness theorem
P
l≥1
|λl || hvl , Sn ul i | ≤
PN
e Bn , where An : H1 → H1 and
lemma D.6. The operators of the form n=1 An ⊗
Bn : H2 → H2 are finite rank operators, and N < ∞, are dense in the Banach space
S1 (H1 ⊗ H2 ).
P
Proof. Let T ∈ S1 (H1 ⊗ H2 ). Then T = n≥1 λn Un ⊗2 Vn , with convergence in trace
norm, where (λn )n≥1 is a summable decreasing sequence of positive numbers, and (Un )n≥1 ⊂
H1 ⊗ H2 and (Vn )n≥1 ⊂ H1 ⊗ H2 are orthonormal sequences. Each Un can be written as
P
(1)
(2)
Un = l≥1 λn,l un,l ⊗ un,l , with convergence in the norm of H1 ⊗ H2 , that we denote by
P
(1)
(2)
k·k2 . Similarly, Vn = l≥1 γn,l vn,l ⊗ vn,l . Let
UnM =
M
X
(1)
(2)
λn,l un,l ⊗ un,l
VnM =
and
l=1
N
X
(1)
(2)
γn,l vn,l ⊗ vn,l .
l=1
T−
Fix ε > 0, and choose N such that
we have
T−
X
λn UnM ⊗2 VnM
n=1
≤
T−
PN
n=1
N
X
λn Un ⊗2 Vn
1
N
X
+
≤ ε/2. For M ≥ 1 fixed,
λn Un ⊗2 Vn
n=1
1
1
λn (Un − UnM ) ⊗2 Vn + UnM ⊗2 (Vn − VnM )
n=1
≤ ε/2 +
N
X
λn
1
(Un − UnM ) ⊗2 Vn
1
UnM ⊗2 (Vn − VnM )
+
n=1
≤ ε/2 +
N
X
λn
Un − UnM
2
kVn k2 + UnM
2
Vn − VnM
2
n=1
Take M ≥ 1 such that
max
n=1,...,N
Un − UnM
, Vn − VnM
2
Then, since Un , Vn have unit length, and UnM
1, . . . , N , we have
T−
N
X
n=1
λn UnM
⊗2 VnM
≤ ε/2 + 3 max
n=1,...,N
2
Un −
1
≤ ε/2 + 3
PN
Since n=1 λn UnM ⊗2 VnM
the proof is finished.
n
o
ε
,1 .
6 Tr C
≤ 1 + Un − UnM
2
≤ min
UnM 2 ,
Vn −
≤ 2 for n =
VnM 2
·
N
X
!
λn
n=1
ε
Tr C
6 Tr C
= ε.
PN PK
(1)
(1)
(2)
e u(2)
⊗
=
n=1
j,l=1 λn λn,l γn,j un,l ⊗ vn,j
n,l ⊗ vn,j
31
.
1
If H = L2 (Ds × Dt , R), we can approximate certain integral operators in a stronger
sense:
lemma D.7. Let Ds ⊂ Rp , Dt ⊂ Rq be compact subsets, and C ∈ S1 (L2 (Ds × Dt , R))
be a positive definite integral operator with symmetric continuous kernel c = c(s, t, s0 , t0 ),
i.e. c(s, t, s0 , t0 ) = c(s0 , t0 , s, t) for all s, s0 ∈ Ds , t, t0 ∈ Dt .
PN
e Bn , where An : L2 (Ds , R) →
For any ε > 0, there exists an operator C 0 = n=1 An ⊗
2
2
2
L (Ds , R), Bn : L (Dt , R) → L (Dt , R) are finite rank operators with continuous kernels
an , respectively bn , such that
1. |||C − C 0 |||1 ≤ ε,
2. sups,s0 ∈Ds ,t,t0 ∈Dt |c(s, t, s0 , t0 ) − c0 (s, t, s0 , t0 )| ≤ ε, where c0 is the kernel of the
operator C 0 ,
Proof. By Mercer’s Theorem, there exists continuous orthonormal functions (Un )n≥1 ⊂
L2 (Ds ×Dt , R) and (λn )n≥1 ⊂ R is a summable decreasing sequence of positive numbers,
such that
X
c(s, t, s0 , t0 ) =
λn Un (s, t)Un (s0 , t0 ),
(D.2)
n≥1
where the convergence is uniform in (s, t, s0 , t0 ).
PN
N
Let C N =
denote its kernel. Fix ε > 0, and let
n=1 λn Un ⊗2 Un , and let c
kgk∞ = supx |g(x)|. We have that for N large enough, both C − C N 1 and c − cN ∞
are bounded by ε/2, since C is positive and (D.2) is also its singular value decomposition.
We can now approximate each of the continuous functions Un (s, t), n = 1, . . . , N,
PM (1)
(2)
by tensor products of continuous functions (Cheney 1986). Let UnM = l=1 un,l ⊗ un,l ,
(1)
(2)
where un,l ∈ L2 (Ds , R), un,l ∈ L2 (Dt , R), l ≥ 1, are continuous functions such that
Un − UnM
∞
n
≤ min
o
ε
,κ ,
6κ Tr C
where κ = maxn=1,...,N kUn k∞ (notice that κ < ∞ since each Un is continuous). Writing
PN
C N,M = n=1 λn UnM ⊗2 UnM , and denoting by cN,M its kernel, we have
cN − cN,M
∞
≤
N
X
λn
Un − UnM
∞
kUn k∞ + UnM
∞
Un − UnM
∞
n=1
≤ 3κ max
n=1,...,N
Un − UnM
∞
· Tr(C)
≤ ε/2.
Furthermore, we also have
C N − C N,M
1
≤
N
X
λn
λn
λn
(Un − UnM ) ⊗2 Un
1
+
UnM ⊗2 (Un − UnM )
1
n=1
≤
N
X
Un − UnM
2
kUn k2 + UnM
Un − UnM
∞
2
Un − UnM
2
n=1
≤
N
X
n=1
32
kUn k∞ + UnM
∞
Un − UnM
∞
≤ ε/2.
Since C N,M =
E
PN
n=1
PK
j,l=1
(1)
(1)
λn un,l ⊗ un,j
(2)
e u(2)
⊗
⊗
u
n,j , the proof is finished.
n,l
Implementation details
All the implementation details described here are implemented in the R package covsep
(Tavakoli 2016).
In practice, random elements of H1 ⊗ H2 are first projected onto a truncated basis of
H1 ⊗ H2 . We shall assume that the truncated basis is of the form (ei ⊗ fj )i=1,...,d1 ;j=1,...,d2 ,
for some d1 , d2 < ∞, where (ei )i≥1 ⊂ H1 , respectively (fj )j≥1 ⊂ H2 , is an orthonormal
basis of H1 , respectively H2 . In this way, one can encode (and approximate) an element
X ∈ H1 ⊗ H2 by a d1 × d2 matrix X ∈ Rd1 ×d2 , whose (k, l)-th coordinate is given by
X(k, l) = hX, ek ⊗ fl i , k = 1, . . . , d1 ; l = 1, . . . , d2 . We therefore assume from now on
that only need to describe the implementation of TN (r, s) for H1 ⊗ H2 = Rd1 ⊗ Rd2 =
Rd1 ×d2 . In this case, X is a random element of Rd1 ×d2 , i.e. a random d1 × d2 matrix, and
i.i.d.
we observe X1 , . . . , Xn ∼ X. We have
e 2,N (l, l0 )
C
C2,N (l, l0 ) = q
,
e 2,N )
Tr(C
e 1,N (k, k 0 )
C
,
C1,N (k, k 0 ) = q
e 1,N )
Tr(C
and
d2
N X
X
e 1,N (k, k 0 ) = 1
Xi (k, l) − X(k, l) Xi (k 0 , l) − X(k 0 , l)
C
N i=1
l=1
!
d2
N
X
T
1 X
Xi − X Xi − X
CN (k, l, k 0 , l),
=
=
N i=1
0
l=1
k,k
e 2,N (l, l0 ) = 1
C
N
=
X(k, l) =
CN (k, l, k 0 , l0 ) =
d1
N X
X
Xi (k, l) − X(k, l)
i=1 k=1
N
T
1 X
Xi − X
Xi − X
N i=1
1
N
Xi (k, l0 ) − X(k, l0 )
N
X
!
=
l,l0
d1
X
CN (k, l, k, l0 ),
k=1
Xi (k, l),
i=1
N
1 X
Xi (k, l) − X(k, l) Xi (k 0 , l0 ) − X(k 0 , l0 ) ,
N i=1
for all k, k 0 = 1, . . . , d1 ; l, l0 = 1, . . . , d2 .
e 1,N (k, k 0 ) using the above formula is not efficient in R, when
The computation of C
implemented using a double for loop. However, if we denote by Ak the N × d2 matrix
with (Ak )il = Xi (k, l)−X(k, l), n = 1, . . . , N ; k = 1, . . . , d1 ; l = 1, . . . , d2 , by Vec(Ak )
the vector obtained by stacking the columns of Ak into a vector of length N d2 , and by
33
Yn the n-th row of the N d2 × d1 matrix A = (Vec(A1 ), . . . , Vec(Ad1 )), we get Y =
PN d
(N d2 )−1 n=12 = 0 and
(N d2 )−1
N
d2
X
(Yn )k (Yn )k0 = (N d2 )−1
n=1
d2
n X
X
Xi (k, l) − Xi (k, l) .
i=1 l=1
Therefore C̃1,N = N dN2 −1 cov(A), where cov is the standard R function returning the covariance, and the computation is very fast. The computation of C̃2,N can be done similarly.
If we denote by (λ̂r , ûr ), respectively (γ̂s , v̂s ), the r-th eigenvalue/eigenvector pair of
C1,N , respectively the s-th eigenvalue/eigenvector pair of C2,N , we have
"
#
N
√
2
1 X T
TN (r, s) = TN (r, s|X1 , . . . , XN ) = N
û (Xi − X)v̂s − λ̂r γ̂s ,
N i=1 r
where ·T denotes matrix transposition. The variance of TN (r, s|X1 , . . . , XN ) is estimated
by
σ̂ 2 (r, s|X1 , . . . , XN ) =
2
2
2λ̂2r γ̂s2 Tr(C1,N )2 + |||C1,N |||2 − 2λ̂r Tr(C1,N ) Tr(C2,N )2 + |||C2,N |||2 − 2γ̂s Tr(C2,N )
Tr(C1,N )2 Tr(C2,N )2
,
(E.1)
Pd1 Pd2
2
2
where |||A|||2 = i=1
j=1 [(A)ij ] for a d1 × d2 matrix A.
If I = {1, . . . , p} × {1, . . . , q}, then (TN (r, s))(r,s)∈I is asymptotically a mean zero
Gaussian random p × q matrix, with separable covariance. Its left (row) covariances are
consistently estimated by the p × p matrix Σ̂L,I = Σ̂L,I (X1 , . . . , XN ) with entries
p
2
2
b
b
b
0
0
0
2
λ̂
λ̂
C
−
(
λ̂
+
λ̂
)
Tr(
C
)
δ
Tr(
C
)
+
r
1,N
r r
1,N
r
rr
1,N
2
=
, (E.2)
Σ̂L,I
b1,N ) Tr(C
b2,N )
r,r 0
Tr(C
r, r0 ∈ {1, . . . , p}, and it right (column) covariances are consistently estimated by the q × q
matrix Σ̂R,I = Σ̂R,I (X1 , . . . , XN ) with entries
2
√
2
b
b
b
2γ̂s γ̂s0 δss0 Tr(C2,N ) + C2,N
− (γ̂s + γ̂s0 ) Tr(C2,N )
2
Σ̂R,I
=
(E.3)
b1,N ) Tr(C
b2,N )
s,s0
Tr(C
s, s0 ∈ {1, . . . , q},
The computation of |||DN |||2 can be done without storing the full covariance CN in
memory. The following pseudo-code returns |||DN |||2 :
I. Compute and store C1,N and C2,N , and set s = 0.
II. Replace Xn by Xn − X for each n = 1, . . . , N .
III. For i, k = 1, . . . , d1 ; j, l = 1, . . . , d2 ,
PN
(a) Compute y = N −1 n=1 Xn (i, j)Xn (k, l).
2
(b) Set s = s + (y − C1,N (i, k)C2,N (j, l)) .
34
IV. Return s.
∗
The computation of |||DN
− DN |||2 requires a slight modification of the pseudo-code. Given
∗
∗
X1 , . . . , XN and X1 , . . . , XN
,
I. Compute and store C1,N and C2,N , C∗1,N and C∗2,N , and set s = 0.
II. Replace Xn by Xn − X, and X∗n by X∗n − X∗ for each n = 1, . . . , N .
III. For i, k = 1, . . . , d1 ; j, l = 1, . . . , d2 ,
PN
(a) Compute y = N −1 n=1 (Xn (i, j)Xn (k, l) − X∗n (i, j)X∗n (k, l)).
2
(b) Set s = s + y − C∗1,N (i, k)C∗2,N (j, l) + C1,N (i, k)C2,N (j, l) .
IV. Return s.
Finally, Algorithms 1 and 2 describe the procedure to approximate the p-values for the tests
based on parametric and empirical bootstrap, respectively.
Algorithm 1 Parametric Bootstrap p-value approximation for HN
Given X1 , . . . , XN ,
e C2,N , and HN = HN (X1 , . . . , XN ).
I. compute X, C1,N ⊗
II. For b = 1, . . . , B,
b }, where
(a) Create bootstrap samples Xb = {X1b , . . . , XN
i.i.d.
e C2,N .
Xib ∼ F X, C1,N ⊗
b = H (Xb ),
(b) Compute HN
N
III. Compute the estimated bootstrap p-value
p=
B
1 X
1{H b >HN } ,
N
B
b=1
where 1{A} = 1 if A is true, and zero otherwise.
F
Additional results from the simulation studies
e N (I) for
Figure 6 shows the empirical powers empirical bootstrap version of the tests G
increasing projection subspaces, i.e. for I = Il , l = 1, 2, 3, where I1 = {(1, 1)}, I2 =
{(i, j) : i, j = 1, 2} and I3 = {(i, j) : i = 1, . . . , 4; j = 1, . . . , 10}, when data are generated from a multivariate t distribution with 6 degrees of freedom (the Non-Gaussian scenario
in the paper). Figure 9 shows the empirical power for the asymptotic test, the parametric
e N (I2 ), as well as parametric and
and empirical bootstrap tests based on the test statistic G
a
e
bootstrap tests based on the test statistics GN (I), GN (I2 ) where I2 = {(i, j) : i, j = 1, 2}.
Figure 10 shows the analogous results for the projection set I3 . Tables 2, 3 and 4 give the
true levels of the tests for I = I1 , I2 , and I = I3 , respectively.
Figure 7 shows the empirical size and power of the separability test, in the Gaussian
scenario, as functions of the projection set
Ir,s = {(i, j) : 1 ≤ i ≤ r, 1 ≤ j ≤ s} ,
35
Algorithm 2 Empirical Bootstrap p-value approximation for HN
Given X = {X1 , . . . , XN },
e C2,N , and HN = HN (X).
I. Compute C1,N ⊗
II. For b = 1, . . . , B,
b } by drawing with repetition from
(a) Create the bootstrap sample Xb = {X1b , . . . , XN
X1 , . . . , XN .
(b) For each bootstrap sample, compute ∆bN = ∆N (Xb ; X).
III. Compute the estimated bootstrap p-value
B
1 X
1{∆b >HN } ,
p=
N
B
b=1
where 1{A} = 1 if A is true, and zero otherwise.
e N (I), with distribution approximated by
for all possible choices of (r, s). The test used is G
the empirical bootstrap with B = 1000. Figure 8 is analogous plot for the Non-Gaussian
scenario.
Acknowledgments
We wish to thank the editor, associate editor, and the referees for their comments that have
led to an improved version of the paper. We also wish to thank Victor Panaretos for interesting discussions.
36
Table 2: Empirical size of the testing procedures (with α = 0.05), for I = I1 .
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
0.17
0.22
0.04
0.02
0.20
0.10
0.10
0.07
0.11
N=25
0.09
0.10
0.04
0.04
0.11
0.05
0.07
0.08
0.05
N=50
0.08
0.08
0.06
0.05
0.08
0.05
0.06
0.07
0.03
N=100
0.06
0.08
0.05
0.05
0.07
0.06
0.06
0.07
0.04
N=25
0.20
0.21
0.13
0.12
0.07
0.06
0.04
0.51
0.01
N=50
0.18
0.18
0.14
0.14
0.06
0.04
0.03
0.55
0.01
N=100
0.15
0.17
0.15
0.14
0.08
0.04
0.03
0.63
0.01
(a) Gaussian scenario
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
0.29
0.31
0.08
0.08
0.20
0.07
0.06
0.37
0.06
(b) Non-Gaussian scenario
37
Table 3: Empirical size of the testing procedures (with α = 0.05), for I = I2 .
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
0.43
0.17
0.04
0.02
0.12
0.01
0.00
0.07
0.11
N=25
0.19
0.09
0.05
0.04
0.08
0.04
0.01
0.08
0.05
N=50
0.11
0.08
0.06
0.05
0.07
0.04
0.02
0.07
0.03
N=100
0.09
0.07
0.05
0.04
0.07
0.05
0.04
0.07
0.04
N=25
0.37
0.19
0.12
0.16
0.04
0.03
0.01
0.51
0.01
N=50
0.32
0.17
0.14
0.20
0.06
0.03
0.01
0.55
0.01
N=100
0.28
0.16
0.14
0.22
0.07
0.03
0.01
0.63
0.01
(a) Gaussian scenario
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
0.61
0.26
0.09
0.10
0.10
0.00
0.00
0.37
0.06
(b) Non-Gaussian scenario
38
Table 4: Empirical size of the testing procedures (with α = 0.05), for I = I3 .
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
1.00
0.18
0.01
0.01
0.10
0.00
0.00
0.07
0.11
N=25
0.98
0.09
0.01
0.02
0.08
0.00
0.00
0.08
0.05
N=50
0.79
0.08
0.04
0.03
0.07
0.01
0.00
0.07
0.03
N=100
0.40
0.07
0.05
0.05
0.07
0.01
0.00
0.07
0.04
N=25
1.00
0.19
0.19
0.34
0.04
0.00
0.00
0.51
0.01
N=50
0.98
0.17
0.23
0.53
0.05
0.00
0.00
0.55
0.01
N=100
0.94
0.17
0.30
0.64
0.07
0.00
0.00
0.63
0.01
(a) Gaussian scenario
Asymptotic Distribution
Gaussian parametric bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Empirical bootstrap (non-Studentized)
(diag Studentized)
(full Studentized)
Gaussian parametric Hilbert–Schmidt
Empirical Hilbert–Schmidt
N=10
1.00
0.28
0.03
0.07
0.09
0.00
0.00
0.37
0.06
(b) Non-Gaussian scenario
39
1.0
N=10
N=25
●
●
●
●
●
●
●
●
0.8
●
●
0.6
●
●
●
●
●
●
0.4
●
●
●
●
●
●
●
●
●
0.2
Empirical power
Empirical
power
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1.00.0
●
●
●
●
●
●
N=50
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
●
●
●
●
●
●
●
●
N=100
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
0.6
●
0.4
●
●
●
0.2
●
0.0
Empirical power
Empirical
power
●
●
●
●
●
●●
●
●
●●
●
●●
0.00
0.02
0.04
0.06
0.08
0.10 0.00
γ
0.02
0.04
0.06
0.08
0.10
γ
γ
γ
e N (Il ), for l = 1 (solid line),
Figure 6: Empirical power of the empirical bootstrap version of G
l = 2 (dashed line) and l = 3 (dash-dotted line), in the non-Gaussian scenario. The horizontal dotted line indicates the nominal level (5%) of the test. Note that the points have been
horizontally jittered for better visibility.
40
gaussian scenario
6
5
15
20
25
30
5
15
20
r
gamma = 0, N = 25
gamma = 0.005, N = 25
25
30
25
30
25
30
6
1
2
3
s
4
5
6
5
4
1
2
3
15
20
25
30
5
10
15
20
r
r
gamma = 0, N = 50
gamma = 0.005, N = 50
6
5
3
2
1
1
2
3
s
4
5
6
7
10
7
5
4
s
10
r
7
10
7
5
s
4
1
2
3
s
4
1
2
3
s
5
6
7
gamma = 0.005, N = 10
7
gamma = 0, N = 10
5
10
15
20
25
30
5
r
0.0
10
15
20
r
0.2
0.4
0.6
0.8
1.0
Figure 7: Empirical size (left column) and power (right column) of the separability test as funce N (I), with distribution approximated by the emtions of the projection set I. The test used is G
pirical bootstrap with B = 1000. The left plots, respectively the right plots, were simulated from
the Gaussian scenario with γ = 0, respectively γ = 0.005. Each row corresponds to a different
sample size: N = 10 (top), N = 25 (middle),41N = 50 (bottom). Each (r, s) rectangle represents the level/power of the test based on the projection set I = {(i, j) : 1 ≤ i ≤ r, 1 ≤ j ≤ s}.
student scenario
6
5
15
20
25
30
5
15
20
r
gamma = 0, N = 25
gamma = 0.005, N = 25
25
30
25
30
25
30
6
1
2
3
s
4
5
6
5
4
1
2
3
15
20
25
30
5
10
15
20
r
r
gamma = 0, N = 50
gamma = 0.005, N = 50
6
5
3
2
1
1
2
3
s
4
5
6
7
10
7
5
4
s
10
r
7
10
7
5
s
4
1
2
3
s
4
1
2
3
s
5
6
7
gamma = 0.005, N = 10
7
gamma = 0, N = 10
5
10
15
20
25
30
5
r
0.0
10
15
20
r
0.2
0.4
0.6
0.8
1.0
Figure 8: Empirical size (left column) and power (right column) of the separability test as
e N (I), with distribution approximated
functions of the projection set I. The test used is G
by the empirical bootstrap with B = 1000. The left plots, respectively the right plots, were
simulated from the Non-Gaussian scenario with γ = 0, respectively γ = 0.005. Each row
corresponds to a different sample size: N =4210 (top), N = 25 (middle), N = 50 (bottom). Each (r, s) rectangle represents the level/power of the test based on the projection set
I = {(i, j) : 1 ≤ i ≤ r, 1 ≤ j ≤ s}.
●
0.8
1.0
●
●
●
N=25
●
●
●
●
●
●
●
●
●
●
0.6
●
0.4
●
●
●
●
●
●
●
●
●
●
N=50●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
N=100
●
●
●
●
0.6
0.4
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
0.2
0.0
●
●
●
1.00.0
●
●
●
Empirical power
●
●
●
0.2
Empirical power
N=10
●
●
●
●
●
●
●
●
0.00
●
●
●
●
●
●
●
●
●
●
●
●
●
0.02
0.04
0.06
0.08
●
●
●
0.10 0.00
●●
●●
0.02
0.04
γ
0.06
0.08
0.10
γ
0.8
N=10
●
●
●
●
●
●
●
●
●
0.6
0.4
●
●
N=25●
●
●
●
N=50●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
N=100
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.2
0.4
0.6
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1.00.0
●
●
●
●
●
Empirical power
●
●
0.2
Empirical power
1.0
(a) Gaussian scenario
0.00
0.02
0.04
0.06
0.08
0.10 0.00
0.02
γ
0.04
0.06
0.08
0.10
γ
(b) Non-Gaussian scenario
Figure 9: Empirical power of the testing procedures in the Gaussian scenario (panel (a)) and
non-Gaussian scenario (panel (b)), for N = 43
10, 25, 50, 100 and I = I3 . The results shown
correspond to the test (3.1) based on its asymptotic distribution (····+····), the Gaussian parametric
e N (I2 ) (dash-dotted line with empty circles), G
e a (I2 ) (dashed line with empty
bootstrap test G
N
circles), and GN (I2 ) (solid line with empty circles), the empirical bootstrap projection tests
e N (I2 ) (– · –4– · –), G
e a (I2 ) (– –4– –), and GN (I2 ) (—4—), the Gaussian parametric
G
N
Hilbert–Schmidt test (dash-dotted line with filled circles) and the empirical Hilbert–Schmidt
test (dash-dotted line with filled triangles). The horizontal dotted line indicates the nominal
level (5%) of the test. Note that the points have been horizontally jittered for better visibility.
1.0
N=25●
N=10●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
0.8
●
●
●
●
0.6
●
●
●
●
0.4
●
0.2
Empirical power
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
N=50●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
0.8
1.00.0
●
●
●
●
●
●
●
●
●
●
●
●
●
N=100
●●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
0.6
●
●
●
●
●
●
●
●
●
0.4
●
●
●
●
●●
0.0
●
●
●
●
0.2
Empirical power
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
0.00
0.02
0.04
0.06
0.08
●
●
●
●
●
0.10 0.00
0.02
0.04
γ
0.06
0.08
0.10
γ
1.0
(a) Gaussian scenario
N=10
0.8
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
N=25●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.2
●
●
●
●
N=50●●
●●
●●
●●
●
0.8
●
●
0.6
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●
●
●
●
●●
N=100
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
0.4
●
●
0.2
●
●
●
●
●
0.0
Empirical power
●
●
●
●
●●●
●
●
0.6
0.4
●
●
●
1.00.0
Empirical power
●
0.00
0.02
0.04
0.06
0.08
0.10 0.00
0.02
γ
0.04
0.06
0.08
0.10
γ
(b) Non-Gaussian scenario
Figure 10: Empirical power of the testing procedures in the Gaussian scenario (panel (a)) and
non-Gaussian scenario (panel (b)), for N = 44
10, 25, 50, 100 and I = I3 . The results shown
correspond to the test (3.1) based on its asymptotic distribution (····+····), the Gaussian parametric
e N (I3 ) (dash-dotted line with empty circles), G
e a (I3 ) (dashed line with empty
bootstrap test G
N
circles), and GN (I3 ) (solid line with empty circles), the empirical bootstrap projection tests
e N (I3 ) (– · –4– · –), G
e a (I3 ) (– –4– –), and GN (I3 ) (—4—), the Gaussian parametric
G
N
Hilbert–Schmidt test (dash-dotted line with filled circles) and the empirical Hilbert–Schmidt
test (dash-dotted line with filled triangles). The horizontal dotted line indicates the nominal
level (5%) of the test. Note that the points have been horizontally jittered for better visibility.
References
Aston, J. A. D. & Kirch, C. (2012), ‘Evaluating stationarity via change-point alternatives
with applications to fMRI data’, The Annals of Applied Statistics 6(4), 1906–1948.
Billingsley, P. (1999), Convergence of probability measures, Wiley Series in Probability and
Statistics: Probability and Statistics, second edn, John Wiley & Sons, Inc., New York. A
Wiley-Interscience Publication.
Chen, K., Delicado, P. & Müller, H.-G. (2015), ‘Modeling function-valued stochastic processes, with applications to fertility dynamics’, Technical report .
Cheney, E. W. (1986), Multivariate approximation theory: Selected topics, SIAM.
Constantinou, P., Kokoszka, P. & Reimherr, M. (2015), ‘Testing separability of space–time
functional processes’, ArXiv e-prints .
Cressie, N. & Huang, H.-C. (1999), ‘Classes of nonseparable, spatio-temporal stationary
covariance functions’, Journal of the American Statistical Association 94(448), 1330–
1339.
Ferraty, F. & Vieu, P. (2006), Nonparametric Functional Data Analysis: Theory and Practice, Springer.
Fuentes, M. (2006), ‘Testing for separability of spatial–temporal covariance functions’,
Journal of statistical planning and inference 136(2), 447–466.
Garcia, D. (2010), ‘Robust smoothing of gridded data in one and higher dimensions with
missing values’, Computational Statistics & Data Analysis 54(4), 1167–1178.
Genton, M. G. (2007), ‘Separable approximations of space-time covariance matrices’, Environmetrics 18(7), 681–695.
Gneiting, T. (2002), ‘Nonseparable, stationary covariance functions for space–time data’,
Journal of the American Statistical Association 97(458), 590–600.
Gneiting, T., Genton, M. G. & Guttorp, P. (2007), ‘Geostatistical space-time models, stationarity, separability, and full symmetry’, Monographs On Statistics and Applied Probability 107, 151.
Gohberg, I. C. & Krejn, M. G. (1971), Introduction à la théorie des opérateurs linéaires
non auto-adjoints dans un espace hilbertien, Dunod, Paris. Traduit du russe par Guy
Roos, Monographies Universitaires de Mathématiques, No. 39.
Gohberg, I., Goldberg, S. & Kaashoek, M. A. (1990), Classes of linear operators. Vol. I,
Operator Theory: Advances and Applications, Birkhäuser Verlag, Basel.
Hall, P. & Hosseini-Nasab, M. (2006), ‘On properties of functional principal components
analysis’, Journal of the Royal Statistical Society: Series B (Statistical Methodology)
68(1), 109–126.
Horváth, L. & Kokoszka, P. (2012a), Inference for functional data with applications,
Springer series in statistics, Springer, New York, NY.
45
Horváth, L. & Kokoszka, P. (2012b), Inference for functional data with applications, Vol.
200, Springer Science & Business Media.
Kadison, R. & Ringrose, J. (1997a), Fundamentals of the Theory of Operator Algebras:
Elementary theory, number v. 1 in ‘Fundamentals of the Theory of Operator Algebras’,
American Mathematical Society.
Kadison, R. V. & Ringrose, J. R. (1997b), Fundamentals of the theory of operator algebras.
Vol. II, Vol. 16 of Graduate Studies in Mathematics, American Mathematical Society,
Providence, RI. Advanced theory, Corrected reprint of the 1986 original.
Lindquist, M. A. (2008), ‘The statistical analysis of fMRI data’, Statistical Science
23(4), 439–464.
Liu, C., Ray, S. & Hooker, G. (2014), ‘Functional Principal Components Analysis of Spatially Correlated Data’, ArXiv e-prints 1411.4681.
Lu, N. & Zimmerman, D. L. (2005), ‘The likelihood ratio test for a separable covariance
matrix’, Statistics & probability letters 73(4), 449–457.
Mas, A. (2006), ‘A sufficient condition for the CLT in the space of nuclear operators—
Application to covariance of random functions’, Statistics & probability letters
76(14), 1503–1509.
Mitchell, M. W., Genton, M. G. & Gumpertz, M. L. (2005), ‘Testing for separability of
space–time covariances’, Environmetrics 16(8), 819–831.
Pigoli, D., Aston, J. A. D., Dryden, I. L. & Secchi, P. (2014), ‘Distances and inference for
covariance operators’, Biometrika 101(2), 409–422.
Rabiner, L. R. & Schafer, R. W. (1978), Digital processing of speech signals, Vol. 100,
Prentice-hall Englewood Cliffs.
Ramsay, J., Graves, S. & Hooker, G. (2009), Functional Data Analysis with R and MATLAB,
Springer, New York.
Ramsay, J. O. & Silverman, B. W. (2002), Applied functional data analysis, Springer Series
in Statistics, Springer-Verlag, New York. Methods and case studies.
Ramsay, J. O. & Silverman, B. W. (2005), Functional data analysis, Springer Series in
Statistics, second edn, Springer, New York.
Ringrose, J. R. (1971), Compact non-self-adjoint operators, Van Nostrand Reinhold Co.,
London.
Rudin, W. (1991), Functional analysis. 2nd ed., 2nd ed. edn, New York, NY: McGraw-Hill.
Ryan, R. A. (2002), Introduction to tensor products of Banach spaces, Springer Monographs in Mathematics, Springer-Verlag London, Ltd., London.
Secchi, P., Vantini, S. & Vitelli, V. (2015), ‘Analysis of spatio-temporal mobile phone data:
a case study in the metropolitan area of Milan’, Statistical Methods & Applications . in
press.
46
Simpson, S. L. (2010), ‘An adjusted likelihood ratio test for separability in unbalanced
multivariate repeated measures data’, Statistical Methodology 7(5), 511–519.
Simpson, S. L., Edwards, L. J., Styner, M. A. & Muller, K. E. (2014), ‘Separability tests
for high-dimensional, low-sample size multivariate repeated measures data’, Journal of
applied statistics 41(11), 2450–2461.
Tang, R. & Müller, H.-G. (2008), ‘Pairwise curve synchronization for functional data’,
Biometrika 95(4), 875–889.
Tavakoli, S. (2016), covsep: Tests for Determining if the Covariance Structure of 2Dimensional Data is Separable. R package version 1.0.0.
URL: https://CRAN.R-project.org/package=covsep
Tibshirani, R. J. (2014), In praise of sparsity and convexity, in X. Lin, C. Genest, D. L.
Banks, G. Molenberghs, D. W. Scott & J.-L. Wang, eds, ‘Past, Present, and Future of
Statistical Science’, Chapman and Hall, pp. 497–506.
Wang, J.-L., Chiou, J.-M. & Mueller, H.-G. (2015), ‘Review of Functional Data Analysis’.
URL: http://arxiv.org/abs/1507.05135v1
Worsley, K. J., Marrett, S., Neelin, P., Vandal, A. C., Friston, K. J. & Evans, A. C. (1996),
‘A unified statistical approach for determining significant signals in images of cerebral
activation’, Human brain mapping 4(1), 58–73.
Yao, F., Müller, H.-G. & Wang, J.-L. (n.d.), ‘Functional Data Analysis for Sparse Longitudinal Data’, Journal of the American Statistical Association 100(470), 577–590.
Zhu, K. (2007), Operator theory in function spaces, Mathematical Surveys and Monographs, second edn, American Mathematical Society, Providence, RI.
47
| 10math.ST
|
Using Conservative Estimation for Conditional
Probability instead of Ignoring Infrequent Case
Masato Kikuchi∗ , Eiko Yamamoto† , Mitsuo Yoshida∗ , Masayuki Okabe‡ , Kyoji Umemura∗
∗ Department
arXiv:1709.08309v1 [cs.IT] 25 Sep 2017
of Computer Science and Engineering
Toyohashi University of Technology, Toyohashi 441-8580, Japan
{m143313@edu, yoshida@cs}.tut.ac.jp, [email protected]
† Department of Economics and Information
Gifu Shotoku Gakuen University, Gifu 500-8288, Japan
[email protected]
‡ Faculty of Management and Information System
Prefectural University of Hiroshima, Hiroshima 734-8558, Japan
[email protected]
Abstract—There are several estimators of conditional probability from observed frequencies of features. In this paper,
we propose using the lower limit of confidence interval on
posterior distribution determined by the observed frequencies to
ascertain conditional probability. In our experiments, this method
outperformed other popular estimators.
Keywords—Conservative estimation, Conditional probability,
Confidence interval, Lower limit.
I. Introduction
Estimating conditional probability from observed frequencies of features is the fundamental operation of natural language processing (NLP) and its practical application [1]. When
we need to analyze the relationship between two features, we
sometimes need to estimate conditional probability from the
frequencies of the occurrence of these features. For example,
we may need to know what is the chance a document contains
the word A under the condition that the document contains the
word B. One problem in estimating conditional probability
could be low frequency. i.e, in this case when only a few
documents may contain word B.
When we need to estimate the probability from observed
frequencies or samples, we usually use maximum likelihood
estimator (MLE), which is an unbiased estimator, and asymptotically converges to true probability when we have infinite
number of observations. When only a few observations are
available, we need to use various smoothing methods [2], [3].
One of the classical smoothing methods is additive smoothing
whose background is Bayesian framework. In this framework, we assume prior distribution (Prior) of the features,
and compute posterior distribution (Posterior) based on the
observations, regardless of the number of observations.
When the number of observations is small, Posterior has
large variance. Therefore, we need to pay attention to deciding the estimated value from Posterior of the conditional
probability. When we choose the value that gives maximum
probability in Posterior and we assume that Prior is uniform
distribution, the value is same as MLE. When we choose the
978–1–5090–1636–5/16/$31.00 c 2016 IEEE
expected value of the probability from Posterior, and assume
that Prior is uniform distribution, the value is same as Laplace
smoothing estimator.
In this paper, we focus our attention on this classical
framework with a novel viewpoint. We propose to form
confidence interval of Posterior, using the lower limit of
confidence interval for the estimator for judging the strength
of the relationship between two features. We conducted experiments where the true conditional probability is the true
strength between features, and found that it outperformed other
estimators.
II. Related Work
Church and Hanks [4], proposed the concept of mutual
information to measure the association between words. This
method solves low-frequency problem, because infrequent
features cannot sum up to a large mutual information quantity.
However, if the true strength of the relationship between two
features can be regarded as conditional probability, it may not
be an appropriate measure to use because mutual information
is symmetric towards two features, whereas conditional probability is asymmetric.
The Good-Turing estimator [5], [6] is a popular method
to adjust frequency for the low frequency case. This method
needs to assume that the distribution of frequency obeys Zip’s
law. We need to verify the distribution of features, and if it
does not obey Zip’s law, this method is not appropriate.
Conditional probability is more popular in the database
field than NLP. Apriori [7] is the most practically used
method for finding the relationship between features, when
the true strength between features is known to be conditional
probability. It uses MLE as the measure for the relationship.
To overcome the problem of low frequency, Apriori ignores
the rare features using a threshold value, which is called
minimum supports. Although Apriori is efficient by ignoring
low-frequency features, we sometimes need to compare a
high frequency but low MLE value relationship with a low
III. Posterior from Observed Frequencies
Let θ be the true value of conditional probability P(A|B),
which we need to estimate. Let n be the frequency of event B.
Let x be the frequency of event A and B. When we assume that
Prior is uniform distribution π(θ), Posterior for n and x is as
follows, where L is normalization constant so that the integral
from negative infinity to positive infinity should become 1.0.
P(Θ | N = n, X = x) =
L × θ x (1 − θ)n−x (0 < θ < 1),
0
Otherwise.
P(Θ | N = 0, X = 0) is uniform distribution (Fig. 1), and
P(Θ | N = 1, X = 0), P(Θ | N = 4, X = 1) as shown in Fig. 2,
and 3 respectively.
IV. Confidence Interval of Posterior
The confidence intervals are in the range of θ, where the
probability that θ fall into this range is the value of the
1
0.8
!(")
0.6
0.4
0.2
0
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
!
Fig. 1. Uniform Distribution π(θ), which is P(Θ | N = 0, X = 0) that is used
in experiment. This Prior is usually used when we have no knowledge of θ.
We have chosen this distribution because Apriori does not utilize the prior
knowledge.
2.5
p(!|N=1,X=0)
2
1.5
1
0.5
0
0
0.5
1
!
Confidence Interval
Fig. 2. Posterior Distribution P(Θ | N = 1, X = 0), and its confidence interval
[θlb , 1].
2.5
2
p(!|N=4,X=1)
frequency but high MLE value relationship. Apriori simply
ignores relation of the low frequency but high MLE value
relationship.
To overcome this problem, Predictive Apriori [8] is proposed, where Posterior of conditional probability of two features is computed based on Prior, which is decided by the
distribution of the two features, and the expected value of
conditional probability is used in stead of MLE. We have
examined Predictive Apriori, and found that the actual Prior
and the Prior that shows the best performance are different.
This suggests that we need another parameter or viewpoints
to decide the estimated value of strength of a relation.
When we need to control of the quality of product, we form
the confidence interval of the probability of the chance that
the product is defective. Then, we use the upper limit of the
confidence interval to estimate the ratio of defectiveness. We
use the upper limit because it causes more trouble if a defective
product is judged as normal than if a normal product is judged
as defective. In the case of finding relationships, we need to
be careful about the precision of the results found; a precision
of 50% is not considered satisfactory. This suggests that it
causes much more trouble if a false relationship is judged as
true, than if a true relationship is judged as false. In this case,
the lower limit is natural choice for measuring the strength of
a relationship.
Another issue in forming these confidence intervals is the
low frequency of occurrence. We need to form the confidence
interval from less frequent features. We find well-known
approximation of confidence interval is not usable because
it assumes that there are enough samples. Though we may
check enough samples to form the confidence interval by
approximation formula for detecting defectiveness, we cannot
increase the number of observation for detecting relationship.
Moreover, we have found that the so-called “exact formula” [9]
of interval has considerable errors. We have found that we need
to numerically compute these lower limits of confidence.
1.5
1
0.5
0
0
0.5
1
!
Confidence Interval
Fig. 3. Posterior Distribution P(Θ | N = 4, X = 1), and its confidence interval
[θlb , 1]. It has same expected value as P(Θ | N = 1, X = 0), but larger θlb ,
since its variance is smaller.
Minsup=1
Minsup=2
Minsup=3
Proposed Method
1
0.8
RECALL
Algorithm 1 Generation algorithm of the synthetic dataset
D∗ B φ; k B 0;
while k < 1000 do
j B 0; tk B φ;
while j < 2 do
extract hS l , Cm i at random from R;
tk B tk ∪ S l ∪ Cm ;
jB j+1
end while
D∗ B D∗ ∪ {tk };
k Bk+1
end while
0.6
0.4
0.2
0
RANK
Fig. 4. Algorithm for preparing dataset. R is hierarchical relation, and we
have chosen actual names of prefectures (corresponds to state) and cities.
confidence level. Although there may be many choices of
intervals, the interval that we form is [θlb , 1], where the lower
bound θlb is defined as follows:
P(Θ > θlb | N, X) = α,
where α is confidence level. Please note that even for P(Θ |
N = 1, X = 0), θlb is positive. Although usual confidence
interval of θ contains the value 0, the proposed confidence
interval will never contains the value 0. Therefore θlb can be
regarded as a conservative smoothing value. We have chosen
confidence level α by considering the required precision of
estimation.
The value θlb is determined by n, x, α and Prior. Therefore,
we can have a table of θlb before judging the strength of
the relation. For making the check experiment easier the
table of θlb by n, x for α = 0.99, using uniform distribution as Prior is available on the Web, whose URL is
“http://www.ss.cs.tut.ac.jp/CI-Laplace”. For n < 7, the value
is shown in the Appendix.
V. Experimental Setting
We synthesized the dataset to clearly compare the
estimators. In this synthesis, the task is to separate
hierarchical relationships from their mixed data. Let
R be the actual hierarchical relationship of prefectures (or states) and cities. For example, R could
be {hS 1 , C1 i, hS 1 , C2 i, hS 1 , C3 i, hS 2 , C4 i, hS 2 , C5 i, hS 3 , C5 i} ,
where S i and Ci are the name of prefectures, and cities,
respectively. We randomly selected two relationships to synthesize the dataset. For example, the synthesized data could
be {{C5 , C1 , S 2 , S 1 }, {C3 , C4 , S 1 , S 2 }, ...}. As different cities may
TABLE I
Statistics of the Synthesized Data
Number of transactions
Kinds of pairs of candidate pairs
Number of occurrences of candidates pairs
Kinds of right pairs
Number of occurrences of right pairs
1,000
4,469
5,934
975
2,000
Fig. 5. Recall rate by θ̂ and θlb . Minsup means minimum supports. Minsup=1
corresponds to maximum likelihood estimator θ̂. Setting the appropriate value
for minimum support benefits of the high (small) rank range with cost of the
low (large) rank range. The proposed method θlb always better than maximum
likelihood estimator θ̂ with any Minsup value.
belong to the same prefecture (or state), there is nothing wrong
if the prefecture name of one city appears in the data of
another city. Therefore, the correct measure for estimating the
relationship between a city and its prefecture is the conditional
probability of the prefecture under the condition that the city
name appears. The algorithm and statistics of the synthesized
data are shown in Fig. 4 and TABLE I respectively.
For the generated dataset, we first estimate the conditional
probabilities of all the pairs of city/prefecture names. We then
rank the list of name pairs from the largest estimation value to
the smallest. Lastly, we assess whether each pair exists in R
and compute the recall and precision using the pairs in order
from the first to the current one. The obtained list could be
(hS 1 , C1 i, hS 3 , C5 i, hS 1 , C5 i, hS 2 , C4 i, ...). In this example, only
the hS 1 , C5 i is not in R, and there are five relationships in R.
Then the values of recall at the ith pair are (1/5, 2/5, 2/5, 3/5,
...).
We observe the estimator performance by plotting the recall
by rank. In this plot, we can also see the precision of each rank
as the slope of the line passing through the origin and each
plotted point. The line of the best estimator will appear above
the other lines. The proposed method has a single parameter,
α (confidence level). In our experiment, we choose α = 0.99
as the desired precision for high (small) rank.
A. Comparison with MLE (Apriori)
It is a common practice to ignore the low-frequency case.
Apriori [7] does this, and calls the minimum frequency as
minimum support (minsup for short). We conducted an experiment changing the minimum support. When minimum support
is 1, it is equivalent to use ordinal MLE θ̂ as estimator. As
shown in Fig. 5, setting the appropriate value for minimum
support benefits of the high (small) rank range with cost of the
low (large) rank range. This result is obtained by completely
ignoring the doubtful relation due to low frequency.
In the case of θlb , we can obtain a similar improvement
in the high rank range by discounting the probability of low
Histgram
Beta(0.17,1.06)
1800
10
1600
Frequency
1200
6
1000
800
4
600
400
Probability density
8
1400
2
200
0
0
0
0.25
0.5
Value
0.75
1
Fig. 6. Prior Distribution of θ in the dataset for all name pairs (blue),
and Prior distribution used in the experiment (red). In determining the Prior
distribution, note that we ignore low-frequency data, e.g., 1/1, 1/2, 1/3, 2/3,
1/4, 3/4, 1/5, 2/5, 3/5 and 4/5.
Predictive Apriori
Proposed Method
1
RECALL
0.8
0.6
0.4
0.2
0
0
400 800 1200 1600 2000 2400 2800 3200 3600 4000
RANK
Fig. 7. Recall rate by θ¯p and θlb . Predictive Apriori θ¯p uses the beta
distribution as the Prior distribution. The parameters of the beta distribution
are determined by examining the dataset. Although we get approximately the
same results, Predictive Apriori needs to estimate the Prior distribution, which
is not always straightforward to do.
frequency. In our case, low frequency cases still have some
positive value and may have chance of being included in the
output. Therefore, we lose nothing in the low rank range, as
shown in Fig. 5.
Both Apriori and θlb have one parameter to choose, minimum support and confidence level. Our result suggests θlb with
the appropriate confidence level, always outperforms Apriori,
regardless with the value of its minimum support.
B. Comparison with expected value using Posterior and Prior
The Predictive Apriori algorithm [8] uses θ¯p , the expected
value of θ in the Posterior distribution. Predictive Apriori
needs the Prior of θ to compute the Posterior. Fig. 6 shows
the histogram (in blue) of θ for all name pairs. Spikes can
be observed at 1/1, 1/2, and 1/3, corresponding to the lowfrequency pairs. Although Predictive Apriori usually uses this
histogram of data as its Prior distribution, assuming that the
Prior distribution has this shape would be wrong. There is no
reason that θ is likely to be a particular number, such as 1/2.
Generally speaking, the estimation of the Prior distribution
is not an easy task. This estimation requires a multitude of
considerations in an actual situation.
By observing the histogram, we choose to ignore the data
of 0, 1/1, 1/2, 1/3, 2/3, 1/4, 3/4, 1/5, 2/5, 3/5, 4/5. The
modified the histogram is then smoothed as a ?? distribution,
allowing the remaining case to be observed in the resulting
distribution. The beta distribution has two parameters. We
usually determine these on the basis of observed mean and
variance. :
θa−1 (1 − θ)b−1
β(a, b) = R 1
ta−1 (1 − t)1−b dt
0
(0 < θ < 1).
The parameters are determined by inspecting the dataset. As a
result, we use??(0.17, 1.06), which gives the red curve shown
in Fig. 7.
Please note that, this Prior distribution has a relatively
large probability for a relative small value of θ. Using this
distribution, θ¯p also behaves like the proposed method. When
we have only a few name occurrences, the shape of Posterior
is close to that of Prior. Thus, θ¯p becomes less than θ̂. For
the evaluation of θ¯p , we obtained almost the same curve with
the proposed method. Although we can get similar results,
estimating the Prior distribution is a more complex operation
and less intuitive than the proposed method. If we choose
the beta distribution as the Prior distribution, the number
of parameters are less than histograms. Still there are two
parameters a and b, whereas θlb has only one parameter α
(confidence level).
VI. Conclusion
We have proposed to use θlb instead of θ̂ or θ̄. We chose the
interval as [θlb , 1], and the confidence level α as the required
precision of the output. Using a synthesized dataset, we have
found that θlb , outperformed θ̂, with or without minimum
support. We have also compared θlb and θ¯p . Although θlb and
θ¯p show almost the same results, we need to know the Prior
distribution of θ to use θ¯p , and determing the Prior distribution
is not always an easy task.
Even in the case where Prior distribution is well known, and
θ¯p seems appropriate, there is a possibility of forming a confidence interval for this estimator. Even when the distribution
of features obeys Zip’s law and the Good-Turing estimator
seems appropriate, there again lies a possibility of forming
a confidence interval. Our proposal is to consider how the
estimator is treated, and to provide an additional viewpoint on
selecting the estimator to use.
Acknowledgment
This work was supported by 2015 Gifu Shotoku Gakuen
University Research Grant.
References
[1] A. Jimeno-Yepes and R. B. Llavori, “Knowledge based word-concept
model estimation and refinement for biomedical text mining,” Journal of
Biomedical Informatics, vol. 53, pp. 300–307, 2015.
[2] A. Hazem and E. Morin, “A comparison of smoothing techniques for
bilingual lexicon extraction from comparable corpora,” in Proceedings of
the 6th Workshop on Building and Using Comparable Corpora, 2013, pp.
24–33.
[3] S. F. Chen and J. Goodman, “An empirical study of smoothing techniques
for language modeling,” in Proceedings of the 34th Annual Meeting of
the Association for Computational Linguistics, 1996, pp. 310–318.
[4] K. W. Church and P. Hanks, “Word association norms, mutual information, and lexicography,” Computational Linguistics, vol. 16, no. 1, pp.
22–29, 1990.
[5] I. J. Good, “The population frequencies of species and the estimation of
population parameters,” Biometrika, vol. 40, no. 3-4, pp. 237–264, 1953.
[6] W. A. Gale and G. Sampson, “Good-turing frequency estimation without
tears*,” Journal of Quantitative Linguistics, vol. 2, no. 3, pp. 217–237,
1995.
[7] R. Agrawal and R. Srikant, “Fast Algorithms for Mining Association
Rules,” in Proceedings of the 20th International Conference on Very Large
Data Bases, 1994, pp. 487–499.
[8] T. Scheffer, “Finding association rules that trade support optimally against
confidence,” Intelligent Data Analysis, vol. 9, no. 4, pp. 381–395, 2005.
[9] C. J. Clopper and E. S. Pearson, “The Use of Confidence or Fiducial
Limits Illustrated in the Case of the Binomial,” Biometrika, vol. 26, no. 4,
pp. 404–413, 1934.
Appendix
TABLE II
Table of Various Estimators, using the Uniform Distribution as the Prior
Distribution of θ
n
1
1
2
2
2
3
3
3
3
4
4
4
4
4
5
5
5
5
5
5
6
6
6
6
6
6
6
x
0
1
0
1
2
0
1
2
3
0
1
2
3
4
0
1
2
3
4
5
0
1
2
3
4
5
6
θ̂
0.00000
1.00000
0.00000
0.50000
1.00000
0.00000
0.33333
0.66667
1.00000
0.00000
0.25000
0.50000
0.75000
1.00000
0.00000
0.20000
0.40000
0.60000
0.80000
1.00000
0.00000
0.16667
0.33333
0.50000
0.66667
0.83333
1.00000
θ̄
0.33333
0.66667
0.25000
0.50000
0.75000
0.20000
0.40000
0.60000
0.80000
0.16667
0.33333
0.50000
0.66667
0.83333
0.14286
0.28571
0.42857
0.57143
0.71429
0.85714
0.12500
0.25000
0.37500
0.50000
0.62500
0.75000
0.87500
θlb
0.00501
0.10000
0.00334
0.05890
0.21544
0.00251
0.04200
0.14087
0.31623
0.00201
0.03268
0.10564
0.22207
0.39811
0.00167
0.02676
0.08473
0.17307
0.29431
0.46416
0.00144
0.02267
0.07080
0.14227
0.23632
0.35664
0.51795
| 7cs.IT
|
Towards Neural Knowledge DNA
Haoxi Zhang 1, Cesar Sanin 2, Edward Szczerbicki 3,
1
Chengdu University of Information Technology, Chengdu, China
2
3
The University of Newcastle, NSW, Australia
Gdansk University of Technology, Gdansk, Poland
[email protected]
[email protected]
[email protected]
Abstract: In this paper, we propose the Neural Knowledge DNA, a
framework that tailors the ideas underlying the success of neural networks
to the scope of knowledge representation. Knowledge representation is a
fundamental field that dedicate to representing information about the world
in a form that computer systems can utilize to solve complex tasks. The
proposed Neural Knowledge DNA is designed to support discovering,
storing, reusing, improving, and sharing knowledge among machines and
organisation. It is constructed in a similar fashion of how DNA formed:
built up by four essential elements. As the DNA produces phenotypes, the
Neural Knowledge DNA carries information and knowledge via its four
essential elements, namely, Networks, Experiences, States, and Actions.
___________________________________________
Keywords: Neural Knowledge DNA, Neural Networks, Knowledge
Representation.
1
INTRODUCTION
Knowledge representation is a fundamental field that dedicate to representing
information about the world in a form that computer systems can utilize to solve
complex tasks (Davis et al. 1993). It is the study of thinking as a computational process.
Then, what is knowledge? This is a question that has been discussed by philosophers
since the ancient Greeks, and it is still not totally demystified. Drucker P. F. (2011)
defines it as “information that changes something or somebody - either by becoming
grounds for actions, or by making an individual (or an institution) capable of different
or more effective action”. While the Oxford Dictionary (2015) defines Knowledge as
“facts, information, and skills acquired through experience or education; the
theoretical or practical understanding of a subject”. O'Dell and Hubert (2011) claim
that Knowledge is not knowledge until the information inside itself has been taken and
used by people. And for scientists and researchers in the AI field, we can argue it as
“knowledge is not knowledge until the information inside itself has been taken and
used by computers, machines, and agents”.
Consequently, a good knowledge representation shall easy to be used by different
systems to allow storing, reusing, improving, and sharing knowledge among these
systems. A survey carried out by Liao (2003) found that there were generally seven
categories of knowledge-based technologies and applications developed until 2002. In
another study (Matayong & Mahmood 2012), after analysing 30 published articles
between 2003 and 2010 from high quality journals, found nine core theories in the
2
knowledge-based area. However, there are limitations to these technologies: most of
them are designed for one specific kind of product; they don‟t have standard knowledge
presentation; most systems lack the capability for information sharing and exchange
and most of these systems only focus on supporting a particular stage of a product
lifecycle (Li, Xie, & Xu 2011).
Recent studies (LeCun et al. 2015; Gallant 2015; Lescroart et al. 2013) in artificial
neural networks (ANN) and psychology have found that the image representations in
ANN are very similar to those in bio brains; which inspires us that why do not organise
and store knowledge as or close to the way how it exists in the human brain?
In this paper, we propose the Neural Knowledge DNA (NK-DNA), a framework
adapting ideas underlying the success of neural networks to the scope of knowledge
representation for neural network-based knowledge discovering, storing, reusing,
improving, and sharing.
THE NEURAL KNOWLEDGE DNA
The Neural Knowledge DNA (NK-DNA) is designed to store and represent the
knowledge captured by its domain. It is constructed in a similar fashion of how DNA
formed (Sinden 1994): built up by four essential elements. As the DNA produces
phenotypes, the Neural Knowledge DNA carries information and knowledge via its
four essential elements, namely, States, Actions, Experiences, and Networks (Figure 1.).
3
Figure 1. Structure of the NK-DNA
The NK-DNA‟s four-element combination is also inspired by reinforcement
learning and Markov Decision Processes (Sutton & Barto 1998; Michael 2015): States
are situations in which a decision or a motion can be made or performed, Actions are
used to store the decisions or motions the domain can select. While Experiences are
domain‟s historical operation segments with feedbacks from outcomes, which stored as
et = (st, at, rt, st+1) at each time-step t. And Networks store the detail of neural networks
for training and using such knowledge, such as network structure, deep learning
framework used (if a third-party deep learning framework is used, like MxNet, Caffe,
etc.), and weights. Figure 2 shows a concept of the NK-DNA-carried knowledge.
4
Figure 2. Concept of the NK-DNA-carried knowledge
INITIAL EXPERIMENT
We examined our NK-DNA in a very simple maze problem. The NK-DNA supported
the agent to store the knowledge of the maze (Figure 3.).
In this initial experiment, the agent is expected to find the shortest path to block 8
from block 1. And the agent uses reinforcement learning (Sutton & Barto 1998;
Michael 2015) and ANN to learn the maze.
5
Figure 3.
The very simple maze
First, the States and Actions are pre-defined, for example, there are 8 states, and for
state 1, the agent can either “go to block 2” or “go to block 4” (i.e. the actions of state
1). Then, the agent starts from state 1, and randomly picks an action of its current state
to explore the maze as long as it reaches the state 8, and the agent gets a reward for
reaching state 8. Meanwhile, the agent stores every single movement (i.e. from one
state to another state) with reward from it as an Experience (st, at, rt, st+1) during its
exploring of the maze, and perform gradient descent to train the neural network that
indicates how wise an action is for a state. For more information about algorithms and
methods used in the agent, please refer to works (Lillicrap et al., 2016; Mnih et al.,
2015). Finally, after exploring the maze, the agent is trained, and its knowledge about
the maze is stored in the NK-DNA as Actions, States, Experiences, and Networks. This
allows the agent for sharing and reusing such knowledge in the future.
6
CONCLUSIONS AND FUTURE WORK
In this paper, we proposed the Neural Knowledge DNA, a framework adapting ideas
underlying the success of neural networks to knowledge representation for neural
network-based knowledge discovering, storing, reusing, improving, and sharing. By
taking advantages of neural networks and reinforcement learning, the NK-DNA stores
the knowledge learnt through domain‟s daily operation, and provides an easy way for
future accessing, reusing, and sharing such knowledge. At the end of this paper, we
tested our proposal idea in an initial experiment, and the results show that the NK-DNA
is very promising for knowledge representation, reuse, and sharing among neural
network-based AI systems.
For further work, we will do: 1) Refinement and further development of the Deep
Learning Engine; 2) Further design and development of the NK-DNA framework,
especially, for supporting a range of third-party deep learning frameworks; 3)
Exploring and developing methods for reasoning tasks.
REFERENCES
Davis, Randall, Howard Shrobe, and Peter Szolovits. "What is a knowledge
representation?" AI magazine 14.1 (1993): 17.
Drucker, Peter F. The new realities. Transaction publishers, 2011.
Gallant, Jack. "Brain reading: Mapping, modeling, and decoding the brain under
naturalistic conditions." 2015 AAAS Annual Meeting (12-16 February 2015). AAAS,
2015.
7
Lecun, Yann, et al. "Backpropagation applied to handwritten zip code recognition."
Neural Computation, 1.4(1989):541-551.
LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton, "Deep learning." Nature 521.7553
(2015): 436-444.
Lescroart, Mark D., Shinji Nishimoto, and Jack L. Gallant. "Representation of object
contour features in intermediate visual areas in the human brain." Journal of Vision 13.9
(2013): 1000-1000.
Li, B. M., Xie, S. Q., & Xu, X., 2011, „Recent development of knowledge-based
systems, methods and tools for one-of-a-kind production‟, Knowledge-Based Systems,
24(7), 1108-1119
Liao, S.H., 2003, 'Knowledge Management Technologies and Applications—
Literature Review from 1995 to 2002', Expert Systems with Applications, vol. 25, pp.
155-164.
Lillicrap, Timothy P., et al. "Continuous control with deep reinforcement learning."
arXiv preprint arXiv:1509.02971 (2015).
M. I. Jordan, and T. M. Mitchell. "Machine learning: Trends, perspectives, and
prospects." Science, 349(2015):255-260.
Matayong, S., & Mahmood, A. K., 2012, „The studies of Knowledge Management
System in organization: A systematic review‟, In Computer & Information Science
(ICCIS), 2012 International Conference, IEEE, Vol. 1, pp. 221-226
Michael A. Nielsen, "Neural Networks and Deep Learning", Determination Press, 2015
Michael L. Littman, "Reinforcement learning improves behavior from evaluative
feedback." Nature 521.7553 (2015): 445-451.
Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning."
Nature 518.7540 (2015): 529-533.
O'Dell, Carla S., and Cindy Hubert. The new edge in knowledge: how knowledge
management is changing the way we do business. John Wiley & Sons. 2011.
Oxford Dictionaries, 2015, „knowledge‟, Oxford University Press. Retrieved October
2015 from http://www.oxforddictionaries.com/definition/english/knowledge
Schmidhuber, Jürgen. "Deep learning in neural networks: An overview." Neural
Networks 61 (2015): 85-117.
Sinden, R.R., (1994) DNA Structure and Function. Academic Press: San Diego.
Sutton, R. S. & Barto, A. G. Reinforcement Learning: An Introduction. MIT Press,
1998.
8
| 9cs.NE
|
arXiv:1804.00489v2 [cs.PL] 3 Apr 2018
Robustly Safe Compilation
or, Efficient, Provably Secure Compilation
Marco Patrignani1
1
2
CISPA
MPI-SWS
Deepak Garg2
Saarbrücken, Germany
{marcopat | dg} @mpi-sws.org
Abstract
Secure compilers generate compiled code that withstands many targetlevel attacks such as alteration of control flow, data leaks or memory corruption. Many existing secure compilers are proven to be fully abstract,
meaning that they reflect and preserve observational equivalence. While
this is a strong property, it comes at the cost of requiring expensive runtime constructs in compiled code that may have no relevance for security,
but are needed to accommodate differences between the source language
and the target language.
As an alternative, this paper explores a different compiler security
criterion called robustly safe compilation or RSC . Briefly, this criterion
means that the compiled code preserves relevant safety properties of the
source program against all adversarial contexts. We show that RSC can be
attained easily and results in code that is much more efficient than that
generated by fully abstract compilers. We also develop two illustrative
RSC -attaining compilers and, through them, develop two different proof
techniques for establishing that a compiler attains RSC .
To better explain and clarify notions, this paper uses colours. For a better
experience, please print or view this paper in colours.
1
Introduction
Secure compilers generate target (low-level) code that can withstand common
target-level attacks such as alteration of control flow, improper stack manipulation and unauthorized reading/writing of private memory. To achieve secure compilation, different mechanisms have been employed: cryptographic
primitives [3, 4, 16, 20], types [7, 8], address space layout randomisation [5, 30],
protected module architectures [6, 42, 43] (also know as isolated enclaves [36]),
tagged architectures [32], etc. Once designed, secure compilers should be formally shown to be secure, by proving that they conform to some criterion of
compiler security. The most widely-used criterion for compiler security is fully
abstract compilation (FAC ) [2, 29, 41], which has been shown to preserve many
1
interesting security properties like confidentiality, integrity, invariant definitions,
well-bracketed control flow and local state [30, 42].
Recent work started questioning FAC as a criterion for compiler security.
In fact, it was argued that FAC is not well-suited for source languages with
undefined behaviour and a strong notion of compartments [32]. Moreover, if
used naively, FAC can fail to preserve even simple safety properties [44] (though,
fortunately, no existing work falls prey to this naivety). Additionally, because
FAC is concerned with preserving all source abstractions, it preserves also those
that are security irrelevant. This affects compiled code efficiency, as can be seen
in existing fully abstract compilers [5, 6, 30, 42, 43]. For example, compiled code
must often perform dynamic typechecks on all function arguments, even when
they play no security role, or ‘wrap’ objects in proxies in order to hide their
allocation addresses.
The reason for these limitations can be found in the definition of FAC .
Informally, a compiler attains FAC if it preserves and reflects observational
equivalence. Most existing work instantiates observational equivalence with
contextual equivalence: co-divergence of two programs in any program context
(i.e., in any larger program they interact with). This is a very strong property.
While it covers many relevant security properties, it also covers a plethora of
other non-security ones, and in order to preserve these, inefficient checks must
be inserted into compiled code by fully abstract compilers.
To overcome these limitations, recent work started investigating alternative
criteria in the form of preservation of hyperproperties or classes of hyperproperties, such as hypersafety properties or safety properties [27]. This paper proposes
a concrete instance of one of these criteria, namely, Robustly Safe Compilation
(RSC ). Informally, a compiler attains RSC if it is correct and if it preserves
robust safety of source programs. A program is robustly safe (RS) if it always respects relevant safety properties even in the presence of an adversary
actively trying to disrupt or attack it. A compiler attains RSC if it maps any
robustly safe source program to a robustly safe compiled program. We choose
preservation of the single-trace, unary robust safety as it has clear security implications [14]. Specifically, robust safety can express data integrity, weak secrecy
and even (approximations of) non-interference in the presence of active attacks.
A RSC -attaining compiler focuses only on preserving security (as captured
by RS) instead of contextual equivalence (typically captured by full abstraction).
So, such a compiler can produce code that is more efficient than code compiled
with a fully abstract compiler. Additionally as robust safety scales naturally to
thread-based concurrency, so does RSC (unlike FAC ) and we demonstrate this.
RSC is a very recently proposed criterion for secure compilers. Garg et
al. [27] define RSC abstractly in terms of preservation of program behaviours,
but their development is limited to the definition only. Our goal in this paper is
to examine how RSC can be realized and established, and show that it leads to
compiled code that is more efficient than what FAC leads to. Specifically, we
present two RSC -attaining compilers in different settings. The first compiler
compiles an untyped source language to an untyped target language with support for fine-grained memory protection via so-called capabilities [51]. Here, we
2
guarantee that if a source program is robustly safe, then so is its compilation.
On the other hand, the second compiler starts from a typed source language
where the types already enforce robust safety, and compiles to a target language similar to that of the first compiler. Here, we guarantee that all compiled
target programs are robustly safe. In doing so, we develop two different proof
techniques for establishing RSC for compilers, and show how types simplify the
proof of RSC in the second compiler, relative to the first.
The rest of the paper is organized as follows. After presenting FAC and
discussing its advantages and limitations (Section 2), we give the formal definition of RSC , comparing RSC and FAC for efficiency, providing details of the
security properties RSC preserves through compilation, and discussing proof
techniques for RSC (Section 3). Next, we present our two RSC -attaining compilers (Section 4 and Section 5). We discuss closely related work and further
directions for research in Section 6. Finally, we discuss remaining related work
(Section 7) and conclude (Section 8).
The interested reader can find formal details, full formalisations and proofs
in the appendix.
Conventions We use a blue, sans-serif font for source elements, an orange, bold
font for target elements and a black , italic font for elements common to both
languages (to avoid repeating similar definitions twice). Thus, C is a sourcelevel program, C is a target-level program and C is generic notation for either a
source-level or a target-level program. Note that we distinguish whole programs
or closed programs from partial programs or components. A partial program is a
unit of compilation, and must be linked with other partial program(s) to form a
whole program which can then be executed. The second partial program mentioned in the previous sentence, which is linked to a partial program of interest,
is often called a “context” (this is standard terminology). The word “program”
without qualification refers to a partial program, not a whole program.
2
Fully Abstract Compilation
This section formalises and discusses FAC . Then it explains why FAC results
in inefficient code, often pointlessly (Section 2.1), and why proof techniques for
establishing FAC are difficult to apply (Section 2.2).
As stated in Section 1, FAC amounts to showing the preservation and reflection of observational equivalence, and most existing work instantiates observational equivalence with contextual equivalence (≃ctx ). Contextual equivalence
and FAC are defined below. Informally, two components C1 and C2 are contextually equivalent if no context (linked component) C interacting with them can
tell them apart, i.e., they are indistinguishable. Contextual equivalence can encode security properties such as confidentiality, integrity, invariant maintenance
and non-interference [5, 6, 42, 44].
Definition 1 (Contextual equivalence ≃ctx ).
def
C1 ≃ctx C2 = ∀C. C [C1 ] ⇑ ⇐⇒ C [C2 ] ⇑
3
where C represents a context, i.e., a partial program with a hole that can be
filled by a component by linking the two together. Filling the hole in C with
component C is denoted as C [C ], and it results in a whole program that can be
run. Also, we indicate divergence, i.e., the execution of an unbounded number
of reduction steps, as ⇑.
S
In the following, we write J·KT for a compiler from language S to language T.
S
Informally, a compiler J·KT is fully abstract if it translates (only) contextuallyequivalent source components into contextually-equivalent target ones.
Definition 2 (Fully abstract compilation).
def
J·KST ∈ FAC = ∀C1 , C2 . C1 ≃ctx C2 ⇐⇒ JC1 KST ≃ctx JC2 KST
The security-relevant part of FAC is the ⇒ implication [22]. This part is
security-relevant because the proof thesis concerns target contextual equivalence
(≃ctx ). By unfolding the definition of ≃ctx , the universal quantification over
all possible target contexts C captures malicious attackers. In fact, there are
contexts C that can interact with compiled code in ways that may not be possible
in the source. A compiler that attains FAC towards untyped target languages
often inserts checks in compiled code that detect these interactions and respond
to them securely [44], often by halting execution [5, 6, 22, 30, 32, 33, 42, 43].
2.1
FAC and Inefficient Compiled Code
We introduce a running example to illustrate various ways in which FAC forces
inefficiencies in compiled code. Consider a password manager written in an
object-oriented source language that is being compiled to an assembly-like language. We elide most code details and focus only on the relevant aspects.
1
private db: Database;
2
3
4
5
6
7
8
9
public testPwd( user: Char[8], pwd: BitString): Bool{
if( db.contains( user )){
return db.get( user ).getPassword() == pwd;
}
}
...
private class Database{ ... }
The source program exports the function testPwd to check whether a user
stored password matches a given password pwd. The stored password is in
a local database, which is represented by a piece of local state in the variable db.
The details of db are not important here, but the database is marked private, so
it is not directly accessible to the context of this program in the source language.
user’s
Example 1 (Extensive checks). A fully-abstract compiler for the program
above must generate code that checks that the arguments passed to testPwd
by the context are of the right type [6, 26, 32, 42, 43]. More precisely, a fully
abstract compiler will generate code similar to the following for testPwd (we
assume that arrays are passed as pointers into the heap).
4
1
2
3
4
5
6
label testpwd
for i = 0; i< 8; i++
add r0 i
load the memory word stored at address r0 into r1
test that r1 is a valid char encoding
...
Basically, this code dynamically checks that the first argument is a character
array. Such a check can be very inefficient. The problem here is that FAC forces
these checks on all arguments, even those that have no security relevance.
Example 2 (Component size in memory). Let us now consider two different
ways to implement the Database class: as a List and as a RedBlackTree . As
the class is private , its internal behaviour and representation of the database is
invisible to the outside. Let Clist be the program with the List implementation
and Ctree be the program with the RedBlackTree implementation; in the source
language, these are equivalent.
However, a subtlety arises when considering the assembly-level, compiled
counterparts of Clist and Ctree : the code of a RedBlackTree implementation consumes much more memory than the code of a List implementation. Thus, a
target-level context can distinguish Clist from Ctree by just inspecting the sizes
of the code segments. So, in order for J·KST to be fully abstract, it must produce
code of a fixed size [6, 42, 43]. This wastes memory and makes it impossible to
compile some components. An alternative would be to spread the components
in an overly-large memory at random places (i.e., use ASLR) so that distinguishing based on size has negligible chance of success [5, 30]. However, ASLR
is now known to be broken [10, 31].
Again, we see that FAC introduces an inefficiency in compiled code (pointless
memory consumption) even though this has no security implication here.
Example 3 (Wrappers for heap resources). Assume that the Database class is
implemented as a List. Shown below are two implementations of the newList
method inside List which we call Cone and Ctwo . The only difference between
Cone and Ctwo is that Ctwo allocates two lists internally; one of these (shadow ) is
used for internal purposes only.
1
2
3
4
1
2
3
4
5
6
public newList(): List{
ell = new List();
return ell;
}
public newList(): List{
shadow = new List();
ell = new List();
...
return ell;
}
Again, Cone and Ctwo are equivalent in the source. To get FAC , the pointers
returned by newList in the two implementations must be the same, but this is
5
very difficult to ensure since the second implementation does more allocations.
A simple solution to this problem is to wrap ell in a proxy object and return the
proxy [6, 37, 42, 43]. Compiled code needs to maintain a lookup table mapping
the proxy to the original object. Proxies must have allocation-independent
addresses.
While proxies work, they are inefficient due to the need to look up the table
every time a list is accessed. The overhead becomes more prominent when the
proxied object is of a primitive type that could be accessed directly via low-level
reads and writes much more efficiently (e.g., if the object is a boolean).
Another way to regain FAC is to weaken the source language, introducing
an operation to distinguish object identities in the source [41]. However, this
is a widely discouraged praxis, as it changes the source language from what it
really is and the implication of such a change may be difficult to fathom for
programmers and verifiers.
Example 4 (Strict termination vs divergence). Consider a source language that
is strictly-terminating while a target language that is not. Below is an extension
of the password manager to allow database encryption via an externally-defined
function. As the database is not directly accessible from external code, the two
implementations below Cenc (which does the encryption) and Cskip which skips
the encryption are equivalent in the source.
1
2
3
4
1
2
3
public encryptDB( func : Database
func( this.db );
return;
}
→Bitstring) : void {
public encryptDB( func : Database
return;
}
→Bitstring) : void {
If we compile Cenc and Cskip to an assembly language, the compiled counterparts cannot be equivalent, since the target-level context can detect which
function is compiled by passing a func that diverges. Calling the compilation
of Cenc with such a func will cause divergence, while calling the compilation of
Cskip will not cause divergence.
This case presents a situation where FAC is outright impossible. The only
way to get FAC is to make the source language artificially non-terminating [22].
(See [23] for more details of this particular problem.)
Remark It is worth noting that many of the inefficiencies above could be
resolved by just replacing contextual equivalence with a different equivalence
in the statement of FAC . However, it is not known how to do this generally
for arbitrary sources of inefficiency and, further, it is unclear what the security
consequences of such instantiations of FAC would be. On the other hand, RSC
is uniform and it does address all these inefficiencies.
An issue that can normally not be added just by tweaking equivalences is side
channels, as they are by definition not expressible in the language. Both FAC
6
and RSC do not deal with side channels, incorporating side channel defences in
secure compilers is future work.
2.2
Proving FAC
Most papers that define fully-abstract compilers also prove that they attain
FAC , as a means to ensure that the compiler is secure. When proving that
a compiler is fully abstract, reasoning about target contexts C is key: the behaviour of all target contexts must be shown to be expressible in the source
too. If this is possible, then a target-level attacker does not intuitively have any
power that no source context has. This proof strategy is often called back translation, and it is a part of all common proof techniques (bisimulations, logical
relations, trace semantics, etc.) to establish full abstraction [22, 26, 32, 39, 42].
Let us consider the hard part of FAC , the ⇒ implication. The statement
S
S
C1 ≃ctx C2 ⇒ JC1 KT ≃ctx JC2 KT
can be stated in the contrapositive as
S
S
JC1 KT 6≃ctx JC2 KT ⇒ C1 6≃ctx C2 .
By unfolding the definition of 6≃ctx we see that, given a target context C that
S
S
distinguishes JC1 KT from JC2 KT , it is necessary to show that there exists a source
context C that distinguishes C1 from C2 . That source context C must be built
(back-translated) starting from the already given target context C.
Example 4 shows the limitations of FAC when it comes to back-translation
because it is not possible to translate an arbitrary target context C, which could
possibly be divergent, to a source context C, which must be terminating. This
is not the only possible instance where back-translation will not work. The
existence of any major differences between the source and target languages can
also cause the back-translation technique (and thus FAC ) to fail or become very
difficult to use. For example, if the source language does not have a heap but
the target language does, then too the back translation will be impossible or
very difficult to define.
Having described how FAC can force inefficiency into compiled code (sometimes pointlessly from the perspective of security), and how proof techniques
for FAC can be difficult, we now present RSC , which does not force similar
inefficiency into compiled code and which can be easier to establish.
3
Robustly Safe Compilation
This section presents an overview of RSC . First, it discusses robust safety (RS )
as a language (not a compiler) property (Section 3.1). Then, it presents RSC
along with techinques to attain and prove it for a compiler (Section 3.2).
7
3.1
Safety and Robust Safety
Safety is a general security notion which asserts that bad behaviour does not
happen. Historically, safety was devised to reason about the absence of program
bugs in whole programs. Robust safety lifts the notion of safety to partial
programs that are co-linked with contexts, which may possibly be malicious
attackers [9,12,17,25,28,35]. A component (note the change from whole program
before) is robustly safe if its composition with an arbitrary attacker is safe.
Here, an attacker is any context that never has bad behaviour in its own code.
Thus, robust safety captures all safety properties even in the presence of an
active attacker. This includes some forms of integrity, weak secrecy and even
approximations of information flow properties [14].
Our formalization of robust safety We need some representation of safety
properties for our technical development. We find it convenient to follow Schneider’s representation of safety properties as security automata [45]. Briefly, a security automaton is a monitor that co-executes with a program. The automaton
is a state machine that transitions states in response to specific security-relevant
events of the program it monitors. By definition, the program is in violation
of the safety property coded by the security automaton if the automaton is unable to make a transition in response to an event. Schneider argues that all
properties codable this way are safety properties and that all enforceable safety
properties can be coded this way.
To keep things simple, we only consider properties of the heap of the monitored program. For example, if the program has an internal timestamp counter,
the property of interest might be that the counter only increases in value over
time. Our security automaton, also called the monitor, is an abstract state
machine that gets to transition its internal state whenever the program relinquishes control via a special command called monitor . Whenever this command
is executed, the monitor transitions its internal state based on the contents of
the current heap. A property violation happens when the monitor has no valid
transition.
We use H to denote a heap, which for now is a map from locations l to
values v (Section 5 uses other instances of heaps). Formally, a monitor M
consists of a set of absract states {s}, the transition relation , an initial state
s0 , the set of heap locations that matter for the monitor, {l }, and the current
state sc . The transition relation
is a set of triples of the form (ss , H , sf )
consisting of a starting state ss , a final state sf and a heap H . The transition
(ss , H , sf ) is interpreted as “The state ss transitions to sf when the heap is H ”.
The heap H occurring in a transition must contain in its domain only locations
from {l } (the set that the monitor cares about). We assume that the transition
relation is deterministic: for any ss and H , there is at most one sf such that
(ss , H , sf ) ∈ .
A program C executing with a monitor M and a heap H is written as
M , H ⊲ C . Program execution is represented as usual with a small step “stepsto” relation → (and →∗ is the reflexive-transitive closure of this relation). Pro8
grams can have a special instruction monitor that atomically invokes the monitor. Based on the current state of the monitor and the current heap, the monitor
either has a (unique) valid transition or it does not have any transition. In the
former case, formalized by Rule Abstract-monitor below, the internal state of
the monitor is updated. In the latter case, the whole system enters a special
state fail , with no possibility of recovery (Rule Abstract-monitor-fail).
(Abstract-monitor)
M = ({s} , , s0 , {l } , sc )
sc , H {l}
sf
M ′ = ({s} , , s0 , {l } , sf )
M , H ⊲ monitor −→ M ′ , H ⊲ skip
(Abstract-monitor-fail)
M = ({s} ,
, s0 , {l } , sc )
sc , H
{l}
6
_
M , H ⊲ monitor −→ fail
With this setup in place, we can give abstract definitions of safety (Definition 8), attackers (Definition 4) and robust safety (Definition 10). For the sake
of readability, we elide and simplify some notation. The precise definitions can
be found in the appendix.
Definition 3 (Safety). Let H0 be the initial heap for C .
def
⊢ C, M : safe = ⊢ C : whole and M , H0 ⊲ C 6→∗ fail
A program C is safe wrt monitor M if C is whole and its execution with M
never reaches fail (i.e., M never gets stuck during the execution).
Definition 4 (Attacker).
def
M ⊢ A : attacker = no monitor inside A and no location
of M reachable or stored in A
An attacker is any program (context) with no occurrence of monitor in it.
Definition 5 (Robust Safety).
def
⊢ C, M : rs = ∀A. if M ⊢ A : attacker
then ⊢ A [C] , M : safe
A component C is robustly safe wrt monitor M if C composed with any
attacker is safe wrt M .
3.2
Robustly Safe Compilation
A compiler is robustly safe if (i) it is correct, and (ii) given any source component that is robustly safe, its compiled counterpart is robustly safe (wrt target
contexts). This informal definition consists of two parts, one regarding correctness and one regarding security. For the sake of simplicity, in this paper when
9
we talk about RSC we mostly refer to the second part, the first part being a
given, since it makes little sense to talk about the security of an incorrect compiler. (Of course, we prove both properties for all the compilers that we define
in this paper.)
3.2.1
Compiler Correctness and Cross-Language Relations
A compiler is correct if it compiles whole programs to whole programs that have
the same behaviour. Unlike most of the literature, we formalise having the same
behaviour as terminating with heaps that contain related values. In order to do
this, we assume a relation ≈ : v × v between source and target values that is
total, so it maps any source value v to a target value v: ∀v.∃v.v ≈ v. This value
relation is used to define a relation between heaps: H ≈ H, which intuitively
holds when related locations point to related values.
Next, we assume a relation M ≈ M between source and target monitors,
which means that the source monitor M and the target monitor M code the
same safety property, modulo the relation ≈ on values assumed above. The
precise definition of this relation depends on the source and target languages;
specific instances are shown in Sections 4.3.2 and 5.3.
Equipped with these two ≈ relations, we can define compiler correctness.
Definition 6 (Correct Compilation [34]).
S
def
⊢ J·KT : CC = if M ≈ M
S
and M, ∅ ⊲ JsKT −→∗ M′ , H ⊲ skip
then M, ∅ ⊲ s −→∗ M′ , H ⊲ skip
and H ≈ H
S
A compiler J·KT is correct (CC ) if whenever a compiled whole program terminates with a heap H, the source counterpart reduces to a related heap H.
3.2.2
Compiler Security
The security-relevant part of robustly safe compilation is defined below.
Definition 7 (Robustly safe Compilation).
S
def
⊢ J·KT : RSC = if M ≈ M
and ⊢ C, M : rs
S
S
then ⊢ JCKT , M : rs
A compiler J·KT attains RSC , if it maps any program C that is robustly safe
wrt M to a program C that is robustly safe wrt M, where M ≈ M.
Note that RSC is defined as a single implication (⇒), unlike FAC that is
a bi-implication ( ⇐⇒ ). This is because the backwards direction (⇐) does
10
not convey security aspects. It would only describe what to do in case of good
behaviours, but this is already described by compiler correctness.
Next, we discuss at a high-level how RSC can be established for a compiler.
We discuss two techniques, that we illustrate on concrete instances in later
sections. The later sections also show how an RSC -attaining compiler can
generate more efficient code than a FAC -attaining compiler.
Proving RSC Proving that a compiler attains RSC can be done by either
adapting existing techniques for establishing FAC , or with simpler ones. As
discussed in Section 2.2, FAC proofs often rely on back-translations of contexts.
If we unfold Definition 15 and then the two instances of Definition 10 inside it,
we see that we have the following assumptions
1. if M, ∅ ⊲ A [P] →∗ M′ , H ⊲ A [monitor] then M′ , H ⊲ A [monitor] → M′ , H ⊲ A [skip]
(for all A).
2. For some A: M, ∅ ⊲ A [P] →∗ M′ , H ⊲ A [monitor]
From these assumptions, we need to prove that
• M′ , H ⊲ A [monitor] → M′ , H ⊲ A [skip]
One way to prove this is by back-translation. By contrapositivity, assume
that the goal does not hold for a given target context A. Then, by backtranslating A, we can obtain a source context A that violates (1). We illustrate
this proof strategy in Section 4.3.2. However, as for FAC , back-translation
based proofs can be quite difficult for RSC .
However, a simpler proof strategy is also viable for RSC when we compile
only those source programs that have been verified to be robustly safe (e.g., using
a type system). The idea is this: using the verification of the source program,
we establish an invariant which is always maintained by the target code, and
which, in turn, implies the robust safety of the target code. For example, if the
safety property is that values in the heap always have their expected types, then
the invariant can simply be that values in the target heap are always related
to the source ones (which have their expected types). This is tantamount to
proving type preservation in the target in the presence of an active adversary.
This is harder than standard type preservation (because of the active adversary)
but is still much easier than back-translation as there is no need to relate target
constructs to source ones. We illustrate this proof technique in Section 5.
3.2.3
Compiling monitors
In our development, we assume that a source and a target monitor are related
but do not actually compile a source monitor to obtain a related target monitor.
While such compilation is feasible, it is at odds with our view of monitors
as specifications of safety properties. Compiling monitors and, in particular,
compiling monitors with the same compiler that we want to prove security
of, leads to a circularity—we must understand the compiler to understand the
11
target safety property, which, in turn, acts as the specification for the compiler!
Consequently, we choose not to compile monitors and talk only of an abstract,
compiler-independent relation between source and target monitors.
4
RSC : First instance
This section presents an instance of an RSC -attaining compiler. First, we
present our source language LU , a simply typed imperative language. Its heap
addresses are abstract and cannot be guessed by an adversary, which provides
support for hidden local state in components. LU includes monitors that enforce
safety properties over the heap (Appendix E). Then, we present our target language LP , an untyped imperative target language with a concrete heap, whose
addresses are natural numbers. LP provides hidden local state via a fine-grained
capability mechanism on heap accesses (Appendix B). Finally, we present the
LU
compiler J·KLP and prove that it attains RSC (Section 4.3).
The languages of this section are deliberately simple. RSC for richer languages is discussed in Section 6. Due to space constraints, only excerpts of
syntax, semantics, the compiler definition, etc. are presented here, while proofs
and helper lemmas are elided entirely. Some aspects of the formalisation are
also simplified for presentation. All the missing details are in the appendix.
4.1
The Source Language LU
LU is an untyped imperative “while” language [40] (Figures 1 and 2). Components C are collections of function definitions F together with a monitor M and
the interfaces I. An interface is a context-defined function that the component
relies on (similar to C’s extern functions). Attackers A (also called program
contexts) represent untrusted code that a component interacts with. Function
bodies include standard boolean and arithmetic operations, generically written
⊕, heap manipulation statements, (recursive) function calls, etc. Expression
evaluation is defined by the small-step reduction relation ֒→
→.
A program state Ω includes the monitor, the heap and a statement being
executed. State evaluation, the main program semantics, are defined by the
reduction relation ( −→ ). Monitors M work as explained in Section 3.1, except
that the description of a monitor includes a single location of interest ℓroot . All
locations reachable directly or indirectly by pointer chasing starting from ℓroot
are assumed to be of interest to the monitor and function reach(ℓroot , H) returns
them for a heap H. In practice, ℓroot may, for instance, be the root of a data
structure that the component being monitored wants to have a safety property
on.1
Heaps H are maps from abstract locations ℓ to values v. The reduction semantics of states is labelled (labels are the annotations on −→ ). The labels
1 The use of a single root location is arbitrary and motivated by simplicity. This can be
easily extended to a set of locations.
12
Components C ::= M; F; I
Attackers A ::= H; F [·]
Interfaces I ::= f
Functions F ::= f(x) 7→ s; return;
Values v ::= true | false | n ∈ N | hv, vi | ℓ
Expressions e ::= x | v | e ⊕ e | e ⊗ e | he, ei | e.1 | e.2 | !e
Statements s ::= skip | s; s | let x = e in s | monitor
| call f e | let x = new e in s
| x := e | if e then s else s
Prog. States Ω ::= C, H ⊲ s | fail
Monitors M ::= ({σ} ,
Mon. States σ ∈ S
Mon. Trans.
::= ∅ |
, σ0 , ℓroot , σc )
; (s, H, s)
Figure 1: Syntax of LU (excerpts). A list of element e1 , · · · , en is indicated as
e.
record function calls and returns that occur between a component and the context implementing its interfaces. This is a standard technical device that is used
in the proofs (see Section 4.3.2). Readers not interested in understanding our
proofs may ignore these labels.
Ω0 (C) denotes the initial state of a component C. In this state, the heap is
empty, the monitor M is in its initial state and the executing statement is the
first line of a designated entry point (akin to C’s main function).
4.2
The Target Language LP
LP is an untyped, imperative language that follows the structure of LU (Figures 3
and 4); it has reduction for expressions ( ֒→
→ ) and a small-step semantics ( −→ ).
However, there are critical differences (to make the compiler interesting!).
The main difference is that heap locations in LP are concrete natural numbers. An adversarial context can guess locations used as private state by a
component and clobber them. To support hidden local state, a location can be
“hidden” explicitly via the expression hide n. This expression allocates a new
capability k, an abstract token that grants access to the location n [46]. Subsequently, all reads and writes to n must be authenticated with the capability.
Unlike locations, capabilities cannot be guessed. To hide a location, a component hides the capability of the location. To bootstrap this hiding process,
we assume that each component has one location that can only be accessed by
it, a priori in the semantics (in our formalization, we always focus on only one
component and we assume that, for this component, this special location is at
13
(ELU -op)
′
′′
(ELU -val)
n⊕n =n
H ⊲ n ⊕ n′ ֒→
→ n′′
H ⊲ v ֒→
→ v
(ELU -dereference)
H ⊲ e ֒→
→ ℓ
ℓ 7→ v ∈ H
H⊲!e ֒→
→ v
(ELU -letin)
H ⊲ e ֒→
→ v
C, H ⊲ let x = e in s −→ C, H ⊲ s[v / x]
(ELU -alloc)
H ⊲ e ֒→
→ v
ℓ∈
/ dom(H)
C, H ⊲ let x = new e in s −→ C, H; ℓ 7→ v ⊲ s[ℓ / x]
(ELU -update)
H ⊲ e ֒→
→ v
H = H1 ; ℓ 7→ v′ ; H2
H′ = H1 ; ℓ 7→ v; H2
C, H ⊲ ℓ := e −→ C, H′ ⊲ skip
(ELU -call)
f(x) 7→ s; return; ∈ C.funs
H ⊲ e ֒→
→ v
C, H ⊲ call f e −→ C, H ⊲ s; return;[v / x]
(ELU -monitor)
(ELU -monitor-fail)
M = ({s} , , s0 , ℓroot , sc )
C = M; F; I
C′ = M′ ; F; I
M; H
M′
C, H ⊲ monitor −→ C′ , H ⊲ skip
M = ({s} , , s0 , ℓroot , sc )
C = M; F; I
M; H 6
C, H ⊲ monitor −→ fail
(LU -Monitor Step)
M = ({σ} , , σ0 , ℓroot , σc )
M′ = ({σ} , , σ0 , ℓroot , σf )
(σc , H′ , σf ) ∈
H′ ⊆ H
dom(H′ ) = reach(ℓroot , H)
M; H
M′
Figure 2: Semantics of LU (excerpts). A list of element e1 , · · · , en is indicated
as e.
address 0).
In detail, LP heaps H are maps from natural numbers n to values v and
a tag η. The tag η can be ⊥, which means that n is globally available (not
protected) or a capability k, which protects n. A globally available location
can be freely read and written (Rule ELP -assign-top, Rule ELP -deref-top) but
one that is protected by a capability requires the capability to be supplied at
the time of read/write for the operation to go through (Rule ELP -assign-k,
14
Values v ::= n ∈ N | hv, vi | k
Expressions e ::= x | v | e ⊕ e | e ⊗ e | he, ei
| e.1 | e.2 | !e with e
Statements s ::= skip | s; s | let x = new e in s
| x := e with e | let x = e in s
| let x = hide e in s | call f e
| monitor | ifz e then s else s
Heaps H ::= ∅ | H; n 7→ v : η | H; k
Tag η ::= ⊥ | k
Monitors M ::= ({σ} ,
, σ0 , σc )
Figure 3: Syntax of LP (excerpts).
(ELP -step)
(ELP -sequence)
C, H ⊲ s −→ C, H ⊲ s′
C, H ⊲ s; s′′ −→ C, H ⊲ s′ ; s
C, H ⊲ skip; s −→ C, H ⊲ s
(ELP -deref-any)
(ELP -deref-k)
n 7→ v : ⊥ ∈ H
H⊲!n with _ ֒→
→ H⊲v
n 7→ (v, k) ∈ H
H⊲!n with k ֒→
→ H⊲v
(ELP -new)
H = H1 ; n 7→ (v, η)
H ⊲ e ֒→
→ v
H′ = H; n + 1 7→ v : ⊥
C, H ⊲ let x = new e in s −→ C, H′ ⊲ s[n + 1 / x]
(ELP -hide)
H ⊲ e ֒→
→ n
k∈
/ dom(H)
H = H1 ; n 7→ v : ⊥; H2
H′ = H1 ; n 7→ v : k; H2 ; k
C, H ⊲ let x = hide e in s −→ C, H′ ⊲ s[k / x]
(ELP -assign-any)
H ⊲ e ֒→
→ v
H = H1 ; n 7→ _ : ⊥; H2
H′ = H1 ; n 7→ v : ⊥; H2
C, H ⊲ n := e with _ −→ C, H′ ⊲ skip
(ELP -assign-k)
H ⊲ e ֒→
→ v
H = H1 ; n 7→ _ : k; H2
H′ = H1 ; n 7→ v : k; H2
C, H ⊲ n := e with k −→ C, H′ ⊲ skip
Figure 4: Semantics of LP (excerpts).
Rule ELP -deref-k). The locations of interest to a monitor are all those that can
be reached from the address 0. 0 itself is protected with a capability kroot that
15
is assumed to occur only in the code of the component in focus.
A second difference between LP and LU is that LP has no booleans, while
U
L has them.
4.3
Compiler from LU to LP
LU
This section presents the compiler J·KLP from LU to LP , detailing how it uses
LU
the capabilities of LP to achieve RSC (Section 4.3.1). Then, it proves that J·KLP
LU
attains RSC (Section 4.3.2). Finally, it explains why code compiled with J·KLP
does not have the inefficiencies that FAC would force (Section 4.3.3).
4.3.1
LU
The compiler J·KLP
U
The compiler J·KLLP takes as input a source component C and a target monitor
LU
M that is related to the monitor in C, so J·KLP : C × M → C. The compiler
performs a simple pass on the structure of functions, statements and expressions (Figure 5). Each LU location is encoded as a pair of a LP location and
LU
the capability to access the location (Rule (J·KLP -New)). Location update and
LU
LU
dereference are compiled accordingly (Rules (J·KLP -Assign) and (J·KLP -Deref)).
4.3.2
Proof of RSC
LU
The compiler J·KLP is robustly safe (Theorem 7 and Theorem 8).
LU
LU
Theorem 1 (J·KLP is cc). ⊢ J·KLP : CC
In order to set up this theorem, we need to instantiate the cross-language
relation for values, which we write as ≈β . The relation is parametrised by a
partial bijection β : ℓ × n× η from source heap locations to target heap locations
such that:
• if (ℓ1 , n, η) ∈ β and (ℓ2 , n, η) then ℓ1 = ℓ2 ;
• if (ℓ, n1 , η1 ) ∈ β and (ℓ, n2 , η2 ) then n1 = n2 and η1 = η2 .
The bijection determines when a source location and a target location are related.
The partial bijection β grows as we consider successive steps of program execution in our proof. For example, if executing let x = new e in s creates location
some source ℓ, then executing its compiled counterpart will create some target
location n. At this point we add (ℓ, n, ⊥) to β.
More generally, ≈β is defined as follows:
• true ≈β 0;
• false ≈β n such that n 6= 0;
16
q
yL U
q yLU q yLU
M; F; I, M LP = M; F LP ; I LP
LU
LU
if M ≈β M
(J·KLP -Comp)
LU
LU
(J·KLP -Fun)
Jf(x) 7→ s; return;KLP = f (x) 7→ JsKLP ; return;
LU
LU
(J·KLP -True)
JtrueKLP = 0
LU
LU
JfalseKLP = 1
(J·KLP -False)
JxKLP = x
(J·KLP -Var)
U
JnKLLP = n
U
(J·KLLP -nat)
LU
LU
LU
LU
LU
LU
(J·KLP -Deref)
J!eKLP = !JeKLP .1 with JeKLP .2
U
U
U
U
Jlet x = e in sKLLP = let x = JeKLLP in JsKLLP
LU
LU
L
(J·KL
P -Letin)
LU
LU
Jif e then st else se KLP = ifz JeKLP then Jst KLP else Jse KLP
t
|LU
let x = new e
LU
= let xloc = new JeKLP in
in s
LP
let xcap = hide xloc in
LU
(J·KLP -If)
LU
(J·KLP -New)
LU
LU
let x = hxloc , xcap i in JsKLP
Jx := e′ KLP = let xloc = x.1 in
LU
(J·KLP -Assign)
let xcap = x.2 in
LU
U
xloc := Je′ KLP with xcap
U
Jcall f eKLLP = call f JeKLLP
LU
JmonitorKLP = monitor
U
(J·KLLP -call)
LU
(J·KLP -Mon)
LU
• ℓ ≈β hn, vi if
(
Figure 5: J·KLP excerpts.
v = k and (ℓ, n, k) ∈ β
v 6= k and (ℓ, n, ⊥) ∈ β
• hv1 , v2 i ≈β hv1 , v2 i if v1 ≈β v1 and v2 ≈β v2 .
This relation is used to define the heap, monitor and state relations. Heaps
are related, written H ≈β H, when related locations in β point to related values.
States are related, written Ω ≈β Ω, when they have related monitors and heaps.
Monitor Relation The monitors in these languages are more concrete than
what is described in Section 3.2. Two monitors are related when they can
simulate each other on related heaps. Given a monitor-specific relation σ ≈ σ
on monitor states, we say that a relation R on source and target monitors is
17
a bisimulation if the following hold whenever M = ({σ} ,
M = ({σ} , , σ0 , σc ) are related by R:
, σ0 , ℓroot , σc ) and
1. σ0 ≈ σ0 , and
2. σc ≈ σc , and
3. For all β containing (ℓroot , 0, kroot ) and all H, H with H ≈β H the following
hold:
(a) (σc , H, _) ∈
′
(b) (σc , H, σ ) ∈
iff (σc , H, _) ∈
′
and (σc , H, σ ) ∈
, and
imply ({σ},
, ℓroot , σ ′ )R({σ},
, σ0 , σ ′ ).
In words, R is a bisimulation only if MRM implies that M and M simulate each
other on heaps related by any β.2 Note that the union of any two bisimulations
is a bisimulation. Hence, there is a largest bisimulation, which we denote as ≈.
Intuitively, M ≈ M implies that M and M encode the same safety property (up
to the relation on values).
LU
LU
Theorem 2 (J·KLP is rs-pres ). ⊢ J·KLP : RSC
As a technical device for this proof, both LU and LP are equipped with a
α
labelled semantics, written ==⇒ . Given a component C and a context A, the
labelled semantics tracks transfers of control from the context to the component
(decorated with a ? as in α?) and transfers of control from the component to
the context (decorated with a ! as in α!). These boundary-crossing transfers of
controls are usually called “actions”. We denote actions with α.
Actions α ::= call f v H ? | call f v H !
| ret H ! | ret H ?
We call the actions above call, callback, return and returnback, respectively. The
H in an action is the heap at the time of the reduction on which the action is
recorded. Finite lists of actions are called traces, written α. Well-formed traces
have alternations of ! and ? decorated actions, starting with ? since execution
starts in the context.
We now outline our proof of Theorem 8. As mentioned in Section 3.2.2,
unfolding the definition of RSC and then of robust safety in LU and LP we
obtain the following two assumptions.
α
1. if M, ∅ ⊲ A [C] ==⇒ M′ , H ⊲ A [monitor] then M′ , H ⊲ A [monitor] −→ M′ , H ⊲ A [skip]
(for all A).
i
h
α
LU
2. For some A: M, ∅ ⊲ A JCKLP ==⇒ M′ , H ⊲ A [monitor].
2 In particular, this means that neither M nor M can be sensitive to the specific addresses
allocated during the run of the program. However, they can be sensitive to the “shape” of the
heap or the values stored in the heap.
18
From these assumptions, we need to prove that
• M′ , H ⊲ A [monitor] −→ M′ , H ⊲ A [skip]
LP
To complete this proof, we define a back-translation hh·iiLU that takes a target
trace α and builds a set of source contexts such that one of them when linked
C, produces a related trace α in the source (Theorem 9).
LP
Theorem 3 (hh·iiLU is correct).
i
h
α
LU
if A JCKLP ==⇒ Ω
P
α
=⇒ Ω and α ≈β α and Ω ≈β Ω.
then ∃A ∈ hhαiiL
LU . A [C] =
In prior work, back-translations return a single context [7,8,15,21,39,42,43].
This is because they all, explicitly or implicitly, assume that ≈ is injective from
source to target. Under this assumption, the back-translation is unique: a target
value v will be related to at most one source value v. We do away with this
assumption and thus there can be multiple source values related to any given
target value. This results in a set of back-translated contexts, of which at least
one will reproduce the trace as we need it.
To prove Theorem 8, we apply the back-translation to the target trace in the
assumption (2). For the specific source context A which reproduces the trace
in the source, the states in the source remain in sync (related) with states in
the target (by Theorem 9). Hence, M′ , H ⊲ A [monitor] ≈β M′ , H ⊲ A [monitor]
for some β. This implies that M′ ≈ M′ , and H ≈β H. By definition of M′ ≈ M′ ,
this implies that either both monitors must step or neither will. Since, we know
from (1) that M′ steps, it follows that M′ must step too, which immediately
yields the required statement.
Comparison to proofs of FAC Conceptually, our back-translation is similar
to that typically needed for FAC [6,32,42,43]. Overall, our proof is simpler than
a typical proof of FAC since establishing full abstraction requires showing that
the trace semantics of the target are complete with respect to observational
equivalence, whereas RSC does not need this (often difficult) step.
4.3.3
Examples of RSC not forcing inefficiency
LU
Having established that J·KLP attains RSC , we now take a look at how code
LU
generated by J·KLP does not have the inefficiencies of a fully-abstract compiler
(Section 2.1).
Extensive checks We saw in Example 1 that FAC forces dynamic type checks
on the parameters of all exported functions. In contrast, RSC does not need
LU
these checks. Indeed, J·KLP does not insert them. Note that any robustly safe
source program will have programmer-inserted checks for all parameters that
19
are relevant to the safety property of interest, and these checks will be compiled
to the target. For other parameters, the checks are irrelevant, both in the source
and the target, so there is no need to insert them.
Component size In Example 2, we saw that FAC forces all components to
have sizes independent of their code. In contrast, RSC does not require this
unless the safety property(ies) of interest care about the size of the code (which
is very unlikely in a security context). In particular, the monitors considered
here cannot depend on code size.
Wrappers for heap resources In Example 3, we explained that if the target
language allows address comparison, then FAC forces all privately allocated
locations to be wrapped in proxies. RSC does not require this. Our target
language, LP , supports address comparison (addresses are natural numbers in
LU
LP ) but J·KLP uses just capabilities to attain security efficiently. On the other
hand, for attaining FAC , capabilities alone would be insufficient since they do
not hide addresses; proxies would still be required.
Strict termination vs divergence We saw in Example 4 that FAC is usually impossible to attain when the source language is higher-order and strictly
terminating but the target language is not strictly terminating. On the other
hand, RSC can be easily attained even in such settings since it is completely
independent of termination in the languages (note that program termination
and nontermination are both different from the monitor getting stuck, which
is what RSC cares about). Indeed, if our source language LU were restricted
to terminating programs only, the same compiler and the same proof of RSC
would still work.
5
RSC : Second Instance with Simpler Proofs
When the source language has a verification system that enforces robust safety,
proving that a compiler attains FAC is even simpler than that in Section 4—it
does not require back translation. To demonstrate this, we consider a specific
class of monitors, namely those that enforce type invariants on a specific set of
locations. Our source language, Lτ , is similar to LU but it has a type system that
Lτ
ensures that the monitor never fails in the source. Our compiler J·KLπ is directed
by typing derivations, and its proof of FAC follows by establishing a specific
cross-language invariant on program execution, rather than a back-translation.
A second, independent goal of this section is to show that FAC is compatible with concurrency. Consequently, our source and target languages include
constructs for forking threads.
20
5.1
The Source Language Lτ
Lτ extends LU with concurrency—it has a fork statement ({k s}), processes and
process soups [13]—and with a more extensive type system (Figures 6 and 7).
Statements s ::= · · · | {k s} | endorse x = e as ϕ in s
Types τ ::= Bool | Nat | τ × τ | Ref τ | UN
Superf . Types ϕ ::= Bool | Nat | UN × UN | Ref UN
Heaps H ::= ∅ | H; ℓ 7→ v : τ
Monitors M ::= ({σ} , , σ0 , ∆, σc )
Mon. Trans.
::= ∅ | ; (s, s)
Envs. Γ ::= ∅ | Γ; (x : τ )
Store Env . ∆ ::= ∅ | ∆; (ℓ : τ )
Processes π ::= (s)
Soups Π ::= ∅ | Π k π
Prog. States Ω ::= C, H ⊲ Π | fail
Figure 6: Syntax of Lτ (excerpts).
The type system for Lτ is a standard type system that provides robust type
safety [1,9,25,28,35]: no context can cause the static types of values in the heap
to be violated at runtime.
The type system includes a type called UN, which stands for “untrusted”.
A location has type UN if the component of interest does not care about the
contents of the location. A value has type UN if it is impossible to extract a
location that the component cares about from the value. Overall, UN contains all
values that the context (adversary) may possess or manipulate safely. We write
τ ⊢ ◦ to mean that values of type τ are safe to share with the adversary, i.e., τ
is a subtype of UN. (See Rule TLτ -coercion.) The language Lτ further contains
an endorsement statement that dynamically checks the top-level constructor of
a value of type UN (potentially coming from the adversary) and gives it a more
precise superficial type ϕ [18]. This is similar to existing type casts [38] but it
only inspects one structural layer of the type (for simplicity of compilation).
The goal of a monitor is to dynamically check that a certain set of heap locations have values of their intended (static) types. Accordingly, the description of
the monitor includes a list of locations and their expected types (in the form of
a context ∆). All these types τ should satisfy τ 6⊢ ◦. To facilitate checks of the
monitor, every heap location carries a type at runtime (in addition to a value).
As Theorem 10 shows, the monitor can never fail in a well-typed component.
Theorem 4 (Typability Implies Robust Safety in Lτ ).
if ⊢ C : UN then ⊢ C : rs
21
τ ⊢◦
(TLτ -bool-pub)
(TLτ -nat-pub)
Bool ⊢ ◦
Nat ⊢ ◦
(TLτ -pair-pub)
τ ⊢ ◦ τ′ ⊢ ◦
τ × τ′ ⊢ ◦
(TLτ -un-pub)
UN ⊢ ◦
(TLτ -references-pub)
Ref UN ⊢ ◦
⊢ C : UN
τ ⊢◦
Component C is well-typed.
Type τ is insecure.
∆, Γ ⊢ e : τ
C, ∆, Γ ⊢ s
Expression e has type τ in Γ and ∆.
Statement s is well-typed in C, Γ and ∆.
(TLτ -coercion)
∆, Γ ⊢ e : τ
τ ⊢◦
∆, Γ ⊢ e : UN
(TLτ -endorse)
∆, Γ ⊢ e : UN
C, ∆, Γ; (x : ϕ) ⊢ s
C, ∆, Γ ⊢ endorse x = e as ϕ in s
(ELτ -endorse)
H ⊲ e ֒→
→ v
∆, ∅ ⊢ v : ϕ
∆ = {ℓ : τ | ℓ 7→ v : τ ∈ H}
C, H ⊲ endorse x = e as ϕ in s −→ C, H ⊲ s[v / x]
(ELτ -fork)
Π = Π1 k {k s}; s′ k Π2
Π′ = Π1 k skip; s′ k Π2 k s
C, H ⊲ Π −→ C, H ⊲ Π′
Figure 7: Judgements, static and dynamic semantics of Lτ (excerpts).
5.2
The Target Language Lπ
Our target language, Lπ , extends the previous target language, LP , with support for concurrency (forking, processes and process soups), atomic co-creation
of a protected location and its protecting capability (statement newhide) and
for examining the top-level construct of a value according to a pattern, B (Figure 8).
A monitor explicitly mentions the entire list of locations it cares about.
These are represented together as a heap, written H0 .
22
Patterns B ::= nat | pair
Statements s ::= · · · | {k s} | let x = newhide e in s
| destruct x = e as B in s or s
Monitors M ::= ({σ} , , σ0 , H0 , σc )
(ELπ -destruct-nat)
H ⊲ e ֒→
→ n
C, H ⊲ destruct x = e as nat in s or s′
−→ C, H ⊲ s[n / x]
(ELπ -new)
H = H1 ; n 7→ (v, η)
H ⊲ e ֒→
→ v
k∈
/ dom(H)
C, H ⊲ let x = newhide e in s
−→ C, H; n + 1 7→ v : k; k ⊲ s[hn + 1, ki / x]
Figure 8: Lπ additions to syntax and semantics (excerpts).
5.3
Compiler from Lτ to Lπ
Lτ
The high-level structure of the compiler, J·KLπ , is similar to that of our earlier
LU
Lτ
compiler J·KLP (Section 4.3). However, J·KLπ is defined by induction on the
type derivation of the component to be compiled (Figure 9). Additionally, some
Lτ
cases such as Rule (J·KLπ -New) explicitly use type information to achieve security
efficiently.
Lτ
The compiler J·KLπ is both correct and robustly safe.
Lτ
Lτ
Theorem 5 (Compiler J·KLπ is cc). ⊢ J·KLπ : cc
LU
The proof of this theorem is similar to that for J·KLP , but the cross-language
monitor relation is different. We describe that relation below.
Monitor Relation Informally, a source and a target monitor are related if
the target monitor can always step whenever the target heap satisfies the types
specified in the source monitor (up to renaming by the partial bijection β).
Formally, given a typing ∆ for the part of the source heap being monitored
and a partial bijection β from source to target locations, we say that a target
monitor M = ({σ} , , σ0 , H0 , σc ) is good, written ⊢ M : β, ∆, if for all σ ∈ {σ}
and all H ≈β H such that ⊢ H : ∆, there is a σ ′ such that (σ, H, σ ′ ) ∈ .
For a fixed partial bijection β0 between the domains of ∆ and H0 , we say that
the source monitor M and the target monitor M are related, written M ≈ M, if
⊢ M : β0 , ∆ for the ∆ in M.
τ
τ
Theorem 6 (Compiler J·KLLπ is rs-pres ). ⊢ J·KLLπ : rs-pres
Lτ
To prove that J·KLπ attains RSC we do not rely on a back-translation. Here,
we know statically which locations can be monitor-sensitive: they must all be
23
u
v
(TLτ -new)
∆, Γ ⊢ e : τ
C, ∆, Γ; x : Ref τ ⊢ s
C, ∆, Γ ⊢ let x = newτ e in s
Lτ
let xo = new J∆, Γ ⊢ e : τ KLπ
in let x = hxo, 0i
in JC, ∆, Γ; x : Ref τ ⊢ sKLτπ
L
=
Lτ
let
x
=
newhide
J∆,
Γ
⊢
e
:
τ
K
Lπ
τ
in JC, ∆, Γ; x : Ref τ ⊢ sKLLπ
u
v
(TLτ -coercion)
∆, Γ ⊢ e : τ
τ ⊢◦
∆, Γ ⊢ e : UN
u
(TLτ -fork)
v
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ {k s}
}Lτ
~
~
Lπ
if τ = UN
Lτ
(J·KLπ -New)
otherwise
Lτ
Lτ
(J·KLπ -Coerce)
Lπ
= J∆, Γ ⊢ e : τ KLπ
Lπ
= {k JC, ∆, Γ ⊢ sKLπ }
}Lτ
~
}Lτ
Lτ
Lτ
(J·KLπ -Fork)
Lτ
Figure 9: J·KLπ excerpts.
safe, i.e., must have a type τ satisfying τ 0 ◦. Using this, we set up a simple
cross-language relation and show it to be an invariant on runs of source and
compiled target components. The relation captures the following:
• Heaps (both source and target) can be partitioned into two parts, a high
part and a low part;
• The high source heap contains only locations whose type is safe;
• The high target heap contains only locations related to source high locations and these point to related values; more importantly, every target
high location is protected by a capability;
• In the target, any capability protecting a high location does not occur in
attacker code, nor is it stored in a low heap location.
We need to prove that this relation is preserved by reductions both in the
compiled code and in the attacker code. The former follows from correctness and
source robust safety (Theorems 10 and 11). The latter is simple since all high
Lτ
locations are protected with capabilities (Rule (J·KLπ -New)), attackers have no
Lτ
access to high locations (Rule (J·KLπ -Coerce)) and capabilities are unforgeable
and unguessable (by the semantics of Lπ ). At this point, knowing that monitors
are related, and that source heaps never make source monitors fail, we can
conclude that monitors do not fail in the target either.
24
Note that this kind of an argument requires all compilable source programs
LU
to be robustly safe and is, therefore, impossible for our first compiler J·KLP .
Avoiding the back-translation here simplifies the proof significantly.
6
Discussion
In this section, we discuss some loose ends and possible extensions of our development.
Richer Source Monitors In Lτ , source language monitors only enforce the
weak property of type safety on specific memory locations (robustly). This
can be generalized substantially to enforce arbitrary invariants other than types
on locations. The only requirement is to find a type system (e.g., based on
refinements or Hoare logics [48]) that can enforce robust safety in the source.
Our compilation and proof strategy should work with little modification. We
did not make this generalization here to keep the type system simple and to
retain focus on the compilation and the proof technique.
Realizing Target Capabilites Both our target languages rely on capabilities
for restricting access to sensitive locations from the context. These capabilities
can be realized in many different ways. First, we could use a capability machine
such as Cheri that supports capabilities in hardware [51]. Capability machines
have been advocated as a target for efficient secure compilation [24] and preliminary work has been made to compile C-like languages to them, but using
FAC , not RSC , as the criterion for compiler security [49]. Second, we could use
memory isolation (via software-fault isolation [50] or an architecture like Intel’s
SGX [36]) to isolate high locations.
Non-Atomic Allocation of Capabilities In Section 5, we assumed a new
target language construct, newhide, that simultaneously allocates a new location and protects it with a capability. This atomic construct simplifies our proof
of security in the concurrent setting at hand: if allocation and protection were
not atomic, then a concurrent adversary thread could protect a location that
had just been allocated and acquire the capability to it before the allocating
thread could do so. This would break the cross-language relation we use in our
proof. However, note that this is not really an attack since it does not give
the adversary any additional power to break the safety property enforced by
the monitor (the allocating thread just gets stuck when it tries to acquire the
capability itself). Consequently, it is possible to do away with the newhide
construct, at the cost of a slightly more involved cross-language relation. The
appendix presents this alternate development as well.
Security criteria In their preliminary work, Garg et al. [27] present new
criteria for secure compilation that ensure preservation of arbitrary classes of
25
hyperproperties. Hyperproperties [19] are a formal representation of predicates
on programs, i.e., they are predicates on sets of traces. Hyperproperties capture
many security-relevant properties including not just conventional safety and
liveness, which are predicates on traces, but also properties like non-interference,
which is a predicate on pairs of traces.
Modulo technical differences, our definition of RSC coincides with the criterion of safety property preservation in [27]. We show, through concrete instances, that this criterion can be easily realized by compilers, and develop two
proof techniques for establishing it. We further show that the criterion leads to
more efficient compiled code than does FAC .
More interestingly, [27] also proposes stronger criterion that imply preservation of larger classes of hyperproperties. One such criterion is the preservation
of all robust hypersafety properties. We believe that this criterion is particularly interesting because it will also lead to efficient compiled code, and it should
be possible to generalize both the proof techniques we presented here to this
criterion.
The criteria in [27] assume that behaviours in the source and target are
represented using the same grammar. Hence, the definitions (somewhat unrealistically or ideally) do not require a translation of source properties to target
properties. In contrast, we consider differences in the representation of behaviour in the source and in the target. This difference is accounted for in
our monitor relation M ≈ M. A slightly different account of this difference is
presented by Patrignani and Garg [44] in the context of reactive black-box programs. Investigating the definition of precise instances of this relation that
account for preservation of specific security properties is left for future work.
7
Other Related Work
We have already discussed a lot of closely related work. We focus here on the
remaining related work only.
ASLR [5, 30], protected module architectures [6, 33, 42, 43], tagged architectures [32], capability machines [49] and cryptographic primitives [3, 4, 16, 20]
have been used as targets for FAC . We believe all of these can also be used
as targets of RSC -attaining compilers. In fact, some targets such as capability
machines seem to be better suited to RSC than FAC .
Ahmed et al. prove full abstraction for several compilers between typed
languages [7, 8, 39]. As compiler intermediate languages are often typed, and as
these types often serve as the basis for complex static analyses, full abstraction
seems like a reasonable goal for (fully typed) intermediate compilation steps.
In the last few steps of compilation, where the target languages are unlikely to
be typed, one could establish robust safety preservation and combine the two
properties (vertically) to get an end-to-end security guarantee.
There are three main other criteria for secure compilation that we would
like to mention: securely compartmentalised compilation (SCC) [32], tracepreserving compilation (TPC) [44] and non-interference-preserving compilation
26
(NIPC) [11]. SCC is a re-statement of the “hard” part of full abstraction that
is better suited for languages with undefined behaviour and a strict notion of
components. Thus SCC suffers from much of the same efficiency drawbacks as
FAC that RSC does not.
TPC is a stronger criterion than FAC and that most existing fully abstract
compilers also attain. Again, compilers attaining TPC also suffer from the
drawbacks of compilers attaining FAC .
NIPC is a criterion that preserves a single property: non-interference (NI).
It has been described in the setting where NI is achieved by typing in both
source and target languages and where the target uses the same security lattice
as the source. Since noninterference is not a safety property, it is difficult to
compare NIPC to RSC directly. However, NIPC seeks to preserve information
flow properties, which can also be approximated as safety properties [14]. So,
in principle, RSC can be applied to the same end-goals as NIPC.
Swamy et al. [47] embed an F∗ model of a gradually and robustly typed variant of JavaScript into an F∗ model of JavaScript. Their type-directed compiler is
proven to attain memory isolation as well as static and dynamic memory safety.
However, they do not consider general safety properties, nor a specific, general
criterion for compiler security. Gradual typing supports constructs similar to
our endorsement construct in Lτ .
8
Conclusion
This paper has examined robustly safe compilation (RSC ), a soundness criterion
for compilers with direct relevance to security, and shown that the criterion is
easily realizable and leads to more efficient code than does the current standard
for secure compilers, fully abstract compilation wrt contextual equivalence. We
have also presented two different techniques for establishing that a compiler attains RSC . One is an adaptation of an existing technique, back-translation, and
the other is based on inductive invariants. In ongoing work, we are both extending our work to the robust preservation of properties stronger than safety, and
deepening our work by developing secure compilers for more realistic languages.
Acknowledgements The authors would like to thank Dominique Devriese, Frank
Piessens, David Swasey, Catalin Hritcu and Marco Stronati for discussions and comments on the subject. This work was partially supported by the German Federal Ministry of Education and Research (BMBF) through funding for the CISPA-Stanford
Center for Cybersecurity (FKZ: 13N1S0762).
References
[1] Martín Abadi. Secrecy by typing in security protocols. In TACS, pages
611–638, 1997.
27
[2] Martín Abadi. Protection in programming-language translations. In Secure
Internet programming, pages 19–34. Springer-Verlag, London, UK, 1999.
[3] Martín Abadi, Cédric Fournet, and Georges Gonthier. Authentication primitives and their compilation. In Proceedings of the 27th ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
POPL ’00, pages 302–315, New York, NY, USA, 2000. ACM.
[4] Martín Abadi, Cédric Fournet, and Georges Gonthier. Secure implementation of channel abstractions. Information and Computation, 174:37–83,
2002.
[5] Martín Abadi and Gordon D. Plotkin. On protection by layout randomization. ACM Transactions on Information and System Security, 15:8:1–8:29,
July 2012.
[6] Pieter Agten, Raoul Strackx, Bart Jacobs, and Frank Piessens. Secure
compilation to modern processors. In 2012 IEEE 25th Computer Security
Foundations Symposium, CSF 2012, pages 171–185. IEEE, 2012.
[7] Amal Ahmed and Matthias Blume. Typed closure conversion preserves observational equivalence. In Proceedings of the 13th ACM SIGPLAN International Conference on Functional Programming, ICFP ’08, pages 157–168,
New York, NY, USA, 2008. ACM.
[8] Amal Ahmed and Matthias Blume. An equivalence-preserving CPS translation via multi-language semantics. In Proceedings of the 16th ACM SIGPLAN International Conference on Functional Programming, ICFP ’11,
pages 431–444, New York, NY, USA, 2011. ACM.
[9] Michael Backes, Catalin Hritcu, and Matteo Maffei. Union, intersection and
refinement types and reasoning about type disjointness for secure protocol
implementations. Journal of Computer Security, 22(2):301–353, 2014.
[10] Antonio Barresi, Kaveh Razavi, Mathias Payer, and Thomas R. Gross.
CAIN: Silently breaking ASLR in the cloud. In 9th USENIX Workshop
on Offensive Technologies (WOOT 15), Washington, D.C., 2015. USENIX
Association.
[11] Gilles Barthe, Tamara Rezk, and Amitabh Basu. Security types preserving
compilation. Computer Languages, Systems and Structures, 33:35–59, 2007.
[12] Jesper Bengtson, Karthikeyan Bhargavan, Cédric Fournet, Andrew D. Gordon, and Sergio Maffeis. Refinement types for secure implementations.
ACM Trans. Program. Lang. Syst., 33(2):8:1–8:45, February 2011.
[13] Gérard Berry and Gérard Boudol. The chemical abstract machine. Theor.
Comput. Sci., 96(1):217–248, 1992.
28
[14] Gérard Boudol. Secure information flow as a safety property. chapter
Formal Aspects in Security and Trust, pages 20–34. Springer-Verlag, Berlin,
Heidelberg, 2009.
[15] William J. Bowman and Amal Ahmed. Noninterference for free. In Proceedings of the 20th ACM SIGPLAN International Conference on Functional
Programming, ICFP ’15, New York, NY, USA, 2015. ACM.
[16] Michele Bugliesi and Marco Giunti. Secure implementations of typed
channel abstractions. In Proceedings of the 34th annual ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, POPL ’07,
pages 251–262, New York, NY, USA, 2007. ACM.
[17] Luca Cardelli, Giorgio Ghelli, and Andrew D. Gordon. Secrecy and group
creation. Inf. Comput., 196(2):127–155, 2005.
[18] Stephen Chong. Expressive and Enforceable Information Security Policies.
PhD thesis, Cornell University, August 2008.
[19] Michael R. Clarkson and Fred B. Schneider. Hyperproperties. J. Comput.
Secur., 18(6):1157–1210, September 2010.
[20] Ricardo Corin, Pierre-Malo Deniélou, Cédric Fournet, Karthikeyan Bhargavan, and James Leifer. A secure compiler for session abstractions. Journal
of Computer Security, 16:573–636, 2008.
[21] Dominique Devriese, Marco Patrignani, Steven Keuchel, and Frank
Piessens. Modular, Fully-Abstract Compilation by Approximate BackTranslation. Logical Methods in Computer Science, Volume 13, Issue 4,
October 2017.
[22] Dominique Devriese, Marco Patrignani, and Frank Piessens. Secure Compilation by Approximate Back-Translation. In POPL 2016, 2016.
[23] Dominique Devriese, Marco Patrignani, and Frank Piessens. Parametricity versus the universal type. In Proceedings of the 45th Annual ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
POPL 2018, Los Angeles, CA, USA, 2016, 2018.
[24] Akram El-Korashy. A Formal Model for Capability Machines – An Illustrative Case Study towards Secure Compilation to CHERI. Master’s thesis,
Universitat des Saarlandes, 2016.
[25] Cédric Fournet, Andrew D. Gordon, and Sergio Maffeis. A type discipline
for authorization policies. ACM Trans. Program. Lang. Syst., 29(5), August
2007.
[26] Cedric Fournet, Nikhil Swamy, Juan Chen, Pierre-Evariste Dagand, PierreYves Strub, and Benjamin Livshits.
Fully abstract compilation to
JavaScript. In Proceedings of the 40th annual ACM SIGPLAN-SIGACT
29
Symposium on Principles of Programming Languages, POPL ’13, pages
371–384, New York, NY, USA, 2013. ACM.
[27] D. Garg, C. Hritcu, M. Patrignani, M. Stronati, and D. Swasey. Robust
Hyperproperty Preservation for Secure Compilation (Extended Abstract).
ArXiv e-prints, October 2017.
[28] Andrew D. Gordon and Alan Jeffrey. Authenticity by typing for security
protocols. J. Comput. Secur., 11(4):451–519, July 2003.
[29] Daniele Gorla and Uwe Nestman. Full abstraction for expressiveness: History, myths and facts. Math Struct Comp Science, 2014.
[30] Radha Jagadeesan, Corin Pitcher, Julian Rathke, and James Riely. Local memory via layout randomization. In Proceedings of the 2011 IEEE
24th Computer Security Foundations Symposium, CSF ’11, pages 161–174,
Washington, DC, USA, 2011. IEEE Computer Society.
[31] Yeongjin Jang, Sangho Lee, and Taesoo Kim. Breaking kernel address
space layout randomization with intel tsx. In Proceedings of the 2016 ACM
SIGSAC Conference on Computer and Communications Security, CCS ’16,
pages 380–392, New York, NY, USA, 2016. ACM.
[32] Yannis Juglaret, Cătălin Hriţcu, Arthur Azevedo de Amorim, and Benjamin C. Pierce. Beyond good and evil: Formalizing the security guarantees
of compartmentalizing compilation. In 29th IEEE Symposium on Computer
Security Foundations (CSF). IEEE Computer Society Press, July 2016. To
appear.
[33] Adriaan Larmuseau, Marco Patrignani, and Dave Clarke. A secure compiler
for ML modules. In Programming Languages and Systems - 13th Asian
Symposium, APLAS 2015, Pohang, South Korea, November 30 - December
2, 2015, Proceedings, pages 29–48, 2015.
[34] Xavier Leroy. A formally verified compiler back-end. J. Autom. Reasoning,
43(4):363–446, 2009.
[35] Sergio Maffeis, Martín Abadi, Cédric Fournet, and Andrew D. Gordon.
Code-Carrying Authorization, pages 563–579. Springer Berlin Heidelberg,
Berlin, Heidelberg, 2008.
[36] Frank McKeen, Ilya Alexandrovich, Alex Berenzon, Carlos V. Rozas,
Hisham Shafi, Vedvyas Shanbhogue, and Uday R. Savagaonkar. Innovative instructions and software model for isolated execution. In HASP ’13,
pages 10:1–10:1. ACM, 2013.
[37] James H. Morris, Jr. Protection in programming languages. Commun.
ACM, 16:15–21, 1973.
30
[38] Georg Neis, Derek Dreyer, and Andreas Rossberg. Non-parametric parametricity. SIGPLAN Not., 44(9):135–148, August 2009.
[39] Max S. New, William J. Bowman, and Amal Ahmed. Fully abstract compilation via universal embedding. In Proceedings of the 21st ACM SIGPLAN
International Conference on Functional Programming, ICFP 2016, pages
103–116, New York, NY, USA, 2016. ACM.
[40] Flemming Nielson, Hanne R. Nielson, and Chris Hankin. Principles of
Program Analysis. Springer-Verlag New York, USA, 1999.
[41] Joachim Parrow. General conditions for full abstraction. Math Struct Comp
Science, 2014.
[42] Marco Patrignani, Pieter Agten, Raoul Strackx, Bart Jacobs, Dave Clarke,
and Frank Piessens. Secure Compilation to Protected Module Architectures. ACM Trans. Program. Lang. Syst., 37:6:1–6:50, April 2015.
[43] Marco Patrignani, Dominique Devriese, and Frank Piessens. On Modular
and Fully Abstract Compilation. In Proceedings of the 29th IEEE Computer
Security Foundations Symposium, CSF 2016, 2016.
[44] Marco Patrignani and Deepak Garg. Secure Compilation and Hyperproperties Preservation. In Proceedings of the 30th IEEE Computer Security
Foundations Symposium CSF 2017, Santa Barbara, USA, CSF 2017, 2017.
[45] Fred B. Schneider. Enforceable security policies. ACM Trans. Inf. Syst.
Secur., 3(1):30–50, 2000.
[46] Ian Stark. Names and Higher-Order Functions. PhD thesis, University
of Cambridge, December 1994. Also available as Technical Report 363,
University of Cambridge Computer Laboratory.
[47] Nikhil Swamy, Cedric Fournet, Aseem Rastogi, Karthikeyan Bhargavan,
Juan Chen, Pierre-Yves Strub, and Gavin Bierman. Gradual typing embedded securely in javascript. SIGPLAN Not., 49(1):425–437, January
2014.
[48] David Swasey, Deepak Garg, and Derek Dreyer. Robust and compositional verification of object capability patterns. In Proceedings of the 2017
ACM SIGPLAN International Conference on Object-Oriented Programming, Systems, Languages, and Applications, OOPSLA 2017. October 22
- 27, 2017, 2017.
[49] Stelios Tsampas, Akram El-Korashy, Marco Patrignani, Dominique Devriese, Deepak Garg, and Frank Piessens. Towards Automatic Compartmentalization of C Programs on Capability Machines. In Workshop on
Foundations of Computer Security 2017 August 21, 2017, FCS 2017, 2017.
31
[50] Robert Wahbe, Steven Lucco, Thomas E. Anderson, and Susan L. Graham. Efficient software-based fault isolation. SIGOPS Oper. Syst. Rev.,
27(5):203–216, December 1993.
[51] Jonathan Woodruff, Robert N.M. Watson, David Chisnall, Simon W.
Moore, Jonathan Anderson, Brooks Davis, Ben Laurie, Peter G. Neumann,
Robert Norton, and Michael Roe. The CHERI Capability Model: Revisiting RISC in an Age of Risk. In Proceeding of the 41st Annual International
Symposium on Computer Architecuture, ISCA ’14, pages 457–468, Piscataway, NJ, USA, 2014. IEEE Press.
32
A
The Untyped Source Language: LU
This is a sequential while language with monitors.
A.1
Syntax
Whole Programs P ::= M; H; F; I
Components C ::= M; F; I
Contexts A ::= H; F [·]
Interfaces I ::= f
Functions F ::= f(x) 7→ s; return;
Operations ⊕ ::= + | −
Comparison ⊗ ::= == | < | >
Values v ::= b ∈ {true, false} | n ∈ N | hv, vi | ℓ
Expressions e ::= x | v | e ⊕ e | e ⊗ e | he, ei | e.1 | e.2 | !e
Statements s ::= skip | s; s | let x = e in s | if e then s else s
| monitor | call f e | let x = new e in s | x := e
Eval . Ctxs. E ::= [·] | e ⊕ E | E ⊕ n | e ⊗ E | E ⊗ n
| he, Ei | hE, vi | E.1 | E.2 | !E
Heaps H ::= ∅ | H; ℓ 7→ v
Monitors M ::= ({σ} ,
Mon. States σ ∈ S
, σ0 , ℓroot , σc )
Mon. Reds.
::= ∅ | ; (s, H, s)
Substitutions ρ ::= ∅ | ρ[v / x]
Prog. States Ω ::= C, H ⊲ (s)f | fail
Labels λ ::= ǫ | α
Actions α ::= call f v H? | call f v H! | ret H! | ret H?
Traces α ::= ∅ | α · α
A.2
Dynamic Semantics
Rules LU -Jump-Internal to LU -Jump-OUT dictate the kind of a jump between
two functions: if internal to the component/attacker, in(from the attacker to the
component) or out(from the component to the attacker). Rule LU -Plug tells how
to obtain a whole program from a component and an attacker. Rule LU -Whole
tells when a program is whole. Rule LU -Initial State tells the initial state of a
whole program. Rule LU -Monitor Step tells when a monitor makes a single step
given a heap.
Helpers
33
(LU -Jump-Internal)
′
(LU -Jump-IN)
′
((f ∈ I ∧ f ∈ I)∨
(f ′ ∈
/ I∧f ∈
/ I))
(LU -Jump-OUT)
′
f ∈I∧f ∈
/I
I ⊢ f, f ′ : in
I ⊢ f, f ′ : internal
A ≡ H; F [·]
C
M ≡ ({σ} , , σ0 , ℓroot , σ0 )
f∈
/ I∧f ∈I
I ⊢ f, f ′ : out
(LU -Plug)
≡ M; F′ ; I
⊢ C, F : whole
main ∈ names(F)
ℓroot ∩ H = ∅
A [C] = M; H; ℓroot 7→ 0; F; F′ ; I
U
(L -Whole)
C ≡ M; F′ ; I
names(F) ∩ names(F′ ) = ∅
names(I) ⊆ names(F) ∪ names(F′ )
fv(F) ∪ fv(F′ ) = ∅
(LU -Initial State)
P ≡ M; H; F; I
C ≡ M; F; I
Ω0 (P) = C; H ⊲ call main 0
⊢ C, F : whole
Let reach(ℓo , H) return a set of locations {ℓ} such that it is possible to reach
any ℓ ∈ {ℓ} from ℓo just by dereferencing and projection expressions.
To ensure monitor transitions have a meaning, they are assumed to be closed
under bijective renaming of locations.
M′
M; H
(LU -Monitor Step)
, σ0 , ℓroot , σc )
M′ = ({σ} ,
′
′
M = ({σ} ,
(σc , H′ , σf ) ∈
, σ0 , ℓroot , σf )
H ⊆H
dom(H ) = reach(ℓroot , H)
M; H
M′
(LU -Monitor No Step)
M = ({σ} , , σ0 , ℓroot , σc )
H′ ⊆ H
(σc , H′ , σf ) ∈
M; H 6
A.2.1
Component Semantics
H ⊲ e ֒→
→ e′
ǫ
Expression e reduces to e′ .
C, H ⊲ s −−→ C′ , H′ ⊲ s′
α
Ω ==⇒ Ω′
Statement s reduces to s′ and evolves the rest accordingly,
emitting label λ.
Program state Ω steps to Ω′ emitting trace α.
H ⊲ e ֒→
→ e′
34
(ELU -ctx)
H ⊲ e ֒→
→ e′
H ⊲ E [e] ֒→
→ E [e′ ]
(ELU -p2)
H ⊲ hv, v′ i .1 ֒→
→ v′
(ELU -val)
(ELU -p1)
H ⊲ v ֒→
→ v
H ⊲ hv, v′ i .1 ֒→
→ v
(ELU -op)
(ELU -comp)
n ⊕ n′ = n′′
H ⊲ n ⊕ n′ ֒→
→ n′′
n ⊗ n′ = b
H ⊲ n ⊗ n′ ֒→
→ b
(ELU -dereference)
H ⊲ e ֒→
→ ℓ ℓ 7→ v ∈ H
H⊲!e ֒→
→ v
λ
C, H ⊲ s −−→ C′ , H′ ⊲ s′
(ELU -step)
λ
(ELU -sequence)
C, H ⊲ s −−→ C, H ⊲ s′
ǫ
C, H ⊲ skip; s −−→ C, H ⊲ s
λ
C, H ⊲ s; s′′ −−→ C, H ⊲ s′ ; s′′
(ELU -if-true)
H ⊲ e ֒→
→ true
ǫ
C, H ⊲ if e then s else s′ −−→ C, H ⊲ s
(ELU -if-false)
H ⊲ e ֒→
→ false
ǫ
C, H ⊲ if e then s else s′ −−→ C, H ⊲ s
(ELU -letin)
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = e in s −−→ C, H ⊲ s[v / x]
(ELU -alloc)
H ⊲ e ֒→
→ v
ℓ∈
/ dom(H)
ǫ
C, H ⊲ let x = new e in s −−→ C, H; ℓ 7→ v ⊲ s[ℓ / x]
(ELU -update)
H ⊲ e ֒→
→ v
H′ = H1 ; ℓ 7→ v; H2
H = H1 ; ℓ 7→ v′ ; H2
ǫ
C, H ⊲ ℓ := e −−→ C, H′ ⊲ skip
U
(EL -monitor)
M = ({s} ,
, s0 , ℓroot , sc )
C = M; F; I
M; H
M′
C′ = M′ ; F; I
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
(ELU -monitor-fail)
M = ({s} ,
, s0 , ℓroot , sc )
M; H 6
C = M; F; I
ǫ
C, H ⊲ monitor −−→ fail
(ELU -call-internal)
′
C.intfs ⊢ f, f : internal
f(x) 7→ s; return; ∈ C.funs
ǫ
f ′ = f ′′ ; f ′
H ⊲ e ֒→
→ v
C, H ⊲ (call f e)f ′ −−→ C, H ⊲ (s; return;[v / x])f ′ ;f
35
(ELU -callback)
f ′′ ; f ′
f′
=
f(x) 7→ s; return; ∈ F
C.intfs ⊢ f ′ , f : out H ⊲ e ֒→
→ v
call f v H!
C, H ⊲ (call f e)f ′ −−−−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELU -call)
f ′′ ; f ′
f′
=
f(x) 7→ s; return; ∈ C.funs
C.intfs ⊢ f ′ , f : in
H ⊲ e ֒→
→ v
call f v H?
C, H ⊲ (call f e)f ′ −−−−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELU -ret-internal)
f′
=
f ′′ ; f ′
C.intfs ⊢ f, f ′ : internal
ǫ
C, H ⊲ (return;)f ′ ;f −−→ C, H ⊲ (skip)f ′
(ELU -retback)
f′
=
f ′′ ; f ′
C.intfs ⊢ f, f ′ : in
ret H?
C, H ⊲ (return;)f ′ ;f −−−−−−→ C, H ⊲ (skip)f ′
(ELU -return)
f′
=
f ′′ ; f ′
C.intfs ⊢ f, f ′ : out
ret H!
C, H ⊲ (return;)f ′ ;f −−−−−→ C, H ⊲ (skip)f ′
α
Ω ==⇒ Ω′
(ELU -single)
Ω=
⇒ Ω′′
′′ α
′
Ω −−→ Ω
α
Ω ==⇒ Ω′
(ELU -silent)
ǫ
′
Ω −−→ Ω
Ω=
⇒ Ω′
36
(ELU -trans)
α
′′
Ω ==⇒ Ω
α′
Ω′′ ==⇒ Ω′
α·α′
Ω ====⇒ Ω′
The Target Language: LP
B
B.1
Syntax
Whole Programs P ::= M; F; I
Components C ::= M; F; I; kroot
Contexts A ::= F [·]
Interfaces I ::= f
Functions F ::= f (x) 7→ s; return;
Operations ⊕ ::= + | −
Comparison ⊗ ::= == | < | >
Values v ::= n ∈ N | hv, vi | k
Expressions e ::= x | v | e ⊕ e | e ⊗ e | he, ei | e.1 | e.2 | !e with e
Statements s ::= skip | s; s | let x = e in s | ifz e then s else s | call f e
| x := e with e | let x = new e in s | let x = hide e in s
| monitor
Eval . Ctxs. E ::= [·] | e ⊕ E | E ⊕ n | e ⊗ E | E ⊗ n | !E with v | !e with E
| he, Ei | hE, vi | E.1 | E.2
Heaps H ::= ∅ | H; n 7→ v : η | H; k
Tag η ::= ⊥ | k
Monitors M ::= ({σ} ,
Mon. States σ ∈ S
, σ0 , σc )
Mon. Reds.
::= ∅ | ; (s, H, s)
Substitutions ρ ::= ∅ | ρ[v / x]
Prog. States Ω ::= C, H ⊲ (s)f | fail
Labels λ ::= ǫ | α
Actions α ::= call f v H? | call f v H! | ret H! | ret H?
Traces α ::= ∅ | α · α
B.2
Operational Semantics of LP
Helpers
(LP -Jump-Internal)
′
((f ∈ I ∧ f ∈ I)∨
(f ′ ∈
/ I∧f ∈
/ I))
I ⊢ f , f ′ : internal
(LP -Jump-IN)
′
f ∈I∧f ∈
/I
′
I ⊢ f , f : in
37
(LP -Jump-OUT)
′
f∈
/ I∧f ∈I
I ⊢ f , f ′ : out
(LP -Plug)
A ≡ F [·]
⊢ C, F : whole
M ≡ ({σ} ,
C ≡ M; F′ ; I; kroot
main(x) 7→ s; return; ∈ F
, σ0 , σc )
kroot ∈
/ fn(A)
A [C] = M; F; F′ ; I
(LP -Whole)
M; F′ ; I; kroot
C≡
names(F) ∩ names(F′ ) = ∅
names(I) ⊆ names(F)
⊢ C, F : whole
(LP -Initial State)
P ≡ M; F; I C ≡ M; F′ ; I; kroot
Ω0 (P) = C, kroot ; 0 7→ 0 : kroot ⊲ call main 0
Define reach(nr , kr , H) as the set of locations {n} such that it is possible
to reach any n ∈ {n} from nr using any expression and relying on capability kr
as well as any capability reachable from nr . Formally:
(
)
H ⊲ e ֒→
→ !n with v ֒→
→ v′
reach(nr , kr , H) = n
fv(e) = nr ∪ kr
M; H
M′
(LP -Monitor Step)
, σ0 , σc )
M′ =
′
′
M = ({σ} ,
({σ} , , σ0 , σf )
(sc , H′ , sf ) ∈ H ⊆ H dom(H ) = reach(0, kroot , H)
M; H; kroot
M′
(LP -Monitor No Step)
∄H′ .H′ ⊆ H
B.2.1
M = ({σ} , , σ0 , σc )
(sc , H′ , sf ) ∈
dom(H′ ) = reach(0, kroot , H)
M; H; kroot 6
Component Semantics
H ⊲ e ֒→
→ e′
ǫ
Expression e reduces to e′ .
C, H ⊲ s −−→ C′ , H′ ⊲ s′
Statement s reduces to s′ and evolves the rest accordingly,
emitting label λ.
α
Ω ==⇒ Ω
′
Program state Ω steps to Ω′ emitting trace α.
H ⊲ e ֒→
→ e′
38
(ELP -val)
(ELP -p1)
(ELP -p2)
H ⊲ v ֒→
→ v
H ⊲ hv, v′ i .1 ֒→
→ v
H ⊲ hv, v′ i .1 ֒→
→ v′
P
P
(EL -op)
(EL -comp)
n ⊕ n′ = n′′
H ⊲ n ⊕ n′ ֒→
→ n′′
if n ⊗ n′ = true then n′′ = 0 else n′′ = 1
H ⊲ n ⊗ n′ ֒→
→ n′′
(ELP -deref-top)
(ELP -deref-k)
n 7→ v : ⊥ ∈ H
H⊲!n with _ ֒→
→ H⊲v
n 7→ (v, k) ∈ H
H⊲!n with k ֒→
→ H⊲v
(ELP -ctx)
H ⊲ e ֒→
→ e′
H ⊲ E [e] ֒→
→ H ⊲ E [e′ ]
λ
H ⊲ s −−→ H′ ⊲ s′
(ELP -step)
λ
(ELP -sequence)
C, H ⊲ s −−→ C, H ⊲ s′
ǫ
C, H ⊲ skip; s −−→ C, H ⊲ s
λ
C, H ⊲ s; s′′ −−→ C, H ⊲ s′ ; s
(ELP -if-true)
H ⊲ e ֒→
→ 0
ǫ
C, H ⊲ ifz e then s else s′ −−→ C, H ⊲ s
(ELP -if-false)
H ⊲ e ֒→
→ n
n 6≡ 0
ǫ
C, H ⊲ ifz e then s else s′ −−→ C, H ⊲ s′
P
(EL -letin)
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = e in s −−→ C, H ⊲ s[v / x]
(ELP -new)
H = H1 ; n 7→ (v, η)
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = new e in s −−→ C, H; n + 1 7→ v : ⊥ ⊲ s[n + 1 / x]
(ELP -hide)
H ⊲ e ֒→
→ n
H = H1 ; n 7→ v : ⊥; H2
k∈
/ dom(H)
H′ = H1 ; n 7→ v : k; H2 ; k
ǫ
C, H ⊲ let x = hide e in s −−→ C, H′ ⊲ s[k / x]
(ELP -assign-top)
H ⊲ e ֒→
→ v
H′ = H1 ; n 7→ v : ⊥; H2
H = H1 ; n 7→ _ : ⊥; H2
ǫ
C, H ⊲ n := e with _ −−→ C, H′ ⊲ skip
P
(EL -assign-k)
H ⊲ e ֒→
→ v
H = H1 ; n 7→ _ : k; H2
H′ = H1 ; n 7→ v : k; H2
ǫ
C, H ⊲ n := e with k −−→ C, H′ ⊲ skip
39
(ELP -monitor)
M = ({s} ,
, σ0 , σc )
C = M; F; I; kroot
M; H; kroot
M′
C′ = M′ ; F; I; kroot
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
(ELP -monitor-fail)
C = M; F; I; kroot
M; H; kroot 6
ǫ
C, H ⊲ monitor −−→ fail
(ELP -call-internal)
′
C.intfs ⊢ f , f : internal f ′ = f ′′ ; f ′
f (x) 7→ s; return; ∈ C.funs
ǫ
C, H ⊲ (call f v)f ′ −−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELP -callback)
f′
=
f ′′ ; f ′
f (x) 7→ s; return; ∈ F
C.intfs ⊢ f ′ , f : out
call f v H!
C, H ⊲ (call f v)f ′ −−−−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELP -call)
f ′ = f ′′ ; f
′
f (x) 7→ s; return; ∈ C.funs
C.intfs ⊢ f ′ , f : in
call f v H?
C, H ⊲ (call f v)f ′ −−−−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELP -ret-internal)
C.intfs ⊢ f , f ′ : internal
f ′ = f ′′ ; f ′
ǫ
C, H ⊲ (return;)f ′ ;f −−→ C, H ⊲ (skip)f ′
(ELP -retback)
′
C.intfs ⊢ f , f : in
f ′ = f ′′ ; f ′
ret H?
C, H ⊲ (return;)f ′ ;f −−−−−−→ C, H ⊲ (skip)f ′
(ELP -return)
′
C.intfs ⊢ f , f : out
f ′ = f ′′ ; f ′
ret H!
C, H ⊲ (return;)f ′ ;f −−−−−→ C, H ⊲ (skip)f ′
α
Ω ==⇒ Ω′
(ELP -single)
Ω=
⇒ Ω′′
′′ α
′
Ω −−→ Ω
α
Ω ==⇒ Ω′
P
(EL -silent)
ǫ
Ω −−→ Ω′
Ω=
⇒ Ω′
40
(ELP -trans)
α
′′
Ω ==⇒ Ω
α′
Ω′′ ==⇒ Ω′
α·α′
Ω ====⇒ Ω′
C
Language and Compiler Properties
C.1
Safety, Attackers and Robust Safety
These properties hold for both languages are written in black and only once.
Definition 8 (Safety).
def
⊢ C : safe = ∀C .
if ⊢ C , ∅ : whole
α
then Ω0 (C ) ==
6 ⇒ fail
A source component is safe if whenever its monitor runs, it does not fail.
Definition 9 (Attacker).
def
M ⊢ A : attacker = no monitor inside A and no location
of M reachable or stored in A
An attacker is valid if it does not invoke monitors.
Definition 10 (Robust Safety).
def
⊢ C : rs = ∀A
if C .M ⊢ A : attacker
then ⊢ A [C ] : safe
A program is robustly safe if it is safe for any attacker it is composed with.
C.2
Cross-language Relations
Assume a partial bijection β : ℓ × n × η from source to target heap locations
such that
• if (ℓ1 , n, η) ∈ β and (ℓ2 , n, η) then ℓ1 = ℓ2 ;
• if (ℓ, n1 , η1 ) ∈ β and (ℓ, n2 , η2 ) then n1 = n2 and η1 = η2 .
we use this bijection to parametrise the relation so that we can relate meaningful
locations.
For compiler correctness we rely on a β0 which relates initial locations of
monitors.
Assume a relation ≈β : v × β × v that is total so it maps any source value
to a target value v.
• ∀v.∃v.v ≈β v.
This relation is used for defining compiler correctness. By inspecting the
semantics of LU , Rules ELP -sequence and ELP -if-true let us derive that
41
• true ≈β 0;
• false ≈β n where n 6= 0;
(
v=k
• ℓ ≈β hn, vi where
v 6= k
if (ℓ, n, k) ∈ β
otherwise, so (ℓ, n, ⊥) ∈ β
• hv1 , v2 i ≈β hv1 , v2 i iff v1 ≈β v1 and v2 ≈β v2 .
We overload the notation and use the same notation to indicate the (assumed) relation between monitor states: σ ≈ σ.
We lift this relation to sets of states point-wise and indicate it as follows:
{σ} ≈ {σ}. In these cases the bijection β is not needed as states do not have
locations inside.
Function names are related when they are the same: f ≈β f .
Variables names are related when they are the same: x ≈β x.
Substitutions are related when they replace related values for related variables: [v / x] ≈β [v / x] iff v ≈β v and x ≈β x.
α ≈β α
(Call relation)
(Callback relation)
f ≈f
v ≈β v
H ≈β H
call f v H? ≈β call f v H?
f≈f
v ≈β v
H ≈β H
call f v H! ≈β call f v H!
(Return relation)
(Returnback relation)
H ≈β H
ret H! ≈β ret H!
H ≈β H
ret H? ≈β ret H?
(Epsilon relation)
ǫ ≈β ǫ
Definition 11 (MRM). Given a monitor-specific relation σ ≈ σ on monitor
states, we say that a relation R on source and target monitors is a bisimulation if the following hold whenever M = ({σ} , , σ0 , ℓroot , σc ) and M =
({σ} , , σ0 , σc ) are related by R:
1. σ0 ≈ σ0 , and
2. σc ≈ σc , and
3. For all β containing (ℓroot , 0, kroot ) and all H, H with H ≈β H the following
hold:
(a) (σc , H, _) ∈
iff (σc , H, _) ∈
(b) (σc , H, σ ′ ) ∈
and (σc , H, σ ′ ) ∈
, and
imply ({σ},
, ℓroot , σ ′ )R({σ},
, 0, σ ′ ).
Definition 12 (M ≈ M). M ≈ M is the union of all bisimulations MRM, which
is also a bisimulation.
H ≈β H
42
(Heap relation)
(Empty relation)
H ≈ β H1 ; H2
ℓ ≈β hn, _i v ≈β v
H = H1 ; n 7→ v : η; H2
H; ℓ 7→ v ≈β H
∅ ≈β k
The heap relation is crucial. A source heap H is related to a target heap
H if for any location pointing to a value in the former, a related location
points to a related value in the target (Rule Heap relation). The base case
(Rule Empty relation) considers that in the target heap we may have keys,
which are not related to source elements.
As additional notation for states, we define when a state is stuck as follows
(Stuck state)
Ω = M;F;I;H ⊲ s
λ
∄Ω ′ , λ.Ω −−→
→ Ω′
s 6≡ skip
Ω×
A state that terminated is defined as follows; this definition is given for a concurrent version of the language too (this is relevant for languages defined later):
(Terminated state)
(Terminated state - fail)
Ω = M ; F ; I ; H ⊲ skip
Ω = fail
Ω
Ω⇓
⇓
(Terminated soup)
Ω = M;F;I;H ⊲ Π
∀π ∈ Π . M ; F ; I ; H ⊲ π
⇓
Ω⇓
To define compiler correctness, we rely on a cross-language relation for program states. Two states are related if their monitors are related and if their
whole heap is related (Rule Related states – Whole) or if they are both fails
(Rule Related states – Fail).
Ω ≈β Ω
(Related states – Whole)
Ω = M; F, F′ ; I; H ⊲ s
q yL U
Ω = M; F, F′ LP ; I; H ⊲ s
M ≈β M
H ≈β H
Ω ≈β Ω
C.3
(Related states – Fail)
Ω = fail Ω = fail
Ω ≈β Ω
Correct and Robustly-safe Compilation
S
Consider a compiler to be a function of this form: J·KT : C × M → C, taking a
source component and a target monitor and producing a target component (the
target monitor is to be related to the monitor inside the source component).
43
Definition 13 (Correct Compilation).
S
def
⊢ J·KT : cc = ∀C ≡ (M; F; I), M...∃β.
if ⊢ C : whole
M ≈β M
ǫ
Ω0 (C) =⇒ Ω
Ω⇓
S
Ω0 (C) ≈β0 Ω0 JC, MKT
ǫ
then Ω0 JC, MKST =⇒ Ω
β0 ⊆ β
and Ω ≈β Ω
Ω⇓
There are no labels because the component is whole already, so I is empty.
Definition 14 (Robust Safety Preserving Compilation).
S
def
⊢ J·KT : rs-pres = ∀C ≡ (M; F; I), M.
if ⊢ C : rs
M ≈β M
then ⊢ JC, MKST : rs
Definition 15 (Safe Compilation).
S
C.3.1
def
S
S
⊢ J·KT : safecomp = ⊢ J·KT : rs-pres ∧ ⊢ J·KT : cc
Compiling Monitors
We can change the definition of compiler to also compile the monitor. Consider
q yS
this compiler to have this type and this notation: · T : C → C.
Definition 16 (Robustly-safe Compilation with Monitors).
⊢
q yS
def
· T : rs-pres(M) = ∀C.
if ⊢ C : rs
r zS
then ⊢ C
: rs
T
44
D
Compiler from LU to LP
U
L
Definition 17 (Compiler LU to LP ). J·KL
P : C × M → C
U
JC, MKLLP is defined as follows:
q
yL U
q yL U q
yLU
M; F; I, M LP = M; F LP ; I; kroot LP
LU
if M ≈β M
LU
LU
(J·KLP -Comp)
LU
(J·KLP -Function)
Jf(x) 7→ s; return;KLP = f (x) 7→ JsKLP ; return;
LU
LU
(J·KLP -Interfaces)
JfKLP = f
Expressions
LU
JtrueKLP = 0
U
JfalseKLLP = 1
LU
LU
if true ≈β 0
(J·KLP -True)
if false ≈β 1
(J·KLLP -False)
U
LU
if n ≈β n
JnKLP = n
(J·KLP -nat)
LU
LU
(J·KLP -Var)
JxKLP = x
U
LU
L
JℓKLP = hn, vi if ℓ ≈β hn, vi
E
D
U
U
U
Jhe1 , e2 iKLLP = Je1 KLLP , Je2 KLLP
LU
LU
LU
LU
Je.1KLP = JeKLP .1
U
(J·KLLP -Pair)
LU
(J·KLP -P1)
LU
(J·KLP -P2)
Je.2KLP = JeKLP .2
U
(J·KLP -Loc)
U
U
U
J!eKLLP = !JeKLLP .1 with JeKLLP .2
LU
LU
LU
LU
LU
LU
(J·KLLP -Deref)
LU
Je ⊕ e′ KLP = JeKLP ⊕ Je′ KLP
(J·KLP -op)
LU
Je ⊗ e′ KLP = JeKLP ⊗ Je′ KLP
(J·KLP -cmp)
Statements
LU
JskipKLP = skip
U
LU
(J·KLP -Skip)
U
U
Jsu ; sKLLP = Jsu KLLP ; JsKLLP
LU
LU
U
(J·KLLP -Seq)
LU
LU
(J·KLP -Letin)
Jlet x = e in sKLP = let x = JeKLP in JsKLP
LU
LU
LU
LU
Jif e then st else se KLP = ifz JeKLP then Jst KLP else Jse KLP
LU
LU
Jlet x = new e in sKLP = let xloc = new JeKLP in
let xcap = hide xloc in
LU
let x = hxloc , xcap i in JsKLP
45
LU
(J·KLP -If)
LU
(J·KLP -New)
LU
LU
Jx := e′ KLP = let x1 = x.1 in
let x2 = x.2 in
(J·KLP -Assign)
LU
x1 := JeKLP with x2
LU
LU
Jcall f eKLP = call f JeKLP
LU
LU
(J·KLP -call)
LU
(J·KLP -Mon)
JmonitorKLP = monitor
U
Note that the case for Rule (J·KLLP -New) only works because we are in a sequential setting. In a concurrent setting an adversary could access xloc before it
Lτ
is hidden, so the definition would change. See Rule (J·KLπ -New) for a concurrent
correct implementation.
. i
h
LU
LU
J[v / x]KLP = JvKLP x
LU
Optimisation We could optimise Rule (J·KLP -Deref) as follows:
• rename the current expressions except dereferencing to b;
• reform expressions both in Lτ and LP as e ::= b | let x = b in e | !b. In
the case of LP it would be · · · | !b with b.
This allows expressions to compute e.g., pairs and projections.
LU
• rewrite the Rule (J·KLP -Deref) case for compiling !b into:
LU
let x = JbKLP in let x1 = x.1 in let x2 = x.2 in !x1 with x2.
• as expressions execute atomically, this would also scale to the compiler for
concurrent languages defined in later sections.
We do not use this approach to avoid nonstandard constructs.
D.1
U
Properties of the J·KLLP Compiler
LU
LU
Theorem 7 (Compiler J·KLP is cc). ⊢ J·KLP : cc
LU
LU
Theorem 8 (Compiler J·KLP is rs-pres ). ⊢ J·KLP : rs-pres
D.2
D.2.1
Back-translation from LP to LU
Values Backtranslation
Here is how values are back translated.
LP
hh·iiLU : v → v
46
P
hh0iiL
LU = true
LP
if n 6= 0
hhniiLU = false
L
P
where n ≈β n
hhniiLU = n
P
hhkiiL
LU = 0
D
E
LP
LP
LP
hhhv, v′ iiiLU = hhviiLU , hhv′ iiLU
LP
where ℓ ≈β hn, vi
hhhn, viiiLU = ℓ
The backtranslation is nondeterministic, as ≈β is not injective. In this case
we cannot make it injective (in the next compiler we can index it by types and
make it so but here we do not have them). This is the reason why the backtranslation algorithm returns a set of contexts, as backtranslating an action
that performs call f v H? could result in either call f true H? or call f 0 H?.
Now depending on f’s body, which is the component to be compiled, supplying
true or 0 may have different outcomes. Let us assume that the compilation of f,
when receiving call f v H? does not get stuck. If f contains if x then s else s′ ,
supplying 0 will make it stuck. However, because we generate all possible contexts, we know that we generate also the context that will not cause f to be
stuck. This is captured in Lemma 1 below.
Lemma 1 (Compiled code steps imply existence of source steps).
∀
if Ω′′ ≈β Ω′′
U
α?
Ω′′ ==⇒ C, H ⊲ JsKLLP ; s′ ρ
LU
λ!
C, H ⊲ JsKLP ; s′ ρ ==⇒ Ω′
{α?} = {α? | α? ≈β α?}
{ρ} = {ρ | ρ ≈β ρ}
then ∃αj ? ∈ {α?} , ρy ∈ {ρ} , Cj , Hj , sj ; s′j ρ′ .
αj ?
if Ω′′ ===⇒ Cj , Hj ⊲ sj ; s′j ρ′
LU
then Cj , Hj ⊲ sj ; s′j ρy ≈β C, H ⊲ JsKLP ; s′ ρ
λ!
Cj , Hj ⊲ sj ; s′j ρy ==⇒ Ω′
λ! ≈β λ!
Ω′ ≈ β Ω ′
D.2.2
Skeleton
LP
hh·iiLU : I → F
47
P
P
hhf iiL
LU = f(x) 7→ incrementCounter(); return;
(hh·iiL
LU -fun)
Functions call incrementCounter() before returning to ensure that when a
returnback is modelled, the counter is incremented right before returning and
not beforehand, as doing so would cause the possible execution of other bactranslated code blocks. Its implementation is described below.
LP
hh·iiLU : I → A
I
LP
LU
LP
(hh·iiLU -skel)
= ℓi 7→ 1;
ℓglob 7→ 0
main(x) 7→ incrementCounter(); return;
incrementCounter() 7→ see below
register(x) 7→ see below
update(x) 7→ see below
LP
hhf iiLU
∀f ∈ I
We assume compiled code does not implement functions incrementCounter,
register and update, they could be renamed to not generate conflicts if they were.
The skeleton sets up the infrastructure. It allocates global locations ℓi , which
is used as a counter to count steps in actions, and ℓglob , which is used to keep
track of attacker knowledge, as described below. Then it creates a dummy for all
functions expected in the interfaces I as well as a dummy for the main. Dummy
functions return their parameter variable and they increment the global counter
before that for reasons explained later.
D.2.3
Single Action Translation
We use the shortcut ak to indicate a list of pairs of locations and tag to access
them hn, ηi that is what the context has access to. We use functions .loc to
access obtain all locations of such a list and .cap to obtain all the capabilities
(or 0 when η = ⊥) of the list.
We use function incrementCounter to increment the contents of ℓi by one.
incrementCounter( ) 7→
let c = !ℓi in let l = ℓi in l := c + 1
Starting from location ℓg we keep a list whose elements are pairs locationsnumbers, we indicate this list as Lglob .
We use function register(hℓ, ni) which adds the pair hℓ, ni to the list Lglob .
Any time we use this we are sure we are adding a pair for which no other
existing pair in Lglob has a second projection equal to n. This function can be
48
defined as follows:
register(x) 7→
let xl = x.1
in let xn = x.2
in Lglob :: hxl, xni
Lglob is a list of pair elements, so it is implemented as a pair whose first projection
is an element (a pair) and its second projection is another list; the empty list
being 0. Where :: is a recursive function that starts from ℓglob and looks for its
last element (i.e., it performs second projections until it hits a 0), then replaces
that second projection with hhxn, xli , 0i
Lemma 2 (register(ℓ, n) does not add duplicates for n). For n supplied as paP
ǫ
−→ C; H′ ⊲ skip and h_, ni ∈
/ Lglob
rameter by hh·iiL
LU , C; H ⊲ register(ℓ, n) −
P
P
L
Proof. Simple analysis of Rules (hh·iiL
LU -call) to (hh·iiLU -ret-loc).
We use function update(n, v) which accesses the elements in the Lglob list,
then takes the second projection of the element: if it is n it updates the first
projection to v, otherwise it continues its recursive call. If it does not find an
element for n, it gets stuck
ǫ
Lemma 3 (update(n, v) never gets stuck). C; H ⊲ update(n, v) −−→ C; H′ ⊲ skip
L
P
for n and v supplied as parameters by hh·iiLU and H′ =H[ℓ 7→ v / ℓ 7→ _] for
ℓ ≈β hn, _i.
LP
LP
Proof. Simple analysis of Rules (hh·iiLU -call) and (hh·iiLU -retback).
We use the meta-level function reachable(H, v, ak) that returns a set of
pairs hn 7→ v : η, ei such that all locations in are reachable from H starting
from any location in ak ∪ v and that are not already in ak and such that e is a
sequence of source-level instructions that evaluate to ℓ such that ℓ ≈β hn, _i.
Definition 18 (Reachable).
reachable(H, v, ak) = hn 7→ v : k, ei
49
and kst ∈ kroot ∪ ak.cap
and n 7→ v : k ∈ H
n ∈ reach(nst , kst , H)
where nst ∈ v ∪ ak.loc
and H⊲!e ֒→
→ !n with k
and ∀H. H ≈β H
P
L
H ⊲ hheiiLU ֒→
→ ℓ
and (ℓ, n, k) ∈ β
Intuitively, reachable( · ) finds out which new locations have been allocated
by the compiled component and that are now reachable by the attacker (the
first projection of the pair, n 7→ v : η). Additionally, it tells how to reach those
locations in the source so that we can register(·) them for the source attacker
(the backtranslated context) to access.
In this case we know by definition that e can only contain one ! and several
·.1 or ·.2. The base case for values is as before.
LP
hh·iiLU : e → e
P
P
L
hh!eiiL
LU = !hheiiLU
LP
LP
LP
LP
hhe.1iiLU = hheiiLU .1
hhe.2iiLU = hheiiLU .2
The next function takes the following inputs: an action, its index, the previous function’s heap, the previous attacker knowledge and the stack of functions
called so far. It returns a set of: code, the new attacker knowledge, its heap, the
stack of functions called and the function where the code must be put. In the
returned parameters, the attacker knowledge, the heap and the stack of called
functions serve as input to the next call.
LP
hh·iiLU : α × n ∈ N × H × n × η × f → s × n × η × H × f × f
**
call f v H?,
n, Hpre , ak, f
++LP
LU
if !ℓi == n then
incrementCounter()
let x1 = new v1 in register(hx1, n1 i)
···
let
xj
=
new
v
in
register(hxj,
n
i)
j
j
=
update(m1 , u1 )
···
update(m
,
u
)
l
l
call
f
v
else skip
, ak′ , H, f; f, f ′
LP
∀
P
L
v1 = hhv1 iiLU
···
LP
v = hhv ii
j
u1 =
···
LP
ul = hhul iiLU
Hn = n1 7→ v1 : η1 , · · · , nj 7→ vj : ηj
and H ∩ Hpre = Hc
Hc = m1 7→ u1 : η1′ , · · · , ml 7→ ul : ηl′
and ak′ = ak, hn1 , η1 i , · · · , hnj , ηj i
50
LP
v = hhviiLU
(hh·iiLU -call)
where H \ Hpre = Hn
j LU
LP
hhu1 iiLU
and f = f ′ f ′
**
call f v H!,
n, Hpre , ak, f
++LP
LU
if !ℓi == n then
incrementCounter()
let l1 = e in register(hl1, n i)
1
1
′
=
,
ak
,
H,
f;
f,
f
· · ·
let lj = ej in register(hlj, nj i)
else skip
LP
(hh·iiLU -callback-loc)
if reachable(H, v, ak) = hn1 7→ v1 : η1 , e1 i , · · · , hnj 7→ vj : ηj , ej i
and ak′ = ak, hn1 , η1 i , · · · , hnj , ηj i
**
ret H?,
n, Hpre , ak, f; f
++LP
LU
if !ℓi == n then
//no incrementCounter() as explained
let x1 = new v1 in register(hx1, n1 i)
···
let xj = new vj in register(hxj, nj i)
=
update(m1 , u1 )
···
update(m
,
u
)
l l
else skip
, ak′ , H, f, f
LP
∀
P
L
v1 = hhv1 iiLU
···
P
L
vj = hhvj iiLU
LP
u1 = hhu1 iiLU
···
P
L
ul = hhul iiLU
(hh·iiLU -retback)
where H \ Hpre = Hn
Hn = n1 7→ v1 : η1 , · · · , nj 7→ vj : ηj
and H ∩ Hpre = Hc
Hc = m1 7→ u1 : η1′ , · · · , ml 7→ ul : ηl′
and ak′ = ak, hn1 , η1 i , · · · , hnj , ηj i
**
ret H!,
n, Hpre , ak, f; f
++LP
LU
if !ℓi == n then
incrementCounter()
let l1 = e in register(hl1, n i)
1
1
′
′
=
,
ak
,
H,
f,
f
···
let
lj
=
e
in
register(hlj,
n
i)
j
j
else skip
P
(hh·iiL
LU -ret-loc)
51
if reachable(H, 0, ak) = hn1 7→ v1 : η1 , e1 i , · · · , hnj 7→ vj : ηj , ej i
and ak′ = ak, hn1 , η1 i , · · · , hnj , ηj i
and f = f ′ f ′
This is the back-translation of functions. Each action is wrapped in an if
statement checking that the action to be mimicked is that one (the same function
may behave differently if called twice and we need to ensure this). After the if,
the counter checking for the action index ℓi is incremented. This is not done
in case of a return immediately, but only just before the return itself, so the
increment is added in the skeleton already. (there could be a callback to the
same function after the return and then we wouldn’t return but execute the
callback code instead)
When back-translating a ?-decorated, we need to set up the heap correctly
before the call itself. That means calculating the new locations that this action
allocated (Hn ), allocating them and registering them in the Lglob list via the
register(·) function. These locations are also added to the attacker knowledge
ak′ . Then we need to update the heap locations we already know of. These
locations are Hc and as we know them already, we use the update(·) function.
When back-translating a !-decorated action we need to calculate what part of
the heap we can reach from there, and so we rely on the reachable( · ) function
to return a list of pairs of locations n and expressions e. We use n to expand
the attacker knowledge ak′ as these locations are now reachable. We use e to
reach these locations in the source heap so that we can register them and ensure
they are accessible through Lglob .
Finally, we use parameter f to keep track of the call stack, so making a call
to f pushes f on the stack (f; f) and making a return pops a stack f; f to f. That
stack carries the information to instantiate the f in the return parameters, which
is the location where the code needs to be allocated.
LP
hh·iiLU : α × n ∈ N × H × n × η × f → s, f
P
hh∅iiL
LU = ∅
LP
αα, n, Hpre , ak, f LU = s, f; s, f
P
(hh·iiL
LU -listact-b)
LP
s, ak′ , H′ , f ′ , f = α, n, Hpre , ak, f LU
LP
s, f ∈ α, n + 1, H′ , ak′ , f ′ LU
LP
(hh·iiLU -listact-i)
This recursive call ensures the parameters are passed around correctly. Note
that each element in a set returned by the single-action back-translation has
the same ak′ , H and f ′ , the only elements that change are in the code s due to
the backtranslation of values. Thus the recursive call can pass those parameters
taken from any element of the set.
52
D.2.4
P
The Back-translation Algorithm hh·iiL
LU
LP
hh·iiLU : I × α → {A}
I, α
LP
LU
=
A
A = Askel ✶ s, f
for all s, f ∈ s, f
P
L
where s, f = hhα, 1, H0 , ∅, mainiiLU
H0 = 0 7→ 0 : kroot
P
L
Askel = I LU
LP
(hh·iiLU -main)
This is the real back-translation algorithm: it calls the skeleton and joins it
with each element of the set returned by the trace back-translation.
✶: A × s, f → A
LP
A✶∅=A
(hh·iiLU -join)
H; F1 ; · · · ; F; · · · ; Fn ✶ s, f; s, f = H; F1 ; · · · ; F′ ; · · · ; Fn ✶ s, f
where F = f(x) 7→ s′ ; return;
F′ = f(x) 7→ s; s′ ; return;
When joining we add from the last element of the list so that the functions
we create have the concatenation of if statements (those guarded by the counter
on ℓi ) that are sorted (guards with a test for ℓi = 4 are before those with a test
ℓi = 5).
D.2.5
Correctness of the Back-translation
LP
Theorem 9 (hh·iiLU is correct).
∀
h
i
α
LU
==⇒ Ω
if Ω0 A JCKLP
ǫ
Ω =⇒ Ω′
I = names(A)
α ≡ α′ · α?
ℓi ; ℓglob ∈
/β
LP
then ∃A ∈ hhI, αiiLU
α
such that Ω0 (A [C]) ==⇒ Ω
and α ≈β α
53
Ω ≈β Ω
Ω.H.ℓi = ||α|| + 1
The back-translation is correct if it takes a target attacker that will reduce
to a state together with a compiled component and it produces a set of source
attackers such that one of them, that together with the source component will
reduce to a related state performing related actions. Also it needs to ensure the
step is incremented correctly.
D.2.6
Remark on the Backtranslation
Some readers may wonder whether the hassle of setting up a source-level representation of the whole target heap is necessary. Indeed for those locations that
are allocated by the context, this is not. If we changed the source semantics to
have an oracle that predicts what a let x = new e in s statement will return as
the new location, we could simplify this. In fact, currently the backtranslation
stores target locations in the list Lglob and looks them up based on their target
name, as it does not know what source name will be given to them. The oracle
would obviate this problem, so we could hard code the name of these locations,
knowing exactly the identifier that will be returned by the allocator. For the
functions to be correct in terms of syntax, we would need to pre-emptively allocate all the locations with that identifier so that their names are in scope and
they can be referred to.
However, the problem still persists for locations created by the component,
as their names cannot be hard coded, as they are not in scope. Thus we would
still require reach to reach these locations, register to add them to the list and
update to update their values in case the attacker does so.
Thus we simplify the scenario and stick to a more standard, oracle-less semantics and to a generalised approach to location management in the backtranslation.
54
E
The Source Language: Lτ
This is an imperative, concurrent while language with monitors.
Whole Programs P ::= M; H; F; I
Components C ::= M; F; I
Contexts A ::= H; F [·]
Interfaces I ::= f
Functions F ::= f(x : τ ) 7→ s; return;
Operations ⊕ ::= + | −
Comparison ⊗ ::= == | < | >
Values v ::= b ∈ {true, false} | n ∈ N | hv, vi | ℓ
Expressions e ::= x | v | e ⊕ e | e ⊗ e | !e | he, ei | e.1 | e.2
Statements s ::= skip | s; s | let x : τ = e in s | if e then s else s
| x := e | let x = newτ e in s | call f e
| monitor | {k s} | endorse x = e as ϕ in s
Types τ ::= Bool | Nat | τ × τ | Ref τ | UN
Superficial Types ϕ ::= Bool | Nat | UN × UN | Ref UN
Eval . Ctxs. E ::= [·] | e ⊕ E | E ⊕ n | e ⊗ E | E ⊗ n
| !E | he, Ei | hE, vi | E.1 | E.2
Heaps H ::= ∅ | H; ℓ 7→ v : τ
Monitors M ::= ({σ} ,
Mon. States σ ∈ S
Mon. Reds.
::= ∅ |
, σ0 , ∆, σc )
; (s, s)
Environments Γ , ∆ ::= ∅ | Γ; (x : τ )
Store Env . ∆ ::= ∅ | ∆; (ℓ : τ )
Substitutions ρ ::= ∅ | ρ[v / x]
Processes π ::= (s)f
Soups Π ::= ∅ | Π k π
Prog. States Ω ::= C, H ⊲ Π | fail
Labels λ ::= ǫ | α
Actions α ::= call f v? | call f v! | ret ! | ret ?
Traces α ::= ∅ | α · α
We highlight elements that have changed from LU .
55
E.1
Static Semantics of Lτ
The static semantics follows these typing judgements.
Component C is well-typed.
Function F takes arguments of type τ under component C.
⊢ C : UN
C⊢F:τ
∆, Γ ⊢ ⋄
Environments Γ and ∆ are well-formed.
∆ ⊢ ok
τ ⊢◦
Environment ∆ is safe.
Type τ is insecure.
∆, Γ ⊢ e : τ
C, ∆, Γ ⊢ s
Expression e has type τ inΓ.
Statement s is well-typed in C and Γ.
C, ∆, Γ ⊢ π
C, ∆, Γ ⊢ Π
Single process π is well-typed in C and Γ.
Soup Π is well-typed in C and Γ.
⊢H:∆
⊢M
Heap H respects the typing of ∆.
Monitor M is valid.
E.1.1
Auxiliary Functions
We rely on these standard auxiliary functions: names( · ) extracts the defined
names (e.g., function and interface names). fv( · ) returns free variables while
fn( · ) returns free names (i.e., a call to a defined function). dom( · ) returns
the domain of a particular element (e.g., all the allocated locations in a heap).
We denote access to the parts of C and P via functions .funs, .intfs and .mon.
We denote access to parts of M with a dot notation, so M.∆ means ∆ where
M = ({σ} , , σ0 , ∆, σc ).
E.1.2
Typing Rules
⊢C
(TLτ -component)
C ≡ M; F; I
C ⊢ F : UN names(F) ∩ names(I) = ∅
M = ({σ} , , σ0 , ∆, σc )
∆ ⊢ ok
⊢M
⊢ C : UN
C ⊢ F : UN
(TLτ -function)
F ≡ f(x : UN) 7→ s; return;
C, ∆; x : UN ⊢ s
∀f ∈ fn(s), f ∈ dom(C.funs) ∨ f ∈ dom(C.intfs)
C ⊢ F : UN
∆, Γ ⊢ ⋄
56
(TLτ -env-e)
∅; ∅ ⊢ ⋄
(TLτ -env-var)
(TLτ -env-loc)
∆, Γ ⊢ ⋄ x ∈
/ dom(Γ)
∆, Γ; (x : τ ) ⊢ ⋄
∆, Γ ⊢ ⋄ l ∈
/ dom(∆)
∆; (l : τ ); Γ ⊢ ⋄
∆, Γ ⊢ ok
(TLτ -safe-loc)
(TLτ -safe-e)
∆ ⊢ ok
∅ ⊢ ok
l∈
/ dom(Γ)
UN ∈
/τ
Γ; (l : τ ) ⊢ ok
∆, Γ ⊢ UN
(TLτ -env-e)
∅ ⊢ UN
(TLτ -env-var)
(TLτ -env-loc)
∆, Γ ⊢ UN x ∈
/ dom(Γ)
Γ, (x : UN) ⊢ UN
∆, Γ ⊢ UN l ∈
/ dom(Γ)
Γ, (l : UN) ⊢ UN
τ ⊢◦
(TLτ -bool-pub)
(TLτ -nat-pub)
Bool ⊢ ◦
Nat ⊢ ◦
(TLτ -pair-pub)
τ ⊢ ◦ τ′ ⊢ ◦
τ × τ′ ⊢ ◦
(TLτ -un-pub)
UN ⊢ ◦
(TLτ -references-pub)
Ref UN ⊢ ◦
∆, Γ ⊢ e : τ
(TLτ -true)
(TLτ -false)
(TLτ -nat)
∆, Γ ⊢ ⋄
∆, Γ ⊢ true : Bool
∆, Γ ⊢ ⋄
∆, Γ ⊢ false : Bool
∆, Γ ⊢ ⋄
∆, Γ ⊢ n : Nat
(TLτ -pair)
(TLτ -var)
(TLτ -loc)
x:τ ∈Γ
∆, Γ ⊢ x : τ
l:τ ∈∆
∆, Γ ⊢ l : Ref τ
(TLτ -proj-2)
(TLτ -proj-1)
∆, Γ ⊢ e : τ × τ
∆, Γ ⊢ e.1 : τ
∆, Γ ⊢ e1 : τ
∆, Γ ⊢ e2 : τ ′
∆, Γ ⊢ he1 , e2 i : τ × τ ′
′
∆, Γ ⊢ e : τ × τ
∆, Γ ⊢ e.2 : τ ′
′
(TLτ -op)
(TLτ -dereference)
∆, Γ ⊢ e : Ref τ
∆, Γ ⊢ !e : τ
(TLτ -cmp)
′
∆, Γ ⊢ e : Nat ∆, Γ ⊢ e′ : Nat
∆, Γ ⊢ e ⊗ e′ : Bool
∆, Γ ⊢ e : Nat
∆, Γ ⊢ e : Nat
∆, Γ ⊢ e ⊕ e′ : Nat
(TLτ -coercion)
C, ∆, Γ ⊢ e : τ
τ ⊢◦
C, ∆, Γ ⊢ e : UN
C, ∆, Γ ⊢ s
57
(TLτ -function-call)
(TLτ -skip)
((f ∈ dom(C.funs)) ∨ (f ∈ dom(C.intfs)))
∆, Γ ⊢ e : UN
∆, Γ ⊢ call f e
C, ∆, Γ ⊢ skip
(TLτ -sequence)
(TLτ -letin)
(TLτ -assign)
C, ∆, Γ ⊢ su
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ su ; s
∆, Γ ⊢ e : τ
C, Γ; x : τ ⊢ s
C, ∆, Γ ⊢ let x : τ = e in s
∆, Γ ⊢ x : Ref τ
∆, Γ ⊢ e′ : τ
C, ∆, Γ ⊢ x := e′
(TLτ -new)
(TLτ -if)
∆, Γ ⊢ e : τ
C, Γ; x : Ref τ ⊢ s
C, ∆, Γ ⊢ let x = newτ e in s
∆, Γ ⊢ e : Bool
C, ∆, Γ ⊢ st
C, ∆, Γ ⊢ se
C, ∆, Γ ⊢ if e then st else se
(TLτ -fork)
(TLτ -monitor)
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ {k s}
C, ∆, Γ ⊢ monitor
(TLτ -endorse)
∆, Γ ⊢ e : UN
C, ∆, Γ; (x : ϕ) ⊢ s
C, ∆, Γ ⊢ endorse x = e as ϕ in s
C, ∆, Γ ⊢ π
(TLτ -process)
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ (s)f
C, ∆, Γ ⊢ Π
(TLτ -soup)
C, ∆, Γ ⊢ π
C, ∆, Γ ⊢ Π
C, ∆, Γ ⊢ π k Π
⊢H:∆
(Lτ -Heap-ok-i)
ℓ :7→ v : τ ∈ H
⊢H:∆
∅⊢v:τ
⊢ H : ∆; ℓ : τ
(Lτ -Heap-ok-b)
⊢H:∅
⊢M
(Lτ -Monitor)
M ≡ ({s} ,
, s0 , ∆, sc )
⊢M
58
∀s∃s′ .(s, s′ ) ∈
Notes Monitor typing just ensures that the monitor is coherent and that it
can’t get stuck for no good reason.
E.1.3
UN Typing
Attackers cannot have monitor nor newτ t terms where τ is different from UN.
∆, Γ ⊢UN e : UN
(TULτ -true)
(TULτ -false)
(TULτ -nat)
∆, Γ ⊢ ⋄
∆, Γ ⊢UN true : UN
∆, Γ ⊢ ⋄
∆, Γ ⊢UN false : UN
∆, Γ ⊢ ⋄
∆, Γ ⊢UN n : UN
(TULτ -pair)
(TULτ -var)
(TULτ -loc)
x:τ ∈Γ
∆, Γ ⊢UN x : UN
l:τ ∈∆
∆, Γ ⊢UN l : UN
∆, Γ ⊢UN e1 : UN
∆, Γ ⊢UN e2 : UN
∆, Γ ⊢UN he1 , e2 i : UN
(TULτ -proj-1)
(TULτ -proj-2)
(TULτ -dereference)
∆, Γ ⊢UN e : UN
∆, Γ ⊢UN e.1 : UN
∆, Γ ⊢UN e : UN
∆, Γ ⊢UN e.2 : UN
∆, Γ ⊢UN e : UN
∆, Γ ⊢UN !e : UN
(TULτ -op)
∆, Γ ⊢UN e : UN ∆, Γ ⊢UN e′ : UN
∆, Γ ⊢UN e ⊕ e′ : UN
(TULτ -cmp)
∆, Γ ⊢UN e : UN
∆, Γ ⊢UN e′ : UN
∆, Γ ⊢UN e ⊗ e′ : UN
C, ∆, Γ ⊢UN s
(TULτ -function-call)
(TULτ -skip)
((f ∈ dom(C.funs)) ∨ (f ∈ dom(C.intfs)))
∆, Γ ⊢UN e : UN
∆, Γ ⊢UN call f e
C, ∆, Γ ⊢UN skip
(TULτ -sequence)
(TULτ -letin)
C, ∆, Γ ⊢UN su
C, ∆, Γ ⊢UN s
C, ∆, Γ ⊢UN su ; s
∆, Γ ⊢UN e : UN
C, Γ; x : UN ⊢UN s
C, ∆, Γ ⊢UN let x : UN = e in s
(TULτ -assign)
(TULτ -new)
∆, Γ ⊢UN x : UN
∆, Γ ⊢UN e′ : UN
C, ∆, Γ ⊢UN x := e′
∆, Γ ⊢UN e : UN
C, Γ; x : UN ⊢UN s
C, ∆, Γ ⊢UN let x = newUN e in s
(TULτ -if)
∆, Γ ⊢UN e : Bool
C, ∆, Γ ⊢UN st
C, ∆, Γ ⊢UN se
C, ∆, Γ ⊢UN if e then st else se
59
(TULτ -fork)
C, ∆, Γ ⊢UN s
C, ∆, Γ ⊢UN {k s}
E.2
Dynamic Semantics of Lτ
Function mon-care( · ) returns the part of a heap the monitor cares for (Rule Lτ -Monitor-related heap).
Rules Lτ -Jump-Internal to Lτ -Jump-OUT dictate the kind of a jump between
two functions: if internal to the component/attacker, in(from the attacker to the
component) or out(from the component to the attacker). Rule Lτ -Plug tells how
to obtain a whole program from a component and an attacker. Rule Lτ -Initial State
tells the initial state of a whole program. Rule Lτ -Initial-heap produces a heap
that satisfies a ∆, initialised with base values. Rule Lτ -Monitor Step tells when
a monitor makes a single step given a heap.
mon-care( · )
(Lτ -Monitor-related heap)
H′ = {ℓ 7→ v : τ | ℓ 7→ v : τ ∈ H}
⊢ mon-care(H, ∆) = H′
Helpers
(Lτ -Jump-Internal)
′
(Lτ -Jump-IN)
′
((f ∈ I ∧ f ∈ I)∨
/ I∧f ∈
/ I))
(f ′ ∈
(Lτ -Jump-OUT)
′
f ∈I∧f ∈
/I
I ⊢ f, f ′ : in
I ⊢ f, f ′ : internal
f∈
/ I∧f ∈I
I ⊢ f, f ′ : out
(Lτ -Plug)
M; F′ ; I
M
A ≡ H; F [·]
C≡
≡ ({σ} , , σ0 , ∆, σ0 )
dom(H) ∩ dom(∆) = ∅
dom(∆) ∩ (fv(F) ∪ fv(H)) = ∅
⊢ C, F : whole ∆ ⊢ H0
main(x : UN) 7→ s; return; ∈ F
A [C] = M; H ∪ H0 ; F; F′ ; I
(Lτ -Whole)
C ≡ M; F′ ; I
(Lτ -Initial State)
P ≡ M; H; F; I
C ≡ M; F; I
Ω0 (P) = C, H ⊲ call main 0
names(F) ∩ names(F′ ) = ∅
names(I) ⊆ names(F) ∪ names(F′ )
⊢ C, F : whole
∆ ⊢ H0
(Lτ -Initial-heap)
∆⊢H ∅⊢v:τ
∆, ℓ : τ ⊢ H; ℓ 7→ v : τ
M; H
M′
(Lτ -Monitor Step)
, σ0 , ∆, σc )
M′ = ({σ} ,
M = ({σ} ,
(σc , σf ) ∈
, σ0 , ∆, σf )
⊢H:∆
M′
M; H
60
E.2.1
Component Semantics
H ⊲ e ֒→
→ e′
Expression e reduces to e′ .
λ
Process π reduces to π ′ and evolves the rest accordingly.
λ
Soup Π reduce to Π′ and evolve the rest accordingly.
C, H ⊲ π −−→ C′ , H′ ⊲ π
C, H ⊲ Π −−→ C′ , H′ ⊲ Π′
α
Program state Ω steps to Ω′ emitting trace α.
Ω ==⇒ Ω′
H ⊲ e ֒→
→ e′
(ELτ -p1)
(ELτ -val)
(ELτ -p2)
H ⊲ hv, v′ i .1 ֒→
→ v
H ⊲ v ֒→
→ v
H ⊲ hv, v′ i .1 ֒→
→ v′
τ
τ
(EL -op)
(EL -comp)
n ⊕ n′ = n′′
H ⊲ n ⊕ n′ ֒→
→ n′′
n ⊗ n′ = b
H ⊲ n ⊗ n′ ֒→
→ b
(ELτ -ctx)
(ELτ -dereference)
H ⊲ e ֒→
→ e′
H ⊲ E [e] ֒→
→ E [e′ ]
H ⊲ e ֒→
→ ℓ
ℓ 7→ v : τ ∈ H
H⊲!ℓ ֒→
→ v
ǫ
C, H ⊲ π −−→ C′ , H′ ⊲ π ′
(ELτ -step)
λ
(ELτ -sequence)
C, H ⊲ s −−→ C, H ⊲ s′
ǫ
C, H ⊲ skip; s −−→ C, H ⊲ s
λ
C, H ⊲ s; s′′ −−→ C, H ⊲ s′ ; s
(EL -if-true)
τ
H ⊲ e ֒→
→ true
ǫ
C, H ⊲ if e then s else s′ −−→ C, H ⊲ s
(ELτ -if-false)
H ⊲ e ֒→
→ false
ǫ
C, H ⊲ if e then s else s′ −−→ C, H ⊲ s′
(ELτ -letin)
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x : τ = e in s −−→ C, H ⊲ s[v / x]
(ELτ -alloc)
ℓ∈
/ dom(H)
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = newτ e in s −−→ C, H; ℓ 7→ v : τ ⊲ s[ℓ / x]
(ELτ -update)
H1 ; ℓ 7→ v′ :
H=
τ ; H2
H′ = H1 ; ℓ 7→ v : τ ; H2
ǫ
C, H ⊲ ℓ := v −−→ C, H′ ⊲ skip
61
(ELτ -monitor)
, s0 , ∆, sc ) C = M; F; I C′ = M′ ; F; I
M; mon-care(H, ∆)
M′
M = ({s} ,
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
τ
(EL -monitor-fail)
, s0 , ∆, sc ) C = M; F; I C′ = M′ ; F; I
¬M; mon-care(H, ∆)
M′
M = ({s} ,
ǫ
C, H ⊲ monitor −−→ fail
(EL -endorse)
τ
∆, ∅ ⊢ v : ϕ
H ⊲ e ֒→
→ v
∆ = {ℓ : τ | ℓ 7→ v : τ ∈ H}
ǫ
C, H ⊲ endorse x = e as ϕ in s −−→ C, H ⊲ s[v / x]
(ELτ -call-internal)
′
C.intfs ⊢ f, f : internal
f(x : τ ) : τ ′ 7→ s; return; ∈ C.funs
f ′ = f ′′ ; f ′
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ (call f e)f ′ −−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELτ -callback)
′
′
f ′ = f ′′ ; f
f(x : τ ) : τ 7→ s; return; ∈ F
C.intfs ⊢ f ′ , f : out
H ⊲ e ֒→
→ v
call f v!
C, H ⊲ (call f e)f ′ −−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELτ -call)
f ′ = f ′′ ; f ′
f(x : τ ) : τ ′ 7→ s; return; ∈ C.funs
C.intfs ⊢ f ′ , f : in
H ⊲ e ֒→
→ v
call f v?
C, H ⊲ (call f e)f ′ −−−−−−−→ C, H ⊲ (s; return;[v / x])f ′ ;f
(ELτ -ret-internal)
′
C.intfs ⊢ f, f : internal
f ′ = f ′′ ; f ′
ǫ
C, H ⊲ (return;)f ′ ;f −−→ C, H ⊲ (skip)f ′
(ELτ -retback)
′
C.intfs ⊢ f, f : in
f ′ = f ′′ ; f ′
ret ?
C, H ⊲ (return;)f ′ ;f −−−−→ C, H ⊲ (skip)f ′
′
C.intfs ⊢ f, f :
(ELτ -return)
out f ′ = f ′′ ; f ′
H ⊲ e ֒→
→ v
ret !
C, H ⊲ (return;)f ′ ;f −−−−→ C, H ⊲ (skip)f ′
λ
C, H ⊲ Π −−→ C′ , H′ ⊲ Π′
(ELτ -par)
(ELτ -fail)
Π = Π1 k (s)f k Π2
Π′ = Π1 k (s′ )f ′ k Π2
λ
Π = Π1 k (s)f k Π2
ǫ
C, H ⊲ (s)f −−→ fail
λ
C, H ⊲ Π −−→ fail
C, H ⊲ (s)f −−→ C′ , H′ ⊲ (s′ )f ′
C, H ⊲ Π −−→ C′ , H′ ⊲ Π′
62
ǫ
(ELτ -fork)
Π = Π1 k ({k s}; s′ )f k Π2
Π′ = Π1 k (skip; s′ )f k Π2 k (s)∅
ǫ
C, H ⊲ Π −−→ C, H ⊲ Π′
α
Ω ==⇒ Ω′
(ELτ -single)
Ω=
⇒ Ω′′
′′ α
′
Ω −−→ Ω
α
Ω ==⇒ Ω′
(ELτ -silent)
ǫ
′
Ω −−→ Ω
Ω=
⇒ Ω′
63
(ELτ -trans)
α
′′
Ω ==⇒ Ω
α′
Ω′′ ==⇒ Ω′
α·α′
Ω ====⇒ Ω′
F
F.1
Lπ : Extending LP with Concurrency and Informed Monitors
Syntax
This extends the syntax of Appendix B.1 with concurrency and a memory allocation instruction that atomically hides the new location.
Statements s ::= · · · | {k s} | destruct x = e as B in s or s
| let x = newhide e in s
Patterns B ::= nat | pair
Monitors M ::= ({σ} , , σ0 , H0 , σc )
Single Process π ::= (s)f
Processes Π ::= ∅ | Π k π
Prog. States Ω ::= C, H ⊲ Π | fail
F.2
Dynamic Semantics
Following is the definition of the mon-care( · ) function for Lπ .
mon-care( · )
(Lπ -Monitor-related heap)
H′ = {n 7→ v : η | n ∈ dom(H0 ) and n 7→ v : η ∈ H}
mon-care(H, H0 ) = H′
Helpers
(Lπ -Plug)
A ≡ F [·]
C ≡ M; F′ ; I
⊢ C, F : whole
main(x) 7→ return;s ∈ F
M ≡ ({σ} , , σ0 , H0 , σ0 )
∀k ∈ H0 .k ∈
/ fv(A)
∀n 7→ v : η ∈ H0 , η = k ∧ k ∈ H0
A [C] = M; F; F′ ; I
(Lπ -Initial State)
P ≡ M; F; I M ≡ ({σ} , , σ0 , H0 , σ0 )
Ω0 (P) = C, H0 ⊲ call main 0
M; H
M = ({σ} ,
M′
(Lπ -Monitor Step)
, σ0 , H0 , σc )
M′ =
({σ} ,
(sc , mon-care(H, H0 ), sf ) ∈
M; H
M′
64
, σ0 , H0 , σf )
F.2.1
Component Semantics
ǫ
C, H ⊲ Π −−→ C′ , H′ ⊲ Π′
Processes Π reduce to Π′ and evolve the rest accordingly.
ǫ
C, H ⊲ s −−→ C′ , H′ ⊲ s′
(ELπ -destruct-nat)
H ⊲ e ֒→
→ n
ǫ
C, H ⊲ destruct x = e as nat in s or s′ −−→ C, H ⊲ s[n / x]
(ELπ -destruct-pair)
′
H ⊲ e ֒→
→ hv, v i
ǫ
C, H ⊲ destruct x = e as pair in s or s′ −−→ C, H ⊲ s[hv, v′ i / x]
(ELπ -destruct-not)
otherwise
ǫ
C, H ⊲ destruct x = e as B in s or s′ −−→ C, H ⊲ s′
(ELπ -new)
H = H1 ; n 7→ (v, η)
H ⊲ e ֒→
→ v
k∈
/ dom(H)
ǫ
C, H ⊲ let x = newhide e in s −−→ C, H; n + 1 7→ v : k; k ⊲ s[hn + 1, ki / x]
C, H ⊲ Π ֒→
→ C′ , H ′ ⊲ Π ′
(ELπ -par)
Π = Π1 k (s)f k Π2
Π′ = Π1 k (s′ )f ′ k Π2
C, H ⊲ (s)f ֒→
→ C′ , H′ ⊲ (s′ )f ′
C, H ⊲ Π ֒→
→ C′ , H ′ ⊲ Π ′
(ELπ -fail)
Π = Π1 k (s)f k Π2
→ fail
C, H ⊲ (s)f ֒→
C, H ⊲ Π ֒→
→ fail
(ELπ -fork)
Π = Π1 k ({k s})f k Π2
Π′ = Π1 k (0)f k Π2 k (s)∅
C, H ⊲ Π ֒→
→ C, H ⊲ Π′
65
G
Extended Language Properties and Necessities
G.1
Properties of Lτ
Definition 19 (Lτ Semantics Attacker).
no monitor in A
def
C ⊢attacker A = ∀ℓ ∈ dom(C.∆), ℓ ∈
/ fn(A)
no let x = newτ e in s in A such that τ 6= UN
This semantic definition of an attacker is captured by typing below, which
allows for simpler reasoning.
Definition 20 (Lτ Attacker).
def
C ⊢att A = ⊢UN A
def
C ⊢att π = π = (s)f;f and f ∈ C.itfs
ǫ
def
C ⊢att Π −−→ Π′ = Π = Π1 k π k Π2 and Π′ = Π1 k π ′ k Π2
and C ⊢att π and C ⊢att π ′
The two notions of attackers coincide.
Lemma 4 (Semantics and typed attackers coincide).
if A [C]
then C ⊢attacker A ⇐⇒ (C ⊢att A)
Theorem 10 (Typability Implies Robust Safety in Lτ ).
∀C
if ⊢ C : UN
then ⊢ C : rs
G.2
Properties of Lπ
Definition 21 (Lπ Attacker).
def
⊢att A = no monitor inside A
def
C ⊢att π = π = (s)f ;f and f ∈ C.itfs
ǫ
def
C ⊢att Π −−→ Π′ = Π = Π1 k π k Π2 and Π′ = Π1 k π ′ k Π2
and C ⊢att π and C ⊢att π ′
66
Compiler from Lτ to Lπ
H
H.1
Assumed Relation between Lτ and Lπ Elements
We can scale the ≈β relation to monitors, heaps, actions and processes as follows.
M≈M
(Ok Mon)
M = ({σ} , , σ0 , H0 , σc )
∀σ ∈ {σ}, ∀H ≈β H. if ⊢ H : ∆ then ∃σ ′ .(σ, H, σ ′ ) ∈
β, ∆ ⊢ M
( Monitor relation )
M = ({σ} , , σ0 , ∆, σc )
M = ({σ} , , σ0 , H0 , σc )
β0 , ∆ ⊢ M
β0 = (dom(∆), dom(H0 ), H0 .η)
M≈M
∆ ⊢ H0 ∆ : τ ⊢ v
(Initial-heap)
∆ ⊢ H ∆, H ⊢ v: τ
ℓ ≈β hn, ki
∆, ℓ : τ ⊢ H; n 7→ v : k
(Initial-value)
∨
(τ ≡ Nat ∧ v ≡ 0)
∨
(τ ≡ Bool ∧ v ≡ 0)
(τ ≡ Ref τ ∧ v ≡ n′ ∧ n′ 7→ v′ : η ∈ H ∧ ℓ′ ≈β hn′ , k′ i ∧ ℓ : τ ∈ ∆, ∆, H ⊢ v′ : τ )
(τ ≡ τ1 × τ2 ∧ v ≡ hv1 , v2 i ∧ ∆, H ⊢ v1 : τ1 ∧ ∆, H ⊢ v2 : τ2 )
∆, H ⊢ v: τ
Π ≈β Π
H.2
(Single process relation)
(Process relation)
f ≈f
(skip)f ≈β (skip)f
Π ≈β Π π ≈β π
Π k π ≈β Π k π
Compiler Definition
The compiler needs a second argument to inform the translation of a monitorτ
sensitive location to hide · that (Rule (J·KLLπ -Loc))
Lτ
Definition 22 (Compiler Lτ to Lπ ). J·KLπ : C × M → C
67
∨
Given that C = M; F; I and M = ({σ} , , σ0 , H0 , σc ) and M = ({σ} ,
Lτ
if ⊢ C : UN and M ≈ M then JC, MKLπ is defined as follows:
}Lτ
u
τ
, σ0 , ∆, σc ),
(TL -component)
w
w
w
w
w
w
w
v
C ≡ M; F; I
C ⊢ F : UN
names(F) ∩ names(I) = ∅
∆ ⊢ ok
⊢M
⊢ C : UN
u
, M
~
(TLτ -function)
w
w
w
w
w
v
F ≡ f(x : UN) 7→ s; return;
C, ∆; x : UN ⊢ s
∀f ∈ fn(s), f ∈ dom(C.funs)
∨f ∈ dom(C.intfs)
C ⊢ F : UN → UN
q yL τ q yL τ
= M; F Lπ ; I Lπ
Lπ
}Lτ
~
Lπ
τ
(J·KLLπ -Component)
Lτ
= f (x) 7→ JC; ∆; x : UN ⊢ sKLπ ; return;
Lτ
(J·KLπ -Function)
Lτ
Expressions
u
(TLτ -true)
v
∆, Γ ⊢ ⋄
∆, Γ ⊢ true : Bool
v
∆, Γ ⊢ ⋄
∆, Γ ⊢ false : Bool
u
u
(TLτ -false)
(TL -nat)
τ
v
∆, Γ ⊢ ⋄
∆, Γ ⊢ n : Nat
t
t
u
w
w
v
(TLτ -var)
x:τ ∈Γ
∆, Γ ⊢ x : τ
(TLτ -loc)
ℓ:τ ∈∆
∆, Γ ⊢ ℓ : τ
(TLτ -pair)
∆, Γ ⊢ e1 : τ
∆, Γ ⊢ e2 : τ ′
∆, Γ ⊢ he1 , e2 i : τ × τ ′
}Lτ
~
if true ≈β 0
(J·KLπ -True)
=1
if false ≈β 1
(J·KLπ -False)
=n
if n ≈β n
Lτ
Lπ
}Lτ
~
Lτ
=0
Lπ
}Lτ
~
Lτ
(J·KLπ -Interfaces)
JfKLπ = f
Lτ
(J·KLπ -Nat)
Lπ
|Lτ
=x
|Lτ
= hn, vi
Lτ
(J·KLπ -Var)
Lπ
if ℓ ≈β hn, vi
Lτ
(J·KLπ -Loc)
Lπ
}Lτ
~
Lπ
D
E
Lτ
Lτ
= J∆, Γ ⊢ e1 : τ KLπ , J∆, Γ ⊢ e2 : τ ′ KLπ
Lτ
(J·KLπ -Pair)
68
u
v
∆, Γ ⊢ e : τ × τ ′
∆, Γ ⊢ e.1 : τ
v
∆, Γ ⊢ e : τ × τ ′
∆, Γ ⊢ e.2 : τ ′
v
∆, Γ ⊢ e : Ref τ
∆, Γ ⊢ !e : τ
u
w
v
u
w
v
u
~
~
Lπ
}Lτ
(TL -dereference)
τ
~
(TLτ -op)
∆, Γ ⊢ e : Nat
∆, Γ ⊢ e′ : Nat
∆, Γ ⊢ e ⊕ e′ : Nat
(TLτ -cmp)
∆, Γ ⊢ e : Nat
∆, Γ ⊢ e′ : Nat
∆, Γ ⊢ e ⊗ e′ : Bool
(TLτ -coercion)
v
Lπ
}Lτ
(TLτ -proj-2)
u
u
}Lτ
(TLτ -proj-1)
∆, Γ ⊢ e : τ
τ ⊢◦
∆, Γ ⊢ e : UN
Lπ
Lπ
Lπ
}Lτ
~
Lτ
(J·KLLπ -P2)
= J∆, Γ ⊢ e : τ × τ ′ KLπ .2
Lτ
τ
Lτ
Lτ
= !J∆, Γ ⊢ e : Ref τ KLπ .1 with J∆, Γ ⊢ e : Ref τ KLπ .2
Lτ
Lτ
Lτ
= J∆, Γ ⊢ e : NatKLπ ⊕ J∆, Γ ⊢ e′ : NatKLπ
Lτ
(J·KLπ -op)
}Lτ
~
(J·KLπ -P1)
(J·KLπ -Deref)
}Lτ
~
Lτ
= J∆, Γ ⊢ e : τ × τ ′ KLπ .1
Lπ
Lτ
Lτ
= J∆, Γ ⊢ e : NatKLπ ⊗ J∆, Γ ⊢ e′ : NatKLπ
Lτ
(J·KLπ -cmp)
τ
(J·KLLπ -Coerce)
= skip
(J·KLπ -Skip)
= J∆, Γ ⊢ e : τ KLLπ
τ
Statements
t
u
w
v
(TLτ -skip)
C, ∆, Γ ⊢ skip
(TLτ -new)
C, ∆, Γ ⊢ e : τ
C, ∆, Γ; x : Ref τ ⊢ s
C, ∆, Γ ⊢ let x = newτ e in s
|Lτ
Lτ
Lπ
}Lτ
~
Lπ
=
Lτ
let xo = new J∆, Γ ⊢ e : τ KLπ
in let x = hxo, 0i
τ
in JC, ∆, Γ; x : Ref τ ⊢ sKLLπ
if τ = UN
Lτ
let x = newhide J∆, Γ ⊢ e : τ KLπ
τ
L
in JC, ∆, Γ; x : Ref τ ⊢ sKLπ
else
Lτ
(J·KLπ -New)
69
u
w
w
w
v
u
w
w
w
v
(TLτ -function-call)
((f ∈ dom(C.funs))
∨(f ∈ dom(C.intfs)))
∆, Γ ⊢ e : UN
∆, Γ ⊢ call f e
(TLτ -if)
∆, Γ ⊢ e : Bool
C, ∆, Γ ⊢ st
C, ∆, Γ ⊢ se
C, ∆, Γ ⊢ if e then st else se
u
w
v
u
w
v
(TLτ -sequence)
C, ∆, Γ ⊢ su
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ su ; s
(TLτ -letin)
∆, Γ ⊢ e : τ
C, ∆, Γ; x : τ ⊢ s
C, ∆, Γ ⊢ let x : τ = e in s
u
(TLτ -assign)
∆, Γ ⊢ x : Ref τ
∆, Γ ⊢ e : τ
C, ∆, Γ ⊢ x := e
w
v
u
v
t
(TLτ -fork)
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ {k s}
(TLτ -monitor)
C, ∆, Γ ⊢ monitor
u
(TLτ -process)
v
C, ∆, Γ ⊢ s
C, ∆, Γ ⊢ (s)f
}Lτ
~
Lπ
}Lτ
~
Lπ
}Lτ
~
Lπ
}Lτ
~
Lπ
}Lτ
~
Lπ
}Lτ
Lτ
= call f J∆, Γ ⊢ e : UNKLπ
τ
(J·KLLπ -call)
Lτ
ifz J∆, Γ ⊢ e : BoolKLπ
τ
= then JC, ∆, Γ ⊢ st KLLπ
τ
else JC, ∆, Γ ⊢ se KLLπ
τ
L
(J·KL
π -If)
Lτ
Lτ
= JC, ∆, Γ ⊢ su KLπ ; JC, ∆, Γ; Γ′ ⊢ sKLπ
Lτ
(J·KLπ -Seq)
Lτ
=
let x=J∆, Γ ⊢ e : τ KLπ
Lτ
in JC, ∆, Γ; x : τ ⊢ sKLπ
Lτ
(J·KLπ -Letin)
let x1 = x.1
= in let x2 = x.2
Lτ
in x1 := J∆, Γ ⊢ e : τ KLπ with x2
Lτ
(J·KLπ -Assign)
Lτ
~
= {k JC, ∆, Γ ⊢ sKLπ }
|Lτ
(J·KLLπ -Fork)
= monitor
(J·KLLπ -Mon)
Lπ
τ
τ
Lπ
}Lτ
~
Lπ
70
Lτ
= (JC, ∆, Γ ⊢ sKLπ )Jf KLτ
Lπ
Lτ
(J·KLπ -Proc)
u
w
w
v
u
w
v
(TLτ -soup)
C, ∆, Γ ⊢ π
C, ∆, Γ ⊢ Π
C, ∆, Γ ⊢ π k Π
(TLτ -endorse)
∆, Γ ⊢ e : UN
C, ∆, Γ; (x : ϕ) ⊢ s
C, ∆, Γ ⊢ endorse x = e as ϕ in s
}Lτ
~
Lπ
}Lτ
~
Lπ
τ
τ
L
= JC, ∆, Γ ⊢ πKLLπ k JC, ∆, Γ ⊢ ΠKL
π
Lτ
(J·KLπ -Soup)
=
Lτ
destruct x = J∆, Γ ⊢ e : UNKLπ as nat in
ifz x then
Lτ
JC, ∆, Γ; (x : ϕ) ⊢ sKLπ
else ifz x − 1 then τ
JC, ∆, Γ; (x : ϕ) ⊢ sKLLπ
else wrong
or wrong
if
ϕ = Bool
Lτ
destruct x = J∆, Γ ⊢ e : UNKLπ as nat in
τ
JC, ∆, Γ; (x : ϕ) ⊢ sKL
Lπ
or wrong
if ϕ = Nat
τ
destruct x = J∆, Γ ⊢ e : UNKLLπ as pair in
Lτ
JC, ∆, Γ; (x : ϕ) ⊢ sKLπ
or wrong
if
ϕ = UN × UN
Lτ
destruct x = J∆, Γ ⊢ e : UNKLπ as pair in
!x.1 with x.2;
τ
JC, ∆, Γ; (x : ϕ) ⊢ sKLLπ
or wrong
if ϕ = Ref UN
Lτ
(J·KLπ -Endorse)
LU
The remark about optimisation for J·KLP in Appendix D is also valid for the
Lτ
Rule (J·KLπ -Deref) case above. As expressions are executed atomically, we are
sure that albeit inefficient, dereferencing will correctly succeed.
We can add reference to superficial types and check this dynamically in the
source, as we have the heap there. But how do we check this in the target? We
only assume that reference must be passed as a pair: location- key from the
Lτ
attacker. Thus the last case of Rule (J·KLπ -Endorse), where we check that we
can access the location, otherwise we’d get stuck.
71
Lτ
NonAtomic Implementation of New-Hide We can also implement Rule (J·KLπ -New)
Lτ
using non-atomic instructions are defined in Rule (J·KLπ -New-nonat) below.
Lτ
let xo = new J∆, Γ ⊢ e : τ KLπ
in let x = hxo, 0i
Lτ
in JC, ∆, Γ; x : Ref τ ⊢ sKLπ
u
}Lτ
if τ = UN
(TLτ -new)
∆, Γ ⊢ e : τ
w
v
~ =
let x = new 0 in
C, ∆, Γ; x : Ref τ ⊢ s
let xk = hide x in
τ
C, ∆, Γ ⊢ let x = newτ e in s
Lπ
let xc = J∆, Γ ⊢ e : τ KLLπ in
x := xc with xk;
Lτ
JC, ∆, Γ; x : Ref τ ⊢ sKLπ
otherwise
τ
(J·KLLπ -New-nonat)
H.3
Properties of the Lτ -Lπ Compiler
Lτ
Lτ
Theorem 11 (Compiler J·KLπ is cc). ⊢ J·KLπ : cc
τ
τ
Theorem 12 (Compiler J·KLLπ is rs-pres ). ⊢ J·KLLπ : rs-pres
72
I
Proofs
∼
We define a more lenient relation on states ∼
∼β analogous to ≈β (Rule Related states – Whole)
but that ensures that all target locations that are related to secure source ones
only vary accordingly: i.e., the attacker cannot change them.
∼
Ω∼
∼β Ω
(Lτ -Secure heap)
′
H = {ℓ 7→ v : τ | ℓ 7→ v : τ ∈ H and τ 0 ◦}
⊢ secure(H) = H′
(Lπ -Low Location)
ℓ ≈β hn, _i
n ∈ dom(H)
∄ℓ ∈ secure(H)
H, H ⊢ low-loc(n)
(Lπ -High Location)
ℓ ∈ secure(H)
ℓ ≈β hn, ki
n 7→ _ : k ∈ H
H, H ⊢ high-loc(n) = ℓ, k
(Lπ -High Capability)
ℓ ∈ secure(H). ℓ ≈β hn, ki
n 7→ _ : k ∈ H
H, H ⊢ high-cap(k)
Ω=
M; F, F′ ; I; H
(Related states – Secure)
q yL τ
Ω = M; F, F′ Lπ ; I; H
⊲Π
⊲Π
M ≈β M
∀k, n, ℓ. if H, H ⊢ high-loc(n) = ℓ, k then
(1) ∀π ∈ Π if C ⊢ π : attacker then k ∈
/ fv(π)
(2) ∀n′ 7→ v : η ∈ H,
(2a) if η = k then n = n′ and ℓ ≈β hn, ki and ℓ 7→ v : τ ∈ H and v ≈β v
(2b) if η 6= k then H, H ⊢ low-loc(n′ ) and ∀k′ . H, H ⊢ high-cap(k′ ), v 6= k′
∼β Ω
Ω∼
∼
There is no secure( · ) function for the target because they would be all locations
that are related to a source location that itself is secure in the source. An
alternative is to define secure( · ) as all locations protected by a key k but the
point of secure( · ) is to setup the invariant to ensure the proof hold, so this
alternative would be misleading.
Rule Lπ -Low Location tells when a target location is not secure. That is,
when there is no secure source location that is related to it. This can be because
the source location is not secure or because the relation does not exist, as in
order for it to exist the triple must be added to β and we only add the triple
for secure locations.
The intuition behind Rule Related states – Secure is that two states are
related if they have related monitors and then: for any target location n that
is high (i.e., it has a related source counterpart ℓ whose type is secure and that
is protected with a capability k that we call a high capability), then we have:
(1) the capability k used to lock it is not in in any attacker code; (2) for any
target level location n′ : (2a) either it is locked with a high capability k (i.e., a
73
capability used to hide a high location) thus n′ is also high, in which case it is
related to a source location ℓ and the values v, v they point to are related; or
(2b) it is not locked with a high capability, so we can derive that n′ is a low
location and its content v is not any high capability k′ .
♠
Lemma 5 (A target location is either high or low).
∀
if H ≈β H
n 7→ v : η ∈ H
then either H, H ⊢ low-loc(n)
or ∃ℓ ∈ dom(H).
H, H ⊢ high-loc(n) = ℓ, η
Proof. Trivial, as Rule Lπ -Low Location and Rule Lπ -High Location are duals.
♠
I.1
U
Proof of Theorem 7 (Compiler J·KLLP is cc)
Proof. The proof proceeds for β0 = (ℓ, 0, kroot ) and and then by Lemma 7 (Generalised compiler correctness for
as initial states are related by definition and the languages are deterministic.
♠
LU
Lemma 6 (Expressions compiled with J·KLP are related).
∀
if H ≈β H
H ⊲ eρ ֒→
→ v
U
U
U
then H ⊲ JeKLLP JρKLLP ֒→
→ JvKLLP
Proof. This proof proceeds by structural induction on e.
Base case: Values
U
U
true By Rule (J·KLLP -True), JtrueKLLP = 0.
As true ≈β 0, this case holds.
74
LU
false Analogous to the first case by Rule (J·KLP -False).
LU
n∈ N Analogous to the first case by Rule (J·KLP -nat).
U
x Analogous to the first case, by Rule (J·KLLP -Var) and by the relatedness
of the substitutions.
LU
ℓ Analogous to the first case by Rule (J·KLP -Loc).
LU
hv, vi By induction on v by Rule (J·KLP -Pair) and then it is analogous to
the first case.
Inductive case: Expressions
LU
e ⊕ e′ By Rule (J·KLP -op) we have that
LU
LU
LU
Je ⊕ e′ KLP = JeKLP ⊕ Je′ KLP
By HP we have that H ⊲ eρ ֒→
→ n and H ⊲ e′ ρ ֒→
→ n′ .
By Rule ELU -op we have that H ⊲ n ⊕ n′ ֒→
→ n′′ .
LU
LU
LU
LU
LU
LU
→ JnKLP and H ⊲ Je′ KLP JρKLP ֒→
→ Jn′ KLP .
By IH we have that H ⊲ JeKLP JρKLP ֒→
LU
LU
LU
By Rule ELP -op we have that H ⊲ JnKLP ⊕ Jn′ KLP ֒→
→ Jn′′ KLP .
So this case holds.
LU
e ⊗ e′ Analogous to the case above by IH, Rule (J·KLP -cmp), Rule ELU -comp
and Rules ELP -op and ELP -if-true.
U
!e Analogous to the case above by IH twice, Rule (J·KLLP -Deref), Rule ELU -dereference
and Rules ELP -p1, ELP -p2 and ELP -letin and a case analysis by
Rules ELP -deref-top and ELP -deref-k.
LU
he, ei Analogous to the case above by IH and Rule (J·KLP -Pair).
LU
LU
LU
e.1 By Rule (J·KLP -P1) Je.1KLP = JeKLP .1.
By HP H ⊲ e.1ρ ֒→
→ hv1 , v2 i ֒→
→ v1 .
LU
LU
LU
By IH we have that H ⊲ JeKLP .1JρKLP ֒→
→ Jhv1 , v2 iKLP .1.
E
D
U
U
U
LU
By Rule (J·KLP -Pair) we have that Jhv1 , v2 iKLLP .1 = Jv1 KLLP , Jv2 KLLP .1.
E
D
LU
LU
LU
→ Jv1 KLP .
Now H ⊲ Jv1 KLP , Jv2 KLP .1 ֒→
So this case holds.
LU
e.2 Analogous to the case above by Rule (J·KLP -P2), Rule ELU -p2 and
Rule ELP -p2.
♠
LU
Lemma 7 (Generalised compiler correctness for J·KLP ).
75
Proof.
∀...∃β ′
if ⊢ C : whole
C = M; F; I
τ
JC, MKLLπ = M; F; I; kroot = C
Lτ
C, H ⊲ s ≈β C, H ⊲ JsKLπ
ǫ
C, H ⊲ sρ −−→ C′ , H′ ⊲ s′ ρ′
LU
Lτ
LU
Lτ
ǫ
then C, H ⊲ JsKLπ JρKLP −−→ C′ , H′ ⊲ Js′ KLπ Jρ′ KLP
C′ = M′ ; F; I; kroot
LU
Lτ
C, H ⊲ s′ ρ′ ≈β ′ C, H ⊲ Js′ KLπ Jρ′ KLP
β ⊆ β′
The proof proceeds by induction on C and the on the reduction steps.
Base case
LU
skip By Rule (J·KLP -Skip) this case follows trivially.
Inductive
let x = new e in s
U
U
By Rule (J·KLLP -New) Jlet x = new e in sKLLP =
U
let xloc = new JeKLLP in
let xcap = hide xloc in
By HP H ⊲ eρ ֒→
→ v
LU
let x = hxloc , xcap i in JsKLP
LU
LU
LU
→ JvKLP and HPV
So by Lemma 6 we have HPE: H ⊲ JeKLP JρKLP ֒→
U
v ≈β JvKLLP .
ǫ
By Rule ELU -alloc: C; H ⊲ let x = new e in s −−→ C; H; ℓ 7→ v ⊲ s[ℓ / x].
So by HPE:
LU
C; H⊲let xloc = new JeKLP in
let xcap = hide xloc in
Rule ELP -new
ǫ
LU
−−→ C; H; n 7→ JvKLP : ⊥⊲ let xcap = hide xloc in
76
LU
let x = hxloc , xcap i in JsKLP ρ
LU
U
≡ C; H; n 7→ JvKLLP : ⊥⊲ let xcap = hide n in
Rule ELP -hide
LU
ǫ
LU
let x = hxloc , xcap i in JsKLP JρKLP [n / xloc ]
LU
LU
let x = hn, xcap i in JsKLP JρKLP
LU
LU
−−→ C; H; n 7→ JvKLP : k; k⊲ let x = hn, xcap i in JsKLP JρKLP [k / xcap ]
LU
LU
≡ C; H; n 7→ JvKLP : k; k⊲ let x = hn, ki in JsKLP ρ
Rule ELP -letin
ǫ
LU
LU
LU
−−→ C; H; n 7→ JvKLP : k; k⊲ JsKLP JρKLP [hn, ki / x]
Let β ′ = β ∪ (ℓ, n, k).
By definition of ≈β ′ and by β ′ we get HPL ℓ ≈β ′ hn, ki.
By a simple weakening lemma for β for substitutions and values apLU
plied to HP and HPV we can get HPVB v ≈β ′ JvKLP .
As H ≈β H by HP, by a simple weakening lemma get that H ≈β ′ H too
and by Rule Heap relation with HPL and HPVB we get H′ ≈β ′ H′ .
LU
We have that ρ′ = ρ[ℓ / x] and ρ′ = JρKLP [hn, ki / x].
So by HPL we get that ρ′ ≈β ′ ρ′ .
U
s; s′ Analogous to the case above by IH, Rule (J·KLLP -Seq) and a case
analysis on what s reduces to, either with Rule ELU -sequence and
Rule ELP -sequence or with Rule ELU -step and Rule ELP -step.
LU
let x = e in s Analogous to the case above by IH, Rule (J·KLP -Letin), Rule ELU -letin
and Rule ELP -letin.
LU
x := e′ Analogous to the case above by Rule (J·KLP -Assign), Rule ELU -update
and Rule ELP -letin (twice), Rules ELP -p1 and ELP -p2 and then a
case analysis by Rules ELP -assign-top and ELP -assign-k.
LU
if e then s else s′ Analogous to the case above by IH, Rule (J·KLP -If) and
then either Rule ELU -if-true and Rule ELP -if-true or Rule ELU -if-false
and Rule ELP -if-false.
LU
LU
monitor By Rule (J·KLP -Mon) JmonitorKLP = monitor.
There are two cases:
• Rule ELU -monitor.
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
We need to prove that
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
By HP H ≈ H, so we know can apply Rule ELP -monitor.
As the only change in C′ and C′ is the new state of the monitor
and they are related by Rule ELP -monitor, this case holds.
77
• Rule ELU -monitor-fail
ǫ
C, H ⊲ monitor −−→ fail
We need to prove that
ǫ
C, H ⊲ monitor −−→ fail
By HP H ≈ H, so we know can apply Rule ELP -monitor-fail and
this case holds.
LU
LU
LU
call f e By Rule (J·KLP -call) Jcall f eKLP = call f JeKLP
By HP H ⊲ eρ ֒→
→ v
LU
LU
LU
So by Lemma 6 we have HPE: H ⊲ JeKLP ρ ֒→
→ JvKLP and HPR v ≈β JvKLP .
So as C is whole, we apply Rule ELU -call-internal
ǫ
C, H ⊲ (call f eρ)f ′ −−→
C, H ⊲ (s; return;ρ[v / x])f ′ ;f
By Rule ELP -call-internal
LU
LU
ǫ
C, H ⊲ (call f JeKLP JρKLP )f ′ −−→
h
. i
LU
LU
C, H ⊲ (s; return;JρKLP JvKLP x )f ′ ;f
By the first induction on C we get
IH1 C, H ⊲ (s; return;ρ′ )f ′ ;f ≈β C, H ⊲ (s; return;ρ′ )f ′ ;f
h
. i
LU
LU
We instantiate ρ′ with ρ[v / x] and ρ′ with JρKLP JvKLP x .
h
. i
LU
LU
So by HP and HPR we have that ρ[v / x] ≈β JρKLP JvKLP x
We we can use IH1 to conclude
h
. i
U
U
C, H ⊲ (s; return;ρ[v / x])f ′ ;f ≈β C, H ⊲ (s; return;JρKLLP JvKLLP x )f ′ ;f
As β ′ = β, this case holds.
♠
I.2
U
Proof of Theorem 8 (Compiler J·KLLP is rs-pres)
Proof. HP1: ⊢ C : rs
U
TH1: ⊢ JC, MKLLP : rs
We can state it in contrapositive form as:
LU
HP2: 0 JC, MKLP : rs
TH2: 0 C : rs
By expanding the definition of rs in HP2 and TH2, we get
78
h
i
α
LU
==⇒ fail
HP21 ∃A.Ω0 A JC, MKLP
α
TH21 ∃A.Ω0 (A [C]) ==⇒ fail
We can expand
get:
hHP21 and
i
α
LU
==⇒ Ω
HPRT1 Ω0 A JC, MKLP
HPRT2 Ω =
⇒ Ch ; Hh ⊲ monitor; s
ǫ
HPRMT1 Ch ; Hh ⊲ monitor; s −−→ fail
P
We can apply Theorem 9 (hh·iiL
LU is correct) with HPRT1 and instantiate A
with a A from hhAii and we get the following unfolded HPs
α
HPRS Ω0 (A [C]) ==⇒ Ω
HPRel Ω ≈β Ω.
We need to show:
⇒ Ch ; Hh ⊲ monitor; s
TH3 Ω =
ǫ
TH4 Ch ; Hh ⊲ monitor; s −−→ fail
Note that reductions here are triggered by compiled components only.
LU
By Lemma 7 (Generalised compiler correctness for J·KLP ) and Lemma 1 (Compiled code steps imply existen
with HPRT2 we get TH3 and
HPMR Ch ; Hh ⊲ monitor; s ≈β Ch ; Hh ⊲ monitor; s
We need to prove TH4.
Assume by contradiction HPBOT: the monitor in the source does not fail
ǫ
HPBOT: Ch ; Hh ⊲ monitor; s −−→ Ch ; Hh ⊲ skip; s
U
By Rule EL -monitor with HPBOT we know that
M′
HPM: M; Hh
We can expand HPM by Rule LU -Monitor Step and get
HPMR: (σc , H′h , σf ) ∈
for a heap H′h ⊆ Hh
By definition of the compiler we have that M ≈β M for initial states.
By Definition 12 (M ≈ M) and the second clause of Definition 11 (MRM)
with HPMR we know that M ≈β M for the current states.
By the first clause of Definition 11 (MRM) we know that
⇐⇒ (σc , H, _) ∈
HPMRBI: (σc , H, _) ∈
By HPMRBI with HPMR we know that
HPMRTC: (σc , H′h , σf ) ∈
However, by HPRMT1 and Rule ELP -monitor-fail we know that
HPNR: M; H 6
so we get
HPCON: ∄(σc , H′h , σf ) ∈
By HPCON and HPMRTC we get the contradiction, so the proof holds.
♠
79
I.3
Proof of Lemma 1 (Compiled code steps imply existence of source steps)
λ
Proof. The proof proceeds by induction on ==⇒ .
λ
λ
Base case: ==⇒ = −−→
We proceed by case analysis on λ
λ=ǫ
The proof proceeds by analysis of the target reductions.
Rule ELP -sequence In this case we do not need to pick and the
thesis holds by Rule ELU -sequence.
Rule ELP -step In this case we do not need to pick and the thesis
holds by Rule ELU -step.
LU
Rule ELP -if-true We have: H ⊲ JeKLP ρ ֒→
→ 0
We apply Lemma 8 (Compiled code expression steps implies existence of source expression steps
and obtain a v ≈β 0
By definition we have 0 ≈β 0 and true ≈β 0, we pick the second.
→ true
So we have H ⊲ eρ ֒→
We can now apply Rule ELU -if-true and this case follows.
Rule ELP -if-false This is analogous to the case above.
Rule ELP -assign-top Analogous to the case above.
Rule ELP -assign-k This is analogous to the case above but for
v = ℓ ≈β hn, ki.
Rule ELP -letin This follows by Lemma 8 and by Rule ELU -letin.
Rule ELP -new This follows by Lemma 8 and by Rule ELU -alloc.
Rule ELP -hide By analisis of compiled code we know this only
happens after a new is executed.
In this case we do not need to perform a step in the source and
the thesis holds.
Rule ELP -monitor In this case we do not need to pick and the
thesis holds by Rule ELU -monitor.
Rule ELP -monitor-fail In this case we do not need to pick and
the thesis holds by Rule ELU -monitor-fail.
Rule ELP -call-internal This follows by Lemma 8 and by Rule ELU -call-internal.
Rule ELP -ret-internal In this case we do not need to pick and the
thesis holds by Rule ELU -ret-internal.
λ=α!
The proof proceeds by case analysis on α!
call f v H! This follows by Lemma 8 and by Rule ELU -callback.
ret H! In this case we do not need to pick and the thesis holds by
Rule ELU -return.
Inductive case: This follows from IH and the same reasoning as for the single
action above.
80
♠
Lemma 8 (Compiled code expression steps implies existence of source expression steps).
∀
LU
if H ⊲ JeKLP ρ ֒→
→ v
and if {ρ} = {ρ | ρ ≈β ρ}
v ≈β v
H ≈β H
then ∃ρj ∈ {ρ} . H ⊲ eρj ֒→
→ v
Proof. This proceeds by structural induction on e.
U
Base case: true This follows from Rule (J·KLLP -True).
LU
false This follows from Rule (J·KLP -False).
LU
n ∈ N This follows from Rule (J·KLP -nat).
x This follows from the relation of the substitutions and the totality of
LU
≈β and Rule (J·KLP -Var).
hv, v′ i This follows from induction on v and v′ .
Inductive case: e ⊕ e′ By definition of ≈β we know that v and v′ could be
either natural numbers or booleans.
We apply the IH with:
IHV1 n ≈β n
IHV2 n′ ≈β n′
By IH we get
LU
IHTE1 H ⊲ JeKLP ρ ֒→
→ n
LU
IHTE2 H ⊲ Je′ KLP ρ ֒→
→ n′
IHSE1 H ⊲ eρj ֒→
→ n
IHSE2 H ⊲ e′ ρj ֒→
→ n′
LU
LU
LU
LU
By Rule (J·KLP -op) we have that Je ⊕ e′ KLP =JeKLP ⊕ Je′ KLP .
LU
LU
By Rule ELP -op with IHTE1 and IHTE2 we have that H ⊲ JeKLP ⊕ Je′ KLP ֒→
→ n′′
′′
′
where IHVT n = n ⊕ n
By Rule ELU -op with IHSE1 and IHSE2 we have that H ⊲ e ⊕ e′ ֒→
→ n′′
′′
′
if n = n ⊕ n
This follows from IHVT and IHV1 and IHV2.
81
LU
e ⊗ e′ As above, this follows from IH and Rule (J·KLP -cmp) and Rule ELU -comp.
LU
he, e′ i As above, this follows from IH and Rule (J·KLP -Pair).
LU
e.1 As above, this follows from IH and Rule (J·KLP -P1) and Rule ELU -p1.
e.2 Analogous to the case above.
LU
!e As above, this follows from IH and Rule (J·KLP -Deref) and Rule ELU -dereference
but with the hypothesis that e evaluates to a v related to a hn, vi.
♠
I.4
P
Proof of Theorem 9 (hh·iiLLU is correct)
h
i
α
LU
Proof. HP1 Ω0 A JCKLP
==⇒ Ω
ǫ
HPF Ω =⇒ Ω′
HPN I = names(A)
HPT α ≡ α′ · α?
HPL ℓi ; ℓglob ∈
/β
LP
THE ∃A ∈ hhI, αiiLU
α
TH1 Ω0 (A [C]) ==⇒ Ω
THA α ≈β α
THS Ω ≈β Ω
THC Ω.H.ℓi = ||α|| + 1
The proof proceeds by induction on α′ .
Base case: α′ ≡ ∅
Case analysis on α?
call f v H?
Given h
i
call f v H?
LU
=========⇒ Ω
HP1 Ω0 A JCKL
P
We need to show that
LP
THE ∃A ∈ hhI, αiiLU
α
TH1 Ω0 (A [C]) ==⇒ Ω
THA call f v H? ≈β call f v H?
THS Ω ≈β Ω
THC Ω.H.ℓi = ||α|| + 1
LP
By Rule (hh·iiLU -call) the back-translated context executes this code
inside main:
82
if !ℓi == n then
incrementCounter()
let x1 = new v1 in register(hx1, n1 i)
···
let xj = new vj in register(hxj, nj i)
call f v
else skip
As Hpre is ∅, no updates are added.
LP
Given that ℓi is initialised to 1 in Rule (hh·iiLU -skel), this code is executed and it generates action call f v H? where H=ℓ1 7→ v1 ; · · · ; ℓn 7→ vn
for all ni ∈ dom(H) such that ℓi ≈β hni , _i and:
HPHR H ≈β H
By HPHR, Lemma 10 (Backtranslated values are related) and Lemma 1 (Compiled code steps imply
with HPF we get THA, THE and TH1.
By Rule Related states – Secure, THS holds too.
Execution of incrementCounter() satisfies THC.
ret H? This cannot happen as by Rule ELP -retback there needs to be a
running process with a non-empty stack and by Rule LP -Initial State
the stack of initial states is empty and the only way to add to the stack
is performing a call via Rule ELP -call, which would be a different
label.
Inductive case:
We know that (eliding conditions HP that are trivially satisfied):
i
h
α
α!
α?
LU
IHP1 Ω0 A JCKLP
==⇒ Ω′ ==⇒ Ω′′ ==⇒ Ω
And we need to prove:
α
α!
α?
LP
ITH1 Ω0 hhI, αiiLU [C] ==⇒ Ω′ ==⇒ Ω′′ ==⇒ Ω
ITHA αα!α? ≈β αα!α?
ITHS Ω ≈β Ω
And the inductive HP is (for ∅ ⊆ β ′ ):
i
h
α
LU
IH-HP1 Ω0 A JCKLP
==⇒ Ω′
P
α
IH-TH1 Ω0 hhI, αiiL
[C]
==⇒ Ω′
U
L
IH-THA α ≈β ′ α
IH-THS Ω′ ≈β ′ Ω′
By IHP1 and HPF we can apply Lemma 1 (Compiled code steps imply existence of source steps)
and so we can apply the IH to get IH-TH1, IH-THA and IH-THS.
83
We perform a case analysis on α!, and show that the back-translated code
performs α!.
LP
By IH we have that the existing code is generated by Rule (hh·iiLU -listact-i):
α, n, Hpre , ak, f
LP
.
LU
The next action α! produces code according to:
HPF
α!, n, Hpre , ak, f
LP
.
LU
LP
By Rule (hh·iiLU -join), code of this action is the first if statement executed.
P
call f v H! By Rule (hh·iiL
LU -callback-loc) this code is placed at function
f so it is executed when compiled code jumps there
if !ℓi == n then
incrementCounter()
let l1 = e1 in register(hl1, n1 i)
···
let lj = ej in register(hlj, nj i)
else skip
By IH we have that ℓi 7→ n, so we get
IHL ℓi 7→ n + 1
By Definition 18 (Reachable) we have for i ∈ 1..j that a reachable
location ni ∈ dom(H) has a related counterpart in ℓi ∈ dom(H) such
that H ⊲ ei ֒→
→ ℓi .
By Lemma 9 (Lτ attacker always has access to all capabilities) we know
all capabilities to access any ni are in ak.
We use ak to get the right increment of the reach.
ret H! In this case from IHF we know that f = f ′ f ′ .
This code is placed at f ′ , so we identify the last called function and
the code is placed there. Source code returns to f ′ so this code is
LP
executed Rule (hh·iiLU -ret-loc)
if !ℓi == n then
incrementCounter()
let l1 = e1 in register(hl1, n1 i)
···
let lj = ej in register(hlj, nj i)
else skip
This case now follows the same reasoning as the one above.
So we get (for β ′ ⊆ β ′′ ):
HP-AC! α! ≈β ′′ α!
84
By IH-THS and Rule Related states – Whole and HP-AC! we get HPOM2:
HP-OM2: Ω′′ ≈β ′′ Ω′′
The next action α? produces code according to:
IHF1
α?, n + 1, H′pre , ak′ , f ′
LP
LU
.
We perform a case analysis on α? and show that the back-translated code
performs α?:
LP
ret H? By Rule (hh·iiLU -retback), after n actions, we have from IHF1 that
f ′ = f ′ f ′′ and inside function f ′ there is this code:
if !ℓi == n then
let x1 = new v1 in register(hx1, n1 i)
···
let xj = new vj in register(hxj, nj i)
update(m1 , u1 )
···
update(ml , ul )
else skip
By IHL, ℓi 7→ n + 1, so the if gets executed.
By definition, forall n ∈ dom(H) we have that n ∈ Hn or n ∈ Hc
(from the case definition).
By Lemma 9 (Lτ attacker always has access to all capabilities) we know
all capabilities to access any n are in ak.
We induce on the size of H; the base case is trivial and the inductive
case follows from IH and the following:
Hn : and n is newly allocated.
In this case when we execute
P
ǫ
′′
C; H′ ⊲ let x1 = new v1 in register(hx1, n1 i) −−→ C; H′ ; ℓ′′ 7→ hhv1 iiL
LU ⊲ register(hℓ , n1 i)
′′
′′
′
′
and we create β by adding ℓ , n, η to β .
By Lemma 2 (register(ℓ, n) does not add duplicates for n) we have
that:
ǫ
LP
LP
C; H′ ; ℓ′′ 7→ hhv1 iiLU ⊲ register(hℓ′′ , n1 i) −−→ C; H′ ; ℓ′′ 7→ hhv1 iiLU ⊲ skip
and we can lookup ℓ′′ via n.
Hc : and n is already allocated.
In this case
ǫ
C; H′ ⊲ update(m1 , u1 ) −−→ C; H′′ ⊲ skip
By Lemma 3 (update(n, v) never gets stuck) we know that H′′ =
H′ [ℓ′′ 7→ _ / ℓ′′ 7→ u1 ]
and ℓ′′ such that (ℓ′′ , m1 , η ′′ ) ∈ β ′ .
By Lemma 10 (Backtranslated values are related) on the values stored
on the heap, let the heap after these reduction steps be H, we can
conclude
85
HPRH H ≈β ′′ H.
As no other if inside f is executed, eventually we hit its return stateLP
LP
ment, which by Rule (hh·iiLU -join) and Rule (hh·iiLU -fun) is incrementCounter(); return;.
Execution of incrementCounter() satisfies THC.
ret H?
So we have Ω′′ =====⇒ Ω (by Lemma 10) and with HPRH.
call f v H? Similar to the base case, only with update bits, which follow
the same reasoning above.
So we get (for β ′′ ⊆ β):
HP-AC? α? ≈β α?
By IH-THA and HP-AC! and HP-AC? we get ITHA.
Now by Rule Related states – Whole again and HP-AC? we get ITHS and
ITH1, so the theorem holds.
♠
Lemma 9 (Lτ attacker always has access to all capabilities).
∀
if
LP
LU
α, n, H, ak, f
k. n 7→ v : k ∈ H
= s, ak′ , H′ , f ′ , f
then n ∈ reach(ak′ .loc, ak′ .cap, H)
LP
LP
Proof. Trivial case analyisis on Rules (hh·iiLU -ret-loc) to (hh·iiLU -callback-loc).
♠
Lemma 10 (Backtranslated values are related).
∀v, β.
LP
hhviiLU ≈β v
Proof. Trivial analysis of Appendix D.2.1.
♠
86
I.5
Proof of Theorem 10 (Typability Implies Robust Safety in Lτ )
Proof.
∀
if ⊢ C : UN
C ⊢att A
⊢ A [C] , ∅ : whole
α
then Ω0 (A [C]) ==
6 ⇒ fail
Since the initial state’s heap satisfies the typing of the monitor, we can apply Lemma 11 (Typing makes monitors not fail in Lτ ). This proof follows as
monitor is the only expression that can reduce to fail and we know all those
expression will reduce to skipinstead.
♠
I.5.1
Proof of Lemma 4 (Semantics and typed attackers coincide)
Proof. This is proven by trivial induction on the syntax of A.
By the rules of Appendix E.1.3, points 1 and 3 follow, point 2 follows from
the HP Rule Lτ -Plug.
♠
Lemma 11 (Typing makes monitors not fail in Lτ ).
∀C, A, M′ , H, E, f, Π, H′ , Π1 .
if ⊢ C : UN
C ≡ M; · · ·
C′ ≡ M′ ; · · ·
M = ({σ} , , σ0 , ∆, σc )
⊢ secure(H, ∆) : ∆
α
C, H ⊲ Πρ ==⇒ C′ , H′ ⊲ Π1 ρ1 k (monitor)f ρ′
ǫ
then C′ , H′ ⊲ Π1 ρ1 k (monitor)f ρ′ −−→ C′ , H′ ⊲ Π1 ρ1 k (skip)f ρ′
α
Proof. Let the state between ==⇒ and =
⇒ be C′′ , H′′ ⊲ s′′ .
τ
By Lemma 12 (L -α reductions respect heap typing) we have ⊢ secure(H′′ , ∆) :
∆ HP2 .
By Lemma 16 (Lτ any ǫ reduction respects heap typing) with HP2 to know
⊢ secure(H′ , ∆) : ∆ HPHP .
87
By Rule ELτ -monitor we need to show that M′ ; secure(H′ , ∆)
M′′ .
τ
′
By Rule L -Monitor Step we need to show that ⊢ secure(H , ∆) : ∆.
This holds by HPHP.
♠
Lemma 12 (Lτ -α reductions respect heap typing).
if C ≡ M; · · ·
M = ({σ} ,
, σ0 , ∆, σc )
⊢ secure(H, ∆) : ∆
α
C, H ⊲ Πρ ==⇒ C′ , H′ ⊲ Π′ ρ′
then ⊢ secure(H′ , ∆) : ∆
Proof. The proof proceeds by induction on α.
Base case This trivially holds by HP.
Inductive case This holds by IH plus a case analysis on the last action:
call f v? This holds by Lemma 17 (Lτ -? actions respect heap typing).
call f v! This holds by Lemma 18 (Lτ -! actions respect heap typing)
ret ! This holds by Lemma 18 (Lτ -! actions respect heap typing)
ret ? This holds by Lemma 17 (Lτ -? actions respect heap typing)
♠
Lemma 13 (Lτ An attacker only reaches UN locations).
∀
if ℓ 7→ v : UN ∈ H
then ∄e
H ⊲ e ֒→
→ ℓ′
ℓ′ 7→ v : τ ∈ H
τ 6= UN
Proof. This proof proceeds by contradiction.
Suppose e exists, there are two cases for ℓ′
88
• ℓ′ was allocated by the attacker:
This contradicts the judgements of Appendix E.1.3.
• ℓ′ was allocated by the compiled code:
The only way this was possible was an assignment of ℓ′ to ℓ, but Rule TLτ -assign
prevents it.
♠
Lemma 14 (Lτ attacker reduction respects heap typing).
if C ≡ M; · · ·
M = ({σ} ,
, σ0 , ∆, σc )
ǫ
C ⊢att Π −−→ Π′
ǫ
C, H ⊲ Πρ −−→ C, H′ ⊲ Πρ′
then secure(H, ∆) = secure(H′ , ∆)
Proof. Trivial induction on the derivation of Π, which is typed with ⊢UN and
by Lemma 13 (Lτ An attacker only reaches UN locations) has no access to locations in ∆ or with a type τ ⊢ ◦ and no monitor.
♠
Lemma 15 (Lτ typed reduction respects heap typing).
if C ≡ M; · · ·
M = ({σ} ,
C, Γ ⊢ s
, σ0 , ∆, σc )
C, Γ ⊢ s′
⊢ secure(H, ∆) : ∆
ǫ
C, H ⊲ sρ −−→ C′ , H′ ⊲ s′ ρ′
then ⊢ secure(H′ , ∆) : ∆
Proof. This is done by induction on the derivation of the reducing statement.
There, the only non-trivial cases are:
89
Rule TLτ -new By IH we have that
H ⊲ eρ ֒→
→ v
So
ǫ
C; H ⊲ let x = newτ e in s −−→ C; Hℓ 7→ v : τ ⊲ s[ℓ / x]
By IH we need to prove that ⊢ secure(ℓ 7→ v : τ , ∆) : ∆
As ℓ ∈
/ dom(∆), by Rule Lτ -Heap-ok-i this case holds.
Rule TLτ -assign By IH we have (HPH) H ⊲ e ֒→
→ v
such that ℓ : Ref τ and v : τ .
So
ǫ
C; H ⊲ x := eρ −−→ C; H′ ⊲ skip
where [x / ℓ] ∈ ρ and
H = H1 ; ℓ 7→ v′ : τ ; H2
H′ = H1 ; ℓ 7→ v : τ ; H2
There are two cases
ℓ ∈ dom(∆) By Rule Lτ -Heap-ok-i we need to prove that ℓ : Ref τ ∈ ∆.
This holds by HPH and Rule Lτ -Initial State, as the initial state
ensures that location ℓ in the heap has the same type as in ∆ .
ℓ∈
/ dom(∆) This case is trivial as for allocation.
Rule TLτ -coercion We have that C, Γ ⊢ e : τ and HPT τ ⊢ ◦.
By IH H ⊲ e ֒→
→ v such that ⊢ secure(H′ , ∆) : ∆.
By HPT we get that secure(H) = secure(H′ ) as by Rule Lτ -Secure heap
function secure( · ) only considers locations whose type is τ 0 ◦, so none
affected by e.
So this case by IH.
Rule TLτ -endorse By Rule ELτ -endorse we have that H ⊲ e ֒→
→ v and that
C, H ⊲ endorse x = e as ϕ in s ֒→
→ C, H ⊲ s[v / x].
So this holds by IH.
♠
Lemma 16 (Lτ any ǫ reduction respects heap typing).
if C ≡ M; · · ·
M = ({σ} ,
90
, σ0 , ∆, σc )
⊢ secure(H, ∆) : ∆
ǫ
C, H ⊲ Πρ −−→ C′ , H′ ⊲ Π′ ρ′
then ⊢ secure(H′ , ∆) : ∆
Proof. By induction on the reductions and by application of Rule ELτ -par. The
base case follows by the assumptions directly. In the inductive case we have the
following:
ǫ
ǫ
C, H ⊲ Πρ −−→ C′′ , H′′ ⊲ Π′′ ρ′′ −−→ C′ , H′ ⊲ Π′ ρ′
This has 2 sub-cases, if the reduction is in an attacker function or not.
ǫ
C ⊢att Π′′ −−→ Π: this follows by induction on Π′′ and from IH and Lemma 14 (Lτ attacker reduction respects he
ǫ
C 6⊢att Π′′ −−→ Π: In this case we induce on Π′′ .
The base case is trivial.
The inductive case is (s)f k Π, which follows from IH and Lemma 15 (Lτ typed reduction respects heap typ
♠
Lemma 17 (Lτ -? actions respect heap typing).
if C ≡ M; · · ·
α?
C, H ⊲ Πρ ==⇒ C, H′ ⊲ v′
M = ({σ} , , σ0 , ∆, σc )
then secure(H, ∆) = secure(H′ , ∆)
Proof. By Lemma 16 (Lτ any ǫ reduction respects heap typing), and a simple
case analysis on α? (which does not modify the heap).
♠
Lemma 18 (Lτ -! actions respect heap typing).
if C ≡ M; · · ·
C′ ≡ M′ ; · · ·
α!
C, H ⊲ Πρ ==⇒ C′ , H′ ⊲ v′
M = ({σ} , , σ0 , ∆, σc )
⊢ secure(H, ∆) : ∆
then ⊢ secure(H′ , ∆) : ∆
91
Proof. By Lemma 16 (Lτ any ǫ reduction respects heap typing) and a simple
case analyis on α! (which does not modify the heap).
♠
I.6
Proof of Theorem 11 (Compiler J·KLLπ is cc)
τ
Proof. By definition initial states have related components, related heaps and
well-typed, related starting processes, for β0 = (dom(∆), dom(H0 ), H0 .η) so we
have:
Lτ
HRS Ω0 (C) ≈β0 Ω0 JC, MKLπ .
As the languages have no notion of internal nondeterminism we can apply
Lτ
Lemma 20 (Generalised compiler correctness for J·KLπ ) with HRS to conclude.
♠
Lτ
Lemma 19 (Expressions compiled with J·KLπ are related).
∀
if H ≈β H
H ⊲ eρ ֒→
→ v
Lτ
Lτ
Lτ
then H ⊲ JeKLπ JρKLπ ֒→
→ JvKLπ
LU
Proof. The proof is analogous to that of Lemma 6 (Expressions compiled with J·KLP are related)
as the compilers perform the same steps and expression reductions are atomic.
♠
Lτ
Lemma 20 (Generalised compiler correctness for J·KLπ ).
∀...∃β ′
if C; Γ ⊢ Π,
⊢ C : whole
C = M; F; I
M = ({σ} , , σ0 , ∆, σc )
M = ({σ} , , σ0 , H0 , σc )
92
Lτ
JC, MKLπ = M; F; I = C
τ
L
C, H ⊲ Π ≈β C, H ⊲ JC; Γ ⊢ ΠKL
π
ǫ
C, H ⊲ Πρ =⇒ C′ , H′ ⊲ Π′ ρ′
Lτ
Lτ
Lτ
ǫ
Lτ
then C, H ⊲ JC; Γ ⊢ ΠKLπ JρKLπ =⇒ C′ , H′ ⊲ JC; Γ ⊢ Π′ KLπ Jρ′ KLπ
C′ = M′ ; F; I
Lτ
Lτ
C, H ⊲ Π′ ρ′ ≈β ′ C, H ⊲ JC; Γ ⊢ Π′ KLπ Jρ′ KLπ
β ⊆ β′
Proof. This proof proceeds by induction on the typing of Π and then of π.
τ
Base Case skip Trivial by Rule (J·KLLπ -Skip).
Inductive Case
In this case we proceed by induction on the typing of s
Inductive Cases Rule TLτ -new There are 2 cases, they are analogous.
τ = UN By HP
Γ⊢e:τ
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = newτ e in sρ −−→ C, H; ℓ 7→ v : τ ⊲ s[ℓ / x]ρ
By Lemma 19 we have:
Lτ
Lτ
Lτ
→ JΓ ⊢ v : τ KLπ
IHR1 H ⊲ JΓ ⊢ e : τ KLπ JρKLπ ֒→
Lτ
By Rule (J·KLπ -New) we get
Lτ
let xo = new JΓ ⊢ e : τ KLπ
in let x = hxo, 0i
τ
in JC, Γ; x : Ref τ ⊢ sKLLπ
So:
τ
L
C, H ⊲ let xo = new JΓ ⊢ e : τ KL
π
in let x = hxo, 0i
Lτ
ǫ
in JC, Γ; x : Ref τ ⊢ sKLπ
τ
−−→ C, H; n 7→ JΓ ⊢ v : τ KLLπ : ⊥ ⊲ let x = hn, 0i
ǫ
Lτ
Lτ
in JC, Γ; x : Ref τ ⊢ sKLπ
Lτ
−−→ C, H; n 7→ JΓ ⊢ v : τ KLπ : ⊥ ⊲ JC, Γ; x : Ref τ ⊢ sKLπ [hn, 0i / x]
For β ′ = β ∪ (ℓ, n, ⊥), this case holds.
else The other case holds follows the same reasoning but
Lτ
for β ′ = β∪(ℓ, n, k) and for H′ =H; n 7→ JC, Γ ⊢ v : τ KLπ : k; k.
Case for Appendix H.2 This follows the case above though with
more reductions.
93
Rule TLτ -sequence By HP
Γ ⊢ s; Γ ⊢ s′
ǫ
C, H ⊲ sρ =⇒ C′ , H′ ⊲ s′′ ρ′′
There are two cases
s′′ =skip Rule ELU -sequence
ǫ
C′ , H′ ⊲ skipρ′′ ; s′ ρ −−→ C′ , H′ ⊲ s′ ρ
By IH
ǫ
Lτ
Lτ
Lτ
Lτ
C, H ⊲ JΓ ⊢ sKLπ JρKLπ =⇒ C′ , H′ ⊲ JΓ ⊢ skipKLπ Jρ′′ KLπ
τ
L
By Rule (J·KLπ -Seq)
τ
L
Lτ
JC, Γ ⊢ sKLπ ; JC, Γ ⊢ s′ KLπ
So
τ
ǫ
Lτ
τ
τ
C, H ⊲ JC, Γ ⊢ sKLLπ JρKLLπ ; JC, Γ ⊢ s′ KLπ JρKLLπ
τ
Lτ
Lτ
τ
=⇒ C′ , H′ ⊲ JΓ ⊢ skipKLLπ Jρ′′ KLπ ; JC, Γ ⊢ s′ KLπ JρKLLπ
ǫ
Lτ
τ
−−→ C′ , H′ ⊲ JC, Γ ⊢ s′ KLπ JρKLLπ
At this stage we apply IH and the case holds.
else By Rule ELU -step we have
ǫ
C, H ⊲ s; s′ =⇒ C′ , H′ ⊲ s′′ ; s′
This case follows by IH and HPs.
Rule TLτ -function-call Analogous to the cases above.
Rule TLτ -letin Analogous to the cases above.
Rule TLτ -assign Analogous to the cases above.
Rule TLτ -if Analogous to the cases above.
Rule TLτ -fork Analogous to the cases above.
Rule TLτ -monitor By HP:
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
IHMON M; mon-care(H, ∆)
M′
τ
By Rule TL -monitor
t
|Lτ
(TLτ -monitor)
= monitor
C, Γ ⊢ monitor
Lπ
So we need to prove that
ǫ
C, H ⊲ monitor −−→ C′ , H ⊲ skip
P
By Rule EL -monitor we need to prove that
M; H
M′
By Rule Lπ -Monitor Step we neet to prove that
(sc , mon-care(H, H0 ), sf ) ∈
By Rule Monitor relation and then Rule Ok Mon we need to
prove that
• heaps states etc are related (By HP)
• ⊢H:∆
By IHMON and Rule Lτ -Monitor Step we get M′ ≈ M′ and
IHMS1 (σc , σf ) ∈
94
IHD ⊢ H : ∆
so this case holds by IHD.
• mon-care(H, ∆) ≈β mon-care(H, H0 )
this holds by definition of compilation.
Having cleared all the cases this case holds.
τ
Rule TLτ -coercion By Rule (J·KLLπ -Coerce), this follows from IH
directly.
Rule TLτ -endorse
This has a number of trivial cases based on
Lτ
Rule (J·KLπ -Endorse) that are analogous to the ones above.
♠
I.7
Proof of Theorem 12 (Compiler J·KLLπ is rs-pres)
τ
Proof. Given:
HP1: ⊢ C : rs
We need to prove:
S
TP1: ⊢ JC, MKT : rs
We unfold the definitions of rs and obtain:
α
fail
HPE1: Ω0 (A
6 ⇒i
h ==
[C])
Lτ
α
==
THE1: Ω0 A JCKLπ
6 ⇒ fail
By definition of the compiler
hwe have
i that
Lτ
∼
HPISR: Ω0 (A [C]) ∼
Ω
A
JCK
π
∼β 0
L
for β = dom(∆), H0 such that M = ({σ} , , σ0 , ∆, σc ) and M = ({σ} ,
We expand HPE1 and we get:
α
HPS1 Ω0 (A [C]) ==⇒ Ω
ǫ
HPS2 Ω =⇒ C′ , H ⊲ Π k (monitor; s)_
ǫ
HPS3 C′ , H ⊲ Π k (monitor; s)_ −−→ C′′ , H ⊲ Π k (skip; s)_
Let us expandTHE1
i
h analogously:
τ
α
HPTHT1 Ω0 A JCKLLπ
==⇒ Ω
, σ0 , H0 , σc )
ǫ
HPTHT2 Ω =⇒ C′ , H ⊲ Π k (monitor; s)_
ǫ
THT3 C′ , H ⊲ Π k (monitor; s)_ −−→ C′′ , H ⊲ Π k (skip; s)_
We now prove this point.
∼
• By Lemma 21 (Related actions preserve ∼
∼) with HPISR IHTHT1 HPS1
′
∼
we know that all related actions preserve ∼
∼β , so for β ⊆ β :
∼
HPOM1: Ω ∼
∼β ′ Ω
95
• Call Π′ = Π k (monitor; s)_ and Π′ = Π k (monitor; s)_ .
We can reduce the possible reductions to two cases, given that Ω =
C; H ⊲ Π′′ and Ω = C; H ⊲ Π′′
ǫ
ǫ
ǫ
ǫ
– C ⊢att Π′′ −−→ Π′ and C ⊢att Π′′ −−→ Π′
– C 6⊢att Π′′ −−→ Π′ and C 6⊢att Π′′ −−→ Π′
∼) and in the
In the first case by Lemma 23 (LP Attacker actions preserve ∼
∼
∼
second case by Lemma 22 (Lτ -compiled actions preserve ∼
∼) with HPOM1,
′
∼
HPTHT2, HPS2 we know that these reductions preserve ∼
∼β , so for β ⊆
′′
β :
′
∼
HPOM2: C′ , H ⊲ Π k (monitor; s)_ ∼
∼β ′′ C , H ⊲ Π k (monitor; s)_
• BY Rule LP -Monitor Step we need to prove σ, H, σ ′ ∈
.
By Rule Monitor relation and Rule Ok Mon we need to prove
– H ≈β ′′ H where for the heaps restricted by mon-care( · ), so dom(H) =
dom(∆) and dom(H) = dom(H0 ).
So by point (2a) of Rule Related states – Secure with HPOM2 we
have ℓ 7→ v : τ ∈ H, n :7→ v : k ∈ H and v ≈β v and ℓ ≈β ′′ hn, ki.
We can apply Rule Heap relation and this case holds.
– ⊢ H : ∆.
This follows from Rule Lτ -Monitor Step, and we need to prove M; H
This holds by HPS3.
M′
♠
∼
Lemma 21 (Related actions preserve ∼
∼).
∀...
α
if Ω ==⇒ Ω′
α
Ω ==⇒ Ω′
α ≈β α
∼
Ω∼
∼β Ω
′∼
∼β ′ Ω′
then Ω ∼
∼
Proof. This is done by trivial case analysis on α and with Lemma 22 (Lτ -compiled actions preserve ∼
∼)
P
∼
∼
and Lemma 23 (L Attacker actions preserve ∼).
96
∼
Lemma 22 (Lτ -compiled actions preserve ∼
∼).
∀...
ǫ
if C, H ⊲ Πρ −−→ C′ , H′ ⊲ Π′ ρ′
Lτ
Lτ
Lτ
ǫ
Lτ
C, H ⊲ JC; Γ ⊢ ΠKLπ JρKLπ =⇒ C′ , H′ ⊲ JC; Γ ⊢ Π′ KLπ Jρ′ KLπ
L
∼
C, H ⊲ Πρ ∼
∼β C, H ⊲ JC; Γ ⊢ ΠKLπ ρ
τ
C; Γ ⊢ Π
Lτ
Lτ
′
′
′
′
∼
then C′ , H′ ⊲ Π′ ρ′ ∼
∼β ′ C , H ⊲ JC; Γ ⊢ Π KLπ Jρ KLπ
Proof. Trivial induction on the derivation of Π, analogous to Lemma 20 (Generalised compiler correctness for J·K
Rule TLτ -new There are 2 cases, they are analogous.
τ = UN By HP
Γ⊢e:τ
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = newτ e in sρ −−→ C, H; ℓ 7→ v : τ ⊲ s[ℓ / x]ρ
Lτ
By Lemma 19 (Expressions compiled with J·KLπ are related) we have:
Lτ
Lτ
Lτ
IHR1 H ⊲ JΓ ⊢ e : τ KLπ JρKLπ ֒→
→ JΓ ⊢ v : τ KLπ
Lτ
By Rule (J·KLπ -New) we get
Lτ
let xo = new JΓ ⊢ e : τ KLπ
in let x = hxo, 0i
Lτ
in JC, Γ; x : Ref τ ⊢ sKLπ
So:
Lτ
C, H ⊲ let xo = new JΓ ⊢ e : τ KLπ
in let x = hxo, 0i
Lτ
ǫ
in JC, Γ; x : Ref τ ⊢ sKLπ
Lτ
−−→ C, H; n 7→ JΓ ⊢ v : τ KLπ : ⊥ ⊲ let x = hn, 0i
ǫ
Lτ
τ
in JC, Γ; x : Ref τ ⊢ sKLLπ
Lτ
−−→ C, H; n 7→ JΓ ⊢ v : τ KLπ : ⊥ ⊲ JC, Γ; x : Ref τ ⊢ sKLπ [hn, 0i / x]
For β ′ = β, this case holds.
else The other case holds follows the same reasoning but
Lτ
for β ′ = β ∪ (ℓ, n, k) and for H′ =H; n 7→ JC, Γ ⊢ v : τ KLπ : k; k.
We need to show that this preserves Rule Related states – Secure,
specifically it preserves point (2a): ℓ ≈β hn, ki and ℓ 7→ v : τ ∈ H and v ≈β v
Lτ
These follow all from the observation above and by Lemma 19 (Expressions compiled with J·KLπ are r
97
∼
Lemma 23 (LP Attacker actions preserve ∼
∼).
∀...
ǫ
if C, H ⊲ Πρ −−→ C′ , H′ ⊲ Π′ ρ′
ǫ
C, H ⊲ Πρ −−→ C′ , H′ ⊲ Π′ ρ′
∼
C, H ⊲ Πρ ∼
∼β C, H ⊲ Πρ
ǫ
C ⊢att Πρ −−→ Π′ ρ′
ǫ
C ⊢att Πρ −−→ Π′ ρ′
′
′
′ ′
∼
then C′ , H′ ⊲ Π′ ρ′ ∼
∼β ′ C , H ⊲ Π ρ
Proof. For the source reductions we can use Lemma 16 (Lτ any ǫ reduction respects heap typing)
to know that secure(H) = secure(H′ ) and that C′ = C, so they don’t change
∼
the interested bits of the ∼
∼β .
By definition we already know that C′ = C.
Suppose this does not hold by contradiction, there can be three clauses that
do not hold based on Rule Related states – Secure:
• violation of (1): ∃π ∈ Π. C ⊢ π : attacker and k ∈ fv(π).
By HP5 this is a contradiction.
• violation of (2a): n 7→ v : k ∈ H and ℓ ≈β hn, ki and ℓ 7→ v : τ ∈ H and ¬(v ≈β v)
To change this value the attacker needs k which contradicts points (1) and
(2b).
• violation of (2b): either of these:
– H, H 0 low-loc(n′ )
Since Rule Lπ -High Location does not hold, by Lemma 5 this is a
contradiction.
– v = k′ for H, H ⊢ high-cap(k′ )
This can follow from another two cases
∗ forgery of k;: an ispection of the semantics rules contradicts this
∗ update of a location to k′ : however k′ is not in the code (contradicts point (1)) and by induction on the heap H we have that
k′ is stored in no other location, so this is also a contradiction.
98
I.8
Proofs for the Non-Atomic Variant of Lτ (Appendix H.2)
The only proof that needs changing is that for Lemma 22: there is this new
case.
∼
For this we weaken ∼
∼β and define ∼β as follows:
Ω ∼β Ω
(Non Atomic State Relation)
∼
Ω∼
∼β Ω
Ω ∼β Ω
(Non Atomic State Relation -stuck)
Ω = C, H ⊲ Π
C = M, F, I
Ω = C, H ⊲ Π
∃π ∈ Π. C 0 π : attacker
∃f ∈ f. f ≈β f
C, H ⊲ π ×
π = (hide n; s)f ;f
∀ℓ. ℓ ∈ dom( ⊢ secure(H))
n 7→ v; k ∈ H ℓ 6∼β hn, ki
ℓ ∼β hn, 0i
Ω ∼β Ω
Two states are now related if:
∼
• either they are related by ∼
∼β
• or the red process is stuck on a hide n where n 7→ v; k but ℓ ∼ hn, ki
does not hold for a ℓ that is secure, and we have that ℓ ∼ hn, 0i (as this
was after the new). And the hide on which the process is stuck is not in
attacker code.
Having this in proofs would not cause problems because now all proofs have an
initial case analysis whether the state is stuck or not, but because it steps it’s
not stuck.
This relation only changes the second case of the proof of Lemma 22 for
Lτ
Rule (J·KLπ -New-nonat) as follows:
Lτ
Proof. new· · is implemented as defined in Rule (J·KLπ -New-nonat).
τ 6= UN By HP
Γ⊢e:τ
H ⊲ e ֒→
→ v
ǫ
C, H ⊲ let x = newτ e in sρ −−→ C, H; ℓ 7→ v : τ ⊲ s[ℓ / x]ρ
By Lemma 19 we have:
Lτ
Lτ
Lτ
IHR1 H ⊲ JΓ ⊢ e : τ KLπ JρKLπ ֒→
→ JΓ ⊢ v : τ KLπ
τ
By Rule (J·KLLπ -New-nonat) we get
99
let x = new 0 in
let xk = hide x in
τ
let xc = J∆, Γ ⊢ e : τ KLLπ in
x := xc with xk;
Lτ
JC, ∆, Γ ⊢ sKLπ
So:
C, H ⊲ let x = new 0 in
let xk = hide x in
Lτ
let xc = J∆, Γ ⊢ e : τ KLπ in
x := xc with xk;
τ
JC, ∆, Γ ⊢ sKLLπ
ǫ
−−→ C, H, n 7→ 0 : ⊥ ⊲ let xk = hide n in
Lτ
let xc = J∆, Γ ⊢ e : τ KLπ in
x := xc with xk;
Lτ
JC, ∆, Γ ⊢ sKLπ
And β ′ = β ∪ (ℓ, n, 0).
Now there are two cases:
• A concurrent attacker reduction performs hide n, so the state changes.
C, H, n 7→ 0 : k; k ⊲ let xk = hide n in
τ
L
let xc = JC, Γ ⊢ e : τ KL
π in
x := xc with xk;
Lτ
JC, ∆, Γ ⊢ sKLπ
At this stage the state is stuck: Rule ELP -hide does not apply.
Also, we have that this holds by the new β ′ : (ℓ ∼β ′ hn, 0i)
And so this does not hold: (ℓ ∼β ′ hn, ki)
As the stuck statement is not in attacker code, we can use Rule Non Atomic State Relation -stuck
to conclude.
• The attacker does not. In this case the proof continues as in Lemma 22.
100
| 6cs.PL
|
arXiv:cs/0311043v1 [cs.PL] 27 Nov 2003
Combining Logi Programs and
Monadi Se ond Order Logi s
by Program Transformation
Fabio Fioravanti1 , Alberto Pettorossi2 , Maurizio Proietti1
(1) IASI-CNR, Viale Manzoni 30, I-00185 Roma, Italy
(2) DISP, University of Roma Tor Vergata, I-00133 Roma, Italy
{fioravanti,adp,proietti}iasi.rm. nr.it
Abstra t We present a program synthesis method based on unfold/fold
transformation rules whi h an be used for deriving terminating de nite
logi programs from formulas of the Weak Monadi Se ond Order theory
of one su essor (WS1S). This synthesis method an also be used as a
proof method whi h is a de ision pro edure for losed formulas of WS1S.
We apply our synthesis method for translating CLP(WS1S) programs
into logi programs and we use it also as a proof method for verifying
safety properties of in nite state systems.
1 Introdu tion
The Weak Monadi Se ond Order theories of k su essors (WSkS) are theories
of the se ond order predi ate logi whi h express properties of nite sets of nite
strings over a k -symbol alphabet (see [?℄ for a survey). Their importan e relies
on the fa t that they are among the most expressive theories of predi ate logi
whi h are de idable. These de idability results were proved in the 1960's [?,?℄,
but they were onsidered as purely theoreti al results, due to the very high
omplexity of the automata-based de ision pro edures.
In re ent years, however, it has been shown that some Monadi Se ond Order
theories an, in fa t, be de ided by using ad-ho , e ient te hniques, su h as
BDD's and algorithms for nite state automata. In parti ular, the MONA system
implements these te hniques for the WS1S and WS2S theories [?℄.
The MONA system has been used for the veri ation of several non-trivial
nite state systems [?,?℄. However, the Monadi Se ond Order theories alone
are not expressive enough to deal with properties of in nite state systems and,
thus, for the veri ation of su h systems alternative te hniques have been used,
su h as those based on the embedding of the Monadi Se ond Order theories
into more powerful logi al frameworks (see, for instan e, [?℄).
In a previous paper of ours [?℄ we proposed a veri ation method for in nite
state systems based on CLP(WSkS), whi h is a onstraint logi programming
language resulting from the embedding of WSkS into logi programs. In order
to perform proofs of properties of in nite state systems in an automati way
a ording to the approa h we have proposed, we need a system for onstraint
1
logi programming whi h uses a solver for WSkS formulas and, unfortunately,
no su h system is available yet.
In order to over ome this di ulty, in this paper we propose a method for
translating CLP(WS1S) programs into logi programs. This translation is performed by a two step program synthesis method whi h produ es terminating
de nite logi programs from WS1S formulas. Step 1 of our synthesis method
onsists in deriving a normal logi program from a WS1S formula, and it is
based on a variant of the Lloyd-Topor transformation [?℄. Step 2 onsists in
applying an unfold/fold transformation strategy to the normal logi program
derived at the end of Step 1, thereby deriving a terminating de nite logi program. Our synthesis method follows the general approa h presented in [?,?℄.
We leave it for future resear h the translation into logi programs starting from
general CLP(WSkS) programs.
The spe i ontributions of this paper are the following ones.
(1) We provide a synthesis strategy whi h is guaranteed to terminate for any
given WS1S formula.
(2) We prove that, when we start from a losed WS1S formula ϕ, our synthesis
strategy produ es a program whi h is either (i) a unit lause of the form f ←,
where f is a nullary predi ate equivalent to the formula ϕ, or (ii) the empty
program. Sin e in ase (i) ϕ is true and in ase (ii) ϕ is false, our strategy is also
a de ision pro edure for WS1S formulas.
(3) We show through a non-trivial example, that our veri ation method
based on CLP(WS1S) programs is useful for verifying properties of in nite state
transition systems. In parti ular, we prove the safety property of a mutual exlusion proto ol for a set of pro esses whose ardinality may hange over time.
Our veri ation method requires: (i) the en oding into WS1S formulas of both
the transition relation and the elementary properties of the states of a transition
system, and (ii) the en oding into a CLP(WS1S) program of the safety property
under onsideration. Here we perform our veri ation task by translating the
CLP(WS1S) program into a de nite logi program, thereby avoiding the use of
a solver for WS1S formulas. The veri ation of the safety property has been
performed by using a prototype tool built on top of the MAP transformation
system [?℄.
2 The Weak Monadi Se ond Order Theory of One
Su essor
We will onsider a rst order presentation of the Weak Monadi Se ond Order
theory of one su essor (WS1S). This rst order presentation onsists in writing
formulas of the form n ∈ S , where ∈ is a rst order predi ate symbol (to be
interpreted as membership of a natural number to a nite set of natural numbers), instead of formulas of the form S(n), where S is a predi ate variable (to
be interpreted as ranging over nite sets of natural numbers).
We use a typed rst order language, with the following two types: nat, denoting the set of natural numbers, and set, denoting the set of the nite sets of
2
natural numbers (for a brief presentation of the typed rst order logi the reader
may look at [?℄). The alphabet of WS1S onsists of: (i) a set Ivars of individual
variables N, N1 , N2 , . . . of type nat, (ii) a set Svars of set variables S, S1 , S2 , . . .
of type set, (iii) the nullary fun tion symbol 0 (zero ) of type nat, and the unary
fun tion symbol s (su essor ) of type nat → nat , and (iv) the binary predi ate
symbols ≤ of type nat × nat , and ∈ of type nat × set . Ivars ∪ Svars is ranged
over by X, X1 , X2 , . . . The syntax of WS1S is de ned by the following grammar:
Individual terms : n ::= 0 | N | s(n)
Atomi formulas : A ::= n1 ≤ n2 | n ∈ S
Formulas :
ϕ ::= A | ¬ϕ | ϕ1 ∧ ϕ2 | ∃N ϕ | ∃S ϕ
When writing formulas we feel free to use also the onne tives ∨, →, ↔ and the
universal quanti er ∀, as shorthands of the orresponding formulas with ¬, ∧,
and ∃. Given any two individual terms n1 and n2 , we will write the formulas
n1 = n2 , n1 6= n2 , and n1 < n2 as shorthands of the orresponding formulas using
≤. Noti e that, for reasons of simpli ity, we have assumed that the symbol ≤ is
primitive, although it is also possible to de ne it in terms of ∈ [?℄.
An example of a WS1S formula is the following formula µ, with free variables
N and S , whi h expresses that N is the maximum number in a nite set S :
µ : N ∈ S ∧ ¬∃N1 (N1 ∈ S ∧ ¬N1 ≤ N )
The semanti s of WS1S formulas is de ned by onsidering the following typed
interpretation N :
(i) the domain of the type nat is the set Nat of the natural numbers and the
domain of the type set is the set Pfin (Nat ) of all nite subsets of Nat ;
(ii) the onstant symbol 0 is interpreted as the natural number 0 and the fun tion
symbol s is interpreted as the su essor fun tion from Nat to Nat ;
(iii) the predi ate symbol ≤ is interpreted as the less-or-equal relation on natural
numbers, and the predi ate symbol ∈ is interpreted as the membership of a
natural number to a nite set of natural numbers.
The notion of a variable assignment σ over a typed interpretation is analogous to the untyped ase, ex ept that σ assigns to a variable an element of the
domain of the type of the variable. The de nition of the satisfa tion relation
I |=σ ϕ, where I is a typed interpretation and σ is a variable assignment is also
analogous to the untyped ase, with the only dieren e that when we interpret
an existentially quanti ed formula we assume that the quanti ed variable ranges
over the domain of its type. We say that a formula ϕ is true in an interpretation
I , written as I |= ϕ, i I |=σ ϕ for all variable assignments σ . The problem
of he king whether or not a WS1S formula is true in the interpretation N is
de idable [?℄.
3 Translating WS1S Formulas into Normal Logi
Programs
In this se tion we illustrate Step 1 of our method for synthesizing de nite programs from WS1S formulas. In this step, starting from a WS1S formula, we de3
rive a strati ed normal logi program [?℄ (simply alled strati ed programs ) by
applying a variant of the Lloyd-Topor transformation, alled typed Lloyd-Topor
transformation. Given a strati ed program P , we denote by M (P ) its perfe t
model (whi h is equal to its least Herbrand model if P is a de nite program) [?℄.
Before presenting the typed Lloyd-Topor transformation, we need to introdu e a de nite program, alled NatSet, whi h axiomatizes: (i) the natural numbers, (ii) the nite sets of natural numbers, (iii) the ordering on natural numbers
(≤), and (iv) the membership of a natural number to a nite set of natural numbers (∈). We represent: (i) a natural number k (≥ 0) as a ground term of the form
sk (0), and (ii) a set of natural numbers as a nite, ground list [b0 , b1 , . . . , bm ]
where, for i = 0, . . . , m, we have that bi is either y or n. A number k belongs
to the set represented by [b0 , b1 , . . . , bm ] i bk = y. Thus, the nite, ground lists
[b0 , b1 , . . . , bm ] and [b0 , b1 , . . . , bm , n, . . . , n] represent the same set. In parti ular,
the empty set is represented by any list of the form [n, . . . , n]. The program
NatSet onsists of the following lauses (we adopt in x notation for ≤ and ∈):
nat(0) ←
0≤N ←
nat(s(N )) ← nat (N )
s(N1 ) ≤ s(N2 ) ← N1 ≤ N2
set([ ]) ←
0 ∈ [y|S] ←
set([y|S]) ← set (S)
s(N ) ∈ [B|S] ← N ∈ S
set([n|S]) ← set (S)
Atoms of the form nat(N ) and set(S) are alled type atoms. Now we will establish
a orresponden e between the set of WS1S formulas whi h are true in N and the
set of the so- alled expli itly typed WS1S formulas whi h are true in the least
Herbrand model M (Nat Set) (see Theorem 1 below).
Given a WS1S formula ϕ, the expli itly typed WS1S formula orresponding
to ϕ is the formula ϕτ onstru ted as follows. We rst repla e the subformulas
of the form ∃N ψ by ∃N (nat (N ) ∧ ψ) and the subformulas of the form ∃S ψ by
∃S (set(S)∧ψ), thereby getting a new formula ϕη where every bound (individual
or set) variable o urs in a type atom. Then, we get:
ϕτ : nat (N1 ) ∧ . . . ∧ nat(Nh ) ∧ set(S1 ) ∧ . . . ∧ set (Sk ) ∧ ϕη
where N1 , . . . , Nh , S1 , . . . , Sk are the variables whi h o ur free in ϕ.
For instan e, let us onsider again the formula µ whi h expresses that N is
the maximum number in a set S . The expli itly typed formula orresponding to
µ is the following formula:
µτ : nat (N ) ∧ set (S) ∧ N ∈ S ∧ ¬∃N1 (nat (N1 ) ∧ N1 ∈ S ∧ ¬N1 ≤ N )
For reasons of simpli ity, in the following Theorem 1 we identify: (i) a natural
number k (≥ 0) in Nat with the ground term sk (0) representing that number,
and (ii) a nite set of natural numbers in Pfin (Nat ) with any nite, ground list
representing that set. By using these identi ations, we an view any variable
assignment over the typed interpretation N also as a variable assignment over
the untyped interpretation M (Nat Set) (but not vi e versa).
Theorem 1. Let ϕ be a WS1S formula and let ϕτ be the expli itly typed
formula orresponding to ϕ. For every variable assignment σ over N ,
N |=σ ϕ i M (Nat Set) |=σ ϕτ
4
Proof. The proof pro eeds by indu tion on the stru ture of the formula ϕ.
(i) Suppose that ϕ is of the form n1 ≤ n2 . By the de nition of the satisfa tion
relation, N |=σ n1 ≤ n2 i the natural number σ(n1 ) is less or equal than
the natural number σ(n2 ). By the de nition of least Herbrand model and by
using the lauses in NatSet whi h de ne ≤, σ(n1 ) is less or equal than σ(n2 )
i M (NatSet) |= σ(n1 ) ≤ σ(n2 ) (here we identify every natural number n with
the ground term sn (0)). It an be shown that M (NatSet) |= nat (σ(n1 )) and
M (NatSet) |= nat (σ(n2 )). Thus, M (NatSet) |= σ(n1 ) ≤ σ(n2 ) i M (NatSet) |=σ
nat (n1 ) ∧ nat (n2 ) ∧ n1 ≤ n2 . Now, the term n1 is either of the form sm1 (0) or
of the form sm1 (N1 ), where m1 is a natural number. Similarly, the term n2 is
either of the form sm2 (0) or of the form sm2 (N ), where m2 is a natural number.
We onsider the ase where n1 is sm1 (N1 ) and n2 is sm2 (N2 ). The other ases
are similar and we omit them. It an be shown that, for all natural numbers m,
M (NatSet) |=σ nat(sm (N )) i M (NatSet) |=σ nat(N ). Thus, M (NatSet) |=σ
nat (sm1 (N1 )) ∧ nat (sm2 (N2 )) ∧ sm1 (N1 ) ≤ sm2 (N2 ) i M (NatSet) |=σ nat (N1 ) ∧
nat (N2 ) ∧ sm1 (N1 ) ≤ sm2 (N2 ), that is, M (NatSet) |=σ (n1 ≤ n2 )τ .
(ii) The ase where ϕ is of the form n ∈ S is similar to Case (i).
(iii) Suppose that ϕ is of the form ¬ψ . By the de nition of the satisfa tion
relation and the indu tion hypothesis, N |=σ ¬ψ i M (NatSet) |=σ ¬(ψτ ).
Sin e ψτ is of the form a1 (X1 ) ∧ . . . ∧ ak (Xk ) ∧ ψη , where X1 , . . . , Xk are the
free variables in ψ and a1 (X1 ), . . . , ak (Xk ) are type atoms, by logi al equivalen e, we get: M (NatSet) |=σ ¬(ψτ ) i M (NatSet) |=σ (a1 (X1 ) ∧ . . . ∧ ak (Xk ) ∧
¬(ψη )) ∨ ¬(a1 (X1 ) ∧ . . . ∧ ak (Xk )). Finally, sin e for all variable assignments σ ,
M (NatSet) |=σ a1 (X1 ) ∧ . . . ∧ ak (Xk ), we have that M (NatSet) |=σ ¬(ψτ ) i
M (NatSet) |=σ (a1 (X1 ) ∧ . . . ∧ ak (Xk ) ∧ ¬(ψη )), that is, M (NatSet) |=σ (¬ψ)τ
(to see this, note that ¬(ψη ) is equal to (¬ψ)η ).
(iv) The ase where ϕ is of the form ψ1 ∧ ψ2 is similar to Case (iii).
(v) Suppose that ϕ is of the form ∃N1 ψ . By the de nition of the satisfa tion
relation and by the indu tion hypothesis, N |=σ ∃N1 ψ i there exists n1 in Nat
su h that M (NatSet) |=σ[N1 7→n1 ] ψτ . Sin e ψτ is of the form nat(N1 ) ∧ . . . ∧
nat (Nh ) ∧ set (S1 ) ∧ . . . ∧ set (Sk ) ∧ ψη , where N1 , . . . , Nh , S1 , . . . , Sk are the free
variables in ψ , we have that:
there exists n1 in Nat su h that M (NatSet) |=σ[N1 7→n1 ] ψτ
i M (NatSet) |=σ ∃N1 (nat (N1 ) ∧ . . . ∧ nat(Nh ) ∧ set(S1 ) ∧ . . . ∧ set (Sk ) ∧ ψη )
i (by logi al equivalen e) M (NatSet) |=σ nat (N2 ) ∧ . . . ∧ nat(Nh ) ∧ set(S1 ) ∧
. . . ∧ set (Sk ) ∧ (∃N1 nat(N1 ) ∧ ψη )
i (by de nition of expli itly typed formula) M (NatSet) |=σ (∃N1 ψ)τ .
✷
(vi) The ase where ϕ is of the form ∃S ψ is similar to Case (v).
As a straightforward onsequen e of Theorem 1, we have the following result.
Corollary 1.
For every losed WS1S formula ϕ, N |= ϕ i M (Nat Set) |= ϕτ .
Noti e that the introdu tion of type atoms is indeed ne essary, be ause there
are WS1S formulas ϕ su h that N |= ϕ and M (NatSet) 6|= ϕ. For instan e, N |=
∀N1 ∃N2 N1 ≤ N2 and M (Nat Set) 6|= ∀N1 ∃N2 N1 ≤ N2 . Indeed, for a variable
assignment σ over M (Nat Set) whi h assigns [ ] to N1 , we have M (NatSet) 6|=σ
5
∃N2 N1 ≤ N2 . (Noti e that σ is not a variable assignment over N be ause [ ] is
not a natural number.)
Now we present a variant of the method proposed by Lloyd and Topor [?℄,
alled typed Lloyd-Topor transformation, whi h we use for deriving a strati ed
program from a given WS1S formula ϕ. We need to onsider a lass of formulas
of the form: A ← β , alled statements, where A is an atom, alled the head of
the statement, and β is a formula of the rst order predi ate al ulus, alled the
body of the statement. In what follows we write C[γ] to denote a formula where
the subformula γ o urs as an outermost onjun t, that is, C[γ] = ψ1 ∧ γ ∧ ψ2
for some subformulas ψ1 and ψ2 .
The Typed Lloyd-Topor Transformation.
We are given in input a set of statements, where: (i) we assume without loss of
generality, that the only onne tives and quanti ers o urring in the body of the
statements are ¬, ∧, and ∃, and (ii) X, X1 , X2 , . . . denote either individual or set
variables.
We perform the following transformation (A) and then the transformation (B):
(A) We repeatedly apply the following rules A.1A.4 until a set of lauses is
generated:
(A.1) A ← C[¬¬γ] is repla ed by A ← C[γ].
(A.2) A ← C[¬(γ ∧ δ)] is repla ed by A ← C[¬newp(X1 , . . . , Xk )]
newp(X1 , . . . , Xk ) ← γ ∧ δ
where newp is a new predi ate and X1 , . . . , Xk are the variables whi h o ur
free in γ ∧ δ .
(A.3) A ← C[¬∃X γ] is repla ed by A ← C[¬newp(X1 , . . . , Xk )]
newp(X1 , . . . , Xk ) ← γ
where newp is a new predi ate and X1 , . . . , Xk are the variables whi h o ur free
in ∃X γ .
(A.4) A ← C[∃X γ] is repla ed by A ← C[γ{X/X1}]
where X1 is a new variable.
(B) Every lause A ← G is repla ed by A ← Gτ .
Given a WS1S formula ϕ with free variables X1 , . . . , Xn , we denote by Cls(f, ϕτ )
the set of lauses derived by applying the typed Lloyd-Topor transformation
starting from the singleton {f (X1 , . . . , Xn ) ← ϕ}, where f is a new n-ary predi ate symbol. By onstru tion, Nat Set ∪ Cls(f, ϕτ ) is a strati ed program. We
have the following theorem.
Theorem 2. Let ϕ be a WS1S formula with free variables X1 , . . . , Xn and let
ϕτ be the expli itly typed formula orresponding to ϕ. For all ground terms
t1 , . . . , tn , we have that:
6
M (Nat Set) |= ϕτ {X1 /t1 , . . . , Xn /tn } i
M (NatSet ∪ Cls(f, ϕτ )) |= f (t1 , . . . , tn )
Proof. It is similar to the proofs presented in [?,?℄ and we omit it.
From Theorems 1 and 2 we have the following orollaries.
Corollary 2. For every WS1S formula ϕ with free variables X1 , . . . , Xn , and
for every variable assignment σ over the typed interpretation N ,
N |=σ ϕ i M (NatSet ∪ Cls(f, ϕτ )) |= f (σ(X1 ), . . . , σ(Xn ))
Corollary 3.
For every losed WS1S formula ϕ,
N |= ϕ i M (NatSet ∪ Cls(f, ϕτ )) |= f
Let us onsider again the formula µ we have onsidered above. By applying the
typed Lloyd-Topor transformation starting from the singleton {max (S , N ) ← µ}
we get the following set of lauses Cls(max , µτ ):
max (S , N ) ← nat (N ) ∧ set(S) ∧ N ∈ S ∧ ¬newp(S, N )
newp(S, N ) ← nat(N ) ∧ nat(N1 ) ∧ set (S) ∧ N1 ∈ S ∧ ¬N1 ≤ N
Unfortunately, the strati ed program NatSet∪Cls(f, ϕτ ) derived from the singleton {f (X1 , . . . , Xn ) ← ϕ} is not always satisfa tory from a omputational point
of view be ause it may not terminate when evaluating the query f (X1 , . . . , Xn )
by using SLDNF resolution. (A tually, the above program Cls(max , µτ ) whi h
omputes the maximum number of a set, terminates for all ground queries, but
in Se tion 5 we will give an example where the program derived at the end of
the typed Lloyd-Topor transformation does not terminate.) Similar termination
problems may o ur by using tabled resolution [?℄, instead of SLDNF resolution.
To over ome this problem, we apply to the program Nat Set ∪ Cls(f, ϕτ )
the unfold/fold transformation strategy whi h we will des ribe in Se tion 5. In
parti ular, by applying this strategy we derive de nite programs whi h terminate
for all ground queries by using LD resolution (that is, SLD resolution with the
leftmost sele tion rule).
4 The Transformation Rules
In this se tion we des ribe the transformation rules whi h we use for transforming
strati ed programs. These rules are a subset of those presented in [?,?℄, and
they are those required for the unfold/fold transformation strategy presented in
Se tion 5.
For presenting our rules we need the following notions. A variable in the
body of a lause C is said to be existential i it does not o ur in the head of
C . The de nition of a predi ate p in a program P , denoted by Def (p, P ), is the
set of the lauses of P whose head predi ate is p. The extended de nition of a
7
predi ate p in a program P , denoted by Def ∗ (p, P ), is the union of the de nition
of p and the de nitions of all predi ates in P on whi h p depends. (See [?℄for
the de nition of the depends on relation.) A program is propositional i every
predi ate o urring in the program is nullary. Obviously, if P is a propositional
program then, for every predi ate p, M (P ) |= p is de idable.
A transformation sequen e is a sequen e P0 , . . . , Pn of programs, where for
0 ≤ k ≤ n−1, program Pk+1 is derived from program Pk by the appli ation of one
of the transformation rules R1R4 listed below. For 0 ≤ k ≤ n, we onsider the set
Defs k of the lauses introdu ed by the following rule R1 during the onstru tion
of the transformation sequen e P0 , . . . , Pk .
When onsidering lauses of programs, we will feel free to apply the following
transformations whi h preserve the perfe t model semanti s:
(1) renaming of variables,
(2) rearrangement of the order of the literals in the body of a lause, and
(3) repla ement of a onjun tion of literals the form L ∧ L in the body of a
lause by the literal L.
Rule R1. De nition. We get the new program Pk+1 by adding to program Pk
a lause of the form newp(X1 , . . . , Xr ) ← L1 ∧ . . . ∧ Lm , where: (i) the predi ate
newp is a predi ate whi h does not o ur in P0 ∪ Defs k , and (ii) X1 , . . . , Xr are
distin t (individual or set) variables o urring in L1 ∧ . . . ∧ Lm .
Rule R2. Unfolding.
Let C be a renamed apart lause in Pk of the form:
H ← G1 ∧ L ∧ G2 , where L is either the atom A or the negated atom ¬A. Let
H1 ← B1 , . . . , Hm ← Bm , with m ≥ 0, be all lauses of program Pk whose head
is uni able with A and, for j = 1, . . . , m, let ϑj the most general uni er of A
and Hj . We onsider the following two ases.
Case 1: L is A. By unfolding lause C w.r.t. A we derive the new program
Pk+1 = (Pk − {C}) ∪ {(H ← G1 ∧ B1 ∧ G2 )ϑ1 , . . . , (H ← G1 ∧ Bm ∧ G2 )ϑm }.
In parti ular, if m = 0, that is, if we unfold C w.r.t. an atom whi h is not uni able
with the head of any lause in Pk , then we derive the program Pk+1 by deleting
lause C .
Case 2: L is ¬A. Assume that: (i) A = H1 ϑ1 = · · · = Hm ϑm , that is, for
j = 1, . . . , m, A is an instan e of Hj , (ii) for j = 1, . . . , m, Hj ← Bj has no
existential variables, and (iii) Q1 ∨ . . . ∨ Qr , with r ≥ 0, is the disjun tive normal
form of G1 ∧¬(B1 ϑ1 ∨. . .∨Bm ϑm )∧G2 . By unfolding lause C w.r.t. ¬A we derive
the new program Pk+1 = (Pk − {C}) ∪ {C1 , . . . , Cm }, where for j = 1, . . . , r, Cj
is the lause H ← Qj .
In parti ular: (i) if m = 0, that is, A is not uni able with the head of any lause
in Pk , then we get the new program Pk+1 by deleting ¬A from the body of lause
C , and (ii) if for some j ∈ {1, . . . , m}, Bj is the empty onjun tion, that is, A is
an instan e of the head of a unit lause in Pk , then we derive Pk+1 by deleting
lause C from Pk .
Rule R3. Folding.
Let C : H ← G1 ∧ Bϑ ∧ G2 be a renamed apart lause
in Pk and D : Newp ← B be a lause in Defs k . Suppose that for every existential variable X of D, we have that Xϑ is a variable whi h o urs neither
8
in {H, G1 , G2 } nor in the term Y ϑ, for any variable Y o urring in B and different from X . By folding lause C using lause D we derive the new program
Pk+1 = (Pk − {C}) ∪ {H ← G1 ∧ Newp ϑ ∧ G2 }.
Rule R4. Propositional Simpli ation.
Let p be a predi ate su h that
Def ∗ (p, Pk ) is propositional. If M (Def ∗ (p, Pk )) |= p then we derive Pk+1 =
(Pk − Def (p, Pk )) ∪ {p ←}. If M (Def ∗ (p, Pk )) |= ¬p then we derive Pk+1 =
(Pk − Def (p, Pk )).
Noti e that we an he k whether or not M (P ) |= p holds by applying program transformation te hniques [?℄ and thus, Rule R4 may be viewed as a derived
rule.
The transformation rules R1R4 we have introdu ed above, are olle tively
alled unfold/fold transformation rules. We have the following orre tness result,
similar to [?℄.
Theorem 3. [Corre tness of the Unfold/Fold Transformation Rules℄
Let us assume that during the onstru tion of a transformation sequen e P0 , . . . ,
Pn , ea h lause of Defs n whi h is used for folding, is unfolded (before or after
its use for folding) w.r.t. an atom whose predi ate symbol o urs in P0 . Then,
M (P0 ∪ Defs n ) = M (Pn ).
Noti e that the statement obtained from Theorem 3 by repla ing `atom' by
`literal', does not hold [?℄.
5 The Unfold/Fold Synthesis Method
In this se tion we present our program synthesis method, alled unfold/fold synthesis method, whi h derives a de nite program from any given WS1S formula.
We show that the synthesis method terminates for all given formulas and also
the derived programs terminate a ording to the following notion of program
termination: a program P terminates for a query Q i every SLD-derivation of
P ∪ {← Q} via any omputation rule is nite.
The following is an outline of our unfold/fold synthesis method.
The Unfold/Fold Synthesis Method.
Let ϕ be a WS1S formula with free variables X1 , . . . , Xn and let ϕτ be the
expli itly typed formula orresponding to ϕ.
Step 1. We apply the typed Lloyd-Topor transformation and we derive a set
Cls(f, ϕτ ) of lauses su h that: (i) f is a new n-ary predi ate symbol, (ii) Nat Set
∪Cls(f, ϕτ ) is a strati ed program, and (iii) for all ground terms t1 , . . . , tn ,
(1) M (NatSet) |= ϕτ {X1 /t1 , . . . , Xn /tn } i
M (NatSet ∪ Cls(f, ϕτ )) |= f (t1 , . . . , tn )
Step 2. We apply the unfold/fold transformation strategy (see below) and from
the program Nat Set∪Cls(f, ϕτ ) we derive a de nite program TransfP su h that,
for all ground terms t1 , . . . , tn ,
9
(2.1) M (NatSet ∪ Cls(f, ϕτ )) |= f (t1 , . . . , tn ) i M (TransfP ) |= f (t1 , . . . , tn );
(2.2) TransfP terminates for the query f (t1 , . . . , tn ).
In order to present the unfold/fold transformation strategy whi h we use for
realizing Step 2 of our synthesis method, we introdu e the following notions of
regular natset-typed lauses and regular natset-typed de nitions.
We say that a literal is linear i ea h variable o urs at most on e in it.
The syntax of regular natset-typed lauses is de ned by the following grammar
(re all that by N we denote individual variables, by S we denote set variables,
and by X, X1 , X2 , . . . we denote either individual or set variables):
Head terms : h ::= 0 | s(N ) | [ ] | [y|S] | [n|S]
Clauses :
C ::= p(h1 , . . . , hk ) ← | p1 (h1 , . . . , hk ) ← p2 (X1 , . . . , Xm )
where for every lause C , (i) both hd(C) and bd(C) are linear atoms, and
(ii) {X1 , . . . , Xm } ⊆ vars(h1 , . . . , hk ) (that is, C has no existential variables). A
regular natset-typed program is a set of regular natset-typed lauses.
The reader may he k that the program NatSet presented in Se tion 3 is
a regular natset-typed program. The following properties are straightforward
onsequen es of the de nition of regular natset-typed program.
Lemma 1.
Let P be a regular natset-typed program. Then:
(i) P terminates for every ground query p(t1 , . . . , tn ) with n > 0;
(ii) If p is a nullary predi ate then Def ∗ (p, P ) is propositional.
The syntax of natset-typed de nitions is given by the following grammar:
Individual terms : n ::= 0 | N | s(n)
Terms :
t ::= n | S
Type atom s:
T ::= nat (N ) | set (S)
Literals :
L ::= p(t1 , . . . , tk ) | ¬p(t1 , . . . , tk )
De nitions :
D ::= p(X1 , . . . , Xk ) ← T1 ∧ . . . ∧ Tr ∧ L1 ∧ . . . ∧ Lm
where for all de nitions D, vars(D) ⊆ vars(T1 ∧ . . . ∧ Tr ).
A sequen e D1 , . . . , Ds of natset-typed de nitions is said to be a hierarhy i for i = 1, . . . , s the predi ate appearing in hd(Di ) does not o ur in
D1 , . . . , Di−1 , bd(Di ). Noti e that in a hierar hy of natset-typed de nitions, any
predi ate o urs in the head of at most one lause.
One an show that given a WS1S formula ϕ the set Cls(f, ϕτ ) of lauses derived by applying the typed Lloyd-Topor transformation is a hierar hy D1 , . . . , Ds
of natset-typed de nitions and the last lause Ds is the one de ning f .
10
The Unfold/Fold Transformation Strategy.
Input : (i) A regular natset-typed program P where for ea h nullary predi ate
p, Def ∗ (p, Transf P ) is either the empty set or the singleton {p ←}, and (ii) a
hierar hy D1 , . . . , Ds of natset-typed de nitions su h that no predi ate o urring
in P o urs also in the head of a lause in D1 , . . . , Ds .
Output : A regular natset-typed program TransfP su h that, for all ground terms
t1 , . . . , t n ,
(2.1) M (P ∪ {D1 , . . . , Ds }) |= f (t1 , . . . , tn ) i M (TransfP ) |= f (t1 , . . . , tn );
(2.2) TransfP terminates for the query f (t1 , . . . , tn ).
TransfP := P ; Defs := ∅;
for i = 1, . . . , s do
Defs := Defs ∪ {Di }; InDefs := {Di };
By the de nition rule we derive the program TransfP ∪ InDefs .
while InDefs 6= ∅ do
(1) Unfolding. From program TransfP ∪InDefs we derive TransfP ∪U by: (i) applying the unfolding rule w.r.t. ea h atom o urring positively in the body of a
lause in InDefs , thereby deriving TransfP ∪ U1 , then (ii) applying the unfolding
rule w.r.t. ea h negative literal o urring in the body of a lause in U1 , thereby
deriving TransfP ∪U2 , and, nally, (iii) applying the unfolding rule w.r.t. ground
literals until we derive a program TransfP ∪U su h that no ground literal o urs
in the body of a lause of U .
(2) De nition-Folding. From program TransfP ∪ U we derive TransfP ∪ F ∪
NewDefs as follows. Initially, NewDefs is the empty set. For ea h non-unit lause
C : H ← B in U ,
(i) we apply the de nition rule and we add to NewDefs a lause of the form
newp(X1 , . . . , Xk ) ← B , where X1 , . . . , Xk are the non-existential variables o urring in B , unless a variant lause already o urs in Defs, modulo the head
predi ate symbol and the order and multipli ity of the literals in the body, and
(ii) we repla e C by the lause derived by folding C w.r.t. B . The folded lause
is an element of F .
No transformation rule is applied to the unit lauses o urring in U and, therefore, also these lauses are elements of F .
(3) TransfP := TransfP ∪ F ;
end while;
Defs := Defs ∪ NewDefs ;
InDefs := NewDefs
Propositional Simpli ation. For ea h predi ate p su h that Def ∗ (p, TransfP ) is
propositional, we apply the propositional simpli ation rule and
if M (TransfP ) |= p
then TransfP := (TransfP − Def (p, TransfP )) ∪ {p ←}
else TransfP := (TransfP − Def (p, TransfP ))
end for
11
The reader may verify that if we apply the unfold/fold transformation strategy starting from the program NatSet together with the lauses Cls(max , µτ )
whi h we have derived above by applying the typed Lloyd-Topor transformation,
we get the following nal program:
max ([y|S], 0) ← new 1(S)
max ([y|S], s(N )) ← max (S, N )
max ([n|S], s(N )) ← max (S, N )
new 1([ ]) ←
new 1([n|S]) ← new 1(S)
To understand the rst lause, re all that the empty set is represented by any list
of the form [n, . . . , n]. A more detailed example of appli ation of the unfold/fold
transformation strategy will be given later.
In order to prove the orre tness and the termination of our unfold/fold
transformation strategy we need the following lemmas whose proofs are mutually
dependent.
Lemma 2. During the appli ation of the unfold/fold transformation strategy,
TransfP is a regular natset-typed program.
Proof. Initially, TransfP is the regular natset-typed program P . Now we assume
that TransfP is a regular natset-typed program and we show that after an exe ution of the body of the for statement, TransfP is a regular natset-typed
program.
First we prove that after the exe ution of the while statement, TransfP is
a regular natset-typed program. In order to prove this, we show that every new
lause E whi h is added to TransfP at Point (3) of the strategy is a regular
natset-typed lause.
Clause E is derived from a lause D of InDefs by unfolding (a ording to
the Unfolding phase) and by folding (a ording to the De nition-Folding phase).
By Lemma 3, D is a natset-typed de nition of the form p(X1 , . . . , Xk ) ← T1 ∧
. . . ∧ Tr ∧ L1 ∧ . . . ∧ Lm . By unfolding w.r.t. the type atoms T1 , . . . , Tr (a ording
to Point (i) of the Unfolding phase) we get lauses of the form p(h1 , . . . , hk ) ←
′
T1′ ∧. . .∧Tr1
∧L′1 ∧. . .∧L′m , where: (a) h1 , . . . , hk are head terms, (b) p(h1 , . . . , hk )
is a linear atom (be ause X1 , . . . , Xk are distin t variables), and ( ) for i =
1, . . . , m, no argument of L′i is a variable. By the indu tive hypothesis TransfP
is a regular natset-typed program and, therefore, by unfolding w.r.t. the literals
L′1 , . . . , L′m (a ording to Points (ii) and (iii) of the Unfolding phase) we get
′
∧ L′′1 ∧ . . . ∧ L′′m1 . Either
lauses of the form D′ : p(h1 , . . . , hk ) ← T1′ ∧ . . . ∧ Tr1
′
D is a unit lause or, by folding a ording to the De nition-Folding phase, it
is repla ed by p(h1 , . . . , hk ) ← newp(X1 , . . . , Xm ) where X1 , . . . , Xm are the
distin t, non-existential variables o urring in bd(D′ ). Hen e, E is either a unit
lause of the form p(h1 , . . . , hk ) ← or a lause of the form p(h1 , . . . , hk ) ←
newp(X1 , . . . , Xm ), where {X1 , . . . , Xm } ⊆ vars(h1 , . . . , hk ). Thus, E is a regular
natset-typed lause.
We on lude the proof by observing that if we apply the propositional simpli ation rule to a natset-typed program, then we derive a natset-typed program,
12
be ause by this rule we an only delete lauses or add natset-typed lauses of the
form p ←. Thus, after an exe ution of the body of the for statement, TransfP
is a regular natset-typed program.
✷
Lemma 3.
During the appli ation of the unfold/fold transformation strategy,
InDefs is a set of natset-typed de nitions.
Proof. Let us onsider the i-th exe ution of the body of the
for statement.
Initially, InDefs is the singleton set {Di } of natset-typed de nitions. Now we
assume that InDefs is a set of natset-typed de nitions and we prove that, after
an exe ution of the while statement, InDefs is a set of natset-typed de nitions.
It is enough to show that every new lause E whi h is added to InDefs at
Point (3) of the strategy, is a natset-typed de nition. By the Folding phase
of the strategy, E is a lause of the form newp(X1 , . . . , Xk ) ← B where B is
the body of a lause derived from a lause D of InDefs by unfolding. By the
indu tive hypothesis, D is a natset-typed de nition of the form p(X1 , . . . , Xk ) ←
T1 ∧ . . . ∧ Tr ∧ L1 ∧ . . . ∧ Lm . By unfolding w.r.t. the type atoms T1 , . . . , Tr
(a ording to Point (i) of the Unfolding phase) we get lauses of the form D′ :
′
′
p(h1 , . . . , hk ) ← T1′ ∧. . .∧Tr1
∧L′1 ∧. . .∧L′m , where vars(D′ ) ⊆ vars(T1′ ∧. . .∧Tr1
).
Sin e, by Lemma 2, TransfP is a regular natset-typed program, by unfolding
w.r.t. the literals L′1 , . . . , L′m (a ording to Points (ii) and (iii) of the Unfolding
′
phase) we get lauses of the form D′′ : p(h1 , . . . , hk ) ← T1′ ∧. . .∧Tr1
∧L′′1 ∧. . .∧L′′m1
′′
′
′
where vars(D ) ⊆ vars(T1 ∧ . . . ∧ Tr1 ). Thus, E is a natset-typed de nition of
′
∧ L′′1 ∧ . . . ∧ L′′m1 with vars(E) ⊆
the form newp(X1 , . . . , Xk ) ← T1′ ∧ . . . ∧ Tr1
′
′
vars(T1 ∧ . . . ∧ Tr1 ).
We on lude the proof by observing that the Propositional Simpli ation
phase does not hange InDefs , and thus, after the exe ution of the body of the
✷
for statement, InDefs is a set of natset-typed de nitions.
Theorem 4.
Let P and D1 , . . . , Ds be the input program and the input hierar hy, respe tively, of the unfold/fold transformation strategy and let TransfP
be the output of the strategy. Then,
(1) TransfP is a natset-typed program;
(2) for every nullary predi ate p, Def ∗ (p, TransfP ) is either ∅ or {p ←};
(3) for all ground terms t1 , . . . , tn ,
(3.1) M (P ∪ {D1 , . . . , Ds }) |= f (t1 , . . . , tn ) i M (TransfP ) |= f (t1 , . . . , tn );
(3.2) TransfP terminates for the query f (t1 , . . . , tn ).
Proof. Point (1) is a straightforward onsequen e of Lemma 2.
For Point (2), let us noti e that, by Lemma 2, at ea h point of the unfold/fold
transformation strategy TransfP is a natset-typed program and therefore, by
Lemma 1, for every nullary predi ate p, Def ∗ (p, TransfP ) is propositional. Sin e
the last step of the unfold/fold transformation strategy onsists in applying to
TransfP the propositional simpli ation rule for ea h predi ate having a propositional extended de nition, Def ∗ (p, TransfP ) is either ∅ or {p ←}.
13
Point (3.1) will be proved by using the orre tness of the transformation rules
w.r.t. the Perfe t Model semanti s (see Theorem 3). Let us rst noti e that the
unfold/fold transformation strategy generates a transformation sequen e (see
Se tion 4), where: the initial program is P , the nal program is the nal value of
TransfP , and the set of lauses introdu ed by the de nition rule R1 is the nal
value of Defs .
To see that our strategy indeed generates a transformation sequen e, let us
observe the following fa ts (A) and (B):
(A) The addition of InDefs to TransfP at the beginning of ea h exe ution of
the body of the for statement is an appli ation of the de nition rule. Indeed,
for i = 1, . . . s, InDefs = {Di } and, by the hypotheses on the input sequen e
D1 , . . . , Ds , we have that the head predi ate of Di does not o ur in the urrent
value of P ∪ Defs .
(B) When we unfold the lauses of U1 w.r.t. negative literals, we have that:
(B.1) Condition (i) of Case (2) of the unfolding rule (see Se tion 4) is satis ed
be ause:
(a) Every lause D of InDefs is a natset-typed de nition (see Lemma 3) and,
thus, for ea h variable X o urring in D there is a type atom of the form a(X)
in bd(D). Sin e we unfold the lauses of InDefs w.r.t. all the atoms whi h o ur
positively in the bodies of the lauses in InDefs , and in parti ular, w.r.t. type
atoms, every argument of a negative literal in the body of a lause of U1 is of
one of the following forms: 0, s(n), [ ], [y|S], [n|S].
(b) For ea h negative literal ¬p(t1 , . . . , tk ) in the body of a lause of U1 , the
de nition of p is a subset of the regular natset-typed program TransfP (see
Lemma 2) and, hen e, the head of a lause in TransfP is a linear atom of the
form p(h1 , . . . , hk ), where h1 , . . . , hk are head terms (see the de nition of regular
natset-typed lauses above).
From (a) and (b) it follows that if p(t1 , . . . , tk ) is uni able with p(h1 , . . . , hk )
then p(t1 , . . . , tk ) is an instan e of p(h1 , . . . , hk ).
(B.2) Condition (ii) of Case (2) of the unfolding rule is satis ed be ause TransfP
is a regular natset-typed program (see Lemma 2) and, thus, no lause in TransfP
has existential variables.
Now, the transformation sequen e onstru ted by the unfold/fold transformation strategy satis es the hypothesis of Theorem 3. Indeed, let us onsider a
lause D whi h is used for folding a lause C . Sin e C has been derived at the
end of the Unfolding phase, no ground literal o urs in bd(C) and, thus, there
is at least one variable o urring in D. Hen e, there is at least one type atom in
bd(D), be ause D is a natset-typed de nition (see Lemma 3). Therefore, during
an appli ation of the unfold/fold transformation strategy (before or after the
use of D for folding), D is unfolded w.r.t. a type atom (see Point (i) of the Unfolding phase). Thus, by Theorem 3, we have that M (P ∪ Defs) = M (TransfP ),
where by Defs and TransfP we indi ate the values of these variables at the end
of the unfold/fold transformation strategy. Observe that Def ∗ (f, P ∪ Defs) =
Def ∗ (f, P ∪ {D1 , . . . , Ds }) and, therefore, M (P ∪ {D1 , . . . , Ds }) |= f (t1 , . . . , tn )
i M (P ∪ Defs) |= f (t1 , . . . , tn ) i M (TransfP ) |= f (t1 , . . . , tn ).
14
Finally, let us prove Point (3.2). We onsider the following two ases:
(n = 0) f is nullary and hen e, by Point (2) of this theorem, Def ∗ (f, TransfP )
is either ∅ or {f ←}. Thus, TransfP terminates for the query f .
(n > 0) By Point (1) of this theorem, TransfP is a natset-typed program and
✷
thus, by Lemma 1, TransfP terminates for the ground query f (t1 , . . . , tn ).
Theorem 5.
The unfold/fold transformation strategy terminates.
Proof. We have to show that the while statement in the body of the for
statement terminates.
Ea h exe ution of the Unfolding phase terminates. Indeed, (a) the number of
appli ations of the unfolding rule at Points (i) and (ii) is nite, be ause InDefs is
a nite set of lauses and the body of ea h lause has a nite number of literals,
and (b) at Point (iii) only a nite number of unfolding steps an be applied
w.r.t. ground literals, be ause the program held by TransfP during the Unfolding
phase terminates for every ground query. To see this latter fa t, let us noti e that,
by Lemma 2, TransfP is a natset-typed program. Thus, by Lemma 1, TransfP
terminates for any ground query p(t1 , . . . , tn ) with n ≥ 1. For a ground query p,
where p is a nullary predi ate, TransfP terminates be ause Def ∗ (p, Transf P ) is
either the empty set or it is the singleton {p ←}. Indeed, this follows from our
assumptions on the input program and from the exe ution of the Propositional
Simpli ation phase after ompletion of the while statement.
Ea h exe ution of the De nition-Folding phase terminates be ause a nite
number of lauses are introdu ed by de nition and a nite number of lauses are
folded.
Thus, in order to show that the strategy terminates, it is enough to show
that after a nite number of exe utions of the body of the while statement,
we get InDefs = ∅. Let Defs j and InDefs j be the values of Defs and InDefs ,
respe tively, at the end of the j -th exe ution of the body of the while statement.
If the while statement terminates after z exe utions of its body, then, for all
j > z , we de ne Defs j to be Defs z and InDefs j to be ∅. We have that, for any
j ≥ 1, InDefs j = ∅ i Defs j−1 = Defs j . Sin e for all j ≥ 1, Defs j−1 ⊆ Defs j ,
the termination of the strategy will follow from the following property:
there exists K > 0 su h that, for all j ≥ 1, |Defs j | ≤ K
(*)
Let TransfP 0 , Defs 0 , and InDefs 0 (⊆ Defs 0 ) be the values of TransfP , Defs , and
InDefs , respe tively, at the beginning of the exe ution of the while statement.
By Lemma 3, for all j ≥ 1, Defs j is a set of natset-typed de nitions. Property (*)
follows from the fa t that, for all D ∈ Defs j , the following holds:
(a) every predi ate o urring in bd(D) also o urs in TransfP 0 ∪ InDefs 0 ;
(b) for every literal L o urring in bd(D),
height (L) ≤ max{height (M ) | M is a literal in the body of a lause in Defs 0 }
where the height of a literal is de ned as the length of the maximal path from
the root to a leaf of the literal onsidered as a tree;
( ) |vars(D)| ≤ max{vars(D′ ) | D′ is a lause in Defs 0 };
(d) no two lauses in Defs j an be made equal by one or more appli ations of the
following transformations: renaming of variables, renaming of head predi ates,
15
rearrangement of the order of the literals in the body, and deletion of dupli ate
literals.
Re all that bd(D) is equal to bd(E ′ ) where E ′ is derived by unfolding (a ording
to the Unfolding phase of the strategy) a lause E in TransfP 0 ∪ InDefs j and E
belongs to InDefs j .
Now Property (a) is a straightforward onsequen e of the de nition of the unfolding rule.
Property (b) an be shown as follows. E is of the form newp(X1 , . . . , Xk ) ← T1 ∧
. . .∧Tr ∧L1 ∧. . .∧Lm . By unfolding w.r.t. the type atoms T1 , . . . , Tr (a ording to
Point (i) of the Unfolding phase) we get lauses of the form newp(h1 , . . . , hk ) ←
′
T1′ ∧ . . . ∧ Tr1
∧ L′1 ∧ . . . ∧ L′m , where h1 , . . . , hk are head terms and, for all
i ∈ {1, . . . , m}, height (L′i ) ≤ height (Li ) + 1. By Lemma 2, TransfP 0 is a regular
natset-typed program and, therefore, by unfolding w.r.t. the literals L′1 , . . . , L′m
(a ording to Point (ii) of the Unfolding phase) we get lauses of the form
′
newp(h1 , . . . , hk ) ← T1′ ∧ . . . ∧ Tr1
∧ L′′1 ∧ . . . ∧ L′′m1 , where for all i ∈ {1, . . . , m1},
there exists i1 ∈ {1, . . . , m}, su h that height (L′′i ) = height (L′i1 )−1. Thus, Property (b) follows from the fa t that E ′ is derived by unfolding w.r.t. ground literals
′
from a lause of the form newp(h1 , . . . , hk ) ← T1′ ∧ . . . ∧ Tr1
∧ L′′1 ∧ . . . ∧ L′′m1 and
every unfolding w.r.t. a ground literal does not in rease the height of the other
literals in a lause.
Property ( ) follows from Lemma 2 and the fa t that by unfolding a lause E
using regular natset-typed lauses we get lauses E ′ where vars(E ′ ) ⊆ vars(E).
To see this, re all that in a regular natset-typed lause C every term has at
most one variable and vars(bd(C)) ⊆ vars(hd(C)) and, thus, by unfolding, a
variable is repla ed by a term with at most one variable and no new variables
are introdu ed.
Finally, Point (d) is a onsequen e of Point (i) of the De nition-Folding phase
✷
of the unfold/fold strategy.
6 De iding WS1S via the Unfold/Fold Proof Method
In this se tion we show that if we start from a losed WS1S formula ϕ, our
synthesis method an be used for he king whether or not N |= ϕ holds and,
thus, our synthesis method works also as a proof method whi h is a de ision
pro edure for losed WS1S formulas.
If ϕ is a losed WS1S formula then the predi ate f introdu ed when onstru ting the set Cls(f, ϕτ ), is a nullary predi ate. Let TransfP be the program
derived by the unfold/fold transformation strategy starting from the program
Nat Set ∪ Cls(f, ϕτ ). As already known from Point (2) of Theorem 4, we have
that Def ∗ (f, TransfP ) is either the empty set or the singleton {f ←}. Thus, we
an de ide whether or not N |= ϕ holds by he king whether or not f ← belongs
to TransfP . Sin e the unfold/fold transformation strategy always terminates, we
have that our unfold/fold synthesis method is indeed a de ision pro edure for
losed WS1S formulas. We summarize our proof method as follows.
16
The Unfold/Fold Proof Method.
Let ϕ be a losed WS1S formula.
Step 1. We apply the typed Lloyd-Topor transformation and we derive the set
Cls(f, ϕτ ) of lauses.
Step 2. We apply the unfold/fold transformation strategy and from the program
Nat Set ∪ Cls(f, ϕτ ) we derive a de nite program TransfP .
If the unit lause f ← belongs to TransfP then N |= ϕ else N |= ¬ϕ.
Now we present a simple example of appli ation of our unfold/fold proof
method.
Example 1. (An appli ation of the unfold/fold proof method.) Let us onsider
the losed WS1S formula ϕ : ∀X ∃Y X ≤ Y . By applying the typed Lloyd-Topor
transformation starting from the statement f ← ϕ, we get the following set of
lauses Cls(f, ϕτ ):
1. h(X) ← nat(X) ∧ nat (Y ) ∧ X ≤ Y
2. g ← nat(X) ∧ ¬h(X)
3. f ← ¬g
Now we apply the unfold/fold transformation strategy to the program Nat Set
and the following hierar hy of natset-typed de nitions: lause 1, lause 2, lause 3.
Initially, the program TransfP is NatSet . The transformation strategy pro eeds
left-to-right over that hierar hy.
(1) Defs and InDefs are both set to { lause 1}.
(1.1) Unfolding. By unfolding, from lause 1 we get:
4. h(0) ←
5. h(0) ← nat(Y )
6. h(s(X)) ← nat (X) ∧ nat(Y ) ∧ X ≤ Y
(1.2) De nition-Folding. In order to fold the body of lause 5 we introdu e the
following new lause:
7. new 1 ← nat (Y )
Clause 6 an be folded by using lause 1. By folding lauses 5 and 6 we get:
8. h(0) ← new 1
9. h(s(X)) ← h(X)
(1.3) At this point TransfP = Nat Set ∪ { lause 4, lause 8, lause 9}, Defs =
{ lause 1, lause 7}, and InDefs = { lause 7}.
(1.4) By rst unfolding lause 7 and then folding using lause 7 itself, we get:
10. new 1 ←
11. new 1 ← new 1
No new lause is introdu ed (i.e., NewDefs = ∅). At this point TransfP =
Nat Set ∪ { lause 4, lause 8, lause 9, lause 10, lause 11}, Defs = { lause 3,
lause 7}, and InDefs = ∅. Thus, the while statement terminates.
Sin e Def ∗ (new 1, TransfP) is propositional and M (TransfP ) |= new 1, by the
propositional simpli ation rule we have:
17
TransfP = Nat Set ∪ { lause 4, lause 8, lause 9, lause 10}.
(2) Defs is set to { lause 1, lause 2, lause 7} and InDefs is set to { lause 2}.
(2.1) Unfolding. By unfolding, from lause 2 we get:
12. g ← nat (X) ∧ ¬h(X)
(Noti e that, by unfolding, lause g ← ¬h(0) is deleted.)
(2.2) De nition-Folding. Clause 12 an be folded by using lause 2 whi h o urs
in Defs . Thus, no new lause is introdu ed (i.e., NewDefs = ∅) and by folding
we get:
13. g ← g
(2.3) At this point TransfP = NatSet ∪ { lause 4, lause 8, lause 9, lause 10,
lause 13}, Defs = { lause 1, lause 2, lause 7}, and InDefs = ∅. Thus, the
while statement terminates.
Sin e Def ∗ (g, TransfP ) is propositional and M (TransfP ) |= ¬g , by the propositional simpli ation rule we delete lause 13 from TransfP and we have:
TransfP = Nat Set ∪ { lause 4, lause 8, lause 9, lause 10}.
(3) Defs is set to { lause 1, lause 2, lause 3, lause 7} and InDefs is set to
{ lause 3}.
(3.1) Unfolding. By unfolding lause 3 we get:
14. f ←
(Re all that, there is no lause in TransfP with head g .)
(3.2) De nition-Folding. No transformation steps are performed on lause 14
be ause it is a unit lause.
(3.3) At this point TransfP = NatSet ∪ { lause 4, lause 8, lause 9, lause 10,
lause 14}, Defs = { lause 1, lause 2, lause 3, lause 7}, and InDefs = ∅.
The transformation strategy terminates and, sin e the nal program TransfP
in ludes the unit lause f ←, we have proved that N |= ∀X ∃Y X ≤ Y .
We would like to noti e that neither SLDNF nor Tabled Resolution (as implemented in the XSB system [?℄) are able to onstru t a refutation of Nat Set ∪
Cls(f, ϕτ ) ∪ {← f } (and thus onstru t a proof of ϕ), where ϕ is the WS1S
formula ∀X ∃Y X ≤ Y . Indeed, from the goal ← f we generate the goal ← ¬g ,
and neither SLDNF nor Tabled Resolution are able to infer that ← ¬g su eeds
✷
by dete ting that ← g generates an in nite set of failed derivations.
We would like to mention that some other transformations ould be applied
for enhan ing our unfold/fold transformation strategy. In parti ular, during the
strategy we may apply the subsumption rule to shorten the transformation proess by deleting some useless lauses. For instan e, in Example 1 we an delete
lause 5 whi h is subsumed by lause 4, thereby avoiding the introdu tion of the
new predi ate new1. In some other ases we an drop unne essary type atoms.
For instan e, in Example 1 in lause 1 the type atom nat (X) an be dropped
be ause it is implied by the atom X ≤ Y . The program derived at the end of the
exe ution of the while statement of the unfold/fold transformation strategy are
nondeterministi , in the sense that an atom with non-variable arguments may be
18
uni able with the head of several lauses. We an apply the te hnique for deriving deterministi program presented in [?℄ for deriving deterministi programs
and thus, obtaining smaller programs.
When the unfold/fold transformation strategy is used for program synthesis,
it is often the ase that the above mentioned transformations also improve the
e ien y of the derived programs.
Finally, we would like to noti e that the unfold/fold transformation strategy
an be applied starting from a program P ∪ Cls(f, ϕτ ) (instead of NatSet ∪
Cls(f, ϕτ )) where: (i) P is the output of a previous appli ation of the strategy,
and (ii) ϕ is a formula built like a WS1S formula, ex ept that it uses prediates o urring in P (besides ≤ and ∈). Thus, we an synthesize programs (or
onstru t proofs) in a ompositional way, by rst synthesizing programs for subformulas. We will follow this ompositional methodology in the example of the
following Se tion 7.
7 An Appli ation to the Veri ation of In nite State
Systems: the Dynami Bakery Proto ol
In this se tion we present an example of veri ation of a safety property of
an in nite state system by onsidering CLP(WS1S) programs [?℄. As already
mentioned, by applying our unfold/fold synthesis method we will then translate
CLP(WS1S) programs into logi programs.
The syntax of CLP(WS1S) programs is de ned as follows. We onsider a set of
user-de ned predi ate symbols. A CLP(WS1S) lause is of the form A ← ϕ ∧ G,
where A is an atom, ϕ is a formula of WS1S, G is a goal, and the predi ates
o urring in A or in G are all user-de ned. A CLP(WS1S) program is a set of
CLP(WS1S) lauses. We assume that CLP(WS1S) programs are strati ed.
Given a CLP(WS1S) program P , we de ne the semanti s of P to be its
perfe t model, denoted M (P ) (here we extend to CLP(WS1S) programs the
de nitions whi h are given for normal logi programs in [?℄).
Our example on erns the Dynami Bakery proto ol, alled DBakery for
short, and we prove that it ensures mutual ex lusion in a system of pro esses
whi h share a ommon resour e, even if the number of pro esses in the system
hanges during a proto ol run in a dynami way. The DBakery proto ol is a
variant of the N-pro ess Bakery proto ol [?℄.
In order to give the formal spe i ations of the DBakery proto ol and its
mutual ex lusion property, we will use CLP(WS1S) as we now indi ate. The
transition relation between pairs of system states, the initial system state, and
the system states whi h are unsafe (that is, the system states where more than
one pro ess uses the shared resour e) are spe i ed by WS1S formulas. However,
in order to spe ify the mutual ex lusion property we annot use WS1S formulas
only. Indeed, mutual ex lusion is a rea hability property whi h is unde idable
in the ase of in nite state systems. The approa h we follow in this example is
to spe ify rea hability (and, thus, mutual ex lusion) as a CLP(WS1S) program
(see the program PDBakery below).
19
Let us rst des ribe the DBakery proto ol. We assume that every pro ess is
asso iated with a natural number, alled a ounter, and two distin t pro esses
have distin t ounters. At ea h instant in time, the system of pro esses is represented by a pair hW, U i, alled a system state, where W is the set of the ounters
of the pro esses waiting for the resour e, and U is the set of the ounters of the
pro esses using the resour e.
A system state hW, U i is initial i W ∪ U is the empty set.
The transition relation from a system state hW, U i to a new system state
hW ′ , U ′ i is the union of the following three relations:
(T1: reation of a pro ess )
if W ∪ U is empty then hW ′ , U ′ i = h{0}, ∅i else hW ′ , U ′ i = hW ∪ {m+1}, U i,
where m is the maximum ounter in W ∪ U ,
(T2: use of the resour e )
if there exists a ounter n in W whi h is the minimum ounter in W ∪ U
then hW ′ , U ′ i = hW −{n}, U ∪ {n}i,
(T3: release of the resour e )
if there exists a ounter n in U then hW ′ , U ′ i = hW, U −{n}i.
The mutual ex lusion property holds i from the initial system state it is not
possible to rea h a system state hW, U i whi h is unsafe, that is, su h that U is
a set of at least two ounters.
Let us now give the formal spe i ation of the DBakery proto ol and its
mutual ex lusion property. We rst introdu e the following WS1S formulas (between parentheses we indi ate their meaning):
empty (X ) ≡ ¬∃x x ∈ X
(the set X is empty)
max (X,m ) ≡ m ∈ X ∧ ∀x (x ∈ X → x ≤ m)
(m is the maximum in the set X )
min (X,m ) ≡ m ∈ X ∧ ∀x (x ∈ X → m ≤ x)
(m is the minimum in the set X )
(Here and in what follows, for reasons of readability, we allow ourselves to use
lower ase letters for individual variables of WS1S formulas.)
A system state hW, U i is initial i N |= init (hW, U i), where:
init (hW, U i) ≡ empty(W ) ∧ empty(U )
The transition relation R between system states is de ned as follows:
hhW, U i , hW ′ , U ′ ii ∈ R i
N |= cre(hW, U i , hW ′ , U ′ i) ∨ use(hW, U i , hW ′ , U ′ i) ∨ rel (hW, U i , hW ′ , U ′ i)
where the predi ates re, use, and rel de ne the transition relations T1, T2, and
T3, respe tively. We have that:
20
cre(hW, U i , hW ′ , U ′ i) ≡ U ′ = U ∧ ∃Z (Z = W ∪ U ∧
((empty(Z) ∧ W ′ = {0}) ∨
(¬empty(Z) ∧ ∃m (max (Z, m) ∧ W ′ = W ∪{s(m)}))))
use(hW, U i , hW ′ , U ′ i) ≡ ∃n (n ∈ W ∧ ∃Z (Z = W ∪ U ∧ min(Z, n)) ∧
W ′ = W −{n} ∧ U ′ = U ∪{n})
rel (hW, U i , hW ′ , U ′ i) ≡ W ′ = W ∧ ∃n (n ∈ U ∧ U ′ = U −{n})
where the subformulas involving the set union (∪), set dieren e (−), and set
equality (=) operators an be expressed as WS1S formulas.
Mutual ex lusion holds in a system state hW, U i i N |= ¬unsafe(hW, U i),
where unsafe(hW, U i) ≡ ∃n1 ∃n2 (n1 ∈ U ∧ n2 ∈ U ∧ ¬(n1 = n2 )), i.e., a system
state hW, U i is unsafe i there exist at least two distin t ounters in U .
Now we will spe ify the system states rea hed from a given initial system state
by introdu ing the CLP(WS1S) program PDBakery onsisting of the following
lauses:
reach(S) ← init (S)
reach(S1) ← cre(S, S1) ∧ reach(S)
reach(S1) ← use(S, S1) ∧ reach(S)
reach(S1) ← rel (S, S1) ∧ reach(S)
where init (S), cre(S, S1), use(S, S1), and rel (S, S1) are the WS1S formulas
listed above.
′
by repla ing the WS1S
From PDBakery we derive a de nite program PDBakery
formulas o urring in PDBakery by the orresponding atoms init (S), cre(S, S1),
use(S, S1), and rel (S, S1), and by adding to the program the lauses (not listed
here) de ning these atoms, whi h are derived from the orresponding WS1S formulas listed above, by applying the unfold/fold synthesis method (see Se tion 5).
Let us all these lauses Init, Cre, Use, and Rel, respe tively.
In order to verify that the DBakery proto ol ensures mutual ex lusion for
every system of pro esses whose number dynami ally hanges over time, we
have to prove that for every ground term s denoting a nite set of ounters,
′
ur (s) 6∈ M (PDBakery
∪ {clause 1}), where lause 1 is the following lause whi h
we introdu e by the de nition rule:
1. ur(S) ← unsafe(S) ∧ reach(S)
and unsafe(S) is de ned by a set, alled Unsafe, of lauses whi h are derived from
the orresponding WS1S formula by using the unfold/fold synthesis method.
In order to verify the mutual ex lusion property for the DBakery proto ol
′
it is enough to show that PDBakery
∪ {clause 1} an be transformed into a new
de nite program without lauses for ur(S). This transformation an be done,
as we now illustrate, by a straightforward adaptation of the proof te hnique
presented for Constraint Logi Programs in [?℄. In parti ular, before performing
folding steps, we will add suitable atoms in the bodies of the lauses to be folded.
We start o this veri ation by unfolding lause 1 w.r.t. the atom rea h. We
obtain the following lauses:
2. ur(S) ← unsafe(S) ∧ init (S)
21
3. ur(S1) ← unsafe(S1) ∧ cre(S, S1) ∧ reach(S)
4. ur(S1) ← unsafe(S1) ∧ use(S, S1) ∧ reach(S)
5. ur(S1) ← unsafe(S1) ∧ rel (S, S1) ∧ reach(S)
Now we an remove lause 2 be ause
M (Unsafe ∪ Init ) |= ¬∃S (unsafe(S) ∧ init (S)).
The proof of this fa ts and the proofs of the other fa ts we state below, are
performed by applying the unfold/fold proof method of Se tion 5. Then, we fold
lauses 3 and 5 by using the de nition lause 1 and we obtain:
6. ur(S1) ← unsafe(S1) ∧ cre(S, S1) ∧ ur (S)
7. ur(S1) ← unsafe(S1) ∧ rel (S, S1) ∧ ur(S)
Noti e that this appli ation of the folding rule is justi ed by the following two
fa ts:
M (Unsafe ∪ Cre) |= ∀S ∀S1 (unsafe(S1) ∧ cre(S, S1) → unsafe(S))
M (Unsafe ∪ Rel ) |= ∀S ∀S1 (unsafe(S1) ∧ rel (S, S1) → unsafe(S))
so that, before folding, we an add the atom unsafe(S) to the bodies of lauses
3 and 5. Now, sin e M (Unsafe ∪ Use) |= ¬∀S ∀S1 (unsafe(S1) ∧ use(S, S1) →
unsafe(S)), lause 4 annot be folded using the de nition lause 1. Thus, we
introdu e the new de nition lause:
8. p 1(S) ← c(S) ∧ reach(S)
where c(hW, U i) ≡ ∃n (n ∈ W ∧ ∃Z (Z = W∪U ∧ min(Z, n))) ∧ ¬empty(U ) whi h
means that: in the system state hW, U i there is at least one pro ess whi h uses
the resour e and there exists a pro ess waiting for the resour e with ounter n
whi h is the minimum ounter in W ∪ U .
Noti e that, by applying the unfold/fold synthesis method, we may derive a
set, alled Busy (not listed here), of de nite lauses whi h de ne c(S).
By using lause 8 we fold lause 4, and we obtain:
9. ur(S1) ← unsafe(S1) ∧ use(S, S1) ∧ p 1(S)
We pro eed by applying the unfolding rule to the newly introdu ed lause 8,
thereby obtaining:
10. p 1(S) ← c(S) ∧ init (S)
11. p 1(S1) ← c(S1) ∧ cre(S, S1) ∧ reach(S)
12. p 1(S1) ← c(S1) ∧ use(S, S1) ∧ reach(S)
13. p 1(S1) ← c(S1) ∧ rel (S, S1) ∧ reach(S)
Clauses 10 and 12 are removed, be ause
M (Busy ∪ Init ) |= ¬∃S (c(S) ∧ init (S))
M (Busy ∪ Use) |= ¬∃S ∃S1 (c(S1) ∧ use(S, S1))
We fold lauses 11 and 13 by using the de nition lauses 8 and 1, respe tively,
thereby obtaining:
14. p 1(S1) ← c(S1) ∧ cre(S, S1) ∧ p1(S)
15. p 1(S1) ← c(S1) ∧ rel(S, S1) ∧ ur(S)
Noti e that this appli ation of the folding rule is justi ed by the following two
fa ts:
22
M (Busy ∪ Cre) |= ∀S ∀S1 ((c(S1) ∧ cre(S, S1)) → c(S))
M (Busy ∪ Rel ) |= ∀S ∀S1 ((c(S1) ∧ rel (S, S1)) → unsafe(S))
′
Thus, starting from program PDBakery
∪{ lause 1} we have derived a new program Q onsisting of lauses 6, 7, 14, and 15. Sin e all lauses in Def ∗ (ur , Q)
are re ursive, we have that for every ground term s denoting a nite set of ounters, ur (s) 6∈ M (Q) and by the orre tness of the transformation rules [?℄, we
on lude that mutual ex lusion holds for the DBakery proto ol.
8 Related Work and Con lusions
We have proposed an automati synthesis method based on unfold/fold program transformations for translating CLP(WS1S) programs into normal logi
programs. This method an be used for avoiding the use of ad-ho solvers for
WS1S onstraints when onstru ting proofs of properties of in nite state multipro ess systems.
Our synthesis method follows the general approa h presented in [?℄ and it
terminates for any given WS1S formula. No su h termination result was given
in [?℄. In this paper we have also shown that, when we start from a losed WS1S
formula ϕ, our synthesis strategy produ es a program whi h is either (i) a unit
lause of the form f ←, where f is a nullary predi ate equivalent to the formula
ϕ, or (ii) the empty program. Sin e in ase (i) ϕ is true and in ase (ii) ϕ is false,
our strategy is also a de ision pro edure for losed WS1S formulas. This result
extends [?℄ whi h presents a de ision pro edure based on the unfold/fold proof
method for the lausal fragment of the WSkS theory, i.e., the fragment dealing
with universally quanti ed disjun tions of onjun tions of literals.
Some related methods based on program transformation have been re ently
proposed for the veri ation of in nite state systems [?,?℄. However, as it is
shown by the example of Se tion 7, an important feature of our veri ation
method is that the number of pro esses involved in the proto ol may hange over
time and other methods nd it problemati to deal with su h dynami hanges.
In parti ular, the te hniques presented in [?℄ for verifying safety properties of
parametrized systems deal with rea tive systems where the number of pro esses
is a parameter whi h does not hange over time.
Our method is also related to a number of other methods whi h use logi
programming and, more generally, onstraint logi programming for the veri ation of rea tive systems (see, for instan e, [?,?,?,?℄ and [?℄ for a survey). The
main novelty of our approa h w.r.t. these methods is that it ombines logi programming and monadi se ond order logi , thereby modelling in a very dire t
way systems with an unbounded (and possibly variable) number of pro esses.
Our unfold/fold synthesis method and our unfold/fold proof method have
been implemented by using the MAP transformation system [?℄. Our implementation is reasonably e ient for WS1S formulas of small size (see the example
formulas of Se tion 7). However, our main on ern in the implementation was not
e ien y and our system should not be ompared with ad-ho , well-established
theorem provers for WS1S formulas based on automata theory, like the MONA
23
system [?℄. Nevertheless, we believe that our te hnique has its novelty and deserves to be developed be ause, being based on unfold/fold rules, it an easily be
ombined with other te hniques for program derivation, spe ialization, synthesis,
and veri ation, whi h are also based on unfold/fold transformations.
24
| 6cs.PL
|
Holographic Algorithms Beyond Matchgates✩
Jin-Yi Caia , Heng Guob,∗, Tyson Williamsc
a University
arXiv:1307.7430v2 [cs.DS] 10 Jan 2018
b University
of Wisconsin - Madison, 1210 West Dayton Street, Madison, WI 53706, US
of Edinburgh, Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, UK
c Blocher Consulting, 206 North Randolph, Champaign, IL 61820, US
Abstract
Holographic algorithms introduced by Valiant are composed of two ingredients: matchgates, which are
gadgets realizing local constraint functions by weighted planar perfect matchings, and holographic reductions,
which show equivalences among problems with different descriptions via certain basis transformations. In
this paper, we replace matchgates in the paradigm above by the affine type and the product type constraint
functions, which are known to be tractable in general (not necessarily planar) graphs. More specifically,
we present polynomial-time algorithms to decide if a given counting problem has a holographic reduction
to another problem defined by the affine or product-type functions. Our algorithms also find a holographic
transformation when one exists. We further present polynomial-time algorithms of the same decision and
search problems for symmetric functions, where the complexity is measured in terms of the (exponentially
more) succinct representations. The algorithm for the symmetric case also shows that the recent dichotomy
theorem for Holant problems with symmetric constraints is efficiently decidable. Our proof techniques are
mainly algebraic, e.g., using stabilizers and orbits of group actions.
Keywords: Counting complexity, holographic algorithms
1. Introduction
Recently a number of complexity dichotomy theorems have been obtained for counting problems. Typically, such dichotomy theorems assert that a vast majority of problems expressible within certain frameworks
are #P-hard, however an intricate subset manages to escape this fate. These exceptions exhibit some rich
mathematical structure, leading to polynomial-time algorithms. Holographic reductions and algorithms, introduced by Valiant [45], play key roles in many recent dichotomy theorems [14, 25, 20, 15, 35, 22, 12, 33].
Indeed, many interesting tractable cases are solvable using holographic reductions. This fascinating fact
urges us to explore the full reach of holographic algorithms.
Valiant’s holographic algorithms [45, 44] have two main ingredients. The first is to encode computation
in planar graphs via gadget construction, called matchgates [43, 42, 9, 17, 10]. The result of the computation
is then obtained by counting the number of perfect matchings in a related planar graph, which can be done
in polynomial time by Kasteleyn’s (a.k.a. the FKT) algorithm [36, 41, 37]. The second one is the notion
of holographic transformations/reductions, which show equivalences of problems with different descriptions
via basis transformations. Thus, in order to apply the holographic algorithm, one must find a suitable
holographic transformation along with matchgates realizing the desired constraint functions. This procedure
has been made algorithmic [9, 17].
In this paper, we replace matchgates in the paradigm above by the affine type or the product type
constraint functions, both of which are known to be tractable over general (i.e. not necessarily planar)
graphs [24]. We present polynomial-time algorithms to decide if a given counting problem has a holographic
✩A
preliminary version has appeared in ICALP 2014 [11].
author
Email addresses: [email protected] (Jin-Yi Cai), [email protected] (Heng Guo), [email protected] (Tyson Williams)
∗ Corresponding
Preprint submitted to Elsevier
January 11, 2018
reduction to another problem defined by affine or product-type functions. Our algorithm also finds a holographic reduction when one exists. Although, conceptually, we do not add new tractable cases, the task of
finding these transformations is often non-trivial. For example, generalized Fibonacci gates [23] are the same
as the product-type via transformations, but at first glance, the former look much more complicated than
the latter.
To formally state the results, we briefly introduce some notation. The counting problems we consider are
those expressible as a Holant problem [23, 21, 19, 24]. A Holant problem is defined by a set F of constraint
functions, which we call signatures, and is denoted by Holant(F ). An instance of Holant(F ) is a tuple
Ω = (G, F , π), called a signature grid, where G = (V, E) is a graph and π labels each vertex v ∈ V and its
incident edges with some fv ∈ F and its input variables. Here fv maps {0, 1}deg(v) to C, where deg(v) is
the degree Q
of v. We consider all possible 0-1 edge assignments. An assignment σ to the edges E gives an
evaluation v∈V fv (σ|E(v) ), where E(v) denotes the incident edges of v and σ|E(v) denotes the restriction
of σ to E(v). The counting problem on the instance Ω is to compute
X
Y
HolantΩ =
fv σ|E(v) .
σ:E→{0,1} v∈V
For example, consider the problem of counting Perfect Matching on G. This problem corresponds to
attaching the Exact-One function at every vertex of G. The Exact-One function is an example of a
symmetric signature, which are functions that only depend on the Hamming weight of the input. We denote
a symmetric signature by f = [f0 , f1 , . . . , fn ] where fw is the value of f on inputs of Hamming weight w.
For example, [0, 1, 0, 0] is the Exact-One function on three bits. The output is 1 if and only if the input is
001, 010, or 100, and the output is 0 otherwise.
Holant problems contain both counting constraint satisfaction problems and counting graph homomorphisms as special cases. All three classes of problems have received considerable attention, which has resulted
in a number of dichotomy theorems (see [39, 34, 29, 2, 28, 4, 27, 1, 5, 31, 32, 8, 13, 14, 6, 30, 3, 7, 24]).
Despite the success with #CSP and graph homomorphisms, the case with Holant problems is more difficult.
Recently, a dichotomy theorem for Holant problems with symmetric signatures was obtained [12], but the
general (i.e. not necessarily symmetric) case has a richer and more intricate structure. The same dichotomy
for general signatures remains open.
Our first main result is an efficient procedure to decide whether a given Holant problem can be solved by
affine or product-type signatures via holographic transformations. In past classification efforts, we have been
in the same situation several times, where one concrete problem determines the complexity of a wide range
of problems. However, the brute force way to check whether this concrete problem already belongs to known
tractable classes is time-consuming. We hope that the efficient decision procedure given here mitigates this
issue, and would help the pursuit towards a general Holant dichotomy.
Theorem 1.1. There is a polynomial-time algorithm to decide, given a finite set of signatures F , whether
Holant(F ) admits a holographic algorithm based on affine or product-type signatures.
The holographic algorithms for Holant(F ) are all polynomial time in the size of the problem input Ω.
The polynomial time decision algorithm of Theorem 1.1 is on another level; it decides based on any specific
set of signatures F whether the counting problem Holant(F ) defined by F has such a holographic algorithm.
Symmetric signatures are an important special case. Because symmetric signatures can be presented
exponentially more succinctly, we would like the decision algorithm to be efficient when measured in terms
of this succinct description. An algorithm for this case needs to be exponentially faster than the one in
Theorem 1.1. In Theorem 1.2, we present a polynomial time algorithm for the case of symmetric signatures.
The increased efficiency is based on several signature invariants under orthogonal transformations.
Theorem 1.2. There is a polynomial-time algorithm to decide, given a finite set of symmetric signatures
F expressed in the succinct notation, whether Holant(F ) admits a holographic algorithm based on affine or
product-type signatures.
2
A dichotomy theorem classifies every set of signatures as defining either a tractable problem or an
intractable problem (e.g. #P-hard). Yet it would be more useful if given a specific set of signatures, one
could decide to which case it belongs. This is the decidability problem of a dichotomy theorem. In [12], a
dichotomy regarding symmetric complex-weighted signatures for Holant problems was proved. However, the
decidability problem was left open. Of the five tractable cases in the dichotomy theorem, three of them are
easy, but the remaining two cases are more challenging, which are (1) holographic algorithms using affine
signatures and (2) holographic algorithms using product-type signatures. As a consequence of Theorem 1.2,
this decidability is now proved.
Corollary 1.3. The dichotomy theorem for symmetric complex-weighted Holant problems in [12] is decidable
in polynomial time.
Previous work on holographic algorithms focused almost exclusively on those with matchgates [45, 44,
16, 25, 17, 18, 33]. (This has led to a misconception in the community that holographic algorithms are
always based on matchgates.) The first example of a holographic algorithm using something other than
matchgates came in [23]. These holographic algorithms use generalized Fibonacci gates. A symmetric
signature f = [f0 , f1 , . . . , fn ] is a generalized Fibonacci gate of type λ ∈ C if fk+2 = λfk+1 + fk holds
for all k ∈ {0, 1, . . . , n − 2}. The standard Fibonacci gates are of type λ = 1, in which case, the entries
of the signature satisfy the recurrence relation of the Fibonacci numbers. The generalized Fibonacci gates
were immediately put to use in a dichotomy theorem [21]. As it turned out, for nearly all values of λ,
the generalized Fibonacci gates are equivalent to product-type signatures via holographic transformations.
Our results provide a systematic way to determine such equivalences and we hope these results help in
determining the full reach of holographic algorithms.
The constraint functions we call signatures are essentially tensors. A group of transformations acting upon
these tensors yields an orbit. Previously, in [12], we have shown that it is sufficient to restrict holographic
transformations to those from or related to the orthogonal group (see Lemma 2.7 and Lemma 2.10). Thus, our
question can be rephrased as the following: given a tensor, determine whether its orbit under the orthogonal
group action (or related transformations) intersects the set of affine or product-type tensors. As showed by
Theorems 1.1 and 1.2, this can be done efficiently, even for a set rather than a single tensor. In contrast,
this orbit intersection problem with the general linear group acting on two arbitrary tensors is NP-hard [38].
In our setting, the actions are much more restricted and we consider an arbitrary tensor against one of the
two fixed sets. Similar orbit problems are central in geometric complexity theory [40].
Our techniques are mainly algebraic. A particularly useful insight is that
an orthogonal transformation
1
in the standard basis is equivalent to a diagonal transformation in the 1i −i
basis. Since diagonal transformations are much easier to understand, this gives us some leverage to understand orbits under orthogonal
transformations. Also, the groups of transformations that stabilize the affine and product-type signatures
play important roles in our proofs. Comparing to similar results for matchgates [17], the proofs are very
different in that each proof relies heavily on distinct properties of matchgates or the affine and product-type
signatures.
In Section 2, we review basic notation and state previous results, many of which come from [12]. In
Section 3, we present some example problems that are tractable by holographic algorithms using affine or
product-type signatures. The proof of Theorem 1.1 spans two sections. The affine case is handled in Section 4
and the product-type case is handled in Section 5. The proof of Theorem 1.2 also spans two sections. Once
again, the affine case is handled in Section 6 and the product-type case is handled in Section 7.
2. Preliminaries
2.1. Problems and Definitions
The framework of Holant problems is defined for functions mapping [q]k to F for a finite q and some
field F. In this paper, we investigate some of the tractable complex-weighted Boolean Holant problems, that
is, all functions are of the type [2]k → C. Strictly speaking, for consideration of models of computation,
functions take complex algebraic numbers.
3
A signature grid Ω = (G, F , π) consists of a graph G = (V, E) and a set of constraint functions (also
called signatures) F , where π labels each vertex v ∈ V and its incident edges with some fv ∈ F and its input
variables. Note that in particular, π specifies an ordering
P Q of edges/variables on each vertex. The Holant
problem on instance Ω is to evaluate HolantΩ =
σ
v∈V fv (σ |E(v) ), a sum over all edge assignments
σ : E → {0, 1}.
A function fv can be represented by listing its values in lexicographical order as in a truth table, which is
deg(v)
a vector in C2
. Equivalently, fv can be treated as a tensor in (C2 )⊗ deg(v) . We also use fx to denote the
value f (x), where x is a binary string. A function f ∈ F is also called a signature. A symmetric signature f
on k Boolean variables can be expressed as [f0 , f1 , . . . , fk ], where fw is the value of f on inputs of Hamming
weight w.
A Holant problem is parametrized by a set of signatures.
Definition 2.1. Given a set of signatures F , we define the counting problem Holant(F ) as:
Input: A signature grid Ω = (G, F , π);
Output: HolantΩ .
A signature f of arity n is degenerate if there exist unary signatures uj ∈ C2 (1 ≤ j ≤ n) such that
f = u1 ⊗ · · · ⊗ un . In a signature grid, it is equivalent to replace a degenerate one by corresponding unary
signatures. A symmetric degenerate signature has the form u⊗n , where the superscript denotes the tensor
power. Replacing a signature f ∈ F by a constant multiple cf , where c 6= 0, does not change the complexity
of Holant(F ). It introduces a global factor to HolantΩ .
We say a signature set F is tractable (resp. #P-hard) if the corresponding counting problem Holant(F )
can be solved in polynomial time (resp. #P-hard). Similarly for a signature f , we say f is tractable (resp. #Phard) if {f } is.
2.2. Holographic Reduction
To introduce the idea of holographic reductions, it is convenient to consider bipartite graphs. We can
always transform a general graph into a bipartite graph while preserving the Holant value, as follows. For
each edge in the graph, we replace it by a path of length two. (This operation is called the 2-stretch of
the graph and yields the edge-vertex incidence graph.) Each new vertex is assigned the binary Equality
signature (=2 ) = [1, 0, 1].
We use Holant (F | G) to denote the Holant problem on bipartite graphs H = (U, V, E), where each
vertex in U or V is assigned a signature in F or G, respectively. An input instance for this bipartite Holant
problem is a bipartite signature grid and is denoted by Ω = (H, F | G, π). Signatures in F are considered
as row vectors (or covariant tensors); signatures in G are considered as column vectors (or contravariant
tensors) [26].
For a 2-by-2 matrix T and a signature set F , define T F = {g | ∃f ∈ F of arity n, g = T ⊗n f }, similarly
for F T . Whenever we write T ⊗n f or T F , we view the signatures as column vectors; similarly for f T ⊗n or
F T as row vectors.
Let T be an element of GL2 (C), the group of invertible 2-by-2 complex matrices. The holographic
transformation defined by T is the following operation: given a signature grid Ω = (H, F | G, π), for the
same graph H, we get a new grid Ω′ = (H, F T | T −1 G, π ′ ) by replacing f ∈ F (or g ∈ G) with T ⊗n f (or
⊗n
T −1
g).
Theorem 2.2 (Valiant’s Holant Theorem [45]). If there is a holographic transformation mapping signature
grid Ω to Ω′ , then HolantΩ = HolantΩ′ .
Therefore, an invertible holographic transformation does not change the complexity of the Holant problem
in the bipartite setting. Furthermore, there is a particular kind of holographic transformation, the orthogonal
transformation, that preserves binary equality and thus can be used freely in the standard setting. Let O2 (C)
be the group of 2-by-2 complex matrices that are orthogonal. Recall that a matrix T is orthogonal if T T T = I.
Theorem 2.3 (Theorem 2.6 in [19]). Suppose T ∈ O2 (C) and let Ω = (H, F , π) be a signature grid. Under
a holographic transformation by T , we get a new grid Ω′ = (H, T F , π ′ ) and HolantΩ = HolantΩ′ .
4
We also use SO2 (C) to denote the group of special orthogonal matrices, i.e. the subgroup of O2 (C) with
determinant 1.
2.3. Tractable Signature Sets without a Holographic Transformation
The following two signature sets are tractable without a holographic transformation [24].
Definition 2.4. A k-ary function f (x1 , . . . , xk ) is affine if it has the form
λ · χAx=0 · i
Pn
j=1 hvj ,xi
,
where λ 6= 0 is in C, x = (x1 , x2 , . . . , xk , 1)T , A is a matrix over F2 , vj is a vector over F2 for each
j = 1, . . . , n, and χ is a 0-1 indicator function such
Pnthat χAx=0 is 1 iff Ax = 0.√Note that the dot product
hvj , xi is calculated over F2 , while the summation j=1 on the exponent of i = −1 is evaluated as a sum
mod 4 of 0-1 terms. We use A to denote the set of all affine functions.
Notice that there is no restriction on the number of rows in the matrix A. It is permissible that A is
the zero matrix so that χAx=0 = 1 holds for all x. An equivalent way to express the exponent of i is as
a quadratic polynomial (evaluated mod 4) where all cross terms have an even coefficient. This equivalent
expression is often easier to use.
Definition 2.5. A function is of product type if it can be expressed as a function product of unary functions,
binary equality functions ([1, 0, 1]), and binary disequality functions ([0, 1, 0]). We use P to denote the set
of product-type functions.
The above two types of functions, when restricted to be symmetric, have been characterized explicitly. It
has been shown (cf. Lemma 2.2 in [35]) that if f is a symmetric signature in P, then f is either degenerate,
binary disequality, or of the form [a, 0, . . . , 0, b] for some a, b ∈ C. It is also known that (cf. [19]) the set of
non-degenerate symmetric signatures in A is precisely the nonzero signatures (λ 6= 0) in F1 ∪ F2 ∪ F3 1
with arity at least 2, where F1 , F2 , and F3 are three families of signatures defined as
o
n
⊗k
⊗k
| λ ∈ C, k = 1, 2, . . . , r = 0, 1, 2, 3 ,
F1 = λ [ 10 ] + ir [ 01 ]
n
o
1 ⊗k
⊗k
F2 = λ [ 11 ] + ir −1
| λ ∈ C, k = 1, 2, . . . , r = 0, 1, 2, 3 , and
o
n
1 ⊗k
⊗k
| λ ∈ C, k = 1, 2, . . . , r = 0, 1, 2, 3 .
F3 = λ [ 1i ] + ir −i
Let F123 = F1 ∪ F2 ∪ F3 be the union of these three sets of signatures. We explicitly list all the signatures
in F123 (as row vectors) up to an arbitrary constant multiple from C:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
[1, 0, . . . , 0, ±1];
[1, 0, . . . , 0, ±i];
[1, 0, 1, 0, . . . , 0 or 1];
[1, −i, 1, −i, . . . , (−i) or 1];
[0, 1, 0, 1, . . . , 0 or 1];
[1, i, 1, i, . . . , i or 1];
[1, 0, −1, 0, 1, 0, −1, 0, . . ., 0 or 1 or (−1)];
[1, 1, −1, −1, 1, 1, −1, −1, . . ., 1 or (−1)];
[0, 1, 0, −1, 0, 1, 0, −1, . . ., 0 or 1 or (−1)];
[1, −1, −1, 1, 1, −1, −1, 1, . . ., 1 or (−1)].
(F1 , r = 0, 2)
(F1 , r = 1, 3)
(F2 , r = 0)
(F2 , r = 1)
(F2 , r = 2)
(F2 , r = 3)
(F3 , r = 0)
(F3 , r = 1)
(F3 , r = 2)
(F3 , r = 3)
1 To be consistent with previous papers, we still use F , F , and F to denote the subclasses of A . They are not to be
1
2
3
confused with A1 , A2 , and A3 that will be introduced in Definition 2.8.
5
2.4. A -transformable and P-transformable Signatures
The tractable sets A and P are still tractable under a suitable holographic transformation. This is
captured by the following definition.
Definition 2.6. A set F of signatures is A -transformable (resp. P-transformable) if there exists a holographic transformation T such that F ⊆ T A (resp. F ⊆ T P) and [1, 0, 1]T ⊗2 ∈ A (resp. [1, 0, 1]T ⊗2 ∈ P).
To refine the above definition, we consider the stabilizer group of A ,
Stab(A ) = {T ∈ GL2 (C) | T A = A }.
Technically what we defined is the left stabilizer group of A , but it turns out that the left and right stabilizer
groups of A coincide [12].
Name
α
D
H2
X
Z
Value
√
πi
√
i = e 4 = 1+i
2
1
0
[0 i ]
√1
2
1 1
1 −1
[ 01 10 ]
1 1
√1
2 i −i
Table 1: Notations for some matrices and numbers
Some matrices and numbers are used extensively
summarize them in Table
1 1 throughout the paper. We
−1 2
1. Note that Z = DH2 and that D2 Z = √12 −i
=
ZX,
hence
X
=
Z
D
Z. It is easy to verify
i
that D, H2 , X, Z ∈ Stab(A ). In fact, Stab(A ) is precisely the set of nonzero scalar multiples of the group
generated by D and H2 [12]. Note that the zero matrix is not a stabilizer since A does not include the zero
function.
The next lemma is the first step toward understanding A -transformable signatures. Recall that O2 (C) is
the group of 2-by-2 orthogonal complex matrices. The lemma shows that to determine A -transformability,
it is necessary and sufficient to consider only the orthogonal transformations and related ones.
Lemma 2.7 ([12]). Let F be a set of signatures. Then F is A -transformable iff there exists an H ∈ O2 (C)
such that F ⊆ HA or F ⊆ H [ 10 α0 ] A .
Non-degenerate symmetric A -transformable signatures are captured by three sets A1 , A2 , and A3 , which
will be defined next (not to be confused with F1 , F2 , and F3 ).
Definition 2.8. A symmetric signature f of arity n is in, respectively, A1 , or A2 , or A3 if there exists an
H ∈ O2 (C) and a nonzero constant
c ∈ C such that f has the following form, respectively:
1 ⊗n
⊗n
, where β = αtn+2r , r ∈ {0, 1, 2, 3}, and t ∈ {0, 1};
• cH ⊗n [ 11 ] + β −1
1 ⊗n
⊗n
;
• or cH ⊗n [ 1i ] + −i
⊗n
⊗n
1
• or cH ⊗n [ α1 ] + ir −α
, where r ∈ {0, 1, 2, 3}.
For i ∈ {1, 2, 3}, when such an orthogonal H exists, we say that f ∈ Ai with transformation H. If f ∈ Ai
with I2 , the identity matrix, then we say f is in the canonical form of Ai . Note that there is no direct
correspondences between (Ai ) and (Fi ).
Lemma 2.9 ([12]). Let f be a non-degenerate symmetric signature. Then f is A -transformable iff f ∈
A1 ∪ A2 ∪ A3 .
6
Analogous results hold for P-transformable signatures. Let the stabilizer group of P be
Stab(P) = {T ∈ GL2 (C) | T P = P}.
The group Stab(P) is generated by (up to nonzero scalars) matrices of the form [ 10 ν0 ] for any ν ∈ C∗ and
X = [ 01 10 ] [12].
Lemma 2.10 ([12]). Let F be a set of signatures. Then F is P-transformable iff there exists an H ∈ O2 (C)
1
such that F ⊆ HP or F ⊆ H 1i −i
P.
Definition 2.11. A symmetric
signature f ofarity n is in P1 if there exist an H ∈ O2 (C) and a nonzero
1 ⊗n
⊗n
c ∈ C such that f = cH ⊗n [ 11 ] + β −1
, where β 6= 0.
It is easy to check that A1 ⊂ P1 . We define P2 = A2 . For i ∈ {1, 2}, when H ∈ O2 (C) exists (in
Definition 2.11 and 2.8, respectively), we say that f ∈ Pi with transformation H. If f ∈ Pi with I2 , then
we say f is in the canonical form of Pi .
Lemma 2.12 ([12]). Let f be a non-degenerate symmetric signature. Then f is P-transformable iff f ∈
P1 ∪ P2 .
3. Some Example Problems
In this section, we illustrate a few problems that are tractable via holographic reductions to affine or
product-type functions. Although the algorithms to solve them follow from a known paradigm, it is often
non-trivial to find the correct holographic transformation. Our main result provides a systematic way to
search for these transformations.
3.1. A Fibonacci-like Problem
Fibonacci gates were introduced in [23]. They define tractable counting problems, and holographic algorithms based on Fibonacci gates work over general (i.e. not necessarily planar) graphs. However, Fibonacci
gates are symmetric by definition. An example of a Fibonacci gate is the signature f = [f0 , f1 , f2 , f3 ] =
[1, 0, 1, 1]. Its entries satisfy the recurrence relation of the Fibonacci numbers, i.e. f2 = f1 +f0 and f3 = f2 +f1 .
For Holant(f ), the input is a 3-regular graph, and the problem is to count spanning subgraphs such that no
vertex has degree 1.
A symmetric signature g = [g0 , g1 , . . . , gn ] is a generalized Fibonacci gate of type λ ∈ C if gk+2 =
λgk+1 + gk holds for all k ∈ {0, 1, . . . , n − 2}. The standard Fibonacci gates are of type λ = 1. An example of
a generalized Fibonacci gate is g = [3, 1, 3, 1], which has type λ = 0. In contrast to Holant(f ), the problem
Holant(g) permits all possible spanning subgraphs. The output is the sum of the weights of each spanning
subgraph. The weight of a spanning subgraph S is 3k(S) , where k(S) is the number of vertices of even degree
in S. Since g = [3, 1, 3, 1] is Fibonacci, the problem Holant(g) is computable in polynomial time [19, 12].
One new family of holographic algorithms in this paper extends Fibonacci gates to asymmetric signatures.
In full notation, the ternary signature g is (3, 1, 1, 3, 1, 3, 3, 1)T. Consider the asymmetric signature h =
(3, 1, −1, −3, −1, −3, 3, 1)T. This signature h differs from g by a negative sign in four entries. Although h is
not a generalized Fibonacci gate or even a symmetric signature,
1 it still defines a tractable Holant problem.
Under a holographic transformation by Z −1 , where Z = √12 1i −i
,
Holant(h) = Holant (=2 | h) = Holant =2 (Z −1 )⊗2 | Z ⊗3 h = Holant [1, 0, −1] | ĥ ,
√
where
√ ĥ = 2i 2(0, 1, 0, 0, 0, 0, 2i, 0). Both [1, 0, −1](x1 , x2 ) = Equality(x1 , x2 )·[1, −1](x1 ) and ĥ(x1 , x2 , x3 ) =
2i 2 · Equality(x1 , x2 ) · Disequality(x2 , x3 ) · [1, 2i](x1 ) are product-type signatures.
It turns out that for all values of λ 6= ±2i, the generalized Fibonacci gates of type λ are P-transformable.
The value of λ indicates under which holographic transformation the signatures become product type. For
λ = ±2i, the generalized Fibonacci gates of type λ are vanishing, which means the output is always zero for
every possible input (see [12] for more on vanishing signatures).
7
0
1
1
1
0
0
0
1
1
1
0
0
1
0
(a) An admissible assignment to this graph fragment.
The circle vertices are assigned ĝ and the square vertices are assigned 6=2 .
(b) The orientation induced by the assignment in (a).
Figure 1: A fragment of an instance to Holant (6=2 | ĝ), which must be a (2, 4)-regular bipartite graph. Note the
saddle orientation of the edges incident to the two vertices with all four edges depicted.
3.2. Some Cycle Cover Problems and Orientation Problems
To express some problems involving asymmetric signatures of arity 4, it is convenient to arrange the 16
outputs into a 4-by-4 matrix. "With a slight abuse#of notation, we also write a function f (x1 , x2 , x3 , x4 ) in
its matrix form, namely f =
f0000
f0100
f1000
f1100
f0010
f0110
f1010
f1110
f0001
f0101
f1001
f1101
f0011
f0111
f1011
f1111
, where the row is indexed by two bits (x1 , x2 ) and the
column is indexed by two bits (x4 , x3 ) in reverse order. We call this the signature matrix.
Consider the problem of counting the number of cycle covers in a given graph. This problem is #Phard even when restricted to planar 4-regular graphs [33]. As a Holant problem, its expression
is Holant(f
),
where f (x1 , x2 , x3 , x4 ) is the symmetric signature [0, 0, 1, 0, 0]. The signature matrix of f is
0
0
0
1
0
1
1
0
0
1
1
0
1
0
0
0
. The six
entries in the support of f , which are all of Hamming weight two (indicating that a cycle cover passes through
each vertex exactly twice), can be divided into two parts, namely {0011, 0110, 1100, 1001} and {0101, 1010}.
In the planar setting, this corresponds to a pairing of consecutive or non-consecutive incident edges. Both
sets are invariant under cyclic permutations.
Suppose we removed the inputs 0101 and 1010 from the support of f , which are the two
1’s on the antidiagonal in the middle of Mf . Call the resulting signature g, which has signature matrix
0
0
0
1
0
1
0
0
0
0
1
0
1
0
0
0
.2 These
new 0’s impose a constraint on the types of cycle covers allowed. We call a cycle cover valid if it satisfies
this new constraint. A valid cycle cover must not pass through a vertex in a “crossing” way. Counting
the number of such cycle covers over 4-regular graphs can be done in polynomial time, even without the
planarity restriction. The signature g(x1 , x2 , x3 , x4 ) = Dis-Equality(x1 , x3 ) · Dis-Equality(x2 , x4 ) is of
the product type P, therefore Holant(g) is tractable.
1
, we obtain the problem
Under a holographic transformation by Z = √12 1i −i
Holant(g) = Holant (=2 | g) = Holant =2 Z ⊗2 | (Z −1 )⊗4 g = Holant (6=2 | ĝ) ,
−1 0 0 0
−1 ⊗4
where ĝ := (Z ) g = 00 01 10 00 . This problem has the following interpretation. It is a Holant problem
0 0 0 −1
on bipartite graphs. On the right side of the bipartite graph, the vertices must all have degree 4 and are
assigned the signature ĝ. On the left side, the vertices must all have degree 2 and are assigned the binary
disequality constraint 6=2 . The disequality constraints suggest an orientation between their two neighboring
vertices of degree 4 (see Figure 1). By convention, we view the edge as having its tail assigned 0 and its head
2 Recall that in general we require the input signature grid to specify the ordering of the edges (namely variables) on each
vertex. This is not necessary for symmetric signatures, but when asymmetric signatures are involved, specifying the ordering
is essential.
8
assigned 1. Then every valid assignment in this bipartite graph naturally corresponds to an orientation in
the original 4-regular graph.
If the four inputs 0011, 0110, 1100, and 1001 were in the support of ĝ, then the Holant sum would be
over all possible orientations with an even number of incoming edges at each vertex. As it is, the sum is over
all possible orientations with an even number of incoming edges at each vertex that also forbid those four
types of orientations at each vertex, as specified by ĝ. The following orientations are admissible by ĝ: The
orientation of the edges are such that at each vertex all edges are oriented out (source vertex), or all edges
are oriented in (sink vertex), or the edges are cyclically oriented in, out, in, out (saddle vertex).
Thus, the output of Holant (6=2 | ĝ) is a weighted sum over of these admissible orientations. Each admissible orientation O contributes a weight (−1)s(O) to P
the sum, where s(O) is the number of source and sink
vertices in an orientation O. We can express this as O∈O(G) (−1)s(O) , where O(G) is the set of admissible
orientations for G, which are those orientations that only contain source, sink, and saddle vertices. In words,
the value is the number of admissible orientations with an even number of sources and sinks minus the
number of admissible orientations with an odd number of sources and sinks. This orientation problem may
seem quite different from the restricted cycle cover problem we started with, but they are, in fact, the same
problem. Since Holant(g) is tractable, so is Holant (6=2 | ĝ).
Now, consider a slight generalization of this orientation problem.
Problem: #λ-SourceSinkSaddleOrientations
Input: AnPundirected 4-regular graph G (equipped with a local edge-ordering on every vertex).
Output: O∈O(G) λs(O) .
For λ = −1, we recover the orientation problem from above. For λ = 1, the problem is also tractable
since, when viewed as a bipartite Holant problem on the (2, 4)-regular bipartite vertex-edge incidence graph,
the disequality constraint on the vertices of degree 2 and the constraint on the vertices of degree 4 are
both product-type functions. As a function of x1 , x2 , x3 , x4 , the constraint on the degree 4 vertices is
Equality(x1 , x3 ) · Equality(x2 , x4 ). Let sk,m (G) be the number of O ∈ O(G) such that s(O) ≡ k
(mod m). Then the output of this problem with λ = 1 is s0,2 (G) + s1,2 (G) and the output of this problem
with λ = −1 is s0,2 (G) − s1,2 (G). Therefore, we can compute both s0,2 (G) and s1,2 (G). However, more is
possible.
For λ = i, the problem is tractable using affine constraints. In the (2, 4)-regular bipartite vertex-edge
incidence graph, the disequality constraint assigned to the vertices of degree 2 is affine. On the vertices
of degree 4, the assigned constraint function is an affine signature since the affine support is defined by
the affine linear system x1 = x3 and x2 = x4 while the quadratic polynomial in the exponent of i is
2x1 x2 + 3x1 + 3x2 + 1. (Recall that in the definition of A , Definition 2.4, we need to evaluate the quadratic
polynomial mod 4 instead of 2, and x2 = x for any x ∈ {0, 1}.) Although the output is a complex number, the
real and imaginary parts encode separate information. The real part is s0,4 (G) − s2,4 (G) and the imaginary
part is s1,4 (G) − s3,4 (G). Since s0,2 (G) = s0,4 (G) + s2,4 (G) and s1,2 (G) = s1,4 (G) + s3,4 (G), we can actually
compute all four quantities s0,4 (G), s1,4 (G), s2,4 (G), and s3,4 (G) in polynomial time.
3.3. An Enigmatic Problem
Some problems may be a challenge for the human intelligence to grasp. But in a platonic view of
computational complexity, they are no less valid problems.
For example, consider the problem Holant((1 + c2 )−1 [1, 0, −i] | f ) where f has the signature matrix
r
r
q
√
√
√
√
√
√
−8i 13+9 2+2 82+58 2
0
(4+4i) 28+20 2+ 2 799+565 2
(4+4i) 28+20 2+ 2 799+565 2
r
q
q
q
√
√
√
√
√
√
√
√
(4+4i) 28+20 2+ 2 799+565 2
8i 18+13 2+4 41+29 2
(−4+4i) 12+8 2+ 274+194 2
−8i 13+9 2+2 82+58 2
r
q
q
q
√
√
√
√
√
√
√
√
−8i 13+9 2+2 82+58 2
(−4+4i) 12+8 2+ 274+194 2
8i 18+13 2+4 41+29 2
(4+4i) 28+20 2+ 2 799+565 2
q
q
q
q
√
√
√
√
√
√
√
√
(−4+4i) 12+8 2+ 274+194 2
−8i 13+9 2+2 82+58 2
(−4+4i) 12+8 2+ 274+194 2
−16 13+9 2+2 82+58 2
q
√
√
and c = 1 + 2 + 2(1 + 2). Most likely no one has ever considered this problem before. Yet this nameless
1 c
problem is A -transformable under T = [ 10 α0 ] −c
1 , and hence it is really the same problem as a more
9
comprehensible problem defined by fˆ = (T −1 )⊗4 f . Namely,
Holant((1 + c2 )−1 [1, 0, −i] | f ) = Holant((1 + c2 )−1 [1, 0, −i]T ⊗2 | (T −1 )⊗4 f ) = Holant([1, 0, 1] | fˆ) = Holant(fˆ),
where fˆ =
1 −1 −1 −1
−1 −1 1 −1
−1 1 −1 −1 .
−1 −1 −1 1
We can express fˆ as fˆ(x1 , x2 , x3 , x4 ) = iQ(x) , where Q(x1 , x2 , x3 , x4 ) = 2(x21 +
x22 + x23 + x24 + x1 x2 + x2 x3 + x3 x4 + x4 x1 ). Therefore, fˆ is affine, which means that Holant(fˆ) as well as
Holant((1 + c2 )−1 [1, 0, −i] | f ) are tractable. Furthermore, notice that fˆ only contains integers even though
(1 + c2 )−1 [1, 0, −i] and f contain many complex numbers with irrational real and imaginary parts. Thus,
Holant((1 + c2 )−1 [1, 0, −i] | f ) is not only tractable, but it always outputs an integer. Apparent anomalies
like Holant((1 + c2 )−1 [1, 0, −i] | f ), however contrived they may seem to be to the human eye, behoove the
creation of a systematic theory to understand and characterize the tractable cases.
4. General A -transformable Signatures
In this section, we give the algorithm to check A -transformable signatures. Our general strategy is
to bound the number of possible transformations by a polynomial in the length of the function, and then
enumerate all of them. There are some cases where this number cannot be bounded, and those cases are
handled separately.
n
Let f be a signature of arity n. It is given as a column vector in C2 with bit length N , which is on the
order of 2n . We denote its entries by fx = f (x) indexed by x ∈ {0, 1}n. The entries are from a fixed degree
algebraic extension of Q and we may assume basic bit operations in the field take unit time.
2
Notice that the number of general affine signatures of arity n is on the order of 2n . Hence a naive check
of the membership of affine signatures would result in a super-polynomial running time in N . Instead, we
present a polynomial-time algorithm.
Lemma 4.1. There is an algorithm to decide whether a given signature f of arity n belongs to A with
running time polynomial in N , the bit length of f .
Proof. We may assume that f is not identically zero. Normalize f so that the first nonzero entry of f is 1.
If there exists a nonzero entry of f after normalization that is not a power of i, then f 6∈ A , so assume that
all entries are now powers of i.
The next step is to decide if the support S 6= ∅ of f forms an affine linear subspace. We try to build a basis
for S inductively. It may end successfully or find an inconsistency. We choose the index of the first nonzero
entry b0 ∈ S as our first basis element. Assume we have a set of basis elements B = {b0 , . . . , bk } ⊆ S.
Consider the affine linear span Span(B). We check if Span(B) ⊆ S. If not, then S is not affine and f 6∈ A ,
so suppose that this is the case. If Span(B) = S, then we are done. Lastly, if S − Span(B) 6= ∅, then pick
the next element bk+1 ∈ S − Span(B). Let B ′ = B ∪ {bk+1 } and repeat with the new basis set B ′ .
Now assume that S is an affine subspace, that we have a linear system defining it, and that every
nonzero entry of f is a power of i. If S has dimension 0, then S is a single point, and f ∈ A . Otherwise,
dim(S) = r ≥ 1, and (after reordering) x1 , . . . , xr are free variables of the linear system defining S. For each
x ∈ {0, 1}r , let y ∈ {0, 1}n−r be the unique extension such that xy ∈ S. For each x, define px ∈ Z4 such
that fxy = ipx 6= 0. We will use the alternative expression for affine functions: namely, we want to decide if
there exists a quadratic polynomial
Q(x) =
r
X
j=1
cj x2j + 2
X
ckℓ xk xℓ + c,
1≤k<ℓ≤r
where c, cj , ckℓ ∈ Z4 , for 1 ≤ j ≤ r and 1 ≤ k < ℓ ≤ r, such that Q(x) ≡ px (mod 4) for all x ∈ {0, 1}r .
Setting x = 0 ∈ {0, 1}r determines c. Setting exactly one xj = 1 and the rest to 0 determines cj . Setting
exactly two xk = xℓ = 1 and the rest to 0 determines ckℓ . Then we verify if Q(x) is consistent with f , and
f ∈ A iff it is so.
10
For later use, we note the following corollary.
Corollary 4.2. There is an algorithm to decide whether a given signature f of arity n belongs to [ 10 α0 ] A
with running time polynomial in N , the bit length of f .
0 ⊗n
f ∈ A by Lemma 4.1.
Proof. For arity(f ) = n, just check if 10 α−1
We can strengthen Lemma 2.7 by restricting to orthogonal transformations within SO2 (C).
Lemma 4.3. Let F be a set of signatures. Then F is A -transformable iff there exists an H ∈ SO2 (C) such
that F ⊆ HA or F ⊆ H [ 10 α0 ] A .
Proof. Sufficiency is obvious by Lemma 2.7.
Assume that F is A -transformable. By Lemma 2.7, there exists an H ∈ O2 (C) such that F ⊆ HA or
F ⊆ H [ 10 α0 ] A . If H ∈ SO2 (C), we are done, so assume that H ∈ O2 (C)
\ SO2 (C). We want to find an
0
H ′ ∈ SO2 (C) such that F ⊆ H ′ A or F ⊆ H ′ [ 10 α0 ] A . Let H ′ = H 10 −1
∈ SO2 (C). There are two cases
to consider.
0
1. Suppose F ⊆ HA . Then since 10 −1
∈ Stab(A ),
0
F ⊆ H 10 −1
A
= H ′A .
2. Suppose F ⊆ H [ 10 α0 ] A . Then since
1
0
0 −1
∈ Stab(A ) commutes with [ 10 α0 ],
0
F ⊆ H [ 10 α0 ] 10 −1
A
1 0 1 0
= H 0 −1 [ 0 α ] A
= H ′ [ 10 α0 ] A .
We now
some properties of a signature under transformations in SO2 (C). Let f be a signature
aobserve
b
2
2
∈
SO
and H = −b
2 (C) where a + b = 1. Notice that v0 = (1, i) and v1 = (1, −i) are row eigenvectors of
a
i
0
. Then Z ′ H = T Z ′ , where T = a−bi
H with eigenvalues a − bi and a + bi respectively. Let Z ′ = 11 −i
0 a+bi .
For an index or a bit-string u = (u1 , . . . , un ) ∈ {0, 1}n of length n, let
vu := vu1 ⊗ vu2 ⊗ . . . ⊗ vun ,
and let wt(u) be the Hamming weight of u. Then vu is a row eigenvector of the 2n -by-2n matrix H ⊗n with
eigenvalue
(a − bi)n−wt(u) (a + bi)wt(u) = (a − bi)n−2wt(u) = (a + bi)2wt(u)−n
(1)
since (a + bi)(a − bi) = a2 + b2 = 1. In this paper, the following Z ′ -transformation plays an important role.
For any function f on {0, 1}n, we define
fˆ = Z ′⊗n f.
Then fˆu = hvu , f i, as a dot product.
Lemma 4.4. Suppose f and g are signatures of arity n and let H =
g = H ⊗n f iff ĝ = T ⊗n fˆ.
Proof. Since Z ′ H = T Z ′ ,
g = H ⊗n f ⇐⇒ Z ′⊗n g = Z ′⊗n H ⊗n f
⇐⇒ Z ′⊗n g = T ⊗n Z ′⊗n f
⇐⇒
ĝ = T ⊗n fˆ.
11
a b
−b a
and T =
a−bi
0
0
a+bi
. Then
We note that vuT is also a column eigenvector of H ⊗n with eigenvalue (a−bi)2wt(u)−n. Now we characterize
the signatures that are invariant under transformations in SO2 (C).
Lemma 4.5. Let f be a signature. Then f is invariant under transformations in SO2 (C) (up to a nonzero
constant) iff the support of fˆ contains at most one Hamming weight.
Proof. This clearly holds when f is identically zero, so assume that f contains a nonzero entry and has
arity n. Such an f is invariant
iff f is a column eigenvector of
a b under any H (up 2to a 2nonzero constant)
⊗n
H ⊗n . Consider H = −b
∈
SO
(C)
where
a
+
b
=
1.
Then
H
has n + 1 distinct eigenvalues
2
a
(a − bi)n−w (a + bi)w , for 0 ≤ w ≤ n. As a consequence, f is a column eigenvector of H ⊗n iff f is a nonzero
linear combination of vuT of the same Hamming weight wt(u). Hence f is invariant under H iff the support
of fˆ contains at most one Hamming weight.
Using Lemma 4.5, we can efficiently decide if there exists an H ∈ SO2 (C) such that H ⊗n f ∈ A .
Lemma 4.6. There is an algorithm to decide in time polynomial in N , for any input signature f of arity
n, whether there exists an H ∈ SO2 (C) such that H ⊗n f ∈ A . If so, either f ∈ A and f is invariant under
any transformation in SO2 (C), or there exist at most 8n many H ∈ SO2 (C) such that H ⊗n f ∈ A , and they
can all be computed in time polynomial in N .
Proof. Compute fˆ = Z ′⊗n f . If the support of fˆ contains at most one Hamming weight, then by Lemma 4.5,
f is invariant under any H ∈ SO2 (C). Therefore we only need to directly decide if f ∈ A , which we do by
Lemma 4.1.
Now assume there are at least two nonzero entries of fˆ with distinct Hamming weights, say u1 , u2 ∈
a b
{0, 1}n. Then fˆu1 and fˆu2 are nonzero, and 0 < wt(u2 ) − wt(u1 ) ≤ n. Suppose there exists an H = −b
a ∈
a−bi 0
⊗n
⊗n ˆ
SO2 (C) such that g = H f ∈ A . Then by Lemma 4.4, we have ĝ = T f , where T =
0 a+bi
√
is a diagonal transformation. Recall H2 and D from Table 1. Since Z ′ = 2H2 D ∈ Stab(A ), we have
ĝ = Z ′⊗n g ∈ A . Also since T is diagonal, both ĝu1 and ĝu2 are nonzero. Therefore, there must exist an
r ∈ {0, 1, 2, 3} such that
ir =
(a + bi)2wt(u2 )−n fˆu2
ĝu2
fˆu
=
= (a + bi)2wt(u2 )−2wt(u1 ) 2 ,
ĝu1
fˆu1
(a + bi)2wt(u1 )−n fˆu1
(2)
where we used (1). Recall that 0 < wt(u2 ) − wt(u1 ) ≤ n. View a + bi as a variable, and then there are at
most 2n solutions to (2), given r and fˆu1 and fˆu2 . There are 4 possible values of r, resulting in at most 8n
many solutions for a, b ∈ C such that a + bi satisfies (2) and a2 + b2 = 1. Each (a, b) solution corresponds to
a distinct H ∈ SO2 (C).
We also want to efficiently decide if there exists an H ∈ SO2 (C) such that H ⊗n f ∈ [ 10 α0 ] A .
Lemma 4.7. There is an algorithm to decide, for any input signature f of arity n, whether there exists an
H ∈ SO2 (C) such that H ⊗n f ∈ [ 10 α0 ] A with running time polynomial in N . If so, either f ∈ [ 10 α0 ] A and
f is invariant under any transformation in SO2 (C), or there exist O(nN 16 ) many H ∈ SO2 (C) such that
H ⊗n f ∈ [ 10 α0 ] A , and they can all be computed in polynomial time in N .
Proof. Compute fˆ = Z ′⊗n f . If the support of fˆ contains at most one Hamming weight, then by Lemma 4.5,
f is invariant under any H ∈ SO2 (C). Therefore we only need to directly decide if f ∈ [ 10 α0 ] A , which we
do by Corollary 4.2.
Now assume there are at least two nonzero entries of fˆ that are of distinct Hamming weight. Let
u1 , u2 ∈ {0, 1}n be such that fˆu1 and fˆu2 are nonzero, and 0 < wt(u2 ) − wt(u1 ) ≤ n. We derive necessary
a b
conditions for the existence of H ∈ SO2 (C) such that H ⊗n f ∈ [ 10 α0 ] A . Thus, assume such an H = −b
a
exists, where a2 + b2 = 1.
12
i 10
Let g = H ⊗n f . Then ĝ = Z ′⊗n g ∈ 11 −i
[ 0 α ] A . By Lemma 4.4, we have ĝ = T ⊗n fˆ, where T =
a−bi 0
2wt(u)−n ˆ
fu for any u ∈ {0, 1}n. Let t = wt(u1 ) − wt(u2 ). Then
0 a+bi . Thus ĝu = (a + bi)
ĝu1
fˆu
(a + bi)2wt(u1 )−n fˆu1
= (a + bi)2t 1 .
=
ĝu2
fˆu2
(a + bi)2wt(u2 )−n fˆu2
Hence
(a + bi)2t =
fˆu2 ĝu1
.
·
fˆu1 ĝu2
We claim that the value of each entry in ĝ as well as the number of possible values is bounded by
a polynomial in N , and hence so are the ratios between them. Let h ∈ A be a signature such that
i ⊗n 1 0 ⊗n
[ 0 α ] h. Every nonzero entry of h is a power of i, up to a constant factor λ. This constant
ĝ = 11 −i
⊗n
factor cancels when taking ratios of entries, so we omit it. Let h′ = [ 10 α0 ] h. Then every entry of h′ is a
1 i ⊗n
is also a power of α. Therefore every entry of ĝ is an
power of α or 0. Moreover, each entry of 1 −i
exponential sum of 2n terms, each a power of α or 0. Recall that α8 = 1 and hence there are 8 possible
values of these powers. Let c0 denote the number of 0 and ci (for 1 ≤ i ≤ 8) denote the number of αi in an
entry ĝu of ĝ. Then we have
c0 +
8
X
ci = 2 n
and
8
X
ci αi = ĝu .
i=1
i=1
Clearly the total number of possible
values
of entries in ĝu is at most the number of possible choices of
n
(c0 , . . . , c8 ). There are at most 2 8+8 = O(N 8 ) choices of (c0 , . . . , c8 ). Thus the number of all possible
ratios is at most O(N 16 ), and can all be enumerated in time polynomial in N .
For any possible value of the ratio
ĝu1
ĝu2
, each possible value of
fˆu2
fˆu
gives at most 2n different transformations
1
H. Therefore, the total number of transformations is bounded by O(nN 16 ), and we can find them in time
polynomial in N .
Now we give an algorithm that efficiently decides if a set of signatures is A -transformable.
Theorem 4.8. There is a polynomial-time algorithm to decide, for any finite set of signatures F , whether
F is A -transformable. If so, at least one transformation can be found.
Proof. By Lemma 4.3, we only need to decide if there exists an H ∈ SO2 (C) such that F ⊆ HA or F ⊆
H [ 10 α0 ] A . To every signature in F , we apply Lemma 4.6 or Lemma 4.7 to check each case, respectively. If no
H exists for some signature, then F is not A -transformable. Otherwise, every signature is A -transformable
for some H ∈ SO2 (C). If every signature in F is invariant under transformations in SO2 (C), then F
is A -transformable. Otherwise, we pick the first f ∈ F that is not invariant under transformations in
SO2 (C). The number of possible transformations that work for f is bounded by a polynomial in the size
of the presentation of f . We simply try all such transformations on all other signatures in F that are not
invariant under transformations in SO2 (C), respectively using Lemma 4.1 or Corollary 4.2 to check if the
transformation works.
5. General P-transformable Signatures
In this section, we give the algorithm to check P-transformable signatures. Once again, our general
strategy is to bound the number of possible transformations (with a few exceptions), and then enumerate
all of them. Indeed, the bound will be a constant in this section. The distinct feature for P-transformable
signatures is that we have to decompose them first.
We begin with the counterpart to Lemma 4.3, which strengthens Lemma 2.10 by restricting to either
orthogonal transformations within SO2 (C) or no orthogonal transformation at all.
13
Lemma 5.1. Let F be a set of signatures. Then F is P-transformable iff F ⊆
H ∈ SO2 (C) such that F ⊆ HP.
1
1
i −i
P or there exists an
Proof. Sufficiency is obvious by Lemma 2.10.
Assume
F is P-transformable. By Lemma 2.10, there exists an H ∈ O2 (C) such that F ⊆ HP or
1 that
P. There are two cases to consider.
F ⊆ H 1i −i
1. Suppose F ⊆ HP. If H ∈ SO2 (C), then we are done, so assume
0 that H ∈ O2 (C) \ SO2 (C). We want
to find an H ′ ∈ SO2 (C) such that F ⊆ H ′ P. Let H ′ = H 10 −1
∈ SO2 (C). Then
1 0
F ⊆ H 0 −1 P
= H ′P
0
since 10 −1
∈ Stab(P).
1
a b
2. Suppose F ⊆ H 1i −i
P. If H = −b
a ∈ SO2 (C), then
1
F ⊆ H 1i −i
P
1 1 a+bi 0
⊆ i −i
0 a−bi P
1 1
⊆ i −i P
a b
a+bi 0
1 a+bi 0
1
= 1i −i
since H 1i −i
0 a−bi ∈ Stab(P). Otherwise, H = b −a ∈ O2 (C) \
0 a−bi and
SO2 (C) and
1
F ⊆ H 1i −i
P
1 1 0 a−bi
⊆ i −i a+bi 0 P
1
P
⊆ 1i −i
1 1 1 0 a−bi
0 a−bi
∈ Stab(P).
since H 1i −i
and a+bi
= i −i a+bi 0
0
The “building blocks” of P are signatures whose support is contained in two entries with complementary
indices. However, for technical convenience that will be explained shortly, in the following definition we
restrict to functions that are either unary, or have support of size exactly two. Recall that two signatures
are considered the same if one is a nonzero multiple of the other.
Definition 5.2. A k-ary function f is a generalized equality if it is a nonzero multiple of [0, 0], [1, 0], [0, 1],
or satisfies
∃x ∈ {0, 1}k , ∀y ∈ {0, 1}k , fy = 0 ⇐⇒ y 6∈ {x, x}.
We use E to denote the set of all generalized equality functions.
For any set F , we let hF i denote the closure under function products without shared variables. It is easy
to show that P = hE i (cf. [20]).
If we view signatures as tensors, then h·i is the closure under tensor products. That is, if f (x1 , x2 ) =
f1 (x1 )f2 (x2 ), then f = f1 ⊗ f2 with a correct ordering of indices. In general, we call such f reducible, defined
next.
Definition 5.3. We call a function f of arity n on variable set x reducible if f has a non-trivial decomposition, namely, there exist f1 and f2 of arities n1 and n2 on variable sets x1 and x2 , respectively, such that
1 ≤ n1 , n2 ≤ n − 1, x1 ∪ x2 = x, x1 ∩ x2 = ∅, and f (x) = f1 (x1 )f2 (x2 ). Otherwise we call f irreducible.
Note that all unary functions, including [0, 0], are irreducible. However, the identically zero function of
arity greater than one is reducible. Recall that we call a function degenerate if it is a tensor product of unary
functions. All degenerate functions of arity ≥ 2 are reducible, but not vice versa — a reducible function may
be decomposable into only non-unary functions. Due to the same reason, degenerate functions are trivially
tractable, but reducible functions are not necessarily so.
14
Definition 5.2 is a slight modification of a similar definition for E that appeared in Section 2 of [20]. For
both definitions of E , it follows that P = hE i. The motivation for our slight change in the definition is so
that every signature in E is irreducible.
Irreducibility is preserved by transformations.
Lemma 5.4. Let f be an irreducible function of arity n, and T be a 2-by-2 non-singular matrix. Then
g = T ⊗n f is also irreducible.
Proof. Suppose g is reducible. By Definition 5.3, there is a non-trivial decomposition g = g1 ⊗ g2 . Hence
⊗n
f = T −1
g also has a non-trivial decomposition.
If a function f is reducible, then we can factor it into functions of smaller arity. This procedure can
be applied recursively and terminates when all components are irreducible. Therefore any function has at
least one irreducible factorization. We show that such a factorization is unique for functions that are not
identically zero.
Lemma 5.5. Let f be a function of arity n on variables x that is not identically zero. Assume there exist
irreducible functions fi and gj , and two partitions {xi } and {yj } of x for 1 ≤ i ≤ k and 1 ≤ j ≤ k ′ , such
that
k′
k
Y
Y
gj (yj ).
fi (xi ) =
f (x) =
j=1
i=1
′
Then k = k , the partitions are the same, and there exists a permutation π on {1, 2, · · · , k} such that
fi = gπ(j) up to nonzero factors.
Proof. Since f is not identically zero, none of the fi or gj is identically zero. Fix an assignment u2 , . . . , uk
Qk
such that c = i=2 fi (ui ) 6= 0. Let zj = yj ∩ x1 , and vj = yj ∩ (∪ki=2 xi ) for 1 ≤ j ≤ k ′ . Let the assignments
u2 , . . . , uk restricted to vj be wj . Then we have
cf1 (x1 ) = f1 (x1 )
k
Y
′
fi (ui ) =
k
Y
gj (zj , wj ).
j=1
i=2
Define new functions hj (zj ) = gj (zj , wj ) for 1 ≤ j ≤ k ′ . Then
′
k
1Y
f1 (x1 ) =
hj (zj ).
c j=1
′
Since f1 is irreducible, there cannot be two zj that are nonempty. And yet, x1 = ∪kj=1 zj , so it follows that
x1 = zj for some 1 ≤ j ≤ k ′ . We may assume j = 1, so x1 ⊆ y1 . By the same argument we have y1 ⊆ xi ,
for some i. But by disjointness of x = ∪ki=1 xi , we must have y1 ⊆ x1 . Thus after a permutation, we have
x1 = y1 . Therefore f1 = g1 up to a nonzero constant.
By fixing some assignment to x1 = y1 such that f1 and g1 are not zero, we may cancel this factor, and
the proof is completed by induction. Therefore we must have that k = k ′ and the two sets {fi } and {gj } are
equal, where we identify functions up to nonzero constants.
In fact, we can efficiently find the unique factorization.
Lemma 5.6. There is an algorithm to compute in time polynomial in N , for any input signature f of arity
n that is not identically zero, the unique factorization of f into irreducible factors. More specifically,
the
P
algorithm computes irreducible f1 , . . . , fk of arities n1 , . . . , nk ∈ Z+ (for some k ≥ 1) such that ki=1 ni = n
Qk
and f (x1 , . . . , xk ) = i=1 fi (xi ).
15
Proof. We may partition the variables x into two sets x1 and x2 of length n1 and n2 , respectively, such that
1 ≤ n1 , n2 ≤ n − 1, x1 ∪ x2 = x, and x1 ∩ x2 = ∅. Define a 2n1 -by-2n2 matrix M such that Mu1 ,u2 = f (u1 , u2 )
for u1 ∈ {0, 1}n1 and u2 ∈ {0, 1}n2 . Then M is of rank at most 1 iff there exist f1 and f2 of arity n1 and n2 ,
such that f (x) = f1 (x1 )f2 (x2 ).
Therefore, in order to factor f , we only need to run through all distinct partitions, and check if there
exists at least one such matrix of rank at most 1. If none exists, then f is irreducible. The total number of
possible such partitions is 2n−1 − 1. Hence the running time is polynomial in 2n ≤ N .
Once we have found f = f1 ⊗ f2 , we recursively apply the above procedure to f1 and f2 until every
component is irreducible. The total running time is polynomial in N .
This factorization algorithm gives a simple algorithm to determine membership in P.
Lemma 5.7. There is an algorithm to decide, for a given signature f of arity n, whether f ∈ P with
running time polynomial in N .
N
Proof. We may assume that f is not identically zero, and we obtain its unique factorization f = i fi by
Lemma 5.6. Then f ∈ P iff for all i, we have fi ∈ E . Since membership in E is easy to check, our proof is
complete.
Let T ∈ GL2 (C) be some transformation and f some signature. To check if f ∈ T P, it suffices to first
factor f and then check if each irreducible factor is in T E .
N
Lemma 5.8. Suppose f = ki=1 fi is not identically zero and that fi is irreducible for all 1 ≤ i ≤ k. Let
T ∈ GL2 (C). Then f ∈ T P iff fi ∈ T E for all 1 ≤ i ≤ k.
Pk
Proof. Suppose f is of arity n and fi is of arity ni so that i=1 ni = n. If fi ∈ T E for all 1 ≤ i ≤ k, then
N
N
N
there exists gi ∈ E such that fi = T ⊗ni gi . Thus f = ki=1 fi = ki=1 T ⊗ni gi = T ⊗n ki=1 gi . Since gi ∈ E ,
Nk
we have i=1 gi ∈ P. Therefore f ∈ T P.
On the other hand, assume f ∈ T P. By the definition of P, there exist g1 , . . . , gk′ ∈ E of arities
Nk′
′
m1 , . . . , mk′ ∈ Z+ , such that f = T ⊗n g, where g =
i=1 gi . Since gi ∈ E , gi is irreducible. Let fi =
′
N
N
k
k
T ⊗mi gi ∈ T E for all 1 ≤ i ≤ k ′ , which is also irreducible by Lemma 5.4. Then i=1 fi′ = f = i=1 fi . By
Lemma 5.5, we have k = k ′ and {fi } and {fi′ } are the same up to a permutation. Therefore each fi ∈ T E .
With Lemma 5.6 and Lemma 5.8 in mind, we focus our attention on membership in E . We show how to
efficiently decide if there exists an H ∈ SO2 (C) such that H ⊗n f ∈ E when f is irreducible.
Lemma 5.9. There is an algorithm to decide, for a given irreducible signature f of arity n ≥ 2, whether
there exists an H ∈ SO2 (C) such that H ⊗n f ∈ E with running time polynomial in N . If so, there exist at
most eight H ∈ SO2 (C) such that H ⊗n f ∈ E unless f = (1, 0, 0, 1)T or f = (0, 1, −1, 0)T.
a b
⊗n
Proof. Assume there exists an H = −b
f ∈ E , where a2 + b2 = 1. Then
a ∈ SO2 (C) such that g = H
i
0
⊗n ˆ
by Lemma 4.4, there exists a diagonal transformation T = a−bi
E . In
f ∈ 11 −i
0 a+bi such that ĝ = T
particular, ĝ and fˆ have the same support. For two vectors u, x ∈ {0, 1}n, the entry indexed by row u and
i ⊗n
is iwt(x) (−1)hx,ui , where wt(·) denotes Hamming weight and h·, ·i is the
column x in the matrix 11 −i
dot product.
Since g ∈ E , g is irreducible. Thus g has two nonzero entries with opposite index, say x and x. Hence
we have
ĝu = iwt(x) (−1)hx,uigx + iwt(x) (−1)hx,ui gx
for any vector u ∈ {0, 1}n.
= iwt(x) (−1)hx,uigx + in−wt(x) (−1)wt(u)−hx,uigx
= (−1)hx,ui iwt(x) gx + in−wt(x) (−1)wt(u) gx
16
For u1 , u2 ∈ {0, 1}n , if wt(u1 ) ≡ wt(u2 ) (mod 2), then
ĝu1 = ±ĝu2 .
(3)
Therefore, if any entry of fˆ with even Hamming weight is 0, then all entries with even Hamming weight
are 0. This also holds for entries with odd Hamming weight. However, fˆ is not identically zero because it is
irreducible and of arity n ≥ 2. Therefore, we know that either all entries of even Hamming weight are not 0
or all entries of odd Hamming weight are not 0. If n ≥ 3, or if n = 2 and all entries of even Hamming weight
are not 0, then we can take two nonzero entries of fˆ whose Hamming weight differ by 2. Their ratio restricts
the possible choices of a + bi, as in the proof of Lemma 4.7, because the only possible ratios for ĝu1 /ĝu2 are
±1 by (3). Together with a2 + b2 = 1, this gives at most 8 possible matrices H ∈ SO2 (C).
The remaining case is when n = 2 and all entries of fˆ with even Hamming weight are 0. By (3), we have
ĝ = λ(0,
1, ±1, 0)T for some λ 6= 0 since ĝ and fˆ have
support.
from fˆ = (T −1 )⊗2 ĝ, where
0the1 same
0 1Then
a+bi
0
−1
−1 T
−1
) = ±1 0 . Hence, up to a nonzero scalar,
T =
±1 0 (T
0 a−bi is diagonal, we calculate that T
fˆ = (0, 1, 1, 0)T or fˆ = (0, 1, −1, 0)T. Finally f = (Z ′−1 )⊗2 fˆ, and we get f = (1, 0, 0, 1)T or f = (0, 1, −1, 0)T,
up to a nonzero scalar.
Now we give an algorithm that efficiently decides if a set of signatures is P-transformable.
Theorem 5.10. There is a polynomial-time algorithm to decide, for any finite set of signatures F , whether
F is P-transformable. If so, at least one transformation can be found.
1
P or if there exists an H ∈ SO2 (C) such that
Proof. By Lemma 5.1, we only need to decide if F ⊆ 1i −i
1 −1
1 1
F ⊆ HP. To check if F ⊆ i −i P, we simply apply Lemma 5.7 to each signature in 1i −i
F.
Now to check if F ⊆ HP. We may assume that no signature in F is identically zero. Now we obtain the
unique factorization of each signature in F using Lemma 5.6. If every irreducible factor is either a unary
signature, or (1, 0, 0, 1)T, or (0, 1, −1, 0)T, then F ⊆ hE i = P. Otherwise,
N let f ∈ F be a signature that is
not of this form. This means that f has a unique factorization f =
i fi where some fi is not a unary
signature, or (1, 0, 0, 1)T , or (0, 1, −1, 0)T. Assume it is f1 .
By applying Lemma 5.8 to f , we get the necessary condition f1 ∈ HE . Then we apply Lemma 5.9 to f1 .
If the test passes, then by the definition of f1 , we have at most eight transformations in SO2 (C) that could
work. For each possible transformation H, we apply Lemma 5.7 to every signature in H −1 F to check if it
works.
6. Symmetric A -transformable Signatures
In the next two sections, we consider the case when the signatures are symmetric. The significant
difference is that a symmetric signature of arity n is given by n + 1 values, instead of 2n values. This
exponentially more succinct representation requires us to find a more efficient algorithm.
6.1. A Single Signature
Recall Definition 2.8. To begin with, we provide efficient algorithms to decide membership in each of
A1 , A2 , and A3 for a single signature. If the signature is in one of the sets, then the algorithm also finds at
least one corresponding orthogonal transformation satisfying Definition 2.8. By Lemma 2.9, this is enough
to check if a single signature is A -transformable.
We say a signature f satisfies a second order recurrence relation, if there exist not all zero a, b, c ∈ C,
such that for all 0 ≤ k ≤ n − 2, afk + bfk+1 + cfk+2 = 0. For a non-degenerate signature of arity at least 3,
these coefficients are unique up to a nonzero scalar.
Lemma 6.1. Let f be a non-degenerate symmetric signature of arity n ≥ 3. If f satisfies a second order
recurrence relation with coefficients a, b, c ∈ C and another one with coefficients a′ , b′ , c′ ∈ C, then there
exists a nonzero k ∈ C such that (a, b, c) = k(a′ , b′ , c′ ).
17
Proof. A function f = [f0 , f1 , . . . , fn ] is degenerate if and only if f0 , . . . , fn forms
a geometric
sequence. As
i
h
f0 f1 ... fn−2
f0 f1 ... fn−1
has rank 2. Let B = f1 f2 ... fn−1 . We claim that
f is non-degenerate, the matrix A = f1 f2 ... fn
f2 f3 ...
fn
rank(B) ≥ 2, which implies that f satisfies at most one second order recurrence relation up to a nonzero
scalar, as desired.
If (f1 , . . . , fn−1 ) = 0, then f0 , fn 6= 0 sinceh rank(A) = i 2, so rank(B)
h = 2 as iwell. Otherwise,
f f ... f
fn−1
(f1 , . . . , fn−1 ) 6= 0. Consider the matrices A1 = f01 f12 ... fn−2
and A2 = ff12 ff23 ...
, which are sub... fn
n−1
matrices of both A and B. Both A1 and A2 have rank at least 1 since (f1 , . . . , fn−1 ) 6= 0. We show that
either rank(A1 ) = 2 or rank(A2 ) = 2, which implies that rank(B) ≥ 2.
For a contradiction, suppose rank(A1 ) = rank(A2 ) = 1. Then there exist λ, µ ∈ C such that (f0 , . . . , fn−2 ) =
λ(f1 , . . . , fn−1 ) and (f2 , . . . , fn ) = µ(f1 , . . . , fn−1 ). If λ = 0, then f0 = f1 = 0 as n ≥ 3. It implies that
rank(A2 ) = rank(A). However, rank(A) = 2, a contradiction. Similarly if µ = 0, then rank(A1 ) = 2, a
contradiction. Otherwise λ, µ 6= 0 and we get fi 6= 0 for all 0 ≤ i ≤ n, and λµ = 1. This implies that
rank(A) = 1, a contradiction.
For a signature with a second order recurrence relation, the quantity b2 − 4ac is nonzero precisely when
the signature can be expressed as the sum of two degenerate signatures that are linearly independent.
Lemma 6.2. Let f be a non-degenerate symmetric signature of arity n ≥ 3. Then f satisfies a second
order recurrence relation with coefficients a, b, c satisfying b2 − 4ac 6= 0 iff there exist a0 , b0 , a1 , b1 (satisfying
a ⊗n a1 ⊗n
.
+ b1
a0 b1 6= a1 b0 ) such that f = b00
Proof. The “only if” direction is straightforward to verify. For the other direction, assume that there exist
a, b, c ∈ C not all zero, such that for all 0 ≤ k ≤ n − 2, afk + bfk+1 + cfk+2 = 0. If c 6= 0, then since
b2 − 4ac 6= 0, we can solve this recurrence with the initial values of f0 and f1 , namely, there exist c0 , c1 6= 0
and λ1 6= λ2 such that for any 0 ≤ k ≤ n,
fk = c0 λk1 + c1 λk2 .
⊗n
⊗n
. Normalizing shows the claim.
+ c1 λ12
In other words, we can express f as f = c0 λ11
The other case of c = 0 implies that b 6= 0. Hence the entries f0 , · · · , fn−1 satisfy a first order recurrence
relation and the recurrence does not involve the last entry fn . Thus there must exist c0 , c1 and λ such that
⊗n
⊗n
f = c0 [ λ1 ] + c1 [ 01 ] . Moreover, if any of c0 or c1 equals 0, then f is degenerate which contradicts the
assumption. The lemma follows from a normalization.
The following definition of the θ function is crucial. A priori, θ(v0 , v1 ) may be not well-defined, but this
is circumvented by insisting that v0 and v1 be linearly independent.
a
a
Definition 6.3. For a pair of linearly independent vectors v0 = b00 and v1 = b11 , we define
2
a0 a1 + b 0 b 1
.
θ(v0 , v1 ) :=
a1 b 0 − a0 b 1
Furthermore, suppose that a signature f of arity n ≥ 3 can be expressed as f = v0⊗n + v1⊗n , where v0 and v1
are linearly independent. Then we define θ(f ) = θ(v0 , v1 ).
Intuitively, this formula is the square of the cotangent of the angle from v0 to v1 . This notion of cotangent
is properly extended to the complex domain. The expression is squared so that θ(v0 , v1 ) = θ(v1 , v0 ).
Let f = v0⊗n + v1⊗n be a non-degenerate signature of arity n ≥ 3. Since f is non-degenerate, v0 and v1
are linearly independent. The next proposition implies that this expression for f via v0 and v1 is unique up
to a root of unity. Therefore, θ(f ) from Definition 6.3 is well-defined.
Proposition 6.4 (Lemma 9.1 in [23]). Let a, b, c, d be four vectors and suppose that c, d are linearly
independent. If for some n ≥ 3, we have a⊗n + b⊗n = c⊗n + d⊗n , then there exist ω0 and ω1 satisfying
ω0n = ω1n = 1 such that either a = ω0 c and b = ω1 d or a = ω0 d and b = ω1 c.
18
For the convenience of future use, we can generalize Proposition 6.4 to the following simple lemma.
Lemma 6.5. Let a, b, c, d be four vectors and suppose that c, d are linearly independent. Furthermore, let
x0 , x1 , y0 , y1 be nonzero scalars. If for some n ≥ 3, we have x0 a⊗n + x1 b⊗n = y0 c⊗n + y1 d⊗n , then there
exist ω0 and ω1 , such that either a = ω0 c, b = ω1 d, x0 ω0n = y0 , and x1 ω1n = y1 ; or a = ω0 d, b = ω1 c,
x0 ω0n = y1 , and x1 ω1n = y0 .
It is easy to verify that θ is invariant under an orthogonal transformation.
Lemma 6.6. For two linearly independent vectors v0 , v1 ∈ C2 and H ∈ O2 (C), let vb0 = Hv0 and vb1 = Hv1 .
Then θ(v0 , v1 ) = θ(vb0 , vb1 ).
Proof. Within the square in the definition of θ, the numerator is the dot product, which is invariant under
any orthogonal transformation. Also, the denominator is the determinant, which is invariant under any
orthogonal transformation up to a sign.
Now we have some necessary conditions for membership in A1 ∪ A2 ∪ A3 . Recall that A1 ⊆ P1 .
Lemma 6.7. Let f be a non-degenerate symmetric signature of arity at least 3. Then
1. f ∈ P1 =⇒ θ(f ) = 0,
2. f ∈ A2 =⇒ θ(f ) = −1, and
3. f ∈ A3 =⇒ θ(f ) = − 21 .
Proof. The result clearly holds when f is in the canonical form of each set. This extends to the rest of each
set by Lemma 6.6.
These results imply the following corollary.
Corollary 6.8. Let f be a non-degenerate symmetric signature f of arity n ≥ 3. If f is A -transformable,
then f is of the form v0⊗n + v1⊗n , where v0 and v1 are linearly independent, and θ(v0 , v1 ) ∈ {0, −1, − 21 }.
The condition given in Lemma 6.7 is not sufficient to
if f ∈ A1 ∪ A2 ∪ A3 . For example, if
1determine
f = v0⊗n + v1⊗n with v0 = [ 1i ] and v1 is not a multiple of −i
, then θ(f ) = −1 but f is not in A2 . However,
this is essentially the only exceptional case. We achieve the full characterization with some extra conditions.
The next lemma gives an equivalent form for membership in A1 , A2 , and A3 using transformations in
O2 (C) \ SO2 (C). Only having to consider transformation matrices in O2 (C) \ SO2 (C) is convenient since
such matrices are their own inverses.
Lemma 6.9. Suppose f is a non-degenerate symmetric signature of arity n ≥ 3 and let F ∈ {A1 , A2 , A3 }.
Then f ∈ F iff there exists an H ∈ O2 (C) \ SO2 (C) such that f ∈ F with H.
Proof. Sufficiency is trivial. For necessity, assume that f ∈ F with H ∈ O2 (C). If H ∈ O2 (C) \ SO2 (C),
then we are done, so further assume that H ∈ SO2 (C). By the definition of F ,
f = cH ⊗n v0⊗n + βv1⊗n ,
0 −1
where c 6= 0 and v0 , v1 , and β depend on F . Let H ′ = 10 −1
H ∈ O2 (C) \ SO2 (C), so it follows that
H ′ T = H ′−1 = H ′ . Then
f = (H ′ H ′ )⊗n f
= cH ′⊗n (H ′ H)⊗n v0⊗n + βv1⊗n
0 ⊗n ⊗n
= cH ′⊗n 10 −1
v0 + βv1⊗n
= cH ′⊗n v1⊗n + βv0⊗n
= cβH ′⊗n v0⊗n + β −1 v1⊗n ,
0
0
where in the fourth step, we use the fact that 10 −1
v0 = v1 and 10 −1
v1 = v0 for any F ∈ {A1 , A2 , A3 }.
To finish, we rewrite β −1 in the form required in Definition 2.8 as follows:
19
• if F = A1 , then β = αtn+2r for some t ∈ {0, 1} and r ∈ {0, 1, 2, 3} and β −1 = α−tn−2r . Pick
′
r′ ∈ {0, 1, 2, 3} such that r′ ≡ −tn − r (mod 4), so β −1 = αtn+2r as required;
• if F = A2 , then β = 1, so β −1 = 1 = β as required;
• if F = A3 , then β = ir for some r ∈ {0, 1, 2, 3}, so β −1 = i−r = i4−r as required.
Before considering A1 , we prove a technical lemma that is also applicable when considering P1 .
a
a
Lemma 6.10. Let f = v0⊗n + v1⊗n be a symmetric signature of arity n ≥ 3, where v0 = b00 and v1 = b11
are linearly independent. If θ(f ) = 0, then there exist an H ∈ O2 (C) and a nonzero k ∈ C satisfying a1 = kb0
and b1 = −ka0 such that
1 ⊗n
⊗n
H ⊗n f = λ [ 11 ] + k n −1
for some nonzero λ ∈ C.
Proof. Since θ(f ) = 0, we have a0 a1 + b0 b1 = 0. By linear independence, we have a1 b0 6= a0 b1 . Thus, there
exists a nonzero k ∈ C such that a1 = kb0 and b1 = −ka0 . (Note that this is clearly true even if one of a0
v0
or b0 , but not both, is zero.) Let c = a20 + b20 , which is nonzero since a1 b0 6= a0 b1 . Also, let u0 = √
and
1 1 c−1
v√
1
1
u1 = k c , so it follows that the matrix M = [u0 u1 ] is orthogonal. Then the matrix H = √2 1 −1 M
is
also orthogonal and what we need. Under a transformation by H, we have
n
⊗n
n n
2
H ⊗n f = H ⊗n c 2 u⊗n
0 + k c u1
1 ⊗n
⊗n
,
= λ [ 11 ] + k n −1
n
where λ = (c/2) 2 6= 0.
Now we give the characterization of A1 .
a
a
Lemma 6.11. Let f = v0⊗n + v1⊗n be a symmetric signature of arity n ≥ 3, where v0 = b00 and v1 = b11
are linearly independent. Then f ∈ A1 iff θ(f ) = 0 and there exist an r ∈ {0, 1, 2, 3} and t ∈ {0, 1} such that
an1 = αtn+2r bn0 6= 0 or bn1 = αtn+2r an0 6= 0.
Proof. Suppose
f ∈ A1 . By Lemma 6.9, after a suitable normalization, there exists a transformation
x y
H = y −x ∈ O2 (C) \ SO2 (C) such that
1 ⊗n
⊗n
,
f = H ⊗n [ 11 ] + β −1
where β = αtn+2r for some r ∈ {0, 1, 2, 3} and some t ∈ {0, 1}. Since H ∈ O2 (C), we have x2 + y 2 = 1. By
Lemma 6.7, θ(f ) = 0.
Now we have two expressions for f , which are
⊗n
x+y ⊗n
a0 ⊗n a1 ⊗n
.
+ β x−y
= f = y−x
+ b1
b0
y+x
Since v0 and v1 are linearly independent, we know that a0 and a1 cannot both be 0. Suppose a0 6= 0. By
Lemma 6.5, we have two cases.
1. Suppose a0 = ω0 (x + y) and b1 = ω1 (x + y) where ω0n = 1 and ω1n = β. Then we have bn1 = β(x + y)n =
βan0 6= 0. Since β = αtn+2r , we are done.
2. Suppose a0 = ω0 (x − y) and b1 = ω1 (y − x) where ω0n = β and ω1n = 1. Then we have an0 = β(x − y)n =
αtn+2r (−1)n (y − x)n = αtn+2r+4n bn1 , so bn1 = α−tn−2r−4n an0 6= 0. Pick r′ ∈ {0, 1, 2, 3} such that
′
r′ ≡ −tn − r − 2n (mod 4). Then α−tn−2r−4n = αtn+2r is of the desired form.
Otherwise, a1 6= 0, in which case, similar reasoning shows that an1 = αtn+2r bn0 6= 0.
For sufficiency, we apply Lemma 6.10, which gives
1 ⊗n
⊗n
H ⊗n f = λ [ 11 ] + k n −1
for some H ∈ O2 (C), some nonzero λ ∈ C, and some nonzero k ∈ C satisfying a1 = kb0 and b1 = −ka0 . The
ratio of these coefficients is k n . We consider two cases.
20
1. Suppose an1 = αtn+2r bn0 6= 0. Then k n = αtn+2r , so f ∈ A1 .
2. Suppose bn1 = αtn+2r an0 6= 0. Then k n = (−1)n αtn+2r . Pick r′ ∈ {0, 1, 2, 3} such that r′ ≡ r + 2n
′
(mod 4). Then k n = αtn+2r , so f ∈ A1 .
Now we give the characterization of A3 .
a
6.12. Let f = v0⊗n + v1⊗n be a symmetric signature of arity n ≥ 3, where v0 = b00 and v1 =
Lemma
a1
are linearly independent. Then f ∈ A3 iff there exist an ε ∈ {1, −1} and r ∈ {0, 1, 2, 3} such that
b1
√
√
√ n
√
n
a1
2a0 + εib0 = b1 εia0 − 2b0 , an1 = ir εia0 − 2b0 , and bn1 = ir
2a0 + εib0 .
Proof. Suppose
f ∈ A3 . By Lemma 6.9, after a suitable normalization, there exists a transformation
x y
H = y −x ∈ O2 (C) − SO2 (C) such that
1 ⊗n
⊗n
f = H ⊗n [ α1 ] + ir −α
for some r ∈ {0, 1, 2, 3}. Since H ∈ O2 (C), we have x2 + y 2 = 1. By Lemma 6.7, θ(f ) = − 12 , which implies
a0 a1 +b0 b1
√i
a0 b1 −a1 b0 = ± 2 . After rearranging terms, we get
a1
√
√
2a0 + εib0 = b1 εia0 − 2b0 ,
for some√ε ∈ {1, −1}. Since v0 √
and v1 are linearly independent,√we know that a1 and b√
1 cannot both be 0.
Also, if 2a0 + εib0 and εia0 − 2b0 are both 0, then we have − 2a0 = εib0 and εia0 = 2b0 , which implies
a0 = b0 = 0, a contradiction. Therefore, we have
√
√
a1 = c(εia0 − 2b0 )
and
b1 = c( 2a0 + εib0 )
(4)
for some c 6= 0. To prove necessity, it remains to show that cn is a power of i.
Now using H −1 = H, we have two expressions for (H −1 )⊗n f , which are
h
xa0 +yb0
ya0 −xb0
i⊗n
+
h
xa1 +yb1
ya1 −xb1
i⊗n
= H ⊗n
a0 ⊗n
b0
+
a1 ⊗n
b1
= H −1
⊗n
f = [ α1 ]
⊗n
+ ir
1 ⊗n
−α
.
By Lemma 6.5, there are two cases to consider, each of which has two more cases depending on ε.
1. Suppose ya0 − xb0 = α(xa0 + yb0 ), ya1 − xb1 = −α(xa1 + yb1 ), (xa0 + yb0 )n = 1, and (xa1 + yb1 )n = ir .
By rearranging the first two equations, we get
(y − αx)a0 = (x + αy)b0
and
(y + αx)a1 = (x − αy)b1 .
(5)
It cannot
√ be the case that a0 = b0 = 0 or y − αx = x + αy = 0. If a0 = 0, then
√ x + αy = 0, so
a1 = − 2ib1 by (5) and y 6= 0 lest x = 0 as well. If b0 = 0, then y − αx = 0, so 2ia1 = b1 , by the
same argument. Now we consider√the different cases√for ε.
√
(a) If ε = 1, then a1 = c(ia0 − 2b0 ) and b1 =√c( 2a0 + ib0 ) by (4). If a0 = 0, then a1 = −c√ 2b0
and b1 = cib0 , which
√ contradicts a1 = − 2ib1 ; if b0 = 0, then a1 = cia0 and b1 = c 2a0 ,
which contradicts 2ia1 = b1 . Thus, (y − αx)a0 = (x + αy)b0 6= 0 by (5). Also from (5),
(y + αx)a1 = (x − αy)b1 . Then since c 6= 0 and using (4) with ε = 1, we get
√
√
(y + αx) ia0 − 2b0 = (x − αy)
2a0 + ib0 .
Using (y − αx)a0 = (x + αy)b0 6= 0, we get
√
√
(y + αx) i(x + αy) − 2(y − αx) = (x − αy)
2(x + αy) + i(y − αx) .
This equation simplifies to x2 + y 2 = 0, which is a contradiction.
21
√
2b0 ) and b1 = c( 2a0 − ib0 ), from (4). Then we get
√
√
xa1 + yb1 = xc −ia0 − 2b0 + yc
2a0 − ib0
√
= c −i(xa0 + yb0 ) + 2(ya0 − xb0 )
(b) If ε = −1, then a1 = c(−ia0 −
√
= c(xa0 + yb0 ),
where in the third step, we used ya0 − xb0 = α(xa0 + yb0 ) from (5). Raising this equation to the
nth power and using (xa0 + yb0 )n = 1 and (xa1 + yb1 )n = ir , we conclude that cn = ir .
2. Suppose ya0 − xb0 = −α(xa0 + yb0 ), ya1 − xb1 = α(xa1 + yb1 ), (xa0 + yb0 )n = ir , and (xa1 + yb1 )n = 1.
Now we consider the different cases
√
√ for ε.
(a) If ε = 1, then a1 = c(ia0 − 2b0 ) and b1 = c( 2a0 + ib0 ) by (4). Using similar reasoning to that
in case 1b leads to (−c)n ir = 1,√so cn is a power of
√i.
(b) If ε = −1, then a1 = c(−ia0 − 2b0 ) and b1 = c( 2a0 − ib0 ) by (4). Using similar reasoning to
that in case 1a leads to a contradiction.
For sufficiency, suppose the three equations hold for some ε ∈ {1, −1} and some r ∈ {0, 1, 2, 3}. Further
assume ε = 1, in which case, the equations are
√
√
a1
(6)
2a0 + ib0 = b1 ia0 − 2b0 ,
as well as
From (6), we have
√ n
an1 = ir ia0 − 2b0
and
a1 = c(ia0 −
and
bn1 = ir
√
n
2a0 + ib0 .
(7)
√
b1 = c( 2a0 + ib0 )
(8)
√
√
for some c ∈ C. In (6), a1 , b1 cannot be both zero. Similarly, 2a0 + ib0 , ia0 − 2b0 cannot be both zero.
Thus at least one equation in (8) has both sides nonzero and we can always find some c even if one factor is
zero. We can write (8) as
h
√ i
a1
a0
i − 2
b0 .
b1 = c √2 i
√
2b0 )
This implies that a0 a1 + b0 b1 = ci(a20 + b20 ). Using (7) or (8), whichever equation is not zero on both sides,
1
2
2
n
r
wehave
c = i . Since (6) implies θ(f ) = − 2 , we know that a0 + b0 6= 0 because otherwise v0 is a multiple
1
of ±i , which makes θ(f ) = −1 regardless of v1 .
1 α
b0
1
We now define two orthogonal matrices T1 = √1+i
and T2 = √ 21 2 ab00 −a
. Also let T =
−α 1
0
a0 +b0
a0 ⊗n a1 ⊗n
, we want to calculate T ⊗n f . First,
+ b1
T1 T2 ∈ O2 (C). For f = b0
T2
where γ =
q
a20 +b20
1+i .
b0
=
q
a20 + b20 [ 10 ]
Furthermore, a1 b0 − a0 b1 =
T2
It follows that
T
Thus
a0
a1
b1
a1
b1
=p
= cγ
1
a20 + b20
1 α
−α 1
h
T ⊗n f = γ n
and
a0
b0
=γ
1
−α
,
√
√
2i(a0 a1 + b0 b1 ) = −c 2(a20 + b20 ) by (6) and (8). Then
a0 a1 +b0 b1
a1 b0 −a0 b1
i
√
− 2
T
i
= cγ
1 ⊗n
−α
22
=c
h
q
i
h
i
a20 + b20 −√
2 .
i
√
i− 2α
√
−iα− 2
+ (−c)n [ α1 ]
= −cγ [ α1 ] .
⊗n
.
So T transforms
x y f into the canonical form of A3 . If we write out the orthogonal transformation T explicitly,
then T = y −x where
b0 − αa0
y= p
.
(i + 1) (a20 + b20 )
√
√
When ε = −1, the argument is similar. In this case, a1 = c(−ia0 − 2b0 ) and b1 = c( 2a0 − ib0 ) for
some c ∈ C satisfying cn = ir and the entries of T are
a0 + αb0
x= p
(i + 1) (a20 + b20 )
a0 − αb0
x= p
(i + 1) (a20 + b20 )
and
and
b0 + αa0
y= p
.
(i + 1) (a20 + b20 )
√
√
√
√
Remark: Notice that either a1 ( 2a0 + ib0 ) = b1 (ia0 − 2b0 ) or a1 ( 2a0 − ib0 ) = b1 (−ia0 − 2b0 ) implies
a
a
θ(f ) = − 12 , unless det( b00 b11 ) = 0.
As mentioned before, A2 = P2 requires a stronger condition than just θ. If f ∈ A2 = P2 , then
θ(f ) = −1, but the reverse is not true. If f = v0⊗n + v1⊗n with v0 = [1, i] and v1 is not a multiple of [1, −i],
then θ(f ) = −1 but f is not in A2 = P2 , since any orthogonal H fixes {[1, i], [1, −i]} set-wise, up to a scalar
multiple.
The next lemma, which appeared in [12], gives a characterization of A2 . It says that any signature in A2
is essentially in canonical form. For completeness, we include its proof.
Lemma
6.13 ([12]). Let f be a non-degenerate symmetric signature. Then f ∈ A2 iff f is of the form
1 ⊗n
⊗n
1
for some c, β 6= 0.
c [ i ] + β −i
1 ⊗n
⊗n
Proof. Assume that f = c [ 1i ] + β −i
for some c, β 6= 0. Consider the orthogonal transformation
1
1
a b
1
1
1
1
−
β 2n − β − 2n . We pick a and b in this way so that
H = b −a , where a = 2 β 2n + β 2n and b = 2i
n
1
1
a+bi
a + bi = β 2n , a − bi = β − 2n , and (a + bi)(a − bi) = a2 + b2 = 1. Also a−bi
= β. Then
a−bi ⊗n
+ β ai+b
1 ⊗n
⊗n
= c (a + bi)n −i
+ (a − bi)n β [ 1i ]
p ⊗n
⊗n
1
= c β −i
,
+ [ 1i ]
H ⊗n f = c
so f can be written as
f =c
a+bi ⊗n
−ai+b
p
⊗n
1 ⊗n
.
β(H −1 )⊗n −i
+ [ 1i ]
Therefore f ∈ A2 .
1
⊗n
⊗n
On the other hand, the desired form f = c([ 1i ] + β [ 1i ] ) follows from the fact that {[ 1i ] , −i
} is
fixed setwise under any orthogonal transformation up to nonzero constants.
Remark: Notice
that θ(v0 , v1 ) = −1 for linearly independent v0 and v1 if and only if at least one of v0 , v1
1
is [ 1i ] or −i
, up to a nonzero scalar.
We now present the polynomial-time algorithm to check if f ∈ A1 ∪ A2 ∪ A3 .
Lemma 6.14. Given a non-degenerate symmetric signature f of arity at least 3, there is a polynomial-time
algorithm to decide whether f ∈ Ak for each k ∈ {1, 2, 3}. If so, k is unique and at least one corresponding
orthogonal transformation can be found in polynomial time.
23
Proof. First we check if f satisfies a second order recurrence relation. If it does, then the coefficients
(a, b, c) of the second order recurrence relation are unique up to a nonzero scalar by Lemma 6.1. If the
coefficients satisfy b2 − 4ac 6= 0, then by Lemma 6.2, we can express f as v0⊗n + v1⊗n , where v0 and v1 are
linearly independent and arity(f ) = n. All of this must be true for f to be in A1 ∪ A2 ∪ A3 . With this
alternate expression for f , we apply Lemma 6.11, Lemma 6.13, and Lemma 6.12 to decide if f ∈ Ak for each
k ∈ {1, 2, 3} respectively. These sets are disjoint by Lemma 6.7, so there can be at most one k such that
f ∈ Ak .
6.2. Set of Symmetric Signatures
We first show that if a non-degenerate signature f of arity at least 3 is in A1 or A3 , then for any set
F containing f , there are only a small constant number of transformations to check to decide whether F is
A -transformable. If f ∈ A2 , then there can be more than a constant number of transformations to check.
However, this number is at most linear in the arity of f .
Notice that any non-degenerate symmetric signature f ∈ A of arity at least 3 is in F123 (introduced
in Section 2.3), which contains signatures expressed as a sum of two tensor powers. Therefore θ(f ) is welldefined. By Lemma 2.7, to check A -transformability, we may restrict our attention to the sets A and
[ 10 α0 ] A up to orthogonal transformations. In particular,
if f ∈ F1 ∪ F2 ∪ [ 10 α0 ] F1 ,
0
θ(f ) = −1 if f ∈ F3 ,
(9)
1
1
0
− 2 if f ∈ [ 0 α ] (F2 ∪ F3 ).
Lemma 6.15. Let F be a set of symmetric signatures and suppose F contains a non-degeneratesignature
1
f ∈ A1 of arity
n ≥ 3 with H ∈ O2 (C). Then F is A -transformable iff F is a subset of HA , or H 11 −1
A,
1
or H 11 −1
[ 10 α0 ] A .
1
∈ O2 (C).
Proof. Sufficiency follows from Lemma 2.7 and both H, H2 = √12 11 −1
Before we prove necessity, we first claim that without loss of generality, we may assume H ∈ O2 (C) \
e = H [ 0 1 ] ∈ O2 (C) \ SO2 (C). Then f ∈ A1 also with H.
e From
SO2 (C). If H ∈ SO2 (C), we let H
10
1 1 1 0
1 1 1 0 1 0
1 −1 1 0
0
1
0
1
e
[ ] ∈ Stab(A ), it follows that HA = HA . Also [ 1 0 ] 1 −1 [ 0 α ] = 1 1 [ 0 α ] = 1 −1 0 −1 [ 0 α ] =
11 0 1 1 0 1 0
1 0
10
1 1 1 0
e 1 1
1 −1 [ 0 α ] 0 −1 , and 0 −1 ∈ Stab(A ). It follows that H 1 −1 [ 0 α ] A = H 1 −1 [ 0 α ] A .
Suppose F is A -transformable. By Lemma 4.3, there exists an H ′ ∈ SO2 (C) such that F ⊆ H ′ A or
F ⊆ H ′ [ 10 α0 ] A . We only need to
M ∈ Stab(A ), such that H ′ = HM in the first case,
1 show
there exists1 an
1
′
and in the second case H = H 1 −1 M , and M [ 0 α0 ] = [ 10 α0 ] M ′ for some M ′ ∈ Stab(A ).
Since f ∈ A1 with H, after a suitable normalization by a nonzero scalar, we have
1 ⊗n
⊗n
,
f = H ⊗n [ 11 ] + β −1
where β = αtn+2r for some r ∈ {0, 1, 2, 3} and t ∈ {0, 1}. Let g = (H ′−1 )⊗n f and T = H ′−1 H so that
1 ⊗n
⊗n
.
g = T ⊗n [ 11 ] + β −1
′
Note that T ∈ O
H ∈ O2 (C) \ SO2 (C). Thus T = T −1 and HT = H ′ .
2 (C) \ SO2 (C) since H ∈ SO22(C) and
b
Let T = ab −a
for some a, b ∈ C such that a + b2 = 1. There are two possibilities according to whether
F ⊆ H ′ A or F ⊆ H ′ [ 10 α0 ] A .
1. If F ⊆ H ′ A , then g ∈ F123 since g is symmetric and non-degenerate. Since θ(g) = 0, by (9), g ∈ F1
or g ∈ F2 . We discuss the two cases of g separately.
• Suppose g ∈ F1 . Then we have
1 ⊗n
⊗n
⊗n
⊗n
= λ [ 10 ] + it [ 01 ]
T ⊗n [ 11 ] + β −1
24
for some λ 6= 0 and t ∈ {0, 1, 2, 3}. Plugging in the expression for T , we have
a−b ⊗n
a+b ⊗n
1 ]⊗n + it [ 0 ]⊗n .
+
β
=
λ
[
0
1
a+b
b−a
Then by Lemma
b = 0 or a− b =
Together
with a2 + b2 = 1, we can solve
1 1 6.5, we have
0.
a 1+ −1
1
1
1
1 1
0 −1
√
√
√
for T = 2 1 −1 or T = 2 −1 −1 = 2 1 −1 1 0 , up to a constant multiple ±1. Since
0 −1
∈ Stab(A ), we have T ∈ Stab(A ), so we are done.
1 0
• Suppose g ∈ F2 . Then we have
1 ⊗n
1 ⊗n
⊗n
⊗n
= λ [ 11 ] + it −1
T ⊗n [ 11 ] + β −1
for some λ 6= 0 and t ∈ {0, 1, 2, 3}. Plugging in the expression for T , we have
1 ⊗n
a−b ⊗n
a+b ⊗n
1 ]⊗n + it
=
λ
[
+
β
−1
1
a+b
b−a
Then by Lemma
b = a − b or a + b = −(a − b). Therefore either a = 0 or b = 0.
0 6.5,
we have a0 +
Thus T = ± 10 −1
or T = ± [ 1 10 ] and both matrices are in Stab(A ).
2. If F ⊆ H ′ [ 10 α0 ] A , then we have g ∈ [ 10 α0 ] F123 . Since θ(g) = 0, by (9), g ∈ [ 10 α0 ] F1 . That is,
1 ⊗n
⊗n
⊗n
⊗n
⊗n
⊗n
⊗n
= λ [ 10 ] + it αn [ 01 ]
[ 10 ] + it [ 01 ]
= λ [ 10 α0 ]
T ⊗n [ 11 ] + β −1
for some λ 6= 0. This is essentially the same as the case where g ∈ F1 above, except that the coefficients
are different.
this
1 However, the
case that
0 in
coefficients
0 −1 do not affect the argument and our conclusion
−1
1
or T = √12 11 −1
T = √12 11 −1
∈ Stab(A ).
,
up
to
a
constant
multiple
±1.
Notice
that
1 0
1 0
Moreover,
0 −1 1 0
[ 0 α ] = 10 −α
0
1 0
= [ 10 α0 ] α0−1 −α
0
= −α [ 10 α0 ] [ 0i 01 ] ,
and [ 0i 01 ] ∈ Stab(A ).
Lemma 6.16. Let F be a set of symmetric signatures and suppose F contains a non-degenerate signature
f ∈ A2 of arity n ≥ 3. Then there exists a set H ⊆ O2 (C) of size O(n) such that F is A -transformable iff
there exists an H ∈ H such that F ⊆ HA . Moreover H can be computed in polynomial time in the input
length of the symmetric signature f .
Proof. Sufficiency is trivial by Lemma 4.3.
Suppose F is A -transformable. By Lemma 4.3, there exists an H ∈ SO2 (C) such that F ⊆ HA or
F ⊆ H [ 10 α0 ] A . In the first case, we show that the number of choices of H can be limited to O(n). Then we
show that the second case is impossible.
Since f ∈ A2 , after a suitable normalization by a nonzero scalar, we have
f = [ 1i ]
⊗n
+ν
1 ⊗n
−i
for some ν 6= 0 by Lemma 6.13. Let g = (H −1 )⊗n f . Then
1 ⊗n
⊗n
.
g = (H −1 )⊗n [ 1i ] + ν −i
There are two possibilities according to whether F ⊆ HA or F ⊆ H [ 10 α0 ] A .
25
1. Suppose F ⊆ HA . Therefore g ∈ F123 . Since θ(g) = −1, by (9), g ∈ F3 . Then we have
1 ⊗n
1 ⊗n
⊗n
⊗n
= λ [ 1i ] + ir −i
(H −1 )⊗n [ 1i ] + ν −i
−1
−1
is of the form
λ 6= 0 2and 2r ∈ {0, 1, 2, 3}. Because H ∈ SO2 (C), we may assume that H
fora some
b
−b a where a + b = 1. Therefore
1 ⊗n
1 ⊗n a b ⊗n 1 ⊗n
⊗n
[ i ] + ν −i
= −b a
λ [ 1i ] + ir −i
1 ⊗n
⊗n
= (a + bi)n [ 1i ] + ν(a − bi)n −i
.
Comparing the coefficients, by Lemma 6.5, we have
λ = (a + bi)n
and
λir = ν(a − bi)n .
Hence,
ir (a + bi)n = ν(a − bi)n .
Since (a + bi)(a − bi) = a2 + b2 = 1, we know that (a + bi)2n = νi−r . Therefore a + bi = ω2n (νi−r )1/2n ,
where ω2n is a 2n-th root of unity. There are 4 choices for r, and 2n choices for ω2n . However,
1
a − bi = a+bi
, and (a, b) can be solved from (a + bi, a − bi). Hence there are only O(n) choices for H,
depending on f .
2. Suppose F ⊆ H [ 10 α0 ] A . Then g ∈ [ 10 α0 ] F123 . However, θ(g) = −1, which contradicts (9).
Lemma 6.17. Let F be a set of symmetric signatures and suppose F contains a non-degenerate signature
f ∈ A3 of arity n ≥ 3 with H ∈ O2 (C). Then F is A -transformable iff F ⊆ H [ 10 α0 ] A .
Proof. Sufficiency is trivial by Lemma 4.3.
Suppose F is A -transformable. As in the proof of Lemma 6.15, we may assume that H ∈ O2 (C)\SO2 (C).
By Lemma 4.3, there exists an H ′ ∈ SO2 (C) such that F ⊆ H ′ A or F ⊆ H ′ [ 10 α0 ] A . We show the first
case is impossible. Then in the second case, we show that there exists an M such that H ′ = HM , where
M [ 10 α0 ] = [ 10 α0 ] M ′ for some M ′ ∈ Stab(A ).
Since f ∈ A3 with H, after a suitable normalization by a nonzero scalar, we have
1 ⊗n
⊗n
f = H ⊗n [ α1 ] + ir −α
for some r ∈ {0, 1, 2, 3}. Let g = (H ′−1 )⊗n f and T = H ′−1 H so that
1 ⊗n
⊗n
.
g = T ⊗n [ α1 ] + ir −α
′
Note that T ∈ O
H ∈ O2 (C) \ SO2 (C). Thus T = T −1 and HT = H ′ .
2 (C) \ SO2 (C) since H ∈ SO22(C) and
a b
2
Let T = b −a for some a, b ∈ C such that a + b = 1. There are two possibilities according to whether
F ⊆ H ′ A or F ⊆ H ′ [ 10 α0 ] A .
1. Suppose F ⊆ H ′ A . Then g = (H ′−1 )⊗n f ∈ F123 . However, θ(g) = − 21 , which contradicts (9).
2. Suppose F ⊆ H ′ [ 10 α0 ] A . Then g ∈ [ 10 α0 ] F123 , so θ(g) = − 21 and g ∈ [ 10 α0 ] (F2 ∪ F3 ) by (9). We
discuss the these two cases separately.
• Suppose g ∈ [ 10 α0 ] F2 . Then we have
1 ⊗n
1 ⊗n
⊗n
⊗n
⊗n
[ 11 ] + it −1
= λ [ 10 α0 ]
T ⊗n [ α1 ] + ir −α
1 ⊗n
⊗n
= λ [ α1 ] + it −α
for some λ 6= 0 and t ∈ {0, 1, 2, 3}. Plugging in the expression for T , we have
1 ⊗n
⊗n
⊗n
a+αb ⊗n
.
= λ [ α1 ] + it −α
+ ir a−αb
b+αa
b−αa
26
Then by Lemma 6.5, we have either
b − aα = α(a + bα)
and
b + aα = −α(a − bα)
or
b − aα = −α(a + bα)
and
b + aα = α(a − bα).
The first
case
is impossible. In the second case,1 0we have a = ±1 and b = 0. This implies
0
T = ± 10 −1
∈ Stab(A ), which commutes with [ 0 α ].
• Suppose g ∈ [ 10 α0 ] F3 . Then we have
1 ⊗n
1 ⊗n
⊗n
⊗n
⊗n
[ 1i ] + it −i
= λ [ 10 α0 ]
T ⊗n [ α1 ] + ir −α
1 ⊗n
1 ]⊗n + it
= λ [ αi
−αi
for some λ 6= 0 and t ∈ {0, 1, 2, 3}. Plugging in the expression for T , we have
1 ⊗n
a−αb ⊗n
a+αb ⊗n
1 ]⊗n + it
.
= λ [ αi
+ ir b+αa
−αi
b−αa
Then by Lemma 6.5, we have either
b − aα = αi(a + bα)
and
b + aα = −αi(a − bα)
or
b − aα = −αi(a + bα)
and
b + aα = αi(a − bα).
The first case is impossible. In the second
b = ±1. This implies that
0 and
we have
0 aα=
0case,
α
−1 0 i
=
α
[
and
T = ± [ 01 10 ]. Note that [ 01 10 ] [ 10 α0 ] = [ 10 α0 ] α−1
−1
1 0 ] ∈ Stab(A ).
α
0
0
Now we are ready to show how to decide if a finite set of signatures is A -transformable. To avoid
trivialities, we assume F contains a non-degenerate signature of arity at least 3. If every non-degenerate
signature in F has arity at most two, then Holant(F ) is tractable.
Theorem 6.18. There is a polynomial-time algorithm to decide, for any finite input set F of symmetric
signatures containing a non-degenerate signature f of arity n ≥ 3, whether F is A -transformable.
Proof. By Lemma 6.14, we can decide if f is in Ak for some k ∈ {1, 2, 3}. If not, then by Lemma 2.9, F
is not A -transformable. Otherwise, f ∈ Ak for some unique k. Depending on k, we apply Lemma 6.15,
Lemma 6.16, or Lemma 6.17 to check if F is A -transformable.
7. Symmetric P-transformable Signatures
To decide if a signature set is P-transformable, we face the same issue as in the A -transformable case.
Namely, a symmetric signature of arity n is given by n + 1 values, instead of 2n values. This exponentially
more succinct representation requires us to find a more efficient algorithm.
The next lemma tells us how to decide membership in P1 for signatures of arity at least 3.
Lemma 7.1. Let f = v0⊗n + v1⊗n be a symmetric signature of arity n ≥ 3, where v0 and v1 are linearly
independent. Then f ∈ P1 iff θ(f ) = 0.
Proof. Necessity is clear by Lemma 6.7 and sufficiency follows from Lemma 6.10.
Since A2 = P2 , the membership problem for P2 is handled by Lemma 6.13. Using Lemma 7.1 and
Lemma 6.13, we can efficiently decide membership in P1 ∪ P2 .
27
Lemma 7.2. Given a non-degenerate symmetric signature f of arity at least 3, there is a polynomial-time
algorithm to decide whether f ∈ Pk for some k ∈ {1, 2}. If so, k is unique and at least one corresponding
orthogonal transformation can be found in polynomial time.
Proof. First we check if f satisfies a second order recurrence relation. If it does, then the coefficients (a, b, c)
of the second order recurrence relation are unique up to a nonzero scalar by Lemma 6.1. If the coefficients
satisfy b2 − 4ac 6= 0, then by Lemma 6.2, we can express f as v0⊗n + v1⊗n , where v0 and v1 are linearly
independent and arity(f ) = n. All of this must be true for f to be in P1 ∪ P2 . With this alternate
expression for f , we apply Lemma 7.1 and Lemma 6.13 to decide if f ∈ Pk for some k ∈ {1, 2} respectively.
These sets are disjoint by Lemma 6.7, so there can be at most one k such that f ∈ Pk .
Like the symmetric affine case, the following lemmas assume the signature set F contains a non-degenerate
signature of arity at least 3 in P1 or P2 . Unlike the symmetric affine case, the number of transformations
to be checked to decide whether F is P-transformable is always a small constant.
Lemma 7.3. Let F be a set of symmetric signatures and suppose F contains a non-degenerate
signature
1
f ∈ P1 of arity n ≥ 3 with H ∈ O2 (C). Then F is P-transformable iff F ⊆ H 11 −1
P.
Proof. Sufficiency is trivial by Lemma 2.10.
Suppose F is P-transformable. As in the proof of Lemma 6.15, we may assumeH ∈O2 (C) \ SO2 (C).
1
P, where in the
Then by Lemma 5.1, there exists an H ′ ∈ SO2 (C) such that F ⊆ H ′ P or F ⊆ H ′ 1i −i
′
second case
we
can
take
H
=
I
.
In
the
first
case,
we
show
that
there
exists
an
M
∈
Stab(P)
such that
2
1
H ′ = H 11 −1
M . Then we show that the second case is impossible.
Since f ∈ P1 with H, after a suitable normalization by a nonzero scalar, we have
1 ⊗n
⊗n
f = H ⊗n [ 11 ] + β −1
for some β 6= 0. Let g = (H ′−1 )⊗n f and T = H ′−1 H so that
1 ⊗n
⊗n
.
g = T ⊗n [ 11 ] + β −1
Note that T ∈ O2 (C) \ SO2 (C) since H ′ ∈ SO2 (C) and H ∈ O2 (C) \ SO2 (C). Thus T = T −1 and HT = H ′ .
1. Suppose F ⊆ H ′ P. Then g must be a generalized equality since g ∈ P with arity n≥ 3. The only sym
⊗n
⊗n
metric non-degenerate generalized equalities in P with arity n ≥ 3 have the form λ [ 10 ] + β ′ [ 01 ]
,
′
for some λ, β 6= 0. Thus
1 ⊗n
⊗n
⊗n
⊗n
T ⊗n [ 11 ] + β −1
= λ [ 10 ] + β ′ [ 01 ]
.
Let T =
a
b
b −a
for a, b ∈ C such that a2 + b2 = 1. Then
a−b ⊗n
a+b ⊗n
1 ]⊗n + β ′ [ 0 ]⊗n .
=
λ
[
+
β
0
1
a+b
b−a
2
2
= 0 or a + b = 0. Together
By Lemma 6.5 we
solutions are
have either1 a−1b −1
0 −1 with a +1b = 1, 1the 0only
1
1
1
1
1
1
∈ Stab(P),
T = ± √2 1 −1 or T = ± √2 −1 −1 = ± √2 1 −1 1 0 . Since ± √2 I2 , ± √2 1 −1
0
this case is complete.
1
1
1
2. Suppose F ⊆ H ′ 1i −i
Then
g ∈ 1i −i
P, and θ(g) = θ([ 11 ] , −1
) = 0 by Lemma 6.7.
P.
1
However, any h ∈ 1i −i
P that is non-degenerate and has arity at least 3 must have the form
1 ⊗n
⊗n
c [ 1i ] + d −i
for some nonzero c, d ∈ C, which implies that θ(h) = −1. This contradicts θ(g) =
0.
Lemma 7.4. Let F be a set of symmetric signatures and suppose F contains a non-degenerate signature
f ∈ P2 of arity n ≥ 3. Then F is P-transformable iff all non-degenerate signatures in F are contained in
P2 ∪ {=2 }.
28
1
. Then by Lemma 5.1, F ⊆ ZP or there exists an
Proof. Suppose F is P-transformable. Let Z = √12 1i −i
H ∈ SO2 (C) such that F ⊆ HP. In first case, we show that all the non-degenerate symmetric signatures
in ZP are contained in P2 ∪ {=2 }. Then we show that the second case is impossible.
1. Suppose F ⊆ ZP. Let g ∈ ZP be a symmetric non-degenerate signature of arity m. If (Z −1 )⊗2 g =
λ[0, 1, 0] is the binary disequality signature up to a nonzero scalar λ ∈ C, then
0
1
g = λZ ⊗2 11 = λ 00
0
1
is the binary equality signature =2 . Otherwise, we can express g as
⊗m
⊗m
+ β [ 01 ]
g = cZ ⊗m [ 10 ]
1 ⊗m
⊗m
+ β −i
= c [ 1i ]
for some c, β 6= 0 with m ≥ 2. Thus, g ∈ P2 = A2 by Lemma 6.13. We conclude that the symmetric
non-degenerate subset of ZP is contained in P2 ∪ {=2 }. Therefore, the non-degenerate subset of F
is contained in P2 ∪ {=2 }.
2. Suppose F ⊆ HP. By assumption, F contains f ∈ P2 = A2 of arity n ≥ 3. After a suitable
normalization by a scalar, we have
1 ⊗n
⊗n
f = [ 1i ] + β −i
for some β 6= 0 by Lemma 6.13. Let g = (H −1 )⊗n f so that
1 ⊗n
⊗n 1 ⊗n
.
[ i ] + β −i
g = H −1
1
) = −1 since
In particular, f and g have the same arity n ≥ 3. By Lemma 6.7, θ(g) = θ([ 1i ] , −i
⊗n
c ⊗n
−1
0
for some nonzero c, d ∈ C, which
H ∈ O2 (C). However, g ∈ P must be of the form [ 0 ] + [ d ]
has θ(g) = 0. This is a contradiction.
It is easy to see that all of above is reversible. Therefore sufficiency follows.
Now we are ready to show how to decide if a finite set of signatures is P-transformable. To avoid
trivialities, we assume F contains a non-degenerate signature of arity at least 3. If every non-degenerate
signature in F has arity at most two, then Holant(F ) is tractable.
Theorem 7.5. There is a polynomial-time algorithm to decide, for any finite input set F of symmetric
signatures containing a non-degenerate signature f of arity n ≥ 3, whether F is P-transformable.
Proof. By Lemma 7.2, we can decide if f is in Pk for some k ∈ {1, 2}. If not, then by Lemma 2.12, F
is not P-transformable. Otherwise, f ∈ Pk for some unique k. Depending on k, we apply Lemma 7.3 or
Lemma 7.4 to check if F is P-transformable.
Acknowledgements. This research is supported by NSF CCF-1714275. HG was also supported by a Simons
Award for Graduate Students in Theoretical Computer Science from the Simons Foundation when he was a
graduate student. We thank anonymous referees for their valuable comments.
References
[1] Andrei Bulatov, Martin Dyer, Leslie Ann Goldberg, Markus Jalsenius, and David Richerby. The complexity of weighted Boolean #CSP with mixed signs. Theor. Comput. Sci., 410(38-40):3949–3961, 2009.
[2] Andrei Bulatov and Martin Grohe. The complexity of partition functions. Theor. Comput. Sci.,
348(2):148–186, 2005.
29
[3] Andrei A. Bulatov. The complexity of the counting constraint satisfaction problem. J. ACM, 60(5):34:1–
34:41, 2013.
[4] Andrei A. Bulatov and Vı́ctor Dalmau. Towards a dichotomy theorem for the counting constraint
satisfaction problem. Inform. and Comput., 205(5):651–678, 2007.
[5] Jin-Yi Cai and Xi Chen. A decidable dichotomy theorem on directed graph homomorphisms with
non-negative weights. In FOCS, pages 437–446. IEEE Computer Society, 2010.
[6] Jin-Yi Cai and Xi Chen. Complexity of counting CSP with complex weights. J. ACM, 64(3):19:1–19:39,
2017.
[7] Jin-Yi Cai, Xi Chen, and Pinyan Lu. Graph homomorphisms with complex values: A dichotomy
theorem. SIAM J. Comput., 42(3):924–1029, 2013.
[8] Jin-Yi Cai, Xi Chen, and Pinyan Lu. Nonnegative weighted #CSP: An effective complexity dichotomy.
SIAM J. Comput., 45(6):2177–2198, 2016.
[9] Jin-Yi Cai, Vinay Choudhary, and Pinyan Lu. On the theory of matchgate computations. Theory of
Computing Systems, 45(1):108–132, 2009.
[10] Jin-Yi Cai and Aaron Gorenstein. Matchgates revisited. Theory of Computing, 10(868):4001–4030,
2014.
[11] Jin-Yi Cai, Heng Guo, and Tyson Williams. Holographic algorithms beyond matchgates. In ICALP
(1), pages 271–282, 2014.
[12] Jin-Yi Cai, Heng Guo, and Tyson Williams. A complete dichotomy rises from the capture of vanishing
signatures. SIAM J. Comput., 45(5):1671–1728, 2016.
[13] Jin-Yi Cai, Sangxia Huang, and Pinyan Lu. From Holant to #CSP and back: Dichotomy for Holantc
problems. Algorithmica, 64(3):511–533, 2012.
[14] Jin-Yi Cai and Michael Kowalczyk. Spin systems on k-regular graphs with complex edge functions.
Theor. Comput. Sci., 461:2–16, 2012.
[15] Jin-Yi Cai, Michael Kowalczyk, and Tyson Williams. Gadgets and anti-gadgets leading to a complexity
dichotomy. In ITCS, pages 452–467. ACM, 2012.
[16] Jin-Yi Cai and Pinyan Lu. Holographic algorithms with unsymmetric signatures. In SODA, pages
54–63. Society for Industrial and Applied Mathematics, 2008.
[17] Jin-Yi Cai and Pinyan Lu. Holographic algorithms: From art to science. J. Comput. Syst. Sci.,
77(1):41–61, 2011.
[18] Jin-Yi Cai and Pinyan Lu. Signature theory in holographic algorithms. Algorithmica, 61(4):779–816,
2011.
[19] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Computational complexity of Holant problems. SIAM J.
Comput., 40(4):1101–1132, 2011.
[20] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Dichotomy for Holant* problems of Boolean domain. In SODA,
pages 1714–1728. SIAM, 2011.
[21] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Holographic reduction, interpolation and hardness. Computational Complexity, 21(4):573–604, 2012.
[22] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Dichotomy for Holant* problems with domain size 3. In SODA,
pages 1278–1295. SIAM, 2013.
30
[23] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Holographic algorithms by Fibonacci gates. Linear Algebra and
its Applications, 438(2):690–707, 2013.
[24] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. The complexity of complex weighted Boolean #CSP. J. Comput.
System Sci., 80(1):217–236, 2014.
[25] Jin-Yi Cai, Pinyan Lu, and Mingji Xia. Holographic algorithms with matchgates capture precisely
tractable planar #CSP. SIAM J. Comput., 46(3):853–889, 2017.
[26] C. T. J. Dodson and T. Poston. Tensor Geometry, volume 130 of Graduate Texts in Mathematics.
Springer-Verlag, second edition, 1991.
[27] Martin Dyer, Leslie Ann Goldberg, and Mark Jerrum. The complexity of weighted Boolean #CSP.
SIAM J. Comput., 38(5):1970–1986, 2009.
[28] Martin Dyer, Leslie Ann Goldberg, and Mike Paterson. On counting homomorphisms to directed acyclic
graphs. J. ACM, 54(6), 2007.
[29] Martin Dyer and Catherine Greenhill. The complexity of counting graph homomorphisms. Random
Struct. Algorithms, 17(3-4):260–289, 2000.
[30] Martin Dyer and David Richerby. An effective dichotomy for the counting constraint satisfaction problem. SIAM J. Comput., 42(3):1245–1274, 2013.
[31] Leslie Ann Goldberg, Martin Grohe, Mark Jerrum, and Marc Thurley. A complexity dichotomy for
partition functions with mixed signs. SIAM J. Comput., 39(7):3336–3402, 2010.
[32] Heng Guo, Sangxia Huang, Pinyan Lu, and Mingji Xia. The complexity of weighted Boolean #CSP
modulo k. In STACS, pages 249–260. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2011.
[33] Heng Guo and Tyson Williams. The complexity of planar Boolean #CSP with complex weights. In
ICALP, pages 516–527. Springer Berlin Heidelberg, 2013.
[34] Pavol Hell and Jaroslav Nešetřil. On the complexity of H-coloring. J. Comb. Theory Ser. B, 48(1):92–
110, 1990.
[35] Sangxia Huang and Pinyan Lu. A dichotomy for real weighted Holant problems. Computational Complexity, 25(1):255–304, 2016.
[36] P. W. Kasteleyn. The statistics of dimers on a lattice. Physica, 27(12):1209–1225, 1961.
[37] P. W. Kasteleyn. Graph theory and crystal physics. In F. Harary, editor, Graph Theory and Theoretical
Physics, pages 43–110. Academic Press, London, 1967.
[38] Neeraj Kayal. Affine projections of polynomials: Extended abstract. In STOC, pages 643–662. ACM,
2012.
[39] László Lovász. Operations with structures. Acta Math. Hung., 18(3-4):321–328, 1967.
[40] Ketan D. Mulmuley and Milind Sohoni. Geometric complexity theory I: An approach to the P vs. NP
and related problems. SIAM J. Comput., 31(2):496–526, 2001.
[41] H. N. V. Temperley and Michael E. Fisher. Dimer problem in statistical mechanics—an exact result.
Philosophical Magazine, 6(68):1061–1063, 1961.
[42] Leslie G. Valiant. Expressiveness of matchgates. Theor. Comput. Sci., 289(1):457–471, 2002.
[43] Leslie G. Valiant. Quantum circuits that can be simulated classically in polynomial time. SIAM J.
Comput., 31(4):1229–1254, 2002.
31
[44] Leslie G. Valiant. Accidental algorthims. In FOCS, pages 509–517. IEEE Computer Society, 2006.
[45] Leslie G. Valiant. Holographic algorithms. SIAM J. Comput., 37(5):1565–1594, 2008.
32
| 8cs.DS
|
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
1
Event-based, 6-DOF Camera Tracking from
Photometric Depth Maps
arXiv:1607.03468v2 [cs.CV] 31 Oct 2017
Guillermo Gallego, Jon E.A. Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, Davide Scaramuzza
Abstract—Event cameras are bio-inspired vision sensors that output pixel-level brightness changes instead of standard intensity
frames. These cameras do not suffer from motion blur and have a very high dynamic range, which enables them to provide reliable
visual information during high-speed motions or in scenes characterized by high dynamic range. These features, along with a very low
power consumption, make event cameras an ideal complement to standard cameras for VR/AR and video game applications. With
these applications in mind, this paper tackles the problem of accurate, low-latency tracking of an event camera from an existing
photometric depth map (i.e., intensity plus depth information) built via classic dense reconstruction pipelines. Our approach tracks the
6-DOF pose of the event camera upon the arrival of each event, thus virtually eliminating latency. We successfully evaluate the method
in both indoor and outdoor scenes and show that—because of the technological advantages of the event camera—our pipeline works
in scenes characterized by high-speed motion, which are still unaccessible to standard cameras.
Index Terms—Event-based vision, Pose tracking, Dynamic Vision Sensor, Bayes filter, Asynchronous processing, Conjugate priors,
Low Latency, High Speed, AR/VR.
F
S UPPLEMENTARY M ATERIAL
Video of the experiments: https://youtu.be/iZZ77F-hwzs.
1
I NTRODUCTION
T
HE task of estimating a sensor’s ego-motion has important
applications in various fields, such as augmented/virtual reality (AR/VR), video gaming, and autonomous mobile robotics.
In recent years, great progress has been achieved using visual
information to fulfill such a task [1], [2], [3]. However, due to
some well-known limitations of traditional cameras (motion blur
and low dynamic-range), current visual odometry pipelines still
struggle to cope with high-speed motions or high dynamic range
scenarios. Novel types of sensors, called event cameras [4, p.77],
offer great potential to overcome these issues.
Unlike standard cameras, which transmit intensity frames at
a fixed framerate, event cameras, such as the Dynamic Vision
Sensor (DVS) [5], only transmit changes of intensity. Specifically,
they transmit per-pixel intensity changes at the time they occur,
in the form of a set of asynchronous events, where each event
carries the space-time coordinates of the brightness change (with
microsecond resolution) and its sign.
Event cameras have numerous advantages over standard cameras: a latency in the order of microseconds, a very high dynamic
range (140 dB compared to 60 dB of standard cameras), and very
low power consumption (10 mW vs 1.5 W of standard cameras).
Most importantly, since all pixels capture light independently, such
sensors do not suffer from motion blur.
It has been shown that event cameras transmit, in principle,
all the information needed to reconstruct a full video stream
•
The authors are with the Robotics and Perception Group, affiliated
with both the Dept. of Informatics of the University of Zurich and the
Dept. of Neuroinformatics of the University of Zurich and ETH Zurich,
Switzerland: http:// rpg.ifi.uzh.ch/ . This research was supported by the
National Centre of Competence in Research (NCCR) Robotics, the SNSFERC Starting Grant, the Qualcomm Innovation Fellowship, the DARPA
FLA program, and the UZH Forschungskredit.
Fig. 1: Sample application: 6-DOF tracking in AR/VR (Augmented or Virtual Reality) scenarios. The pose of the event camera
(rigidly attached to a hand or head tracker) is tracked from a
previously built photometric depth map (RGB-D) of the scene.
Positive and negative events are represented in blue and red,
respectively, on the image plane of the event camera.
[6], [7], [8], [9], which clearly points out that an event camera
alone is sufficient to perform 6-DOF state estimation and 3D
reconstruction. Indeed, this has been recently shown in [9], [10].
However, currently the quality of the 3D map built using event
cameras does not achieve the same level of detail and accuracy as
that of standard cameras.
Although event cameras have become commercially available
only since 2008 [11], the recent body of literature on these new
sensors1 as well as the recent plans for mass production claimed
by companies, such as Samsung and Chronocam2 , highlight that
1. https://github.com/uzh-rpg/event-based vision resources
2. http://rpg.ifi.uzh.ch/ICRA17 event vision workshop.html
orientation [deg]
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
60
40
20
EB roll
EB pitch
EB yaw
which is obtained by minimizing the Kullback-Leibler divergence
(Section 4.4). The result is a filter adapted to the asynchronous
nature of the event camera, which also incorporates an outlier
detector that weighs measurements according to their confidence
for improved robustness of the pose estimation. The approximation of the posterior distribution allows us to obtain a closedform solution to the filter update equations and has the benefit of
being computationally efficient. Our method can handle arbitrary,
6-DOF, high-speed motions of the event camera in natural scenes.
The paper is organized as follows: Section 2 reviews related literature on event-based ego-motion estimation. Section 3 describes
the operating principle of event cameras. Our proposed eventbased, probabilistic approach is described in Section 4, and it is
empirically evaluated on natural scenes in Section 5. Conclusions
are highlighted in Section 6.
GT roll
GT pitch
GT yaw
0
−20
12.9
13
13.1
13.2
time [s]
2
13.3
13.4
Fig. 2: High-speed motion sequence. Top left: image from a
standard camera, suffering from blur due to high-speed motion.
Top right: set of asynchronous events from a DVS in an interval
of 3 milliseconds, colored according to polarity. Bottom: estimated
poses using our event-based (EB) approach, which provides low
latency and high temporal resolution updates. Ground truth (GT)
poses are also displayed.
there is a big commercial interest in exploiting these new vision
sensors as an ideal complement to standard cameras for mobile
robotics, VR/AR, and video game applications.
Motivated by these recent developments, this paper tackles
the problem of tracking the 6-DOF motion of an event camera
from an RGB-D (i.e., photometric depth) map that has been
previously built via a traditional, dense reconstruction pipeline
using standard cameras or RGB-D sensors (cf. Fig. 1). This
problem is particularly important in both AR/VR and video game
applications, where low-power consumption and robustness to
high-speed motion are still unsolved. In these applications, we
envision that the user would first use a standard sensor to build
a high resolution and high quality map of the room, and then
the hand and head trackers would take advantage of an event
camera to achieve robustness to high-speed motion and low-power
consumption.
The challenges we address in this paper are two: i) event-based
6-DOF pose tracking from an existing photometric depth map; ii)
tracking the pose during very fast motions (still unaccessible to
standard cameras because of motion blur), as shown in Fig. 2. We
show that we can track the 6-DOF motion of the event camera
with comparable accuracy as that of standard cameras and also
during high-speed motion.
Our method is based on Bayesian filtering theory and has three
key contributions in the way that the events are processed: i)
event-based pose update, meaning that the 6-DOF pose estimate
is updated every time an event is generated, at microsecond time
resolution, ii) the design of a sensor likelihood function using a
mixture model that takes into account both the event generation
process and the presence of noise and outliers (Section 4.3),
and iii) the approximation of the posterior distribution of the
system by a tractable distribution in the exponential family,
2 R ELATED WORK
M OTION E STIMATION
ON
E VENT- BASED
E GO -
The first work on pose tracking with a DVS was presented in [12].
The system design, however, was limited to slow planar motions
(i.e., 3 DOF) and planar scenes parallel to the plane of motion
consisting of artificial B&W line patterns. The particle filter
pose tracker was extended to 3D in [13], where it was used in
combination with an external RGB-D sensor (depth estimation)
to build a SLAM system. However, a depth sensor introduces the
same bottlenecks that exist in standard frame-based systems: depth
measurements are outdated for very fast motions, and the depth
sensor is still susceptible to motion blur.
In our previous work [14], a standard grayscale camera was
attached to a DVS to estimate the small displacement between
the current event and the previous frame of the standard camera.
The system was developed for planar motion and artificial B&W
striped background. This was due to the sensor likelihood being
proportional to the magnitude of the image gradient, thus favoring
scenes where large brightness gradients are the source of most of
the event data. Because of the reliance on a standard camera, the
system was again susceptible to motion blur and therefore limited
to slow motions.
An event-based algorithm to track the 6-DOF pose of a DVS
alone and during very high-speed motion was presented in [15].
However, the method was developed specifically for artificial,
B&W line-based maps. Indeed, the system worked by minimizing
the point-to-line reprojection error.
Estimation of the 3D orientation of an event camera was
presented in [6], [16], [17], [18]. However, such systems are
restricted to rotational motions, and, thus, do not account for
translation and depth.
Contrarily to all previous works, the approach we present in
this paper tackles full 6-DOF motions, does not rely on external
sensors, can handle arbitrary fast motions, and is not restricted to
specific texture or artificial scenes.
Other pose tracking approaches have been published as part
of systems that address the event-based 3D SLAM problem. [10]
proposes a system with three interleaved probabilistic filters to
perform pose tracking as well as depth and intensity estimation.
The system is computationally intensive, requiring a GPU for realtime operation. The parallel tracking-and-mapping system in [9]
follows a geometric, semi-dense approach. The pose tracker is
based on edge-map alignment and the scene depth is estimated
without intensity reconstruction, thus allowing the system to run
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
3
log-intensity changes (1) instead of absolute intensity. The spatial
resolution of the DVS is 128×128 pixels, but newer sensors, such
as the Dynamic and Active-pixel VIsion Sensor (DAVIS) [22], the
color DAVIS (C-DAVIS) [23], and the Samsung DVS [24] have
higher resolution (640 × 480 pixels), thus overcoming current
limitations.
4
(a) The Dynamic Vision Sensor (DVS) from iniLabs.
(b) Visualization of the output of a DVS (event stream) while viewing
a rotating scene, which generates a spiral-like structure in space-time.
Events are represented by colored dots, from red (far in time) to blue
(close in time). Event polarity is not displayed. Noise is visible by
isolated points.
Fig. 3: An event camera and its output.
in real-time on the CPU. More recently, visual inertial odometry
systems based on event cameras have also been proposed, which
rely on point features [19], [20], [21].
3
Event-based vision constitutes a paradigm shift from conventional
(e.g., frame-based) vision. In standard cameras, pixels are acquired
and transmitted simultaneously at fixed rates; this is the case
of both global-shutter or rolling-shutter sensors. Such sensors
provide little information about the scene in the “blind time”
between consecutive images. Instead, event-based cameras such
as the DVS [11] (Fig. 3a) have independent pixels that respond
asynchronously to relative contrast changes. If I(u, t) is the
intensity sensed at a pixel u = (x, y)> of the DVS, an event
is generated if the temporal visual contrast (in log scale) exceeds
a nominal threshold Cth :
∆ ln I := ln I(u, t) − ln I(u, t − ∆t) ≷ Cth ,
Consider an event camera moving in a known static scene. The
map of the scene is described by a sparse set of reference images
r Nr
r
{Ilr }N
l=1 , poses {ξ l }l=1 , and depth map(s). Suppose that an initial
guess of the location of the event camera in the scene is also
known. The problem we face is that of exploiting the information
conveyed by the event stream to track the pose of the event camera
in the scene. Our goal is to handle arbitrary 6-DOF, high-speed
motions of the event camera in realistic (i.e., natural) scenes.
We design a robust filter combining the principles of Bayesian
estimation, posterior approximation, and exponential family distributions with a sensor model that accounts for outlier observations.
In addition to tracking the kinematic state of the event camera, the
filter also estimates some sensor parameters automatically (e.g.,
event triggering threshold Cth ) that would otherwise be difficult to
tune manually. 3
The outline of this section is as follows. First, the problem
is formulated as a marginalized posterior estimation problem in a
Bayesian framework. Then, the motion model and the measurement model (a robust likelihood function that can handle both
good events and outliers) are presented. Finally, the filter equations
that update the parameters of an approximate distribution to the
posterior probability distribution are derived.
4.1
E VENT C AMERAS
(1)
where ∆t is the time since the last event was generated at
the same pixel. Different thresholds may be specified for the
cases of contrast increase (Cth+ ) or decrease (Cth− ). An event
e = (x, y, t, p) conveys the spatio-temporal coordinates and sign
(i.e., polarity) of the brightness change, with p = +1 (ON-event:
∆ ln I > Cth+ ) or p = −1 (OFF-event: ∆ ln I < Cth− ). Events
are time-stamped with microsecond resolution and transmitted
asynchronously when they occur, with very low latency. A sample
output of the DVS is shown in Fig. 3b. Another advantage
of the DVS is its very high dynamic range (140 dB), which
notably exceeds the 60 dB of high-quality, conventional framebased cameras. This is a consequence of events triggering on
P ROBABILISTIC APPROACH
Bayesian Filtering
We model the problem as a time-evolving system whose state s
consists of the kinematic description of the event camera as well
as sensor and inlier/outlier parameters. More specifically,
2 >
s = (ξ c , ξ i , ξ j , Cth , πm , σm
) ,
(2)
where ξ c is the current pose of the sensor (at the time of the event,
t in (1)), ξ i and ξ j are two poses along the sensor’s trajectory that
are used to interpolate the pose of the last event at the same pixel
2
(time t − ∆t in (1)), Cth is the contrast threshold, and πm and σm
are the inlier parameters of the sensor model, which is explained
in Section 4.3.2.
Let the state of the system at time tk be sk , and let the
sequence of all past observations (up to time tk ) be o1:k , where
ok is the current observation (i.e., the latest event).
Our knowledge of the system state is contained in the posterior
probability distribution p(sk |o1:k ), also known as belief [25,
p.27], which is the marginalized distribution of the smoothing
problem p(s1:k |o1:k ). The Bayes filter recursively estimates the
system state from the observations in two steps: prediction and
correction. The correction step updates the posterior by:
p(sk |o1:k ) ∝ p(ok |sk )p(sk |o1:k−1 ),
(3)
3. Today’s event-based cameras, such as the DVS [11] or the DAVIS [22],
have almost a dozen tuning parameters that are neither independent nor linear.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
4.2
(t t)
Z(t t)
Reference
4.3
Event camera
(t t)
(t)
Fig. 4: Computation of the contrast (measurement function) by
transferring events from the event camera to a reference image.
For each event, the predicted contrast (13), ∆ ln I , used in
the measurement function (7) is computed as the log-intensity
difference (as in (1)) at two points on the reference image I r : the
points (12) corresponding to the same pixel u on the event camera,
at times of the event (tk and tk − ∆t).
where p(ok |sk ) is the likelihood function (sensor model) and we
used independence of the events given the state. The prediction
step, defined by
Z
p(sk |o1:k−1 ) = p(sk |sk−1 )p(sk−1 |o1:k−1 )dsk−1 ,
(4)
incorporates the motion model p(sk |sk−1 ) from tk−1 to tk .
We incorporate in our state vector not only the current event
camera pose ξ kc but also the other relevant poses for contrast
calculation (poses ξ ki , ξ kj in (2)), so that we may use the filter
to partially correct errors of already estimated poses. Past events
that are affected by the previous pose are not re-evaluated, but
future events that reference back to such time will have better
previous-pose estimates.
To have a computationally feasible filter, we approximate the
posterior (3) by a tractable distribution with parameters ηk−1 that
condense the history of events o1:k−1 ,
p(sk |o1:k ) ≈ q(sk ; ηk ).
(5)
Assuming a motion model with slowly varying zero-mean random
diffusion, so that most updates of the state are due to the events, the
recursion on the approximate posterior becomes, combining (3)(5),
q(sk ; ηk ) ≈ C p(ok |sk )q(sk ; ηk−1 )
Motion Model
The diffusion process leaves the state mean unchanged and propagates the covariance. How much process noise is added to the
evolving state is determined by the trace of the covariance matrix
(sum of the eigenvalues): each incoming event adds white noise
to the covariance diagonal, thus increasing its trace, up to some
allowed maximum. This works gracefully across many motion
speeds. More specifically, we used a maximum standard deviation
of 0.03 for poses parametrized in normalized coordinates (with
translation in units relative to the mean scene depth), to factor out
the metric scale in the diffusion process.
Z(t)
(t)
4
(6)
for some normalizing constant C . The approximate posterior q
is computed by minimization of the Kullback-Leibler (KL) divergence between both sides of (6). As tractable distribution we
choose one in the exponential family because they are very flexible
and have nice properties for sequential Bayes estimation. The KL
minimization gives the update equations for the parameters of the
approximate posterior.
Measurement Model
Here we elaborate on the choice of likelihood function p(ok |sk )
in (6) that is used to model the events. Our contributions are,
starting from an ideal sensor model, i) to define a dimensionless
implicit function based on the contrast residual to measure how
well the event camera pose and the a priori information (e.g., a
map of the scene) explain an event (Section 4.3.1), and ii) to
build upon such measurement function taking into account noise
and outliers, yielding a mixture model for the likelihood function
(Section 4.3.2).
4.3.1 Ideal Sensor Model
In a noise-free scenario, an event is triggered as soon as the
temporal contrast reaches the threshold (1). Such a measurement
would satisfy ∆ ln I − Cth = 0. For simplicity, let us assume
that the polarity has already been taken into account to select
the appropriate threshold Cth+ > 0 or Cth− < 0. Defining the
measurement function by
∆ ln I
− 1,
(7)
Cth
the event-generation condition becomes M = 0 in a dimensionless formulation. Assuming a prediction of the temporal contrast
is generated using the system state, ∆ ln I(sk ), then (7) depends
on both the system state and the observation, M (ok , sk ). More
precisely, denoting by
M :=
s̃ = (ξ c , ξ i , ξ j , Cth )> ,
(8)
the part of the state (2) needed to compute (7), we have
M (ok , s̃k ). The likelihood function that characterizes such an
ideal sensor model is
p(ok |sk ) = δ(M (ok , s̃k )),
(9)
where δ is the Dirac delta distribution.
All deviations from ideal conditions can be collectively modeled by a noise term in the likelihood function. Hence, a more
realistic yet simple choice than (9) that is also supported by
the bell-shaped form of the threshold variations observed in the
DVS [11] is a Gaussian distribution,
2
p(ok |sk ) = N (M (ok , s̃k ); 0, σm
).
(10)
Most previous works in the literature do not consider an
implicit measurement function (7) or Gaussian model (10) based
on the contrast residual. Instead, they use explicit measurement
functions that evaluate the goodness of fit of the event either in
the spatial domain (reprojection error) [12], [15] or in the temporal domain (event-rate error), e.g., image reconstruction thread
of [16], assuming Gaussian errors. Our measurement function (7)
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
is based on the event-generation process and combines in a scalar
quantity all the information contained in an event (space-time and
polarity) to provide a measure of its fit to a given state and a
priori information. However, models based on a single Gaussian
distribution (10) are very susceptible to outliers. Therefore, we opt
for a mixture model to explicitly account for them, as explained
next.
4.3.2 Resilient Sensor Model. Likelihood Function
Based on the empirical observation that there is a significant
amount of outliers in the event stream, we propose a likelihood
function consisting of a normal-uniform mixture model. This
model is typical of robust sensor fusion problems [26], where the
output of the sensor is modeled as a distribution that mixes a good
measurement (normal) with a bad one (uniform):
2
p(ok |sk ) = πm N (M (ok , s̃k ); 0, σm
)
(11)
+ (1 − πm ) U(M (ok , s̃k ); Mmin , Mmax ),
where πm is the inlier probability (and (1 − πm ) is the outlier
probability). Inliers are normally distributed around 0 with vari2
ance σm
. Outliers are uniformly distributed over a known interval
2
[Mmin , Mmax ]. The measurement parameters σm
and πm are
considered unknown and are collected in the state vector sk to
be estimated.
To evaluate M (ok , s̃k ), we need to compute the contrast
∆ ln I(s̃k ) in (7). We do so based on a known reference image I r
(and its pose) and both relevant event camera poses for contrast
calculation, as explained in Fig. 4. Assuming the depth of the
scene is known, the point u0 in the reference image corresponding
to the event location (u, t) in the event camera satisfies the
following equation (in calibrated camera coordinates):
u0 (t) = π TRC (t) π −1 u, Z(t) ,
(12)
where TRC (t) is the transformation from the event camera frame
at time t to the frame of the reference image, Z(t) represents the
scene structure (i.e., the depth of the map point corresponding to
u with respect to the event camera), π : R3 → R2 , (X, Y, Z) 7→
(X/Z, Y /Z) is the canonical perspective projection, and π −1 is
the inverse perspective projection. The transformation TRC (tk ) at
the time of the current event depends on the current estimate of the
event camera pose ξ c ≡ ξ(tk ) in (8); the poses ξ i ≡ ξ(ti ) and
ξ j ≡ ξ(tj ) along the event camera trajectory ξ(t) enclosing the
past timestamp tk −∆t are used to interpolate the pose ξ(t−∆t),
which determines TRC (tk − ∆t). For simplicity, separate linear
interpolations for position and rotation parameters (exponential
coordinates) are used, although a Lie Group formulation with the
SE(3) exponential and logarithm maps (more computationally
expensive) could be used.
Once the corresponding points of the event coordinates (u, tk )
and (u, tk −∆t) have been computed, we use their intensity values
on the reference image I r to approximate the contrast:
∆ ln I ≈ ln I r (u0 (tk )) − ln I r (u0 (tk − ∆t)),
(13)
where tk is the time of the current event and ∆t is the time since
the last event at the same pixel. This approach is more accurate
than linearizing ∆ ln I . We assume that for a small pose change
there is a relatively large number of events from different pixels.
In this case, the information contribution of a new event to an old
pose will be negligible, and the new event will mostly contribute
to the most recent pose.
5
Next, we linearize the measurement function in (11) around
the expected state s̄k = Ep(sk |o1:k−1 ) [sk ], prior to incorporating
the measurement correction:
M (ok , s̃k ) ≈ M (ok , s̃¯k ) + ∇s̃ M (ok , s̃¯k ) · (s̃k − s̃¯k )
= M̄k + Jk · ∆s̃k ,
(14)
where M̄k and Jk are the predicted measurement and Jacobian at
s̄k , respectively. Substituting (14) in (11) we get:
2
p(ok |sk ) = πm N (M̄k + Jk · ∆s̃k ; 0, σm
) + (1 − πm ) U. (15)
We assume that the linearization is a good approximation to the
original measurement function.
Finally, we may re-write the likelihood (15) in a more general
and convenient form for deriving the filter equations, as a sum
of exponential families for the state parameters sk (see the
Appendix):
X
p(ok |sk ) =
h(sk ) exp(ηo,j · T (sk ) − Ao,j ).
(16)
j
4.4
Posterior Approximation and Filter Equations
Our third contribution pertains to the approximation of the posterior distribution using a tractable distribution. For this, we consider
variational inference theory [27], and choose a distribution in the
exponential family as well as conjugate priors, minimizing the
relative entropy error in representing the true posterior distribution
with our approximate distribution, as we explain next.
Exponential families of distributions are useful in Bayesian
estimation because they have conjugate priors [27]: if a given
distribution is multiplied by a suitable prior, the resulting posterior
has the same form as the prior. Such a prior is called a conjugate
prior for the given distribution. The prior of a distribution in the
exponential family is also in the exponential family, which clearly
simplifies recursion. A mixture distribution like (16) does not,
however, have a conjugate prior: the product of the likelihood and
a prior from the exponential family is not in the family. Instead,
the number of terms of the posterior doubles for each new measurement, making it unmanageable. Nevertheless, for tractability
and flexibility, we choose as conjugate prior a distribution in the
exponential family and approximate the product, in the sense of
the Kullback-Leibler (KL) divergence [28], by a distribution of the
same form, as expressed by (6). This choice of prior is optimal if
either the uniform or the normal terms of the likelihood dominates
the mixture; we expect that small deviations from this still gives
good approximations.
Letting the KL divergence (or relative entropy) from a distribution f to a distribution g be
Z
f (x)
dx,
(17)
DKL (f kg) = f (x) ln
g(x)
which measures the information loss in representing distribution
f by means of g , the posterior parameters ηk are calculated by
minimization of the KL divergence from the distribution on the
right hand side of (6) to the approximating posterior (left hand
side of (6)):
ηk = arg min DKL C p(ok |sk )q(sk ; ηk−1 )kq(sk ; η) .
η
It can be shown [27, p.505] that for g in the exponential family,
the necessary optimality condition ∇η DKL (f kg) = 0 gives the
system of equations (in η )
Ef (s) [T (s)] = Eg(s) [T (s)],
(18)
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
Algorithm 1 Event-based pose tracking
Initialize state variables (event camera pose, contrast threshold,
inlier ratio). Then, for each incoming event:
- propagate state covariance (zero-mean random diffusion)
- transfer the event to the map, compute the depth and evaluate the
measurement function M function (14).
- compute Kk in (20), the inlier probability πm , the weight wk
in (21), and the gain wk Kk .
- update filter variables and covariance (e.g., (22)-(23)).
i.e., the expected sufficient statistics must match. Additionally, the
right hand side of (18) is ∇A ≡ ∇η A = Eg(s) [T (s)] since
g is in the exponential family. In our case, g ≡ q(sk ; η), f ∝
p(ok |sk )q(sk ; ηk−1 ) and (18) can also be written in terms of the
parameters of (16) [(3)-(6) in the Appendix], the log-normalizer A
and its gradient:
X
0=
exp A(ηo,j + ηk−1 ) − A(ηk−1 ) − Ao,j )
j
× ∇A(ηo,j + ηk−1 ) − ∇A(η) .
(19)
Equation (19) describes a system of equations that can be solved
for η , yielding the update formula for ηk in terms of ηk−1 and
the current event ok . For a multivariate Gaussian distribution over
the event camera poses, explicit calculation of all update rules has
the simple form of an Extended Kalman Filter (EKF) [25], [29]
weighted by the inlier probability of that event:
2 −1
Kk = Pk Jk> (Jk Pk Jk> + σm
)
2
)
πm N (M̄k ; 0, σm
wk =
2
πm N (M̄k ; 0, σm ) + (1 − πm )U
ξ k+1 = ξ k + wk Kk M̄k
Pk+1 = (1 − wk Kk Jk )Pk ,
(20)
(21)
(22)
(23)
where 1 is the identity, M̄k and Jk are given in (14), ξ are the
6-DOF coordinates (3 for translation and 3 for rotation) of the
event camera pose, P is the pose covariance matrix, and wk Kk
acts as the Kalman gain. A pseudocode of the approach is outlined
in Algorithm 1.
The posterior approximation described in this section allows
us to fuse the measurements and update the state-vector efficiently,
without keeping multiple hypothesis in the style of particle filters,
which would quickly become intractable due to the dimension of
the state-vector.
5
E XPERIMENTAL R ESULTS
Our event-based pose estimation algorithm requires an existing
photometric depth map of the scene. As mentioned at the beginning of Section 4, without loss of generality we describe the map
in terms of depth maps with associated reference frames. These
can be obtained from a previous mapping stage by means of an
RGB-D camera or by classical dense reconstruction approaches
using standard cameras (e.g., DTAM [30] or REMODE [31]),
RGB-D sensors [32], or even using an event camera (future
research). In this work we use an Intel Realsense R200 RGBD camera. We show experiments with both nearly planar scenes
and scenes with large depth variations.
We evaluated the performance of our algorithm on several
indoor and outdoor sequences. The datasets also contain fast
motion with excitations in all six degrees of freedom (DOF). For
6
the interested reader, we would like to point out that sequences
similar to the ones used in these experiments can be found in the
publicly available Event Camera Dataset [33].
Indoor Experiments
First, we assessed the accuracy of our method against ground truth
obtained by a motion-capture system. We placed the event camera
in front of a scene consisting of rocks (Fig. 5) at a mean scene
depth of 60 cm and recorded eight sequences. Fig. 5 shows the position and orientation errors (i.e., difference between the estimated
ones and ground truth)4 for one of the sequences, while Fig. 10
shows the actual values of the estimated trajectory and ground
truth over time. Fig. 6 summarizes the errors of the estimated
trajectories for all sequences. The mean RMS errors in position
and orientation are 1.63 cm and 2.21°, respectively, while the
mean and standard deviations of the position and orientation errors
are µ = 1.38 cm, σ = 0.84 cm, and µ = 1.89°, σ = 1.15°,
respectively. Notice that the RMS position error corresponds to
2.71 % of the average scene depth, which is very good despite the
poor spatial resolution of the DVS.
Outdoor Experiments
For the outdoor experiments, we used structure from motion from
a standard camera as ground truth, more specifically we used
SVO [35]. To this end, we rigidly mounted the DVS and a standard
camera on a rig (see Fig. 7), and the same lens model was mounted
on both sensors. The DVS has a spatial resolution of 128 × 128
pixels and operates asynchronously, in the microsecond scale. The
standard camera is a global shutter MatrixVision Bluefox camera
with a resolution of 752 × 480 pixels and a frame rate of up
to 90 Hz. Both camera and DVS were calibrated intrinsically
and extrinsically. For reference, we measured the accuracy of
the frame-based method against the motion-capture system, in
the same sequences previously mentioned (rocks, as in Fig. 10).
The average RMS errors in position and orientation are 1.08 cm
(i.e., 1.8 % of the mean scene depth) and 1.04°, respectively.
Comparing these values to those of the event-based method, we
note that, in spite of the limited resolution of the DVS, the
accuracy of the results provided by our event-based algorithm
is only slightly worse (2.71 % vs. 1.8 % in position, and 2.21°
vs. 1.04° in orientation) than that obtained by a standard camera
processing 20× higher resolution images. This is made possible
by the DVS temporal resolution being ten thousand times larger
than the standard camera.
The three outdoor sequences (ivy, graffiti, and building)
were recorded with the DVS-plus-camera rig viewing an ivy, a
graffiti covered by some plants, and a building with people moving
in front of it, respectively (see Fig. 8, 1st column and accompanying video submission). The rig was moved by hand with increasing
speed. All sequences exhibit significant translational and rotational
motion. The error plots in position and orientation of all 6-DOFs
are given in Fig. 8. The reported error peaks in the graffiti and
building sequences are due to a decrease of overlap between the
event camera frustum and the reference map, thus making pose
estimation ambiguous for some motions (e.g., Y -translation vs.
X -rotation).
Table 1 summarizes the statistics of the pose tracking error for
the three outdoor sequences. For the ivy dataset, the mean and
4. The rotation error is measured using the angle of their relative rotation
(i.e., geodesic distance in SO(3) [34]).
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
7
15
X
Y
Z
20
15
10
5
0
0
2
4
6
time [s]
8
10
12
x
orientation error [deg]
position error [%]
25
y
10
z
5
0
0
2
4
6
8
10
12
time [s]
Fig. 5: Error plots in position (relative to a mean scene depth of 60 cm) and in orientation (in degrees) for one of the test sequences
with ground truth provided by a motion capture system with sub-millimeter accuracy.
4
Orientation error [deg]
Position error [%]
4
3
2
1
0
RMS
Mean
3
2
1
0
Std
RMS
Mean
Std
Fig. 6: Error in position (relative to a mean scene depth of 60 cm)
and orientation (in degrees) of the trajectories recovered by our
method for all rocks sequences (ground truth is given by a motion
capture system). We provide box plots of the root-mean-square
(RMS) errors, the mean errors and the standard deviation (Std) of
the errors.
which correspond to 3.97 % and 1.84 % of the average scene
depth (2.5 m), respectively. The mean and standard deviation of
the orientation error are 2.0° and 0.94°, respectively. For the
building dataset, which presents the largest errors, the mean and
standard deviation of the orientation error are 3.43° and 2.05°,
respectively, while, in position error, the corresponding figures are
1.94 m and 1.08 m, that correspond to 6.47 % and 3.60 % of the
average scene depth (30 m), respectively.
As reported by the small errors in Table 1, overall our eventbased algorithm is able to accurately track the pose of the event
camera also outdoors. We expect that the results provided by our
approach would be even more accurate with the next generation of
event-based sensors currently being developed [22], [23], which
will have higher spatial resolution (640 × 480 pixels). Finally,
observe that in the building sequence (Fig. 8, bottom row), our
method gracefully tracks the pose in spite of the considerable
amount of events generated by moving objects (e.g., people) in
the scene (see Fig. 9).
5.1 Tracking during High-Speed Motions
Fig. 7: An event camera (DVS) and a standard camera mounted
on a rig. The standard camera was only used for comparison.
In addition to the error plots in Fig. 5, we show in Fig. 10 the
actual values of the trajectories (position and orientation) acquired
by the motion capture system (dashed line) and estimated by
the event-based method (solid line) and the frame-based method
(dash-dot). Notice that they are all are almost indistinguishable
relative to the amplitude of the motion excitation, which gives a
better appreciation of the small errors reported in Figs. 5 and 6.
Figure 11 shows a magnified version of the estimated trajectories during high-speed motions (occurring at t ≥ 7 s in Fig. 10).
The frame-based method is able to track in the shaded region, up
to t ≈ 8.66 s (indicated by a vertical dashed line), at which point
it loses tracking due to motion blur, while our event-based method
continues to accurately estimate the pose.
5.2 Experiments with Large Depth Variation
standard deviation of the position error are 9.93 cm and 4.60 cm,
TABLE 1: Error measurements of three outdoor sequences. Translation errors are relative (i.e., scaled by the mean scene depth).
Position error [%]
RMS
µ
σ
ivy
graffiti
building
4.37
5.88
7.40
3.97
5.23
6.47
1.84
2.70
3.60
Orientation error [°]
RMS
µ
σ
2.21
3.58
3.99
2.00
3.09
3.43
0.94
1.80
2.05
In the following set of experiments, we also assessed the accuracy
of our method on scenes with large depth variation and, therefore
larger parallax than in previous experiments. We recorded seven
sequences with ground truth from a motion-capture system of a
scene consisting of a set of textured boxes (Fig. 12, top row). We
also recorded two outdoor sequences: pipe and bicycles (middle
and bottom rows of Fig. 12). The pipe sequence depicts a creek
going through a pipe, surrounded by rocks and grass; the bicycle
sequence depicts some parked bicycles next to a building; both
outdoor scenes present some occlusions. All sequences exhibit
significant translational and rotational motion.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
X
Y
Z
Outdoor dataset: ivy
20
orientation error [deg]
position error [%]
25
15
10
5
0
0
2
4
8
6
15
3y
10
3z
5
0
0
2
time [s]
orientation error [deg]
position error [%]
Outdoor dataset: graffiti
15
X
Y
Z
10
5
0
0
5
10
15
X
Y
Z
10
5
2
4
6
8
orientation error [deg]
position error [%]
Outdoor dataset: building
0
3x
Outdoor dataset: graffiti
3y
10
3z
5
0
0
5
10
time [s]
25
0
6
15
time [s]
20
4
time [s]
25
20
3x
Outdoor dataset: ivy
15
Outdoor dataset: building
3x
3y
10
3z
5
0
0
time [s]
2
4
6
8
time [s]
Fig. 8: Error plots in position (2nd column, relative to the mean scene depth) and in orientation (3rd column, in degrees) for three
outdoor test sequences (1st column): ivy, graffiti, and building. The mean scene depths are 2.5 m, 3 m, and 30 m, respectively.
TABLE 2: Error measurements of the sequences in Fig. 12.
Translation errors are relative (i.e., scaled by the mean scene
depth).
Position error [%]
RMS
µ
σ
boxes
pipe
bicycles
Fig. 9: The algorithm is able to track the pose of the event camera
in spite of the considerable amount of events generated by moving
objects (e.g., people) in the scene.
Fig. 13 summarizes the position and orientation error statistics
of our algorithm on the boxes sequences (compared with ground
truth from the motion-capture system). The position error is given
relative to the mean scene depth, which is 1.9 m. As it is observed,
the errors are very similar to those in Fig. 6, meaning that our
pose tracking method can handle arbitrary 3D scenes, i.e., not
necessarily nearly planar.
Table 2 reports the numerical values of the trajectory errors
2.50
4.04
2.14
2.23
3.04
1.724
1.17
2.66
1.27
Orientation error [°]
RMS
µ
σ
1.88
2.90
1.46
1.65
2.37
1.19
1.02
1.67
0.84
in both indoors and outdoor sequences. The row corresponding
to the boxes sequences is the average of the errors in the seven
indoor sequences (Fig. 13). For the position error, the mean scene
depths of the pipe and bicycles sequences are 2.7 m and 2.2 m,
respectively. The mean RMS errors in position and orientation are
in the range 2.5 % to 4.0 % of the mean scene depth and 1.4°
to 2.9°, respectively, which are in agreement with the values in
Table 1 for the scenes with mean depths smaller than 3 m. It is
remarkable that the method is able to track despite some lack of
texture (as in the pipe sequence, where there are only few strong
edges), and in the presence of occlusions, which are more evident
in the bicycles sequence.
5.3
Computational Effort
We measured the computational cost of our algorithm on a single
core of an Intel(R) i7 processor at 2.60 GHz. The processing time
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
9
30
0.2
20
orientation [deg]
position [m]
0.1
0
-0.1
-0.2
X
0
Y
Z
2
4
10
0
-10
-20
-30
-40
6
8
10
12
0
2
4
time [s]
6
roll
pitch
8
10
yaw
12
time [s]
Fig. 10: Indoor experiment with 6-DOF motion. Left: Image of the standard camera overlaid with events (during mild motion). Events
are displayed in red and green, according to polarity. Estimated position (center) and orientation (right) from our event-based algorithm
(solid line), a frame-based method (dash-dot line) and ground truth (black line) from a motion capture system.
X
Y
20
Z
0.1
orientation [deg]
position [m]
0.2
0
-0.1
roll
pitch
yaw
10
0
-10
-0.2
8
8.5
9
9.5
10
time [s]
-20
8
8.5
9
9.5
10
time [s]
Fig. 11: Zoom of Fig. 10. Left: Image of the standard camera overlaid with events (red and green points, according to polarity) during
high-speed motion. Center and right: estimated trajectories. Due to the very high temporal resolution, our algorithm can still track
the motion even when the images of the standard camera are sufficiently blurred so that the frame-based method (FB) failed. The
event-based method (EB) provides pose updates even in high-speed motions, whereas the frame-based method loses track (it only
provides pose updates in the region marked with the shaded area, then it fails).
per event is 32 µs, resulting in a processing event rate of 31.000
events per second. Depending on the texture of the scene and
the speed of motion, the data rate produced by an event camera
ranges from tens of thousands (moderate motion) to over a million
events per second (high-speed motion). However, notice that our
implementation is not optimal; many computations can indeed
still be optimized, cached, and parallelized to increase the runtime
performance of the algorithm.
6
C ONCLUSION
We have presented an approach to track the 6-DOF pose of an
arbitrarily moving event camera from an existing photometric
depth map in natural scenes. Our approach follows a Bayesian
filtering methodology: the sensor model is given by a mixturemodel likelihood that takes into account both the event-generation
process and the presence of noise and outliers; the posterior
distribution of the system state is approximated according to the
relative-entropy criterion using distributions in the exponential
family and conjugate priors. This yields a robust EKF-like filter
that provides pose updates for every incoming event, at microsecond time resolution.
We have compared our method against ground truth from a
motion capture system and a state-of-the-art frame-based posetracking pipeline. The experiments revealed that the proposed
method accurately tracks the pose of the event-based camera, both
in indoor and outdoor experiments in scenes with significant depth
variations, and under motions with excitations in all 6-DOFs.
R EFERENCES
[1]
R. Mur-Artal, J. M. M. Montiel, and J. D. Tardós, “ORB-SLAM: a
versatile and accurate monocular SLAM system,” IEEE Trans. Robot.,
vol. 31, no. 5, pp. 1147–1163, 2015. 1
[2] J. Engel, J. Schöps, and D. Cremers, “LSD-SLAM: Large-scale direct
monocular SLAM,” in Eur. Conf. Comput. Vis. (ECCV), 2014. 1
[3] C. Forster, Z. Zhang, M. Gassner, M. Werlberger, and D. Scaramuzza,
“SVO: Semidirect visual odometry for monocular and multicamera
systems,” IEEE Trans. Robot., vol. 33, no. 2, pp. 249–265, 2017. 1
[4] A. N. Belbachir, Smart Cameras. Springer US, 2009. 1
[5] P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128x128 120dB 30mW
asynchronous vision sensor that responds to relative intensity change,” in
IEEE Intl. Solid-State Circuits Conf. (ISSCC), Feb. 2006, pp. 2060–2069.
1
[6] M. Cook, L. Gugelmann, F. Jug, C. Krautz, and A. Steger, “Interacting
maps for fast visual interpretation,” in Int. Joint Conf. Neural Netw.
(IJCNN), 2011, pp. 770–776. 1, 2
[7] P. Bardow, A. J. Davison, and S. Leutenegger, “Simultaneous optical flow
and intensity estimation from an event camera,” in Proc. IEEE Int. Conf.
Comput. Vis. Pattern Recog., 2016. 1
[8] C. Reinbacher, G. Graber, and T. Pock, “Real-time intensity-image reconstruction for event cameras using manifold regularisation,” in British
Machine Vis. Conf. (BMVC), 2016. 1
[9] H. Rebecq, T. Horstschäfer, G. Gallego, and D. Scaramuzza, “EVO: A
geometric approach to event-based 6-DOF parallel tracking and mapping
in real-time,” IEEE Robot. Autom. Lett., vol. 2, pp. 593–600, 2017. 1, 2
[10] H. Kim, S. Leutenegger, and A. J. Davison, “Real-time 3D reconstruction
and 6-DoF tracking with an event camera,” in Eur. Conf. Comput. Vis.
(ECCV), 2016, pp. 349–364. 1, 2
[11] P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128×128 120 dB 15 µs
latency asynchronous temporal contrast vision sensor,” IEEE J. SolidState Circuits, vol. 43, no. 2, pp. 566–576, 2008. 1, 3, 4
[12] D. Weikersdorfer and J. Conradt, “Event-based particle filtering for robot
self-localization,” in IEEE Int. Conf. Robot. Biomimetics (ROBIO), 2012,
pp. 866–870. 2, 4
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
10
40
0.6
X
Y
Z
0.2
0
-0.2
-0.4
20
Orientation [deg]
Position [m]
0.4
0
5
10
15
20
25
roll
pitch
yaw
0
-20
-40
-60
30
0
5
10
15
Time [s]
X
Y
Z
0.4
0.2
0
-0.2
10
0
-10
-20
-30
0
5
10
15
20
-40
25
Time [s]
0
5
10
15
20
25
Time [s]
40
1.5
X
Y
Z
roll
pitch
yaw
20
Orientation [deg]
1
Position [m]
30
roll
pitch
yaw
20
Orientation [deg]
Position [m]
0.6
0.5
0
-0.5
25
30
0.8
-0.4
20
Time [s]
0
5
10
Time [s]
0
-20
-40
-60
0
5
10
Time [s]
Fig. 12: Experiments on scenes with significant depth variation and occlusions. Scene impressions (1st column): boxes, pipe, and
bicycles. Estimated position (2nd column, in meters) and orientation (3rd column, in degrees) from our event-based algorithm (solid
line) compared with ground truth (dashed line). The mean scene depths are 1.8 m, 2.7 m, and 2.3 m, respectively.
4
3
2
1
0
RMS
Mean
Std
Orientation error [deg]
Position error [%]
4
3
2
1
0
RMS
Mean
Std
Fig. 13: Error in position (relative to a mean scene depth of
1.9 m) and orientation (in degrees) of the trajectories recovered
by our method for all boxes sequences (ground truth is given by
a motion-capture system). We provide box plots of the root-meansquare (RMS) errors, the mean errors and the standard deviation
(Std) of the errors.
[13] D. Weikersdorfer, D. B. Adrian, D. Cremers, and J. Conradt, “Eventbased 3D SLAM with a depth-augmented dynamic vision sensor,” in
IEEE Int. Conf. Robot. Autom. (ICRA), Jun. 2014, pp. 359–364. 2
[14] A. Censi and D. Scaramuzza, “Low-latency event-based visual odometry,” in IEEE Int. Conf. Robot. Autom. (ICRA), 2014. 2
[15] E. Mueggler, B. Huber, and D. Scaramuzza, “Event-based, 6-DOF pose
tracking for high-speed maneuvers,” in IEEE/RSJ Int. Conf. Intell. Robot.
Syst. (IROS), 2014, pp. 2761–2768. 2, 4
[16] H. Kim, A. Handa, R. Benosman, S.-H. Ieng, and A. J. Davison,
“Simultaneous mosaicing and tracking with an event camera,” in British
Machine Vis. Conf. (BMVC), 2014. 2, 4
[17] G. Gallego and D. Scaramuzza, “Accurate angular velocity estimation
with an event camera,” IEEE Robot. Autom. Lett., vol. 2, pp. 632–639,
2017. 2
[18] C. Reinbacher, G. Munda, and T. Pock, “Real-time panoramic tracking
for event cameras,” in IEEE Int. Conf. Comput. Photography (ICCP),
2017. 2
[19] E. Mueggler, G. Gallego, H. Rebecq, and D. Scaramuzza, “Continuoustime visual-inertial trajectory estimation with event cameras,” 2017,
arXiv:1702.07389. 3
[20] A. Zhu, N. Atanasov, and K. Daniilidis, “Event-based visual inertial
odometry,” in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recog., 2017.
3
[21] H. Rebecq, T. Horstschäfer, and D. Scaramuzza, “Real-time visualinertial odometry for event cameras using keyframe-based nonlinear
optimization,” in British Machine Vis. Conf. (BMVC), Sep. 2017. 3
[22] C. Brandli, R. Berner, M. Yang, S.-C. Liu, and T. Delbruck, “A 240x180
130dB 3us latency global shutter spatiotemporal vision sensor,” IEEE J.
Solid-State Circuits, vol. 49, no. 10, pp. 2333–2341, 2014. 3, 7
[23] C. Li, C. Brandli, R. Berner, H. Liu, M. Yang, S.-C. Liu, and T. Delbruck,
“An RGBW color VGA rolling and global shutter dynamic and activepixel vision sensor,” in International Image Sensor Workshop (IISW),
Vaals, Netherlands, Jun. 2015. 3, 7
[24] B. Son, Y. Suh, S. Kim, H. Jung, J.-S. Kim, C. Shin, K. Park, K. Lee,
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
J. Park, J. Woo, Y. Roh, H. Lee, Y. Wang, I. Ovsiannikov, and H. Ryu, “A
640x480 dynamic vision sensor with a 9um pixel and 300Meps addressevent representation,” in IEEE Intl. Solid-State Circuits Conf. (ISSCC),
2017. 3
S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. The MIT
Press, Cambridge, MA, 2005. 3, 6
G. Vogiatzis and C. Hernández, “Video-based, real-time multi view
stereo,” Image Vis. Comput., vol. 29, no. 7, pp. 434–441, 2011. 5
C. M. Bishop, Pattern Recognition and Machine Learning. SpringerVerlag New York, Inc., 2006. 5
S. Kullback and R. A. Leibler, “On information and sufficiency,” Ann.
Math. Statist., vol. 22, no. 1, pp. 79–86, 1951. 5
R. Kalman, “A new approach to linear filtering and prediction problems,”
J. Basic Eng., vol. 82, pp. 35–45, 1960. 6
R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “DTAM: Dense
tracking and mapping in real-time,” in Int. Conf. Comput. Vis. (ICCV),
Nov. 2011, pp. 2320–2327. 6
M. Pizzoli, C. Forster, and D. Scaramuzza, “REMODE: Probabilistic,
monocular dense reconstruction in real time,” in IEEE Int. Conf. Robot.
Autom. (ICRA), 2014, pp. 2609–2616. 6
R. A. Newcombe, A. J. Davison, S. Izadi, P. Kohli, O. Hilliges, J. Shotton,
D. Molyneaux, S. Hodges, D. Kim, and A. Fitzgibbon, “KinectFusion:
Real-time dense surface mapping and tracking,” in IEEE ACM Int. Sym.
Mixed and Augmented Reality (ISMAR), Basel, Switzerland, Oct. 2011,
pp. 127–136. 6
E. Mueggler, H. Rebecq, G. Gallego, T. Delbruck, and D. Scaramuzza,
“The event-camera dataset and simulator: Event-based data for pose
estimation, visual odometry, and SLAM,” Int. J. Robot. Research, vol. 36,
pp. 142–149, 2017. 6
D. Q. Huynh, “Metrics for 3D rotations: Comparison and analysis,” J.
Math. Imaging Vis., vol. 35, no. 2, pp. 155–164, 2009. 6
C. Forster, M. Pizzoli, and D. Scaramuzza, “SVO: Fast semi-direct
monocular visual odometry,” in IEEE Int. Conf. Robot. Autom. (ICRA),
2014, pp. 15–22. 6
Guillermo Gallego received the Ph.D. degree
in electrical and computer engineering from the
Georgia Institute of Technology, Atlanta, GA,
USA, in 2011. He received the Ingeniero de
Telecomunicación degree (five-year engineering
program) from the Universidad Politécnica de
Madrid (UPM), Madrid, Spain, in 2004, the M.S.
degree in mathematical engineering (Magı́ster
en Ingenierı́a Matemática) from the Universidad
Complutense de Madrid, Madrid, in 2005, and
the M.S. degrees in electrical and computer engineering and mathematics from the Georgia Institute of Technology,
Atlanta, GA, USA, in 2007 and 2009, respectively. From 2011 to 2014,
he was a Marie Curie Post-Doctoral Researcher with the UPM. Since
2014, he has been with the Robotics and Perception Group, University
of Zurich, Zurich, Switzerland. His current research interests include
computer vision, signal processing, robotics, geometry, optimization,
and ocean engineering. Dr. Gallego was a recipient of the Fulbright
Scholarship to pursue graduate studies at the Georgia Institute of
Technology in 2005. He is a recipient of the Misha Mahowald Prize for
Neuromorphic Engineering (2017).
Jon E.A. Lund is a flight control and video engineer at Prox Dynamics/FLIR UAS in Oslo, Norway. He received his M.Sc. (2015) in Neural Systems and Computation from ETH and University
of Zurich, Switzerland. Before that, he earned a
B.Sc. (2013) in Physics at the University of Oslo,
Norway.
11
Elias Mueggler received his Ph.D. degree from
the University of Zurich, Switzerland in 2017,
where he was working at the Robotics and
Perception Group, lead by Prof. Davide Scaramuzza, in the topics of event-based vision for
high-speed robotics and air-ground robot collaboration. He received B.Sc. (2010) and M.Sc.
(2012) degrees in Mechanical Engineering from
ETH Zurich, Switzerland. He is a recipient of the
KUKA Innovation Award (2014), the Qualcomm
Innovation Fellowship (2016) and the Misha Mahowald Prize for Neuromorphic Engineering (2017). He has been a
visiting researcher with Prof. John Leonard (Massachusetts Institute of
Technology) and Dr. Chiara Bartolozzi (Istituto Italiano di Tecnologia).
His research interests include computer vision and robotics.
Henri Rebecq is a Ph.D. student in the Robotics
and Perception Group at the University of Zurich,
where he is working on on event-based vision
for robotics. In 2014, he received an M.Sc.Eng.
degree from Télécom ParisTech, and an M.Sc.
degree from Ecole Normale Supérieure de
Cachan, both located in Paris, France. Prior to
pursuing graduate studies, he worked as a research and software engineer at VideoStitch,
Paris, France. His research interests include omnidirectional vision, visual odometry, 3D reconstruction and SLAM. He is a recipient of the Misha Mahowald Prize for
Neuromorphic Engineering (2017).
Tobi Delbruck (M’99–SM’06–F’13) received the
B.Sc. degree in physics and applied mathematics from the University of California, San Diego,
CA, USA, and the Ph.D. degree from the California Institute of Technology, Pasadena, CA,
USA, in 1986 and 1993, respectively. He has
been a Professor of Physics with the Institute
of Neuroinformatics, University of Zurich and
ETH Zurich, Switzerland, since 1998. His group
focuses on neuromorphic sensory processing.
He worked on electronic imaging at Arithmos,
Synaptics, National Semiconductor, and Foveon. Dr. Delbruck has coorganized the Telluride Neuromorphic Cognition Engineering summer
workshop and the live demonstration sessions at International Symposium on Circuits and Systems. He is also co-founder of iniLabs and
Insightness. He was the Chair of the IEEE CAS Sensory Systems
Technical Committee, is current Secretary of the IEEE Swiss CAS/ED
Society, and an Associate Editor of the IEEE Transactions on Biomedical
Circuits and Systems. He has received 9 IEEE awards.
Davide Scaramuzza Davide Scaramuzza (born
in 1980, Italian) is Associate Professor of
Robotics and Perception at both departments of
Informatics (University of Zurich) and Neuroinformatics (University of Zurich and ETH Zurich),
where he does research at the intersection of
robotics, computer vision, and neuroscience. He
did his PhD in robotics and computer vision at
ETH Zurich and a postdoc at the University of
Pennsylvania. From 2009 to 2012, he led the
European project sFly, which introduced the PX4
autopilot and pioneered visual-SLAM–based autonomous navigation of
micro drones. For his research contributions, he was awarded the Misha
Mahowald Neuromorphic Engineering Award, the IEEE Robotics and
Automation Society Early Career Award, the SNSF-ERC Starting Grant,
a Google Research Award, the European Young Research Award, and
several conference paper awards. He coauthored the book Introduction
to Autonomous Mobile Robots (published by MIT Press) and more than
100 papers on robotics and perception.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. ?, NO. ?, 2017
A PPENDIX A
R EWRITING THE L IKELIHOOD F UNCTION
A distribution in the exponential family can be written as
p(x; η) = h(x) exp (η · T (x) − A(η)) ,
(24)
where η are the natural parameters, T (x) are the sufficient
statistics of x, A(η) is the log-normalizer, and h(x) is the base
measure.
The likelihood (15) can be rewritten as:
1
p(ok |sk ) = √ exp(ln(πm ) − ln(σm )
(25)
2π "
#
1 i i s̃ik s̃jk
s̃i
M̄ 2
−
Jk Jk 2 + 2M̄k Jki 2k + 2k
2
σm
σm
σm
+ exp(ln((1 − πm )/(Mmax − Mmin ))),
where we use the Einstein summation convention for the indices
of Jk = (Jki ) and s̃k = (s̃ik ). Collecting the sufficient statistics
into
"
#
s̃ik s̃jk s̃ik 1
T (sk ) =
, 2 , 2 , ln(σm ), ln(πm ), ln(1 − πm ) ,
2
σm
σm σm
the likelihood can be conveniently rewritten as a sum of two
exponential families (16), j = 1, 2, with h(s) = 1,
1 2
1 i j
i
(26)
ηo,1 = − Jk Jk , −M̄k Jk , − M̄k , −1, 1, 0
2
2
ηo,2 = [0ij , 0i , 0, 0, 1]
(27)
√
Ao,1 = ln 2π
(28)
Ao,2 = − ln(Mmax − Mmin ).
(29)
12
| 7cs.IT
|
Improving NSGA-II with an Adaptive Mutation Operator
Arthur G. Carvalho
Aluizio F. R. Araujo
School of Computer Science
University of Waterloo
Waterloo, Ontario, Canada
Informatics Center
Federal University of Pernambuco
Recife, Pernambuco, Brazil
arXiv:1305.4947v1 [cs.NE] 21 May 2013
[email protected]
ABSTRACT
The performance of a Multiobjective Evolutionary Algorithm (MOEA) is crucially dependent on the parameter setting of the operators. The most desired control of such parameters presents the characteristic of adaptiveness, i.e., the
capacity of changing the value of the parameter, in distinct
stages of the evolutionary process, using feedbacks from
the search for determining the direction and/or magnitude
of changing. Given the great popularity of the algorithm
NSGA-II, the objective of this research is to create adaptive
controls for each parameter existing in this MOEA. With
these controls, we expect to improve even more the performance of the algorithm.
In this work, we propose an adaptive mutation operator
that has an adaptive control which uses information about
the diversity of candidate solutions for controlling the magnitude of the mutation. A number of experiments considering different problems suggest that this mutation operator
improves the ability of the NSGA-II for reaching the Pareto
optimal Front and for getting a better diversity among the
final solutions.
Categories and Subject Descriptors
I.2.8 [ARTIFICIAL INTELLIGENCE]: Problem Solving, Control Methods, and Search—Heuristic methods
General Terms
Algorithms
Keywords
Evolutionary Multiobjective Optimization, Parameter Control, Adaptive Mutation Operator
1.
INTRODUCTION
The term optimization, in the field of mathematics, refers
to the study of problems in which we are looking for optimal solutions, minimum or maximum, for a given function.
These solutions are obtained through systematic changes in
the values of the variables. When we want to optimize systematically and simultaneously various objective functions
(usually conflicting between themselves), we will have the
process known as multiobjective optimization.
Copyright is held by the author/owner(s).
GECCO’09, July 8–12, 2009, Montréal, Québec, Canada.
ACM 978-1-60558-505-5/09/07.
[email protected]
A good algorithm created for solving multiobjective optimization problems must: 1) find multiple Pareto optimal
solutions and 2) find a good diversity of solutions on the
obtained Pareto front (close to an uniform distribution)[2].
Variations of evolutionary algorithms, known as Multiobjective Evolutionary Algorithms (MOEAs), are the metaheuristic best known for solving multiobjective optimization
problems. Due to the characteristics inherited from the evolutionary computing, these algorithms have operators with
parameters that need to be configured. Moreover, the performance of a MOEA is crucially dependent on the parameter setting of these operators.
The most desired control of such parameters presents the
characteristic of adaptiveness, i.e., the capacity of changing the value of the parameter, in distinct stages of the
evolutionary process, using feedbacks from the search for
determining the direction and/or magnitude of changing.
However, MOEAs usually employs stochastic operators with
static parameters.
According to Eiben and Smith [6] a run of an evolutionary
algorithm is a process intrinsically dynamic and adaptive.
Then, this static approach can result in an inefficient convergence to the Pareto optimal solutions and a failure for
creating an (almost) uniform distribution of final solutions
on the obtained Pareto front.
Given the great popularity of the algorithm Non-dominated
Sorting Genetic Algorithm II (NSGA-II) [4], we propose to
create adaptive controls for each parameter existing in this
MOEA for increasing even more its ability for reaching the
Pareto optimal front and for getting a better diversity among
the final solutions.
Within the context presented, we propose in this work an
adaptive mutation operator which uses information about
the diversity of candidate solutions for controlling the magnitude of the mutation.
The rest of this paper is organized as follows. Section
2 presents the concept of crowding distance [4], a density
estimator that will provide information for controlling the
magnitude of the mutation. Section 3 describes the adaptive mutation operator proposed. The experiments and the
statistical validation of the results are described in Section
4. Finally, Section 5 summarizes the results of this work and
proposes additional topics for further research.
2. CROWDING DISTANCE
The crowding distance is an important concept proposed
by Deb et. al. [4] in his algorithm NSGA-II. It serves for
getting an estimate of the density of solutions surrounding
a particular solution in the population. More specifically,
the crowding distance for a point i (called idistance ) is an
estimate of the size of the largest cuboid enclosing i without
including any other point in the population. It is calculated
by taking the average distance of the two points on either
side of i along each of the objectives. The algorithm used
for calculating the crowding distance for each point in a
population L is:
crowding-distance-assignment(L):
l = |L|
for each i ∈ L
L[i]distance = 0
for each objective m
L = sort(L, m)
L[1]distance = L[l]distance = ∞
for i = 2 to (l − 1)
L[i]distance += fm (i + 1) − fm (i − 1)
max fm − min fm
In the first line, it is assigned the size of the population
L to the variable l. Following this operation, there is a
loop responsible for initializing with 0 the idistance of each
element i of the population L.
In the fourth line, each objective m is selected at a time
and the population is sorted in a ranking according to the
value for m. The idistance value for solutions in the first and
in the last position is assigned as infinity (∞) for preserving
solutions with extreme values.
The inner loop presents in the seventh line updates the
idistance value for each remaining solution i from position 2
to l − 1. First, it is calculated the m-th objective function
value for the neighbors of i. Thereafter, it is calculated the
difference between the highest and the lowest value. Finally,
the idistance value for i is updated by the sum of its previous
value with the normalized result of that subtraction. Figure
1 shows an illustration of this calculation for a given solution
i. In this scenario, the idistance value for i will be r+s where:
s=
f1 (i − 1) − f1 (i + 1)
f1 (a) − f1 (z)
and
r=
f2 (i + 1) − f2 (i − 1)
f2 (z) − f2 (a)
3. ADAPTIVE MUTATION OPERATOR
According to Eiben and Smith [6], an adaptive parameter
control uses feedback from the search for serving as input
to a mechanism used for determining the direction and/or
magnitude of changing. Using the well known static mutation operator proposed by Deb and Goyal [3] together with
an adaptive parameter control for updating its parameter,
this section presents the adaptive mutation operator created
for improving even more the performance of the algorithm
NSGA-II.
In the original (static) version of the mutation operator,
the current value of a continuous variable is changed to a
neighboring value using a polynomial probability distribution. This distribution has its mean at the current value of
the variable and its variance as a function of a parameter n.
This parameter will define the strength of the mutation and
we are interested in adaptively changing its value.
Besides this parameter, the polynomial probability distribution depends on a factor of disturbance δ for calculating
the mutated value as can be seen in the following equation:
P (δ) = 0.5(n + 1)(1 − |δ|)n
(1)
where δ ∈ [−1, 1]. Figure 2 shows this distribution for some
values of n.
Figure 2: Probability distribution for creating a mutated value.
Initially, for creating a mutated value we need to generate
a random number u ∈ [0, 1]. Thereafter, the equation 2
(obtained from equation 1) can be used for calculating the
factor of disturbance δ corresponding to u:
δ=
(
1
(2u) n+1 − 1
1
1 − [2(1 − u)] n+1
if u < 0.5
if u ≥ 0.5
(2)
In the end, the mutated valued is calculated using the
following equation:
c = p + δ∆max
Figure 1: The crowding distance calculation.
(3)
where c is the mutated value, p is the original value and
∆max is the maximum disturbance allowed in the value of p
(it was defined here as the difference between the maximum
and the minimum value for the decision variable).
To change the variance of the probability distribution (the
parameter n in equation 1) in an adaptive way, we will use
two empirical facts observed. First, the initial solutions are
dispersed in the search space and distant from the Pareto optimal front. Furthermore, the difference between the greatest idistance value not infinite and the lowest idistance value
is lifted. In this scenario, it is necessary to apply a strong
mutation for ensuring a quicker convergence to the Pareto
optimal Front and a fast attainment of distinct solutions.
Second, at the end of the evolutionary process it will be expected solutions closer to the Pareto optimal front due to the
efficacy of the NSGA-II. Moreover, the difference between
the greatest (not infinite) and the lowest idistance value is
reduced. Now, it is necessary to apply a soft mutation for
avoiding destroying solutions previously generated and for
trying to approximate them to the Pareto optimal front.
So, the main ideas exploited by the adaptive control are
to use information about the difference between the greatest (not infinite) and the lowest idistance value and about the
current stage of the evolutionary process. Due to the fact
that the NSGA-II calculates the idistance for all individuals in the current population before applying evolutionary
operators, it will not be necessary to re-calculated it again.
So, we just have to calculate ∆, the difference between the
greatest (different of ∞) and the lowest idistance value:
∆
where:
=
max g(idistance ) − min idistance
1≤i≤|L|
1≤i≤|L|
g(x) = 0 if x = ∞
= x otherwise
The next step is to use information about the current generation t of the evolutionary process. For ensuring that it
will have an acceptable weight in the update of the parameter, we applied on it a logistic function. So, the second step
taken by the controller is to calculate the function:
sigm(t) =
1
1 + e−0.07t
sigm(t)
∆
In order for evaluating the performance of the proposed
adaptive mutation operator, this section provides a comparative study among different settings for the NSGA-II. The
first one uses the original mutation operator proposed by
Deb and Goyal [3] with n = 5 (for representing a strong
mutation). The second configuration also uses this mutation operator, but this time with n = 20 (for representing a
smooth mutation). At least, the third configuration is represented by the adaptive mutation operator proposed here.
The remaining parameters are the same for all settings.
We used a population size of 20 individuals (this small value
was chosen for making the mutation more valorous), a crossover
probability of 0.9, a mutation probability of 1/n (where n is
the number of variables). The variables were treated as real
numbers and the simulated binary crossover operator (SBX)
[3] was used. For all experiments, the implementation used
as reference was proposed by Durillo et al [5].
The problems used in experiments were chosen based on
characteristics usually present in real problems [1]: continuous Pareto optimal front vs. discontinuous Pareto optimal
front; convex Pareto optimal front vs. non-convex Pareto
optimal front; uniformly represented Pareto optimal Front
vs. non-uniformly represented Pareto optimal front.
The first problem used was proposed by Fonseca and Fleming [7] (called here as FON2). The next four problems used
(ZDT1, ZDT2, ZDT3, ZDT6) were proposed by Zitzler et al
[9] and belong to a test suite called ZDT.
Due to the fact that the convergence to the Pareto optimal
front and the maintenance of a diverse set of solutions are
two different goals of the multiobjective optimization, it will
be need two different metrics for deciding the performance
of a setting in an absolute manner [2].
The first metric used, called Generational Distance (GD)
[8], is responsible for finding the closeness of the obtained
set of solutions to the Pareto optimal front as follows:
(4)
where t is the current generation. The inspiration for using
such function is the fact that it would fit perfectly into our
proposal because we would like to apply a strong mutation
in the early stages of the evolutionary process and gradually reduce its value during the process. The constant value
−0.07 is used because the value of e−0.07t will be approximately 0 when t → ∞. Actually, when t is greater than
100, the function sigm(t) will practically stop influencing
the mutation because its value will be equals to 1.
It is useful to cite that the new value for the parameter
n has to be inversely proportional to ∆. This happens due
to the fact that for higher values of ∆ it will be necessary
to apply a strong mutation and, consequently, it will be
needed a lower value for n to increase the variance of the
probability distribution. Furthermore, the new value for n
has to be directly proportional to the sigm(t) due to the fact
that for higher values of t it will be needed a soft mutation
and, consequently, it will be needed a higher value for n
to reduce the variance of the probability distribution. In
the end, the last step taken by the controller is to update
n, before applying a mutation in the current generation, as
follows:
n=
4. EXPERIMENTS
(5)
P|Q|
1
d2i ) 2
(6)
|Q|
where Q is the set of the obtained solutions and di is the Euclidean distance between the solution i ∈ Q and the nearest
member of the Pareto optimal front as exhibited below:
GD =
(
i=1
v
u M
2
uX
∗(k)
(i)
di = min t
fm − fm
|P ∗|
k=1
(7)
m=1
∗(k)
where P ∗ is the Pareto optimal front and fm is the m-th
objective function value of the k -th member of P ∗. This
metric has the constraint that it is necessary the Pareto
optimal front. Here, for each problem utilized in the experiments we used the front provided by Coello et al [1]. It is
useful to note that before calculating this distance measure,
it is necessary to normalize the objective function values.
The second metric used measures the spread of the obtained set of solutions calculating the non-uniformity in the
distribution. It was proposed by Deb et al [4] as follows:
P|Q|−1
¯
df + dl + i=1
|di − d|
(8)
¯
df + dl + (|Q| − 1)d
where di is any distance metric between neighboring solutions, d¯ is the mean value of these distance measures and df
∆=
and dl are the distances between boundary solutions from
the set of obtained solutions and the Pareto optimal front.
For both metrics, a lower value implies in a better result.
In the end, we run each configuration 100 independent
times until the 100-th generation in each problem. The obtained results according to the metrics spread and generational distance are shown respectively in Table 1 and Table
2. In each row of these tables, we have the upper cell containing the mean for the 100 runs (the lower value is highlighted
with bold font) and below it a cell containing the standard
deviation. Moreover, for the rows representing the settings
n = 5 and n = 20 we have a bottom cell containing the
results of the use of the statistical test called test t with a
confidence level of 95%.
This test is applicable for comparing two samples of two
populations normally distributed, not necessarily of the same
size, where the mean and the variance of the population
are unknown. We used this test for understanding whether
there is a statistically significant difference between the results produced by the setting n = 5 or n = 20 and the results
obtained by the adaptive approach. The value ”+” indicates
that the adaptive approach will have a lower value with 95%
of confidence, the value ”−” represents that the adaptive approach will have a higher value with 95% of confidence and
the value ”≈” means that there is not statistically significant
difference between the approaches.
Table 1: Results obtained by the metric spread
Setting
n=5
n = 20
adaptive
ZDT1
0.443
0.078
≈
0.609
0.070
+
0.428
0.065
ZDT2
0.619
0.175
+
0.940
0.076
+
0.463
0.077
ZDT3
0.574
0.062
+
0.677
0.063
+
0.561
0.051
ZDT6
0.482
0.201
≈
0.646
0.213
+
0.462
0.140
FON2
0.441
0.086
+
0.412
0.083
≈
0.410
0.091
Table 2: Results obtained by the metric GD
Setting
n=5
n = 20
adaptive
ZDT1
0.012
0.008
+
0.077
0.023
+
0.008
0.005
ZDT2
0.011
0.007
+
0.212
0.159
+
0.005
0.002
ZDT3
0.008
0.004
+
0.052
0.016
+
0.006
0.003
ZDT6
0.016
0.040
+
0.039
0.052
+
0.008
0.019
FON2
0.005
0.001
≈
0.005
0.001
≈
0.005
0.001
As can be seen from the tables, the adaptive mutation operator got the lowest means for both metrics in all problems.
Furthermore, in 3 of 5 problems the adaptive approach obtained the lowest standard deviation for the metric spread
and in all problems it got the lowest standard deviation for
the metric generational distance.
Looking for the results of the test t, the adaptive approach
was superior to the setting with n = 5 in 3 problems for the
metric spread and in 4 problems for the metric generational
distance. In relation to the setting n = 20, the adaptive
approach was better in 4 problems for both metrics.
5. CONCLUSIONS
This paper presented the first step for creating adaptive
controls for each parameter present in the algorithm NSGAII to improve even more its performance. We proposed an
adaptive mutation operator that uses information about the
diversity of the population, through the concept of crowding
distance, for controlling the strength of the mutation.
Running the algorithm NSGA-II on five different problems, we compared the results obtained by the adaptive approach with the results obtained by two different static settings: a setting that applied a strong mutation and a setting
that applied a smooth mutation. The experimental results
have shown that the approach proposed outperformed both
settings in convergence to the Pareto optimal Front and in
diversity of the final solutions. A statistical test was done
to prove the relevance of the results.
While the approach seems interesting, it is clear that more
work will be necessary to understand its impact on the search.
Moreover, a clear empirical study is required to demonstrate
its significance. It is useful to cite that this approach can also
be used for controlling parameters of other operators. For
instance, the parameter that controls the proximity of the
offspring from the parents in the crossover operator (SBX)
proposed by Deb [3] can be controlled in such way that new
solutions staying closer of parents with higher crowding distance. This would help in increasing diversity.
6. REFERENCES
[1] C. A. C. Coello, G. B. Lamont, and D. A. V.
Veldhuizen. Evolutionary Algorithms for Solving
Multi-Objective Problems. Springer, 2007.
[2] K. Deb. Multi-Objective Optimization using
Evolutionary Algorithms. John Wiley and Sons, 2001.
[3] K. Deb and M. Goyal. A combined genetic adaptive
search (geneas) for engineering design. Computer
Science and Informatics, 26(4):30–45, 1996.
[4] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A
fast and elitist multiobjective genetic algorithm:
Nsga-ii. IEEE Transactions on Evolutionary
Computation, 6(2):182–197, April 2002.
[5] J. J. Durillo, A. J. Nebro, F. Luna, B. Dorronsoro, and
E. Alba. jMetal: A Java Framework for Developing
Multi-Objective Optimization Metaheuristics. Technical
Report ITI-2006-10, Departamento de Lenguajes y
Ciencias de la Computación, University of Málaga,
E.T.S.I. Informática, Campus de Teatinos, December
2006.
[6] A. E. Eiben and J. E. Smith. Introduction to
evolutionary computing. Springer, 2003.
[7] C. M. Fonseca and P. J. Fleming. Multiobjective
genetic algorithms made easy: Selection, sharing and
mating restriction. In Proceedings of the First
International Conference on Genetic Algorithms in
Engineering Systems: Innovations and Applications,
pages 45–52, 1995.
[8] D. A. V. Veldhuizen. Multiobjective evolutionary
algorithms: classifications, analyses, and new
innovations. PhD thesis, Air Force Institute of
Technology, Wright Patterson AFB, OH, USA, 1999.
[9] E. Zitzler, K. Deb, and L. Thiele. Comparison of
multiobjective evolutionary algorithms: Empirical
results. Evolutionary Computation, 8(2):173–195, 2000.
| 9cs.NE
|
QUASICONVEXITY AND DEHN FILLING
arXiv:1708.07968v1 [math.GR] 26 Aug 2017
DANIEL GROVES AND JASON FOX MANNING
Abstract. We define a new condition on relatively hyperbolic
Dehn filling which allows us to control the behavior of a relatively
quasiconvex subgroups which need not be full. As an application,
in combination with a recent result of Cooper and Futer [6], we
provide a new proof of the virtual fibering of non-compact finitevolume hyperbolic 3–manifolds, a result first proved by Wise [24].
Additionally, we explain how the results of [2, Appendix A] can be
generalized to the relative setting to control the relative height of
relatively quasiconvex subgroups under appropriate Dehn fillings.
1. Introduction
Dehn filling results for hyperbolic and relatively hyperbolic groups
have been used to great effect in recent years, notably in solving the
isomorphism problem for a broad class of relatively hyperbolic groups
[8], and as part of Agol’s proof of the Virtual Haken Conjecture [2] (see
particularly the Appendix to [2] and [4]). In many of these results a
key ingredient is the control of relatively quasiconvex subgroups under
Dehn filling, building on techniques developed in [3].
In previous work, this control was limited by the requirement that
the fillings be ‘H–fillings’, for a relatively quasiconvex subgroup H.
This requirement is mild when H is ‘full’1 but more restrictive for
general relatively quasiconvex subgroups. One way to avoid this issue
is to apply combination theorems such as those in [17, 16], etc. to
enlarge relatively quasiconvex subgroups to full ones. Even in case this
is possible, the methods of this paper are conceptually simpler as they
avoid this intermediate enlargement step.
Both authors thank the National Science Foundation for support (under grants
DMS-1507067 and DMS-1462263). The first author was partially supported by a
grant from the Simons Foundation (#342049 to Daniel Groves). Thanks also to the
Mathematical Sciences Research Institute, where the second author was in residence
during the conception of this work, and to the American Institute of Mathematics,
where the paper was finished.
1A subgroup of a relatively hyperbolic group is full if each infinite intersection
with a maximal parabolic subgroup is finite index in that parabolic.
1
QUASICONVEXITY AND DEHN FILLING
2
In this paper, we propose a new condition on relatively hyperbolic
Dehn filling, which we call H–wide, which is applicable to relatively
quasiconvex subgroups H in much greater generality than previous
techniques. We prove that under sufficiently long and H–wide fillings,
the same control can be had over the behavior of a relatively quasiconvex subgroup H under filling as could be obtained with sufficiently
long H–fillings in the previous works.
We provide two main applications. The first is to combine with
a recent result of Cooper and Futer [6] to provide a new proof of the
virtual specialness of finite-volume hyperbolic 3–manifolds, a result first
proved by Wise [24, Theorem 14.29].
Theorem A. Suppose that G is the fundamental group of a noncompact finite-volume hyperbolic 3–manifold. Then G is virtually compact special.
The following is an immediate consequence of Theorem A and Agol’s
criterion for fibering [1, Theorem 1.1] (and of Agol’s result [2, Theorem
9.2] in the compact case).
Corollary B. Suppose that M is a finite-volume hyperbolic 3–manifold.
Then M has a finite-sheeted cover which fibers over the circle.
Previous to this paper, Wise’s unpublished manuscript [24] contained
the only proof that a finite-volume hyperbolic 3-manifold virtually
fibers over a circle. Our proof does not rely on any results from [24]
(and neither does the one in [6]).
The second application we provide is to use H–wide fillings to explain
how the results from [2, Appendix A] can be generalized to control the
‘relative height’ of relatively quasiconvex subgroups under Dehn fillings.
We apply this to prove a result (Theorem 7.18) needed by Wilton and
Zalesskii in their work [23] on profinite rigidity of 3–manifold groups.
1.1. On virtual fibering of hyperbolic 3–manifolds. Agol proved
in [2] that fundamental groups of closed hyperbolic 3–manifolds are virtually special, which implies that these manifolds are virtually Haken,
and virtually fibered. He also proved that Kleinian groups are LERF,
and large.
In the non-compact but finite-volume case, the LERF and large results are included in Agol’s result, and these manifolds are well known
to be Haken. However virtual fibering in the non-compact case is not
covered by [2].
QUASICONVEXITY AND DEHN FILLING
3
In Wise’s proof of [24, Theorem 16.16]2, he asserts that Osin [20, Theorem 1.1] proves that the image of a relatively quasiconvex subgroup
remains relatively quasiconvex under long Dehn fillings. Osin does not
prove such a result. As explained above, existing results about controlling relatively quasiconvex subgroups under Dehn filling, such as those
in [3, 2] and also [24, Theorem 15.6], apply to relatively quasiconvex
subgroups which are full, and they do not apply in the setting needed
in [24, Theorem 16.16].
We remark that Wise’s proof of Theorem A relies on [24, Theorem
16.28], the proof of which explicitly relies on [24, Theorem 16.16] in
three places, so this appears to be quite important in Wise’s approach.
We believe the issues in [24, Theorem 16.16] can be fixed using either
a ‘Combination Theorem’ approach or techniques as in the current
paper. However, one of our goals here is to use the advances of the
last five years to give an alternative proof of virtual fibering in the
non-compact setting.
1.2. Outline. In Section 2 we recall the basic concepts about relatively
hyperbolic groups, relatively quasiconvex subgroups, and Dehn filling.
In Section 3 we introduce the notion of H–wide fillings. In Section 4 we
prove that the behavior of a relatively quasiconvex subgroup H under
sufficiently long and H–wide fillings is well-controlled. In Section 5
we prove that in certain circumstances we can ensure the existence of
appropriate H–wide fillings. In Section 6 we provide the application to
virtual specialness of fundamental groups of finite-volume hyperbolic
3–manifolds. Finally, in Section 7, we prove that the results about
height in the hyperbolic setting from [2, Appendix A] can be generalized
to control relative height under sufficiently long and H–wide fillings.
These results may be of independent interest. As an application, we
prove Theorem 7.18, the result required by Wilton and Zalesskii.
Acknowledgments. Thanks to Stefan Friedl for asking for an alternative account of virtual fibering for finite-volume hyperbolic 3–
manifolds, and to Henry Wilton for asking us to prove Theorem 7.18.
2. Background
For background on relatively hyperbolic groups, their associated
cusped spaces, and relatively hyperbolic Dehn filling see [10]. For background on relatively quasiconvex subgroups see [3, 13]. We always work
2Specific
references here are to the version of Wise’s manuscript dated October
29, 2012, which is the most recent to which we have access.
QUASICONVEXITY AND DEHN FILLING
4
in a combinatorial cusped space X = X(G, P, S), where S is some chosen generating set for G which also contains generating sets for the
peripheral groups P ∈ P. This cusped space contains a copy of the
Cayley graph of G with respect to the generators S. The depth of a
vertex of X is its distance to the Cayley graph.
Suppose that (G, P) is relatively hyperbolic. We fix a combinatorial
cusped space X for the pair (G, P) as in [10]. Since (G, P) is relatively
hyperbolic, X is Gromov hyperbolic. Unless otherwise stated, δ is a
hyperbolicity constant for the space X.
Recall that a Dehn filling of (G, P) is determined by a collection
N = {NP }P ∈P of normal subgroups NP P . The Dehn filling is the
quotient
G(N ) = G/hh∪NP ii.
The peripheral groups of G(N ) are the images of the elements of P in
G(N ). We often abbreviate this as (G, P) → (G, P). A statement S
holds for all sufficiently long fillings if there is a finite set B ⊂ ∪P r{1}
so that S holds for any fillings G(N ) so that NP ∩ B = ∅ for all P ∈ P.
If (G, P) is a Dehn filling of (G, P), with G = G/K, then one obtains
a combinatorial cusped space X for (G, P) by taking X equal to K X
with self-loops removed. In fact, since the self-loops do not affect the
metric on the zero-skeleton, we ignore
the issue of removing them and
X
abuse notation by setting X = K
.
The following result is key to any approach to relatively hyperbolic
Dehn filling theorems using the cusped space.
Theorem 2.1. Using the cusped spaces just described, let B be a finite
metric
ball in X. For all sufficiently long fillings the quotient map X →
K X restricts to an isometric embedding whose image is a metric ball.
Proof. This follows immediately from [20, Theorem 1.1].
Using the coarse Cartan-Hadamard Theorem [7, A.1] and the uniform hyperbolicity of combinatorial horoballs [10, Theorem 3.8] we
obtain the following corollary, which was stated in a slightly weaker
form as [3, Proposition 2.3].
Proposition 2.2. For all δ > 0 there is a δ 0 > 0 so that if the combinatorial cusped space X is δ–hyperbolic, then for all sufficiently long
fillings, the combinatorial cusped space X of the Dehn filling is δ 0 –
hyperbolic.
Remark 2.3. Proposition 2.2 implies that (G, P) is relatively hyperbolic, using only Theorem 2.1. The proof of Theorem 2.1 in [10] used
the bicombing of X by preferred paths, whereas the proof that the
QUASICONVEXITY AND DEHN FILLING
5
cusped space of X is Gromov hyperbolic used a homological bicombing
which used an adaptation of results of Mineyev from [19]. Using the
above approach allows one to avoid the homological bicombing in [10]
entirely.
Definition 2.4. Suppose that (G, P) is relatively hyperbolic with associated cusped space X. Let A be a horoball in X, and let R > 0.
A geodesic penetrates A to depth R or R–penetrates A if it contains a
point in A at depth R.
Suppose H ≤ G. Then A is R–penetrated by H if there is a geodesic
γ with endpoints in H which R–penetrates A.
Recall the following result from [16].
Proposition 2.5. [16, Proposition A.6] Let (G, P) be relatively hyperbolic and H ≤ G be relatively quasiconvex. There is a constant R
so that whenever a horoball A of X is R–penetrated by H then the
intersection of H with the stabilizer of the horoball is infinite.
The following is a combination of Lemma 3.3 from [11], and a statement implicit in its proof.
Lemma 2.6. [11, Lemma 3.3] Suppose (G, P) is relatively hyperbolic,
with δ–hyperbolic combinatorial cusped space X. Suppose further that
P1 and P2 are distinct conjugates of elements of P, and that F =
P1 ∩ P2 . Then F acts freely on some set Q in X which lies in the
Cayley graph and has diameter (in X) at most 2δ + 1.
In particular, there is a constant C depending only on δ and the
cardinality of S so that #F ≤ C.
The second part of the Lemma says that if P1 and P2 are distinct
maximal parabolics, then #P1 ∩ P2 ≤ C. In other words, the family P
is C–almost malnormal. In particular, for a parabolic subgroup A of
size more than C, there is no ambiguity about which g ∈ G and which
P ∈ P has A ≤ P g (up to the choice of conjugating element in gP ).
The following result was stated without proof as [11, Proposition
3.4]. The proof that we provide here is more elementary than the one
suggested in [11].
Proposition 2.7. If (G, P) is relatively hyperbolic, and P is C–almost
malnormal, then for all sufficiently long fillings (G, P) of (G, P), the
collection P is C–almost malnormal.
Proof. By Proposition 2.2, there is a δ 0 so that for all sufficiently long
fillings (G, P) of (G, P), the cusped space X for (G, P) is δ 0 –hyperbolic.
Fix a filling (G, P) so that the induced map between cusped spaces is
QUASICONVEXITY AND DEHN FILLING
6
injective on any ball of radius 100δ 0 centered in the Cayley graph of X
(see Theorem 2.1).
Suppose that P 1 and P 2 are distinct conjugates of elements of P,
and let F = P 1 ∩ P 2 . There are horoballs A1 and A2 in X so that
P i stabilizes Ai . As explained in the proof of [11, Lemma 3.3], the
subgroup F acts freely on a subset of the Cayley graph of G in X of
diameter at most 2δ 0 + 1. Moreover, it is clear by considering the F –
orbit of a geodesic between the limit points of A1 and A2 in ∂X that
there are also subsets Q1 and Q2 of diameter at most 2δ 0 + 1 so that F
acts freely on each Qi and Qi is contained in Ai at depth 5δ 0 .
Suppose that a geodesic between Q1 and Q2 10δ 0 –penetrates some
other horoball B. Then let B be the closest such horoball to A1 , and
replace A2 by B and Q2 by an F –invariant subset Q of B at depth 5δ 0
and diameter at most 2δ 0 + 1. In this manner, we may suppose that
any geodesic between Q1 and Q2 stays within a 10δ 0 –neighborhood of
the Cayley graph in X 0 .
We may thus lift Q1 , Q2 and the geodesics between them to X. To
see that this is possible, consider that any pair of points in Q1 and pair
of points in Q2 are the vertices of a geodesic quadrilateral with two sides
of length at most 5δ 0 and so can be filled with a disk which lies entirely
within a 20δ 0 –neighborhood of the Cayley graph. In an entirely similar
way to the proof of [11, Theorem 4.1] (a result whose proof did not
rely on the result we are currently trying to prove), it now follows that
F can be lifted bijectively to a finite subgroup F of G which stabilizes
two distinct horoballs. Because P is C–almost malnormal, it follows
that |F | = |F | ≤ C, which is what we were required to prove.
Definition 2.8. Suppose that (G, P) is relatively hyperbolic and that
X is a cusped space for (G, P) which is δ–hyperbolic. A parabolic
subgroup Q of G is uniquely parabolic if there is a unique conjugate of
an element of P which contains Q.
It follows from [11, Lemma 3.3] that there is a constant C so that any
parabolic subgroup of size more than C, and in particular any infinite
parabolic subgroup, is uniquely parabolic.
It is an immediate consequence of the definition that a uniquely
parabolic subgroup stabilizes a unique horoball in the cusped space.
In order to fix notation, we recall a definition from [3, Section 3],
and slightly adapt the notation from there. Let H ≤ G. Suppose that
D is a collection of representatives of H–conjugacy classes of maximal
uniquely parabolic subgroups of H. Given D ∈ D, there exists PD ∈ P
and cD ∈ G so that D ≤ cD PD c−1
D . We fix such cD , and suppose
that cD is a shortest possible choice. We abuse notation slightly and
QUASICONVEXITY AND DEHN FILLING
7
write (H, D) ≤ (G, P). Let Y be a combinatorial cusped space for
the pair (H, D). The inclusion ι : H ,→ G extends to an H–equivariant
Lipschitz map ι̌ : Y (0) → X as follows:
A vertex in a horoball of Y is determined by a triple (sD, h, n) where
s ∈ H, D ∈ D and n ∈ N. We define
ι̌(sD, h, n) = (scD PD , hcD , n).
It follows from [3, Lemma 3.1] that ι̌ is H–equivariant and α–Lipschitz
for some α. We refer to ι̌ as the induced map on cusped spaces. Whenever we have a pair (H, D) ≤ (G, P) as above, we fix the subgroups
PD ∈ P and the (shortest) elements cD as above.
Definition 2.9. Suppose that (G, P) is relatively hyperbolic and that
H ≤ G is a subgroup. Suppose that D consists of a set of representatives of H–conjugacy classes of maximal uniquely parabolic subgroups
of H. Then (H, D) is relatively quasiconvex in (G, P) if the image of
the 0–skeleton of the cusped space of (H, D) in the cusped space of
(G, P) is λ–quasiconvex for some λ. In this case we say that λ is a
quasiconvexity constant for (H, D) in (G, P).
This definition is slightly different than the one in [3], since we do not
assume that (H, D) is relatively hyperbolic. However, we do assume
that D consists of maximal uniquely parabolic subgroups of H. If the
image of the cusped space of (H, D) in the cusped space of (G, P) is
quasiconvex, then it follows from the proof of [16, Theorem A.10] that
H is relatively quasiconvex in the sense of Hruska [13], and hence that
(H, D) is relatively hyperbolic. Therefore, this definition is equivalent
to others in the literature, by the results in [16, Appendix A].
3. H-wide fillings
Definition 3.1. Let P be a group, B ≤ P a subgroup, and S a finite
set. A normal subgroup N P is (B, S)–wide in P if whenever there
are b ∈ B and s ∈ S so that bs ∈ N we have s ∈ B.
Definition 3.2. Let (G, P) be relatively hyperbolic
and let (H, D) ≤
S
(G, P) be relatively quasiconvex. Let S ⊆ P r {1}. A filling
G → G = G(N )
is (H, S)–wide iffor any D ∈ D
(with D ≤ PDcD as above) the normal
−1
subgroup ND is DcD , S ∩ PD –wide in PD . (To simplify notation, for
D ∈ D, we write ND for NPD .3)
3Since
it is possible that PD1 = PD2 for D1 6= D2 , it is also possible that
ND1 = ND2 . We also remark that ND need not be a subgroup of H.
QUASICONVEXITY AND DEHN FILLING
8
We say that a property P holds for
S all sufficiently long and H-wide
fillings if there is a finite set S ⊆ P r {1} so that P holds for any
(H, S)–wide filling G → G(N ) for which N ∩ S = ∅ for each N ∈ N .
In place of the statement in Definition 3.2 above, we sometimes use
the equivalent formulation that for any D ∈ D (with D ≤ PDcD as
cD
above), any d ∈ D and any w ∈ S ∩ PD , if dcD wc−1
D ∈ ND , then
cD wc−1
D ∈ D.
Remark 3.3. In the definition of (H,S
S)–wide, one should think of S
containing all nontrivial elements of P in a large ball around the
identity. This ensures that, for each D ∈ D, a “big neighborhood” of
cD
D = D/(NDcD ∩ D) embeds in P D = (PD /ND )cD , ruling out behavior
like that pictured in Figure 1.
cD
PD
w cD
1
−1
d2
d1
D
Figure 1. A cartoon of the coset graph for NDcD in PDcD
and the kind of loop forbidden by (H, S)–wideness with
w ∈ S.
Previous quasiconvex Dehn filling results [2, 3, 16] have been in terms
of “H–fillings”, whose definition we now recall.
Definition 3.4. Let (G, P) be relatively hyperbolic, let H < G be
relatively quasiconvex, and let N = {NP }P ∈P be a collection of filling
kernels. The Dehn filling G → G = G(N ) is said to be an H–filling if,
whenever #(P g ∩ H) = ∞, the kernel NPg lies entirely in H.4
4In
[3] the condition ‘P g ∩ H 6= {1}’ was used instead of ‘#(P g ∩ H) = ∞.’
As explained in [16], the formulation in Definition 3.4 is the correct one if there is
torsion, and this is the definition that is used in [2, 16].
QUASICONVEXITY AND DEHN FILLING
9
The following result shows that, at least for long fillings, the notion
of H–wide filling generalizes that of H–filling.
Lemma 3.5. Let (G, P) be relatively hyperbolic, and let H < G be
relatively quasiconvex. For any finite S ⊂ G any sufficiently long H–
filling is (H, S)–wide.
Proof. Let R be the constant from Proposition 2.5, as applied to H.
Let D be the peripheral structure on H consisting of maximal uniquely
parabolic subgroups, and {PD ∈ P} and {cD ∈ G} the elements
described before, so that D < PDcD for each D ∈ D. Let M =
max{dX (1, cD )} + max{dX (1, w) | w ∈ S}. Choose filling kernels
{Nj Pj } determining a sufficiently long H–filling so that any geodesic joining 1 to n ∈ Nj \ {1} must (R + M + 2δ + 2)–penetrate the
horoball stabilized by Pj .
Now suppose that for some w ∈ S and some d ∈ D ∈ D we have
−1
dwcD ∈ NDcD . We must show that wcD ∈ D. Since (dwcD )cD ∈ ND ,
−1
the geodesic from 1 to (dwcD )cD must (R + M + 2δ + 2)–penetrate
−1
the horoball stabilized by PD . In particular, dX (1, (dwcD )cD ) must be
at least 2R + 2M + 4δ + 4. Consider the quadrilateral with vertices
1, cD , dcD w, d. The segment [cD , dcD w] is the translate of a geodesic
−1
[1, (dwcD )cD ] by cD , so it has length at least 2R + 2M + 4δ + 4 and
(R+M +2δ+2)–penetrates the horoball based on cD PD . Since the sides
[1, cD ] and [dcD w, d] have length at most M , the side [1, d] must pass
within 2δ of [cD , dcD w] at its midpoint, which is also its deepest point in
the horoball. In particular [1, d] must R–penetrate the horoball based
on cD PD . Since d ∈ H, Proposition 2.5 implies that H ∩ PDcD = D is
infinite. Since the filling kernels N determine an H–filling, this implies
that ND < D, and in particular, the element dwcD ∈ D. It immediately
follows that wcD ∈ D, so we have established that the filling is (H, S)–
wide.
In any case, if (H, D) is a relatively quasiconvex subgroup of the
relatively hyperbolic pair (G, P), any Dehn filling of (G, P) induces a
Dehn filling of (H, D), which may or may not inject into the filling of
G.
Definition 3.6. Let (G, P) be relatively hyperbolic, and let H < G
be relatively quasiconvex. Let D be the canonical (uniquely parabolic)
peripheral structure on H, so each D ∈ D is contained in some PDcD
for a unique PD ∈ P, and some shortest cD . Let N = {NP }P ∈P be
a collection of filling kernels for (G, P). The induced filling kernels
for (H, D) are the collection NH = {NPcDD ∩ D}D∈D . These define the
QUASICONVEXITY AND DEHN FILLING
10
induced filling
(H, D) −→ (H(NH ), D),
where
S D consists of the images of the elements of D in H(NH ) =
H/hh NH iiH .
There is a natural map from H(NH ) to the filling G(N ).
4. Properties of H–wide fillings
In this section we prove various results which imply that a relatively
quasiconvex subgroup H can be controlled in H–wide fillings. These
results should be compared to those in [3, Section 4], where analogous results are proved for the behavior of a full relatively quasiconvex
subgroup H under sufficiently long H–fillings.
Let (G, P) be relatively hyperbolic. According to Proposition 2.2,
there exists a constant δ so that the cusped space for G is δ–hyperbolic,
and moreover the induced cusped spaces for sufficiently long fillings of
(G, P) are also δ–hyperbolic. In this section, we assume that δ is such
a constant, and that all fillings we perform are long enough so that the
cusped spaces of the filled groups are δ–hyperbolic.
The following lemma is a reformulation of [3, Lemma 4.1].
Lemma 4.1. Suppose (G, P) is relatively hyperbolic, and that L1 , L2 ≥
10δ. For sufficiently long fillings π : G → G = G/K with induced map
between cusped spaces π : X → X, and any geodesic γ in X either:
(1) There is a 10δ–local geodesic in X between the endpoints of π(γ)
which lies in a 2–neighborhood of π(γ) and agrees with π(γ) in
the L1 –neighborhood of the Cayley graph in X; or
(2) There is a horoball A in X so that γ L2 –penetrates A in a
segment [x, y] with x, y ∈ G, and there is some k ∈ K stabilizing
A so that dX (x, k.y) ≤ 2L1 + 3.
The following result is very similar to [3, Lemma 4.2] but for H–wide
fillings rather than H–fillings. The induced filling is defined above in
Definition 3.6.
Lemma 4.2. Suppose that (G, P) is relatively hyperbolic and that H ≤
G is relatively quasiconvex. Let R be the constant from Proposition
2.5. For any L1 ≥ 10δ, L2 > max{2L1 + 3, R}, and all sufficiently
long and H–wide fillings π : G → G, the following holds: suppose that
KH ≤ ker(π) ∩ H is the kernel of the induced filling of H, that h ∈ H
and that γ is a geodesic from 1 to h. If conclusion (2) of Lemma 4.1
holds then there exists k ∈ KH so that dX (1, kh) < dX (1, h).
QUASICONVEXITY AND DEHN FILLING
11
Proof. Let D be a collection of representatives of H–conjugacy classes
of maximal uniquely parabolic subgroups of H, so that (H, D) is relatively quasiconvex in (G, P).
Let γ be a geodesic as in the statement of the lemma, and suppose
that conclusion (2) of Lemma 4.1 holds. Accordingly there is some
horoball A which is L2 –penetrated by γ. Let gP be the coset on which
A is based. According to Proposition 2.5, H ∩ P g is infinite. This
implies that there are r ∈ H, and D ∈ D, so that P = PD and
gP = rcD PD . The intersection of γ with A is the segment [x, y], and
there is an element k ∈ NDrcD so that dX (x, k.y) ≤ 2L1 + 3.
Now, by quasiconvexity of (H, D), there exists some d1 , d2 ∈ D so
that dX (x, rd1 cD ), dX (y, rd2 cD ) are both bounded by some constant L
depending only on the quasiconvexity constant for (H, D).
Let w1 = (rd1 cD )−1 ky and w2 = y −1 rd2 cD . Both dX (1, w1 ) and
dX (1, w2 ) are at most L + 2L1 + 3, and both w1 and w2 lie in PD .
Let S be the set of words in the parabolic subgroups of X–length at
most 2(L + 2L1 + 3). Since k ∈ NDrcD , we can find n ∈ ND so that
−1
k = rcD nc−1
D r .
We have
k = ky.y −1
−1 −1
= rd1 cD w1 w2 c−1
D d2 r .
Therefore,
−1 −1
cD nc−1
D = d1 cD w1 w2 cD d2 ,
and
−1
−1
−1
(d−1
= (d−1
2 cD )n(d2 cD )
2 d1 )cD w1 w2 cD .
However, for an (H, S)–wide filling, there can only be an element of
−1
ND of this form if cD w1 w2 c−1
D ∈ D. This implies that cD ncD ∈ D,
cD
r
which implies that k ∈ D ∩ ND ≤ KH .
Since dX (x, ky) ≤ 2L1 + 3, but dX (x, y) ≥ 2L2 > 2L1 + 3, it is clear
that dX (1, k.h) < dX (1, h), as required.
The following result is an immediate consequence.
Corollary 4.3. Suppose that (G, P) is relatively hyperbolic and that
H ≤ G is relatively quasiconvex. For any L ≥ 10δ and for all sufficiently long and H–wide fillings π : G → G, if h ∈ H is the shortest
element of H ∩ π −1 (π(h)) and γ is a geodesic from 1 to h then there is
a 10δ–local geodesic in X with the same endpoints as π(γ) which lies in
a 2–neighborhood of π(γ) and agrees with π(γ) in an L–neighborhood
of the Cayley graph of G in X.
QUASICONVEXITY AND DEHN FILLING
12
Recall that in a δ–hyperbolic space, 10δ–local geodesics are quite
close to geodesics. In particular, we have the following (see [5, III.H.1.13]
for a more general and precise statement):
Lemma 4.4. Let γ be a 10δ–local geodesic in a δ–hyperbolic space.
Then γ is a (7/3, 2δ)–quasigeodesic, and is Hausdorff distance at most
3δ from any geodesic with the same endpoints.
Lemma 4.2, and its interpretation in the form of Corollary 4.3 are
the key results needed to generalize many results about H–fillings to
sufficiently long and H–wide fillings, as we now explain.
Proposition 4.5 (cf. Proposition 4.3, [3]). Let (G, P) be relatively
hyperbolic and suppose that H is a relatively quasiconvex subgroup of
(G, P), with relative quasiconvexity constant λ. There exists λ0 =
λ0 (λ, δ) so that for all sufficiently long and H–wide fillings π : G → G
the subgroup π(H) is λ0 –relatively quasiconvex in G.
Proof. Recall that at the beginning of the section we fixed a constant
δ so that the cusped space X for (G, P) is δ–hyperbolic and that for
sufficiently long fillings π : G → G the cusped space X for (G, P ) is
also δ–hyperbolic. Suppose that H is λ–relatively quasiconvex.
Let h ∈ π(H) and suppose that h ∈ H is the shortest element of H
so that π(h) = h. Let γ be a geodesic from 1 to h in X. By Corollary
4.3 with L = 10δ, for sufficiently long and H–wide fillings there is a
10δ–local geodesic in X from 1 to h which lies in a 2–neighborhood of
π(γ) and agrees with π(γ) in a 10δ–neighborhood of the Cayley graph.
By Lemma 4.4, any geodesic from 1 to h is contained in an (3δ + 2)–
neighborhood of π(γ), and thus within a (λ + 3δ + 2)–neighborhood of
the image of the cusped space of H in X. This suffices to prove the
result, as in the proof of [3, Proposition 4.3]. (All that remains is to
consider geodesics between points in the image of the cusped space of
H which do not lie at depth 0, and it is straightforward to deal with
these points given what has already been proved.)
Proposition 4.6 (cf. Proposition 4.4, [3]). Let H ≤ G be relatively
quasiconvex. For sufficiently long and H–wide fillings π : G → G the
map from the induced filling of H to G is injective.
Proof. Let X be the cusped space for G and X the cusped space for
G. Suppose that h ∈ H ∩ ker(π) is nontrivial. Let KH be the kernel of
the induced filling on H. We must show that h ∈ KH .
Let γ be a geodesic in X from 1 to h, and note that π(γ) is a loop.
Suppose that condition (1) from Lemma 4.1 holds. Then there is a
QUASICONVEXITY AND DEHN FILLING
13
nontrivial 10δ–local geodesic loop based at 1 in X agreeing with π(γ)
in a 10δ–neighborhood of 1 ∈ X. This is impossible.
Therefore, Lemma 4.2 applies, and there is an element k ∈ KH so
that dX (1, kh) < dX (1, h). Induction on the length of h shows that
h ∈ KH , as required.
Proposition 4.7 (cf. Proposition 4.5, [3]). Let H ≤ G be relatively
quasiconvex and suppose that g ∈ G r H. For sufficiently long and
H–wide fillings π : G → G we have π(g) 6∈ π(H).
Proof. Choose L1 = 3dX (1, g) + 10δ and any L2 > max{2L1 + 3, R},
and suppose that π is sufficiently long and H–wide that Lemmas 4.1
and 4.2 hold for π, and also so that π induces a bijection between the
ball of radius L1 about 1 in X and the ball of radius L1 about the
image of 1 in the cusped space of π(G).
In order to obtain a contradiction, suppose that π(g) ∈ π(H), and
choose h ∈ H ∩ π −1 (π(g)) with dX (1, h) minimal. Let γ be a geodesic
from 1 to h in X, and let σ be a geodesic from 1 to g in X. Note that
π(σ) is a geodesic.
The minimality of h and Lemma 4.2 ensure that condition (1) from
Lemma 4.1 holds for γ.
There are now two cases, depending on whether π(γ) (equivalently
γ) leaves the L1 –neighborhood of the Cayley graph. If γ lies in the L1 –
neighborhood of the Cayley graph, it is a 10δ–local geodesic joining 1
to g. Its length is therefore at most 73 dX (1, g) + 2δ < L1 , by Lemma
4.4. But since π is injective on the L1 –ball about 1, this implies g = h,
a contradiction.
The second case is that π(γ) leaves the L1 –neighborhood of the Cayley graph, in which case there is a 10δ–local geodesic as in Lemma 4.1,
joining 1 to g, which coincides with π(γ) in the L1 –neighborhood of the
Cayley graph, but may differ elsewhere. The length of this 10δ–local
geodesic is at least L1 > 73 dX (1, g) + 2δ, again contradicting Lemma
4.4.
Proposition 4.7 is a kind of weak separability result; it says the kernel
of a nice enough filling misses a chosen coset gH for quasiconvex H, and
given g 6∈ H. The following simple example shows that it is not possible
to replace “H–wide fillings” or “H–fillings” simply by “sufficiently long
fillings” in Proposition 4.7.
Example 4.8. Suppose that G = ha, b | [a, b] = 1i ∗ hc, d | [c, d] = 1i ∼
=
Z2 ∗ Z2 .
QUASICONVEXITY AND DEHN FILLING
14
For some i > 1, let H be the (full, infinite-index) subgroup hai , b, c, di.
Since G is locally relatively quasiconvex, H is a quasiconvex subgroup
of the relatively hyperbolic pair (G, {ha, bi, hc, di}).
For some j > 1 coprime to i, let k = baj . Then it is straightforward
to see that the image of H in
G → Gj = G (hki, {1})
is Gj , for any j coprime to i, even though a 6∈ H. These fillings are not
H–wide fillings. They are however a “cofinal” sequence of fillings, so
any statement which applies to all sufficiently long fillings must apply
to all but finitely many of the Gj .
The fact that the image of a lies in the image of H illustrates the
necessity of restricting to H–wide fillings in Proposition 4.7.
5. Existence of H–wide fillings
In this section, we prove two results which imply that in our applications in Sections 6 and 7 we can find sufficiently long and H–wide
fillings. The key observation is that separability allows us to do this.
Lemma 5.1. Suppose that P is a group and that B is a separable subgroup. For any finite set S there exists a finite-index normal subgroup
KS ≤ P so that for any N P with N ≤ KS , the subgroup N is
(B, S)–wide in P .
Proof. For each s ∈ S r B, choose some
T Ps ≤ P finite index and
satisfying B < Ps and s 6∈ Ps . Let KS = {Ps | s ∈ S r B}, and note
that KS is finite index in P , and contains B.
Suppose N P is contained in KS . We verify that N is (B, S)–wide.
Let b ∈ B and s ∈ S, and suppose bs ∈ N . If s ∈ B there is nothing
to show, so suppose s 6∈ B. The element s is not contained in KS , but
bs ∈ N < KS , so we must have b 6∈ KS . But this contradicts B ≤ KS .
The subgroup KS just constructed may not be normal, but we may
replace KS by its normal core without disturbing the conclusion.
In Lemma 5.2 we consider a finite collection {(H1 , D1 ), . . . , (Hk , Dk )}
of relatively quasiconvex subgroups of a relatively hyperbolic pair (G, P).
For each i and each D ∈ Di we assume that we have fixed PD ∈ P and
cD ∈ G so that D ≤ PDcD , and that cD is a shortest such conjugating
element.
Lemma 5.2. Suppose that (G, P) is relatively hyperbolic, and that
H = {(H1 , D1 ), . . . , (Hk , Dk )} is a collection of relatively quasiconvex
subgroups. Suppose that for each 1 ≤ j ≤ k and for each D ∈ Dj the
subgroup D is separable in PDcD .
QUASICONVEXITY AND DEHN FILLING
15
S
Then for any finite S ⊂ P r {1} there exist finite index subgroups
{KP P | P ∈ P} so that any filling
G → G(N ), with N = {NP < KP | P ∈ P}
is (Hj , S)–wide for each 1 ≤ j ≤ k.
Proof. Fix S ⊂ ∪P r {1} a finite set.
Fix P ∈ P and let SP = P ∩ S. Suppose, for some i, that D ∈ Di is
so that P = PD . By Lemma 5.1 there is a finite-index normal subgroup
cD
KD P so that any N P cD for which N ≤ KD
is (D, SP )–wide in
cD
P .
We choose KP to be the intersection of all KD for which P = PD .
It is straightforward to see that NP ≤ KP then the conclusion of the
lemma holds. This completes the proof.
The following result is an immediate consequence of Lemma 5.2 due
to the fact that all subgroups of finitely generated abelian groups are
normal and separable.
Corollary 5.3. Suppose that (G, P) is relatively hyperbolic, that each
element of P is free abelian, that Q is a finite collection of relatively
quasiconvex subgroups, and that S ⊂ ∪P r {1} is a finite set. There
exist finite-index subgroups {KP P | P ∈ P} so that for any elements
γP ∈ KP , the filling
G → G ({hγP i | P ∈ P}) ,
is (Q, S)–wide for each Q ∈ Q.
6. Application to virtual specialness and virtual fibering
In this section, we explain how the results in the beginning of the
paper, as well as a recent result of Cooper and Futer [6] give a proof of
[24, Theorem 14.29] which is independent of the results from [24].
The following result is due to Cooper and Futer.
Theorem 6.1. [6, Corollary 1.3] Suppose that G is the fundamental
group of a non-compact finite-volume hyperbolic 3–manifold. Then G
acts freely and cocompactly on a CAT(0) cube complex dual to finitely
many immersed quasi-Fuchsian surfaces.
In this section, our main result is that this cubulation is virtually
special.
Theorem A. Suppose that G is the fundamental group of a noncompact finite-volume hyperbolic 3–manifold M . Then G is virtually
compact special.
QUASICONVEXITY AND DEHN FILLING
16
If P is a collection of conjugacy-representatives of maximal parabolic
subgroups of G then (G, P) is relatively hyperbolic. After possibly replacing M by an orientable double-cover of M , each element of P is
free abelian of rank 2. As explained in the proof of Theorem A below,
given the results of [2] and [6] (and the criterion for virtual specialness
[21, Criterion 2.3]) proving Theorem A reduces to establishing separability of certain double cosets of relatively quasiconvex subgroups of
G. In the closed case, such double cosets are separable by results in [2]
and [18]. We reduce to this case by performing orbifold Dehn filling on
M and applying the following ‘weak separability’ criterion for double
cosets.
Proposition 6.2. Suppose that (G, P) is relatively hyperbolic and that
each element of P is free abelian. Suppose further that H is a finite
collection
of relatively quasiconvex subgroups of (G, P) and that S ⊆
S
( P) r {1} and F ⊆ G are finite subsets.
There exist finite-index subgroups {KP P | P ∈ P so that for any
elements γP ∈ KP the filling
G → G ({hγP i | P ∈ P}) ,
is (H, S)–wide for each H ∈ H and furthermore whenever f ∈ F ,
Ψ, Θ ∈ H satisfy 1 6∈ ΨΘf , there is no element of K in ΨΘf .
Proof. By Corollary 5.3 there are finite-index subgroups KP P so
that if γP ∈ KP then the filling is (H, S)–wide for each H. Below,
we find other finite-index subgroups K̂P P so that if γP ∈ K̂P is
sufficiently long then the condition on double cosets holds. We then
choose sufficiently long γP that lie in KP ∩ K̂P . For the remainder of
the proof we concentrate on finding the subgroups K̂P .
Let X be the cusped space for (G, P) and suppose that X is δ–
hyperbolic. We further assume that δ is chosen so that the cusped
spaces of all sufficiently long fillings are δ–hyperbolic. We suppose
that δ ≥ 1. Let λ be a quasiconvexity constant which works for every
element in H. Finally, let M = max{dX (1, f ) | f ∈ F }.
Fix f ∈ F and Ψ, Θ ∈ H so that 1 6∈ ΨΘf , and consider the equation
k ∈ ΨΘf for elements k of the kernel of a filling. After finding conditions on the filling which ensure there is no such element, we consider
a filling appropriate for all f ∈ F simultaneously.
Let D be a collection of representatives of Ψ–conjugacy classes of
maximal uniquely parabolic subgroups of Ψ. For D ∈ D, we have
D ≤ PDcD for some PD ∈ P and some (shortest) cD ∈ G. Similarly, let
E be a collection of representatives of Θ–conjugacy classes of maximal
parabolic subgroups of Θ, and for E ∈ E we have E ≤ PEdE for some
QUASICONVEXITY AND DEHN FILLING
17
PE ∈ P and some (shortest) dE ∈ G. Let XΨ be the cusped space
for the pair (Ψ, D) and let XΘ be the cusped space for (Θ, E) (both
with respect to some choices of generating sets). Let ι̌Ψ : XΨ → X and
ι̌Θ : XΘ → X be the induced maps of cusped spaces, and note that
(0)
(0)
ι̌Ψ (XΨ ) and ι̌Θ (XΘ ) are both λ–quasiconvex subsets of X.
In order to apply the results from Section 4, choose L1 = 10δ and
L2 = max{20δ+M +λ+4, RΨ , RΘ }, where RΨ and RΘ are the constants
from Proposition 2.5 applied to Ψ and Θ, respectively.
For P ∈ P, let
SP ⊇ {p ∈ P | dX (1, p) ≤ 32δ + 2M + 4λ + 2L1 + 3}
be a finite set which is large enough so that for all H ∈ H the (H, S)–
wideness condition of Lemma 4.2 is satisfied with L1 and L2 as above.
Consider the collection of subgroups of P of the form PB1 ,B2 = hB1 , B2 i
where B1 = P ∩Ψg1 for some g1 ∈ G and B2 = P ∩Θg2 for some g2 ∈ G.
There are finitely many such pairs of subgroups of P . For each such
pair (B1 , B2 ), by Lemma 5.1 there exists a finite-index K̂B1 ,B2 P
T
which is (PB1 ,B2 , SP )–wide. We define K̂P = K̂B1 ,B2 and check the
condition on double cosets.
Choose γP ∈ K̂P , and consider the filling
G → G ({hγP i | P ∈ P}) = G/K.
In order to obtain a contradiction suppose that there is an element
g ∈ K, and elements f ∈ F , ψ ∈ Ψ and θ ∈ Θ so that
g = ψθf.
Choose a g so that dX (1, g) is minimal amongst all choices of g for
which there is such an expression.
Consider a geodesic quadrilateral in X with vertices 1, ψ, ψθ, g, and
let ξ1 be the geodesic from 1 to ψ, ξ2 the geodesic from ψ to ψθ, η the
geodesic from ψθ to g and ρ the geodesic from 1 to g, respectively. By
assumption, we know that g 6= 1.
Let π : X → X be the map on cusped spaces induced by the filling
map π : G → G/K. Since g ∈ K r{1} the image of ρ in X is a loop, so
condition (1) from Lemma 4.1 cannot hold. This means that condition
(2) from Lemma 4.1 holds. Let A be a horoball L2 –penetrated by π,
so π meets A in a segment [x, y], and let k ∈ K ∩ Stab(A) be so that
dX (x, k.y) ≤ 2L1 + 3. It is straightforward to see that dX (1, k.g) <
dX (1, g), since dX (x, k.y) ≤ 2L1 + 3 but dX (x, y) ≥ 2L2 > 2L1 + 4. We
arrive at a contradiction by showing that k.g ∈ ΨΘf , contradicting the
choice of g as a shortest element of K with such an expression.
QUASICONVEXITY AND DEHN FILLING
18
The geodesic through A between x and y consists of a vertical segment down from x, a horizontal segment of at most 3 edges, and then a
vertical segment terminating at y. Let x0 be the point on this geodesic
directly below x at depth 3δ + M + λ and let y 0 be the point directly
below y at depth 3δ +M +λ. The quadrilateral ξ1 ∪ξ2 ∪η ∪ρ is 2δ–slim,
so there are points on η ∪ ξ1 ∪ ξ2 within 2δ of x0 and of y 0 . Because
η is a geodesic (of length at most M ) joining two points at depth 0,
no point on η can be within 2δ of either x0 or y 0 . Therefore, there are
points on ξ1 ∪ ξ2 within 2δ of x0 and of y 0 .
The geodesic ξ1 travels between two points in Ψ and ξ2 joins two
points in ψΘ. By quasiconvexity, any point on ξ1 lies within λ of a point
in ι̌Ψ (XΨ ) and any point on ξ2 lies within λ of a point in ψ.ι̌Θ (XΘ ). Let
u0 and v0 be points in ι̌Ψ (XΨ ) ∪ ψ.ι̌Θ (XΘ ) lying within distance 2δ + λ
of x0 and y 0 respectively. Note that u0 and v0 lie in the horoball A.
There are points at depth more than 0 in a horoball A which lie in
ι̌Ψ (XΨ ) exactly when they are of the form (scD PD , hcD , n) for some
s ∈ Ψ, h ∈ sD and n ∈ N, where D ∈ D is so that D ≤ PDcD , as above,
and A is the horoball based on the coset scD PD .
Similarly, there are points at depth more than 0 in A which lie in
ψ.ι̌Θ (XΘ ) exactly when they are of the form (ψ.tdE PE , ψ.gcD , m) for
t ∈ Θ, g ∈ tE and m ∈ N, where E ∈ E and E ≤ PEdE , and A is the
horoball based on ψ.tdE PE .
The points u0 , v0 have one of these forms, and they are at distance
at most 2δ + λ from x0 and y 0 respectively, which implies that the
appropriate n or m is at most 3δ + M + 2λ. Thus, there are points
u, v at depth 0 in A, directly above u0 and v0 respectively, so that
dX (u, x), dX (v, y) ≤ 10δ + 2M + 4λ. All of the points in A directly
above u0 lie in ι̌Ψ (XΨ ) or ψ.ι̌Θ (XΘ ), except possibly the point u at
depth 0. This point u will not lie in ι̌Ψ (XΨ ) unless cD = 1, and similarly
for ι̌Θ (XΘ ). However, certainly u lies within distance 1 of ι̌Ψ (XΨ ) or
ψ.ι̌Θ (XΘ ). Similarly, v lies within distance 1 of ι̌Ψ (XΨ ) or ψ.ι̌Θ (XΘ ).
We deal with four cases, depending on whether each of u0 and v0 are
contained in ι̌Ψ (XΨ ) or ψ.ι̌Θ (XΘ ).
Case 1: Both u0 and v0 are contained in ι̌Ψ (XΨ ).
(The case where they are both contained in ψ.ι̌Θ (XΘ ) is entirely
similar and we omit it.)
In this case, if u0 = (scD PD , ψu cD , n) then u = ψu cD , where ψu ∈ Ψ
and A is the horoball based on scD PD . For ease of notation we write
P = PD and c = cD , and so we have ψu c ∈ scP .
Note that v = ψv c ∈ scP also, for some ψv ∈ Ψ. We have u−1 v =
−1
c−1 (ψu−1 ψv )c ∈ Dc .
QUASICONVEXITY AND DEHN FILLING
19
Now, dX (x, k.y) ≤ 2L1 + 3, and we also have dX (x, u), dX (y, v) ≤ α,
from which it follows that dX (u, k.v) ≤ 2α + 2L1 + 3. Write w =
v −1 k −1 u, a group element of X–length at most 2α + 2L1 + 3 and note
that u−1 k −1 u is in the filling kernel NP P , and so cu−1 k −1 uc−1 is
contained in NPc . On the other hand, we also have
cu−1 k −1 uc−1 =
=
=
=
cu−1 v(v −1 k −1 u)c−1
cu−1 vwc−1
c(c−1 ψu−1 ψv c)wc−1
(ψu−1 ψv )cwc−1 .
Note that w ∈ SP , and that the filling is (Ψ, SP )–wide. Since (ψu−1 ψv ) ∈
D, this implies that cwc−1 ∈ D, so cu−1 k −1 uc−1 = (ψu−1 ψv )cwc−1 ∈ D.
Therefore, ψu−1 kψu = cu−1 kuc−1 ∈ D, which means that
k.ψ = kψu (ψu−1 ψ) = ψu (ψu−1 kψu )ψu−1 ψ ∈ Ψ.
Therefore,
k.g = (k.ψ) θf,
gives an expression for k.g as an element of ΨΘf , contradicting the
fact that g was the shortest element of K with such an expression.
Case 2: u0 is contained in ι̌Ψ (XΨ ) and v0 contained in ψ.ι̌Θ (XΘ ).
We can write u0 = (scD PD , ψu cD , n) and v0 = (ψ.tdE PE , ψ.rv dE , m),
so u = ψu cD and v = ψ.θv dE , where ψu ∈ Ψ, θv ∈ Θ, D ≤ PDcD and
E ≤ PEdE . We clearly have PE = PD , which we write as P . We write
c = cD and d = dE , and note that A is the horoball based on the coset
scP = ψtdP .
We still have dX (u, x), dX (v, y) ≤ α. The geodesic ξ1 intersects A in
a segment [g1 , h1 ], where the entrance point g1 is within 4δ of x, and
within α of u. The exit point h1 , we may similarly argue, is within α of
some group element w = ψw c in the coset scP , with ψw ∈ Ψ. Likewise,
the geodesic ξ2 intersects the horoball A in a segment [g2 , h2 ], where
dX (g2 , h1 ) ≤ 4δ, and there is another point z = ψθz d in scP with
θz ∈ Θ, and satisfying dX (z, g2 ) ≤ α. See Figure 2.
−1
−1
−1
−1
We have u−1 w ∈ Dc and z −1 v ∈ E d . Both Dc and E d are
subgroups of P .
−1
−1
Let B1 = Dc and B2 = E d so u−1 wz −1 v ∈ PB1 ,B2 . Now,
u−1 k −1 u = (u−1 w)(w−1 z)(z −1 v)(v −1 k −1 u)
= (u−1 w)(z −1 v) (w−1 z)(v −1 k −1 u) .
(Note that we use here that P is abelian.)
Since w−1 z has length at most 2α + 4δ and v −1 k −1 u has length at
most 2α + 2L1 + 3, so the last of the three terms above is in SP and
QUASICONVEXITY AND DEHN FILLING
20
ψ
z
w
ψθ
v
u
1
x
y
ky
A
Figure 2. Case 2. The dotted line represents the coset
scP = ψtdP .
we can apply the (PB1 ,B2 , SP )–wideness of the kernel to deduce that
u−1 ku ∈ PB1 ,B2 = B1 B2 . Choose elements b1 ∈ B1 and b2 ∈ B2 so that
u−1 ku = b1 b2 . We now have
kψθ = u(u−1 ku)(u−1 w)(w−1 z)(z −1 v)(v −1 ψθ)
= u(b1 b2 )(u−1 w)(w−1 z)(z −1 v)(v −1 ψθ)
= u b1 (u−1 w) (w−1 z) (z −1 v)b2 (v −1 ψθ)
= u b1 (u−1 w) (w−1 ψ)(ψ −1 z) (z −1 v)b2 (v −1 ψθ)
= ψu c(b1 u−1 w)c−1 (cw−1 ψ)(ψ −1 z) (z −1 v)b2 (v −1 ψθ)
= ψu c(b1 u−1 w)c−1 (ψw−1 ψ) θz d(z −1 vb2 )d−1 (θv−1 θ)
The first three terms of this expression are in Ψ and the last three
terms are in Θ, which proves that kψθ ∈ ΨΘ. Therefore, k.g = kψθf ∈
ΨΘf . Since we know that dX (1, k.g) < dX (1, g), this contradicts the
minimality of g, hence proving the result in Case 2.
It remains to note that the case that u0 is contained in ψ.ι̌Θ (XΘ ) and
v0 is contained in ι̌Ψ (XΨ ) essentially becomes Case 1. Indeed, suppose
that v0 is contained in ι̌Ψ (XΨ ). Then the geodesic from v to 1 lies near
to the geodesic from y to 1, which easily implies (since x lies on the
geodesic from y to 1) that x lies near ι̌Ψ (XΨ ), as required.
We now turn to the proof of Theorem A.
g
QUASICONVEXITY AND DEHN FILLING
21
Proof (of Theorem A). Pass to a finite cover which is orientable, so
that all cusps in M have torus cross-sections. It is well known that if
P is a collection of representatives of G–conjugacy classes of maximal
cusp subgroups of M then (G, P) is relatively hyperbolic.
By Theorem 6.1, there is a CAT(0) cube complex X upon which
G acts freely and cocompactly. By [21, Criterion 2.3] (see also [12,
Section 4]), to prove that the action of G on X is virtually special it
suffices to prove that for a certain finite list of subgroups Qi which
stabilize hyperplanes in X, the subgroups Qi and double cosets Qi Qj
are separable in G. Since G is a Kleinian group, it is LERF by [2,
Corollary 9.4], so it remains to prove double coset separability.
The cube complex X built by Cooper and Futer for Theorem 6.1 is
built using the Sageev construction [22] (see also [15]). The hyperplane
subgroups in G are commensurable to the codimension 1 subgroups,
which are quasi-Fuchsian surface subgroups and therefore geometrically finite. It now follows immediately by [13, Corollary 1.3] that the
subgroups Qi are relatively quasiconvex in G .
Suppose now that h 6∈ Qi Qj . Equivalently, 1 6∈ Qi Qj h−1 . By Proposition 4.5, for sufficiently long and Qi –wide fillings the image of Qi is
relatively quasiconvex, and similarly for Qj . By Proposition 6.2, there
exists a filling
G → G = G ({γP | P ∈ P})
so that the images of Qi and Qj are relatively quasiconvex and there
is no element of K in Qi Qj h−1 . For such a filling, the image of h is
outside the image of Qi Qj .
Possibly replacing the γP by powers, the Orbifold Hyperbolic Dehn
Surgery Theorem [9, Theorem 5.3], implies that the group G is the
fundamental group of a compact hyperbolic orbifold, and so is Kleinian
and word-hyperbolic. Since it is Kleinian, it is LERF by [2, Corollary
9.4], and since it is also word-hyperbolic [18, Theorem 1.1] implies
that all double cosets of quasiconvex subgroups of G are separable.
Therefore, the image of h can be separated from the image of Qi Qj in
a finite quotient of G, which is clearly also a finite quotient of G.
This proves that Qi Qj is separable in G, which proves that the G–
action on X is virtually special, as required.
7. Relative height and relative multiplicity
For quasiconvex subgroups of hyperbolic groups, the height is an important invariant. For full relatively quasiconvex subgroups, it remains
a useful invariant, but because we cannot control the normalizer in P
of an intersection H ∩ P when H is (non-full) relatively quasiconvex
QUASICONVEXITY AND DEHN FILLING
22
and P is a maximal parabolic subgroup, height is not always a useful notion as it is too often infinite. Instead, we should consider the
relative height, defined as follows.
Definition 7.1. (cf. [14, §1.4]) Suppose that (G, P) is relatively hyperbolic and H ≤ G. The relative height of H in (G, P) is the maximum
number n ≥ 0 so that there are distinct cosets {g1 H, . . . , gn H} so that
n
T
gi Hgi−1 is an infinite non-parabolic subgroup.
i=1
In [14], they refer to relative height merely as ‘height’, but we prefer
to keep this term for its traditional meaning.
Remark 7.2. It follows from the classification of groups acting isometrically on δ–hyperbolic spaces that a subgroup of a relatively hyperbolic group is infinite and non-parabolic if and only if it contains a
loxodromic element. We use this equivalent characterization without
further mention.
In this section, we prove results for relative height analogous to those
proved for height in [2, Appendix A]. Specifically, we define a notion
of relative multiplicity (see Definition 7.7) and prove in Theorem 7.8
that relative multiplicity is equal to relative height. This gives a new
proof of a theorem of Hruska and Wise [14] that the relative height of
a relatively quasiconvex subgroup is finite. In Theorem 7.15 we prove
that for sufficiently long and H–wide fillings the relative height of a
relatively quasiconvex subgroup does not increase under Dehn filling.
The definition of a weakly geometrically finite (or WGF ) action is
given in [2, A.27]. We note here that a weakly geometrically finite
action differs from the usual notion of a geometrically finite action (as in
[13, Definition 3.2 (RH-2)]) in allowing horoballs with finite stabilizer.
Definition 7.3. Suppose that (G, P) is relatively hyperbolic and that
(H, D) is relatively quasiconvex. Let X be a cusped space for (G, P)
(considered as containing G as a subset), let e
∗ be the basepoint of X,
and let R ≥ 0. An R–hull for H acting on X is a connected H–invariant
full sub-graph Ze ⊂ X so that
e
(1) e
∗ ∈ Z;
(2) If γ is a geodesic in X with endpoints in Λ(H) then NR (γ) ∩
e and
NR (G) ⊂ Z;
(3) If A is any horoball containing vertices of depth greater than
0 in the image ι̌(XH ), then Z̃ ∩ A(0) contains every vertex of a
maximal vertical ray in A containing a.
QUASICONVEXITY AND DEHN FILLING
23
(4) The action of (H, D) on Ze (with its induced path metric) is
WGF.
Remark 7.4. This definition is not the same as [2, Definition A.32]
unless H is full relatively quasiconvex (as was assumed in [2]). It is
important that we do not include an R–neighborhood of γ, but only
that part of the R–neighborhood near the Cayley graph. The third
condition in [2, Definition A.32] has similarly been modified. Both of
these changes are made so that Lemma 7.12 below is true.
Let XH be the cusped space for (H, D) and ι̌ : XH → X be the
extension of the natural inclusion of H into G on the level of cusped
spaces.
Definition 7.5. Suppose that (H, D) ≤ (G, P) is relatively quasiconvex, that X is the cusped space for (G, P), XH the cusped space for
(H, D) and ι̌ : XH → X the extension of the natural inclusion of H into
G on the level of cusped spaces. For a positive integer D, the restricted
D–neighborhood of ι̌(XH ), denoted NDR (ι̌(XH )), is the full subgraph of
X on the vertices of either of the following two types:
(1) Vertices of ND (ι̌(XH )) ∩ ND (G); and
(2) For any horoball A so ι̌(XH ) contains vertices of arbitrary depth
in A, include all vertices a ∈ A which are connected by a vertical
geodesic to a vertex of the first type.
Lemma 7.6 (cf. Lemma A.41, [2]). Let R ≥ 0. There exists some
D so that the restricted D-neighborhood of ι̌(XH ) is an R–hull for the
action of H on X.
Proof. We first note that if any of the requirements of an R–hull are
satisfied by the restricted D–neighborhood of ι̌(XH ), then they are
satisfied for the restricted D0 –neighborhood of ι̌(XH ), for any D0 ≥ D.
It therefore suffices to consider each of the four requirements separately,
and take D to be the maximum needed for any of the four.
Condition (1) is satisfied for any D, since e
∗ = 1 ∈ H.
Condition (2) is satisfied as soon as D ≥ R + 2δ + λ, where λ is
the quasiconvexity constant for ι̌(XH ). Indeed, Λ(H) ⊂ Λ(ι̌(XH )), so
if γ is a biinfinite geodesic with endpoints in Λ(H), it must lie in a
λ + 2δ–neighborhood of ι̌(XH ). Suppose x ∈ NR (γ) ∩ NR (G); we want
to show that x ∈ NDR (ι̌(XH )). Let z ∈ γ, g ∈ G be vertices at distance
at most R from x. As we have noted, there is a q ∈ ι̌(XH ) satisfying
dX (z, q) ≤ λ + 2δ. Thus x lies in the (R + λ + 2δ)–neighborhood of
ι̌(XH ). If D ≥ R + 2δ + λ, then x ∈ ND (ι̌(XH )) ∩ ND (G).
Condition (3) is built in to the definition of restricted D–hull.
QUASICONVEXITY AND DEHN FILLING
24
Condition (4) (the weak geometric finiteness) follows once we observe
that for large enough D, the restricted D–neighborhood is equivariantly quasi-isometric to the D–neighborhood, and either one is quasiisometrically embedded in X. In particular, the limit set of the restricted D–neighborhood is equivariantly homeomorphic to ∂XH =
∂(H, D). (Though in general ι̌ is not a quasi-isometric embedding if
some D ∈ D is very distorted in PDcD .)
/
Let Ze be an R–hull for the action of H on G, and let Z = H Ze .
Similarly, let Y = G X . Then there is a natural map i : Z → Y which
induces the inclusion of H into G (in the sense described in [2]).
For n > 0, let
Sn = {(z1 , . . . , zn ) ∈ Z n | i(z1 ) = · · · = i(zn )} \ ∆
where ∆ = {(z1 , . . . , zn ) | there exist i 6= j so that zi = zj } is the fat
diagonal.
Points in Sn have a well-defined depth which is the depth of the
image in Y We consider components C of Sn which contain a point
with depth 0.
As in [2], choosing a maximal tree in Z, and a basepoint p at depth
0, a component C of Sn induces well-defined maps τC,i : π1 (C, p) → H,
for i = 1, . . . , n.
Definition 7.7. The relative multiplicity of i : Z → Y is the largest
n so that Sn contains a component C so that for all i ∈ 1, . . . , n the
group τC,i (π1 (C, p)) contains a loxodromic element.
Theorem 7.8 (cf. Theorem A.38, [2]). For sufficiently large R, depending only on δ and the quasi-convexity constant of H, if Ze is an
R–hull for the action of H on X, and i : Z → Y is as described above,
then the relative height of H in G is equal to the relative multiplicity
of i : Z → Y .
Definition 7.9. A geodesic σ in a combinatorial horoball is regular if
it has at most three horizontal edges, and these are at the maximum
depth for σ. A path in a cusped space X(G, P) is regular if every
intersection with a horoball is regular.
A path σ in a combinatorial horoball is super-regular if it has at
most 1 horizontal edge, this edge is at maximum depth for σ, and σ
has minimal length among paths with this property. A path in a cusped
space X(G, P) is super-regular if every intersection with a horoball is
super-regular.
QUASICONVEXITY AND DEHN FILLING
25
Lemma 7.10. Let g be a loxodromic element of the relatively hyperbolic
group pair (G, P). Then for any D > 0, and any sufficiently large
n > 0, there is a bi-infinite quasigeodesic axis σ for g n satisfying:
(1) σ is super-regular;
(2) σ is contained in a (4δ + 3)–neighborhood of any geodesic with
the same endpoints.
(3) dX (p, g n p) > D for any point p ∈ σ.
Proof. Let g ±∞ be the two limit points in ∂X of the cyclic group hgi.
Since X is proper, there is a bi-infinite geodesic γ joining g ±∞ . Note
that g n γ and γ are Hausdorff distance at most 2δ from one another,
for any n. Fix n large enough so that dX (x, g n x) > max{D, 100δ} for
every point x ∈ X.
Since the endpoints of γ are distinct, γ is not contained in a single
horoball. Choose some h ∈ γ in the Cayley graph of G. Choose a
regular geodesic α0 joining h to g n h.
σ0 be the concatenation of the g n –translates of α0 ; namely σ0 =
S Let in
i∈Z g α0 . Let σ be the path obtained my modifying σ0 to be superregular. (This means first ensuring that paths within any horoball
consist of two vertical segments and a single horizontal segment, and
then removing the horizontal subsegments of σ0 inside horoballs, and
replacing them by minimal length super-regular paths with the same
endpoints.)
The path σ0 is a broken geodesic with each breakpoint on g k γ for
some γ. The individual geodesics have length at least 100δ, and the
Gromov products at the vertices are at most 6δ. In particular, σ is
a quasi-geodesic. The local modifications producing σ from σ0 do not
change the fact of quasi-geodesicity (though they do change the constants of quasi-geodesicity).
The path σ0 lies a 2δ–neighborhood of γ. The path σ thus lies in
a (2δ + 3)–neighborhood of γ, and in a (4δ + 3)–neighborhood of any
other geodesic with the same endpoints.
Proof of Theorem 7.8. The proof from [2] works almost as written. Let
λ be the constant of quasiconvexity for ι̌(XH ), and let C = 2(λ + 2δ) +
maxi {dX (1, cD )}, where the elements cD are those elements chosen as
in Section 2. We suppose R > C + λ + 6δ + 4.
We first show the more difficult direction, that relative multiplicity
dominates relative height. Suppose that the relative height is at least n,
so there is some collection of cosets {H, g2 H, . . . , gn H} and loxodromic
elements h1 , . . . hn ∈ H so that h1 = g2 h2 g2−1 = · · · = gn hn gn−1 . Let
σ be the quasi-axis for h1 given by Lemma 7.10, and let γ be any biinfinite geodesic with the same endpoints at infinity as σ. Requirement
QUASICONVEXITY AND DEHN FILLING
26
e
(2) implies that NR (γ) ∩ NR (G) is contained in J = Ze ∩ g2 Ze ∩ · · · ∩ gn Z.
In particular, any points of σ ∩ NR−(4δ+3) (G) are contained in J. We
next need to show that the deeper points of σ are also contained in J.
Choose g ∈ G on the quasi-axis σ. Fix 1 ≤ i ≤ n. We claim that
there is an element ĥi in H so that dX (g, gi ĥi ) ≤ C (taking g1 = 1).
Indeed, gi hi gi−1 = h1 leaves σ invariant, so the endpoints of σ lie in
gi ΛH . Suppose that ρ is a bi-infinite geodesic with the same endpoints
as σ. Then by Lemma 7.10 σ lies in a (4δ + 3)–neighborhood of ρ. On
the other hand, quasi-convexity implies that any point on ρ lies within
distance λ + 2δ of ι̌(XH ). Possibly, the point g lies within λ + 2δ of
a point in ι̌(XH ) which lies within a horoball, but then this point has
depth at most λ+2δ, and so lies within distance λ+2δ+max{dX (1, cD )}
of a point in H. The claim follows.
ĥ1
h1 ĥ1
g2 ĥ2
h1 g2 ĥ2 = g2 h2 ĥ2
g
h1 g
σ
g3 ĥ3
Figure 3. Showing that the points of γ lie in J.
Now, let A be a horoball (R −4δ +3)–penetrated by σ, and note that
R − (4δ + 3) > C + λ + 2δ + 1. Any point on σ lies within 2δ of some
point of geodesics [g, ĥ1 ], [ĥ1 , hĥ1 ], [hĥ1 , h1 g]. However, the first and
third of these geodesics are between points in the Cayley graph and
have length at most C. Therefore, any points of σ at depth greater
than R − (4δ + 3) in A must be within 2δ of the geodesic between ĥ1
and hĥ1 . This implies that there is a geodesic with endpoints in H
which (λ + 1)–penetrates A, which by λ-quasiconvexity implies that A
contains points at depth greater than 0 in the image ι̌(XH ). Condition
(3) from Definition 7.3 (along with the requirement that an R–hull be
a full subgraph) ensures that the intersection of Ze with A consists of a
collection of vertical lines together with any horizontal edges connecting
them. In particular, the (super-regular) subsegment of σ meeting A
e An exactly analogous argument shows that this
is contained in Z.
subsegment is contained in gi Ze for each i, so all of σ is contained in J.
This implies that σ projects to a loop in Sn of the type desired; if C is
the component containing the image of σ, then τC,i (π1 (C, p)) contains
a conjugate of the loxodromic h1 for each i.
QUASICONVEXITY AND DEHN FILLING
27
The other direction, that relative height dominates relatively multiplicity, is almost exactly the same as in [2, Appendix A]. The only
difference is that we assume that the intersection is infinite and nonparabolic, so that it contains a loxodromic element. This loxodromic
element is then the one required by Definition 7.7.
Corollary 7.11. [14, Theorem 1.4] The relative height of a relatively
quasiconvex subgroup of a relatively hyperbolic group is finite.
Proof. If the relative multiplicity is n, then in particular, Sn contains
a loop with a vertex at depth 0. Since Sn avoids the fat diagonal, this
vertex represents an n–tuple of distinct depth 0 vertices of Z. There are
only finitely many such vertices, so the relative height is bounded.
7.1. Non-increasing of height under wide fillings. Suppose that
Ze is an R–hull for the action of H on X. The following is an analog of
[2, Lemma A.45].
Lemma 7.12 (cf. Lemma A.45, [2]). For all sufficiently long and H–
wide fillings φ : G → G(N1 , . . . , Nm ), if K = ker(φ), KH = K ∩ H,
and k ∈ K r KH , then k.Ze ∩ Ze = ∅.
Proof. By conditions (3) and (4) of an R–hull, there is some R0 so that
if g Ze ∩ Ze 6= ∅, and if N is the R0 –neighborhood of H in the Cayley
graph of G, then gN ∩ N 6= ∅. It follows that the set of g for which
g Ze ∩ Ze 6= ∅ is contained in a finite union of double cosets
A=
l
G
Hgi H, with g0 = 1.
i=0
Now let φ be long and H–wide enough to apply Proposition 4.7 to all
the elements g1 , . . . , gl . For such a filling we have φ(gi ) 6∈ φ(H) for each
i. Equivalently, there is no k ∈ K of the form gi h for h ∈ H and i > 0.
Suppose by way of contradiction that k ∈ (K \ KH ) ∩ A. Then we
can write k = h1 gi h2 for some i > 0 and some h1 , h2 ∈ H. Conjugating
we obtain a k 0 ∈ (K \ KH ) ∩ A which lies in gi H. But this contradicts
the last paragraph.
Remark 7.13. Lemma A.45 in [2] is a special case of Lemma 7.12.
The proof given in [2] contains the erroneous assertion that A is a
finite union of left cosets; otherwise the proof given there is similar to
our proof here of Lemma 7.12, but using a theorem about H–fillings
[2, A.43] in place of our Proposition 4.7.
Suppose π : (G, P) → (G, P) is a Dehn filling, and X(G, P) is the
combinatorial cusped space for (G, P). If K is the kernel of the quotient
QUASICONVEXITY AND DEHN FILLING
28
/
map G → G, then the quotient X = K X(G, P) is very nearly equal
to the cusped space for the pair (G, P), differing only in the addition
of some self-loops. In particular, their 0–skeleta are isometric, and we
can safely ignore the difference.
Putting Lemma 7.12 together with uniformity of hyperbolicity and
quasiconvexity after long Dehn fillings, we can prove the following:
Lemma 7.14. Fix (G, P) relatively hyperbolic, and a relatively quasiconvex subgroup H. For all R, there is an R0 satisfying the following:
For all sufficiently long and H–wide fillings φ : G → G(N1 , . . . , Nm ),
e ⊂ X is an R–hull
if K = ker(φ), if Z̃ is an R0 –hull for H, then Z
K
for the image of H in G(N1 , . . . , Nm ).
Proof. Let δ be such that X = K X is δ–hyperbolic whenever K is the
kernel of a sufficiently long filling (see Proposition 2.2). As discussed
above this quotient is essentially equal to the combinatorial cusped
space for the pair (G, P), where G = G(N1 , . . . , Nm ), and P consists
of the images of the elements of P. Let λ0 be the constant from Proposition 4.5, so that φ(H) is λ0 –relatively quasiconvex for a sufficiently
long and H–wide filling. Let R0 = R0 (λ0 , δ) be such that any bi-infinite
geodesic with endpoints in the limit set of a λ0 –quasiconvex subset of
a δ–hyperbolic space is contained in the R0 –neighborhood of that quasiconvex subset. Let C = max{dX (1, cD )}, where cD ranges over the
elements chosen in Section 2. Finally we fix some R0 > 3R + 2R0 + C.
We assume that φ is sufficiently long and H–wide so that the results
from the last paragraph apply.
e ⊂ X
We suppose that Z̃ is an R0 –hull, and show that the image Z
K
is an R–hull for the image H of H in G(N1 , . . . , Nm ).
Conditions (1) and (3) of Definition 7.3 follow easily from the fact
that Z̃ is an R0 –hull. Condition (4) is a fairly straightforward consequence of the fact that H is relatively quasiconvex.
We now establish Condition (2). Suppose that γ is a bi-infinite
geodesic with endpoints in Λ(H). Let p ∈ NR (γ)∩NR (G). Since ι̌(XH )
is λ0 –quasiconvex, we have dX (p, x) ≤ R0 +R for some x ∈ ι̌(XH ). Since
the depth of p is at most R, the depth of x is at most R0 + 2R. Thus
there is some h ∈ H with dX (x, h) ≤ R0 + 2R + C. Choose h ∈ H
projecting to h, and note that there is a bi-infinite geodesic passing
through h with endpoints in Λ(H). 5 In particular, an R0 –ball about
5We
remark that because we assumed that γ exists, Λ(H) contains more than
one point. It follows that Λ(H) contains more than one point, which implies the
existence of such a bi-infinite geodesic.
QUASICONVEXITY AND DEHN FILLING
29
h is contained
in the R0 –hull Z̃. Since R0 > dX (h, p), the image of Z̃
in K X must contain p.
We now prove that the relative height of H does not increase under
sufficiently long and H–wide fillings.
Theorem 7.15. (cf. [2, A.46]) For sufficiently long and H–wide fillings, the relative height of (H, D) in (G, P) is at most the relative height
of (H, D) in (G, P).
Proof. As usual, let δ be a constant so that the cusped space of (G, P)
and also those of sufficiently long fillings, are δ–hyperbolic, and let
λ be a quasi-convexity constant for H, which we also assume (using
Proposition 4.5) is a quasi-convexity constant for the image of H under
sufficiently long and H–wide fillings.
Let R be sufficiently large to apply Theorem 7.8 with these values of
δ and λ. Let R0 the the constant (depending on R) from the conclusion
of Lemma 7.14.
Consider the following commutative diagram, which is equivariant
with respect to the group actions and the natural maps between the
groups (inclusion and quotient maps, as appropriate):
G
H
/
Ze
H/KH
$
X
/
e
Z
Xq
G/K
where X is the cusped space for (G, P), X = K X , Ze is an R–hull
for H, K/H = H ∩ K is the kernel of the induced filling on H, and
e=
e
Z
KH Z .
e embeds in X, and
It follows immediately from Lemma 7.12 that Z
e is an R–hull for H/K in X.
it follows from Lemma 7.14 that Z
H
Taking quotients by the relevant groups we get the diagram,
(1)
Z
Z
i
/
Y
ı
/
Y
where the horizontal maps are immersions inducing the inclusions H →
G and H → G on the level of fundamental group. The vertical maps
from Y to Y and from Z to Z are homeomorphisms. Theorem 7.8
QUASICONVEXITY AND DEHN FILLING
30
implies that the relative multiplicities of Z in Y and of Z in Y measure
the relative heights of H in G and of H/KH in G/K, respectively.
For n > 0, define
Sn = {(z1 , . . . , zn ) ∈ Z n | i(z1 ) = · · · = i(zn )} \ ∆
and
n
S n = (z1 , . . . , zn ) ∈ Z | ı(z1 ) = · · · = ı(zn ) \ ∆,
where ∆ = {(z1 , . . . , zn ) | there exist i 6= j so that zi = zj } is the fat
n
diagonal in Z n and ∆ is the fat diagonal in Z .
For each i ∈ {1, . . . , n} and each component C of Sn the projections
of Z n to its factors induce maps
τC,i : π1 (C) → H,
and
τ C,i : π1 (C) → H.
/
e
e
Since the quotient Z = H Z can also be thought of as (H/KH ) Z ,
the homomorphisms τ C,i all factor as τ C,i = φ|H ◦ τC,i , where φ is the
filling map.
In particular, if γ is a loop in S n so that τ C,i ([γ]) is infinite for each
i ∈ {1, . . . , n} then it must be that τC,i ([γ]) is already infinite for each
i.
In case parabolic subgroups of a relatively hyperbolic group are finite,
the relative height is the same as the height. Therefore, we immediately
obtain the following.
Corollary 7.16. For sufficiently long and H–wide peripherally finite
fillings, the height of H in G is at most the relative height of (H, D) in
(G, P).6
7.2. A result required by Wilton and Zalesskii.
Definition 7.17. [23, Definition 4.1] Suppose that (G, P) is relatively
hyperbolic and H ≤ G. We say that H is relatively malnormal if for
any g 6∈ H the intersection H g ∩ H is conjugate into some element of
P.
If (G, P) is relatively hyperbolic and G is torsion-free, then a relatively malnormal subgroup is either parabolic or a subgroup of relative
height 1.
6Recall
that a filling G → G(N1 , . . . , Nm ) is peripherally finite if for each 1 ≤
j ≤ m the subgroup Nj has finite-index in Pj .
QUASICONVEXITY AND DEHN FILLING
31
For his joint work with Zalesskii [23], Henry Wilton asked us to prove
the following result, which appeared as [23, Theorem 4.4]:
Theorem 7.18. Let G be a toral relatively hyperbolic group with parabolic subgroups {P1 , . . . , Pn } and let H be a subgroup which is relatively
quasi-convex and relatively malnormal. There exist subgroups of finite
index Ki0 ⊂ Pi (for all i) such that, for all subgroups of finite index
Li ⊂ Ki0 , if
η : G → Q = G/hhL1 , . . . , Ln ii
is the quotient map, the quotient Q is word-hyperbolic and the image
η(H) in Q is quasi-convex and almost malnormal.
Proof. We restrict to peripherally finite fillings. For sufficiently long
peripherally finite fillings η : G → Q, the quotient Q is hyperbolic
(since it is hyperbolic relative to finite groups) by [10, Theorem 7.3.(2)].
Since the peripheral subgroups of Q are finite, there is no difference
between quasiconvex and relatively quasiconvex subgroups.
By Proposition 4.5, for sufficiently long and H–wide fillings η : G →
Q the image η(H) is quasi-convex, and by Corollary 7.16, for sufficiently
long and H–wide fillings η(H) is almost malnormal in Q.
It remains only to note that sufficiently long and H–wide peripherally
finite fillings exist by Lemma 5.2, since the peripheral subgroups of G
are free abelian, and hence ERF.
References
[1] I. Agol. Criteria for virtual fibering. J. Topol., 1(2):269–284, 2008.
[2] I. Agol. The Virtual Haken Conjecture. Doc. Math., 18:1045–1087, 2013. With
an appendix by Agol, Daniel Groves, and Jason Manning.
[3] I. Agol, D. Groves, and J. F. Manning. Residual finiteness, QCERF and fillings
of hyperbolic groups. Geom. Topol., 13(2):1043–1073, 2009.
[4] I. Agol, D. Groves, and J. F. Manning. An alternate proof of Wise’s malnormal
special quotient theorem. Forum Math. Pi, 4:e1, 54, 2016.
[5] M. R. Bridson and A. Haefliger. Metric Spaces of Non–Positive Curvature,
volume 319 of Grundlehren der mathematischen Wissenschaften. Springer–
Verlag, Berlin, 1999.
[6] D. Cooper and D. Futer. Ubiquitous quasi-Fuchsian surfaces in cusped hyperbolic 3-manifolds. Preprint, arXiv:1705.03831, 2017.
[7] R. Coulon. On the geometry of Burnside quotients of torsion free hyperbolic
groups. International Journal of Algebra and Computation, 24(3):251–345,
2014.
[8] F. Dahmani and V. Guirardel. Recognizing a relatively hyperbolic group by
its Dehn fillings. Preprint arXiv:1506.03233, 2015.
[9] W. D. Dunbar and G. R. Meyerhoff. Volumes of hyperbolic 3-orbifolds. Indiana
Univ. Math. J., 43(2):611–637, 1994.
QUASICONVEXITY AND DEHN FILLING
32
[10] D. Groves and J. F. Manning. Dehn filling in relatively hyperbolic groups.
Israel Journal of Mathematics, 168:317–429, 2008.
[11] D. Groves and J. F. Manning. Dehn fillings and elementary splittings. Transactions of the American Mathematical Society, to appear, 2015.
[12] F. Haglund and D. T. Wise. Coxeter groups are virtually special. Adv. Math.,
224(5):1890–1903, 2010.
[13] G. C. Hruska. Relative hyperbolicity and relative quasiconvexity for countable
groups. Algebr. Geom. Topol., 10(3):1807–1856, 2010.
[14] G. C. Hruska and D. T. Wise. Packing subgroups in relatively hyperbolic
groups. Geom. Topol., 13(4):1945–1988, 2009.
[15] G. C. Hruska and D. T. Wise. Finiteness properties of cubulated groups. Compos. Math., 150(3):453–506, 2014.
[16] J. F. Manning and E. Martı́nez-Pedroza. Separation of relatively quasiconvex
subgroups. Pacific J. Math., 244(2):309–334, 2010.
[17] E. Martı́nez-Pedroza. Combination of quasiconvex subgroups of relatively hyperbolic groups. Groups Geom. Dyn., 3(2):317–342, 2009.
[18] A. Minasyan. Separable subsets of GFERF negatively curved groups. J. Algebra, 304(2):1090–1100, 2006.
[19] I. Mineyev. Straightening and bounded cohomology of hyperbolic groups.
Geom. Funct. Anal., 11(4):807–839, 2001.
[20] D. V. Osin. Peripheral fillings of relatively hyperbolic groups. Invent. Math.,
167(2):295–326, 2007.
[21] P. Przytycki and D. T. Wise. Mixed 3-manifolds are virtually special. Preprint,
arXiv:1205.6742, 2012.
[22] M. Sageev. Ends of group pairs and non-positively curved cube complexes.
Proc. London Math. Soc. (3), 71(3):585–617, 1995.
[23] H. Wilton and P. Zalesskii. Distinguishing geometries using finite quotients.
Geom. Topol., 21(1):345–384, 2017.
[24] D. T. Wise. The structure of groups with a quasiconvex hierarchy. Unpublished
manuscript.
Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago, 322 Science and Engineering Offices
(M/C 249), 851 S. Morgan St., Chicago, IL 60607-7045
E-mail address: [email protected]
Department of Mathematics, 310 Malott Hall, Cornell University,
Ithaca, NY 14853
E-mail address: [email protected]
| 4math.GR
|
arXiv:cs/0606094v1 [cs.DB] 22 Jun 2006
On Typechecking Top-Down XML
Tranformations: Fixed Input or Output Schemas∗
Wim Martens†
Frank Neven‡
Marc Gyssens§
Hasselt University and
Transnational University of Limburg,
Agoralaan, Gebouw D
B-3590 Diepenbeek, Belgium
Abstract
Typechecking consists of statically verifying whether the output of an
XML transformation always conforms to an output type for documents
satisfying a given input type. In this general setting, both the input and
output schema as well as the transformation are part of the input for the
problem. However, scenarios where the input or output schema can be
considered to be fixed, are quite common in practice. In the present work,
we investigate the computational complexity of the typechecking problem
in the latter setting.
1
Introduction
In a typical XML data exchange scenario on the web, a user community creates
a common schema and agrees on producing only XML data conforming to that
schema. This raises the issue of typechecking: verifying at compile time that
every XML document which is the result of a specified query or document transformation applied to a valid input document satisfies the output schema [32, 33].
The typechecking problem is determined by three parameters: the classes
of allowed input and output schemas, and the class of XML-transformations.
As typechecking quickly becomes intractable [2, 23, 25], we focus on simple but
practical XML transformations where only little restructuring is needed, such
as, for instance, in filtering of documents. In this connection, we think, for
example, of transformations that can be expressed by structural recursion [8]
or by a top-down fragment of XSLT [5]. As is customary, we abstract such
transformations by unranked tree transducers [19, 23]. As schemas, we adopt the
∗ An extended abstract of a part of this paper appeared as Section 3 in reference [22] in the
ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems, 2004.
† Corresponding author. Email: [email protected]
‡ Email: [email protected]
§ Email: [email protected]
1
usual Document Type Definitions (DTDs) and their robust extensions: regular
tree languages [25, 17] or, equivalently, specialized DTDs [28, 3]. The latter
serve as a formal model for XML Schema [30].
Our work studies sound and complete typechecking algorithms, an approach
that should be contrasted with the work on general-purpose XML programming
languages like XDuce [14] and CDuce [4], for instance, where the main objective
is fast and sound typechecking. The latter kind of typechecking is always incomplete due to the Turing-completeness of the considered XML-transformations.
That is, it can happen that type safe transformations are rejected by the typechecker. As we only consider very simple transformations which are by no means
Turing-complete, it makes sense to ask for complete algorithms.
The typechecking scenario outlined above is very general: both the schemas
and the transducer are determined to be part of the input. However, for some
exchange scenarios, it makes sense to consider the input and/or output schema
to be fixed when transformations are always from within and/or to a specific
community. Therefore, we revisit the various instances of the typechecking
problem considered in [23] and determine the complexity in the presence of fixed
input and/or output schemas. The main goal of this paper is to investigate to
which extent the complexity of the typechecking problem is lowered in scenarios
where the input and/or output schema is fixed. An overview of our results is
presented in Table 2.
The remainder of the paper is organized as follows. In Section 2, we discuss
related work. In Section 3, we provide the necessary definitions. In Section 4,
we discuss typechecking in the restricted settings of fixed output and/or input
schemas. The results are summarized in Table 2. We obtain several new cases
for which typechecking is in polynomial time: (i) when the input schema is fixed
and the schemas are DTDs with SL-expressions; (ii) when the output schema
is fixed and the schemas are DTDs with NFAs; and (iii) when both the input
and output schemas are fixed and the schemas are DTDs using DFAs, NFAs, or
SL-expressions. We conclude in Section 5.
2
Related Work
The research on typechecking XML transformations was initiated by Milo, Suciu, and Vianu [25]. They obtained the decidability for typechecking of transformations realized by k-pebble transducers via a reduction to satisfiability of
monadic second-order logic. Unfortunately, in this general setting, the latter
non-elementary algorithm cannot be improved [25]. Interestingly, typechecking
of k-pebble transducers has recently been related to typechecking of compositions of macro tree transducers [12]. Alon et al. [1, 2] investigated typechecking
in the presence of data values and show that the problem quickly turns undecidable. As our interest lies in formalisms with a more manageable complexity for
the typechecking problem, we choose to work with XML transformations that
are much less expressive than k-pebble transducers and that do not change or
use data values in the process of transformation.
2
A problem related to typechecking is type inference [24, 28]. This problem
consists in constructing a tight output schema, given an input schema and a
transformation. Of course, solving the type inference problem implies a solution for the typechecking problem, namely, checking containment of the inferred
schema into the given one. However, characterizing output languages of transformations is quite hard [28]. For this reason, we adopt different techniques for
obtaining complexity upper bounds for the typechecking problem.
The transducers considered in the present paper are restricted versions of
the DTL-programs, studied by Maneth and Neven [19]. They already obtained
a non-elementary upper bound on the complexity of typechecking (due to the
use of monadic second-order logic in the definition of the transducers). Recently, Maneth et al. considered the typechecking problem for an extension of
DTL-programs and obtained that typechecking was still decidable [20]. Their
typechecking algorithm, like the one of [25], is based on inverse type inference.
That is, they compute the pre-image of all ill-formed output documents and
test whether the intersection of the pre-image and the input schema is empty.
Tozawa considered typechecking with respect to tree automata for a fragment
of top-down XSLT [34]. He uses a more general framework, but he was not
able to derive a bound better than double-exponential on the complexity of his
algorithm.
Martens and Neven investigated polynomial time fragments of the typechecking problem by putting syntactical restrictions on the tree transducers,
and making them as general as possible [21]. Here, tractability of the typechecking problem is obtained by bounding the deletion path width of the tree
transducers. The deletion path width is a notion that measures the number of
times that a tree transducer copies part of its input. In particular, it also gives
rise to tractable fragments of the typechecking problem where the transducer is
allowed to delete in a limited manner.
3
Preliminaries
In this section we provide the necessary background on trees, automata, and
tree transducers. In the following, Σ always denotes a finite alphabet.
By N we denote the set of natural numbers. A string w = a1 · · · an is
a finite sequence of Σ-symbols. The set of positions, or the domain, of w is
Dom(w) = {1, . . . , n}. The length of w, denoted by |w|, is the number of
symbols occurring in it. The label ai of position i in w is denoted by labw (i).
The size of a set S, is denoted by |S|.
As usual, a nondeterministic finite automaton (NFA) over Σ is a tuple N =
(Q, Σ, δ, I, F ) where Q is a finite set of states, δ : Q × Σ → 2Q is the transition
function, I ⊆ Q is the set of initial states, and F ⊆ Q is the set of final states. A
run ρ of N on a string w ∈ Σ∗ is a mapping from Dom(w) to Q such that ρ(1) ∈
δ(q, labw (1)) for q ∈ I, and for i = 1, . . . , |w| − 1, ρ(i + 1) ∈ δ(ρ(i), labw (i + 1)).
A run is accepting if ρ(|w|) ∈ F . A string is accepted if there is an accepting
run. The language accepted by N is denoted by L(N ). The size of N is defined
3
P
as |Q| + |Σ| + q∈Q,a∈Σ |δ(q, a)|.
A deterministic finite automaton (DFA) is an NFA where (i) I is a singleton
and (ii) |δ(q, a)| ≤ 1 for all q ∈ Q and a ∈ Σ.
3.1
Trees and Hedges
It is common to view XML documents as finite trees with labels from a finite
alphabet Σ. Figures 1(a) and 1(b) give an example of an XML document together with its tree representation. Of course, elements in XML documents can
also contain references to nodes. But, as XML schema languages usually do
not constrain these nor the data values at leaves, it is safe to view schemas as
simply defining tree languages over a finite alphabet. In the rest of this section,
we introduce the necessary background concerning XML schema languages.
<store>
<dvd>
<title> "Amelie" </title>
<price> 17 </price>
</dvd>
<dvd>
<title> "Good bye, Lenin!" </title>
<price> 20 </price>
</dvd>
<dvd>
<title> "Pulp Fiction" </title>
<price> 11 </price>
<discount> 6 </discount>
</dvd>
</store>
(a) An example XML document.
store
dvd
title
price
dvd
title
dvd
price
title
price discount
“Amelie” 17 “Good bye, Lenin!” 20 “Pulp Fiction” 11
6
(b) Its tree representation with data values.
Figure 1: An example of an XML document and its tree representation.
The set of unranked Σ-trees, denoted by TΣ , is the smallest set of strings over
Σ and the parenthesis symbols “(” and “)” such that, for a ∈ Σ and w ∈ TΣ ∗ ,
a(w) is in TΣ . So, a tree is either ε (empty) or is of the form a(t1 · · · tn ) where
4
store (ε)
dvd (1)
title
(1 1)
price
(1 2)
dvd (2)
title
(2 1)
price
(2 2)
dvd (3)
title
(3 1)
price
(3 2)
discount
(3 3)
(a) The tree of Figure 1(b) without data values. The nodes are
annotated next to the labels, between brackets.
store (1)
dvd (1 1)
title
(1 1 1)
price
(1 1 2)
dvd (1 2)
title
(1 2 1)
price
(1 2 2)
dvd (1 3)
title
(1 3 1)
price
(1 3 2)
discount
(1 3 3)
(b) Tree of Figure 2(a) viewed as a hedge. The nodes are annotated
next to the labels, between brackets.
Figure 2: The document of Figure 1 without data values, viewed as a tree and
as a hedge.
each ti is a tree. In the tree a(t1 · · · tn ), the subtrees t1 , . . . , tn are attached to
a root labeled a. We write a rather than a(). Note that there is no a priori
bound on the number of children of a node in a Σ-tree; such trees are therefore
unranked. For every t ∈ TΣ , the set of tree-nodes of t, denoted by DomT (t), is
the set defined as follows:
(i) if t = ε, then DomT (t) = ∅; and,
(ii) if t = a(t1 · · · tn ), where each ti ∈ TΣ , then DomT (t) = {ε} ∪
DomT (ti )}.
Sn
i=1 {iu
|u∈
Figure 2(a) contains a tree in which we annotated the nodes between brackets.
Observe that the n child nodes of a node u are always u1, . . . , un, from left to
right. The label of a node u in the tree t = a(t1 · · · tn ), denoted by labtT (u), is
defined as follows:
(i) if u = ε, then labtT (u) = a; and,
(ii) if u = iu′ , then labtT (u) = labtTi (u′ ).
We define the depth of a tree t, denoted by depth(t), as follows: if t = ε, then
depth(t) = 0; and if t = a(t1 · · · tn ), then depth(t) = max{depth(ti ) | 1 ≤ i ≤
5
n} + 1. In the sequel, whenever we say tree, we always mean Σ-tree. A tree
language is a set of trees.
A hedge is a finite sequence of trees. Hence, the set of hedges, denoted by
HΣ , equals TΣ∗ . For every hedge h ∈ HΣ , the set of hedge-nodes of h, denoted
by DomH (h), is the subset of N∗ defined as follows:
(i) if h = ε, then DomH (h) = ∅; and,
(ii) if h = t1 · · · tn and each ti ∈ TΣ , then DomH (h) =
DomT (ti )}.
Sn
i=1 {iu
| u ∈
The label of a node u = iu′ in the hedge h = t1 · · · tn , denoted by labhH (u), is
defined as labhH (u) = labtTi (u′ ). Note that the set of hedge-nodes of a hedge
consisting of one tree is different from the set of tree-nodes of this tree. For
example: if the tree in Figure 2(a) were to represent a single-tree hedge, it
would have the set of hedge-nodes {1, 11, 12, 13, 111, 112, 121, 122, 131, 132, 133},
as shown in Figure 2(b). The depth of the hedge h = t1 · · · tn , denoted by
depth(h), is defined as max{depth(ti ) | i = 1, . . . , n}. For a hedge h = t1 · · · tn ,
we denote by top(h) the string obtained by concatenating the root symbols of
all ti s, that is, labtH1 (1) · · · labtHn (n).
In the sequel, we adopt the following conventions: we use t, t1 , t2 , . . . to
denote trees and h, h1 , h2 , . . . to denote hedges. Hence, when we write h =
t1 · · · tn we tacitly assume that all ti ’s are trees. We denote DomT and DomH
simply by Dom, and we denote labT and labH by lab when it is understood
from the context whether we are working with trees or hedges.
3.2
DTDs and Tree Automata
We use extended context-free grammars and tree automata to abstract from
DTDs and the various proposals for XML schemas. We parameterize the definition of DTDs by a class of representations M of regular string languages such
as, for instance, the class of DFAs (Deterministic Finite Automata) or NFAs
(Non-deterministic Finite Automata). For M ∈ M, we denote by L(M ) the set
of strings accepted by M . We then abstract DTDs as follows.
Definition 3.1. Let M be a class of representations of regular string languages
over Σ. A DTD is a tuple (d, sd ) where d is a function that maps Σ-symbols to
elements of M and sd ∈ Σ is the start symbol.
For convenience of notation, we denote (d, sd ) by d and leave the start
symbol sd implicit whenever this cannot give rise to confusion. A tree t satisfies d if (i) labt (ε) = sd and, (ii) for every u ∈ Dom(t) with n children,
labt (u1) · · · labt (un) ∈ L(d(labt (u))). By L(d) we denote the set of trees satisfying d.
Given a DTD d, we say that a Σ-symbol a occurs in d(b) when there exist
Σ-strings w1 and w2 such that w1 aw2 ∈ L(d(b)). We say that a occurs in d if a
occurs in d(b) for some b ∈ Σ.
6
We denote by DTD(M) the class of DTDs where the regular string languages
are represented by elements of M. The size of a DTD is the sum of the sizes of
the elements of M used to represent the function d.
Example 3.2. The following DTD (d, store) is satisfied by the tree in Figure 2(a):
d(store) = dvd dvd∗
d(dvd) = title price (discount + ε)
This DTD defines the set of trees where the root is labeled with “store”; the
children of “store” are all labeled with “dvd”; and every “dvd”-labeled node
has a “title”, “price”, and an optional “discount” child.
In some cases, our algorithms are easier to explain on well-behaved DTDs
as considered next. A DTD d is reduced if, for every symbol a that occurs in
d, there exists a tree t ∈ L(d) and a node u ∈ Dom(t) such that labt (u) = a.
Hence, for example, the DTD (d, a) where d(a) = a is not reduced. Reducing a
DTD(DFA) is in ptime, while reducing a DTD(SL) is in conp (see the Appendix,
Corrollary .9). Here, SL is a logic as defined next.
To define unordered languages, we make use of the specification language
SL inspired by [27] and also used in [1, 2]. The syntax of this language is as
follows:
Definition 3.3. For every a ∈ Σ and natural number i, a=i and a≥i are atomic
SL-formulas; “true” is also an atomic SL-formula. Every atomic SL-formula is
an SL-formula and the negation, conjunction, and disjunction of SL-formulas
are also SL-formulas.
A string w over Σ satisfies an atomic formula a=i if it has exactly i occurrences of a; w satisfies a≥i if it has at least i occurrences of a. Furthermore,
“true” is satisfied by every string. Satisfaction of Boolean combinations of
atomic formulas is defined in the obvious way.1 By w |= φ, we denote that w
satisfies the SL-formula φ.
As an example, consider the SL-formula ¬(discount≥1 ∧ ¬price≥1 ). This
expresses the constraint that a discount can only occur when a price occurs.
The size of an SL-formula is the number of symbols that occur in it, that is,
Σ-symbols, logical symbols, and numbers (every i in a=i or a≥i is written in
binary notation).
We recall the definition of non-deterministic tree automata from [6]. We
refer the unfamiliar reader to [26] for a gentle introduction.
Definition 3.4. A nondeterministic tree automaton (NTA) is a 4-tuple B =
(Q, Σ, δ, F ), where Q is a finite set of states, F ⊆ Q is the set of final states,
∗
and δ : Q × Σ → 2Q is a function such that δ(q, a) is a regular string language
over Q for every a ∈ Σ and q ∈ Q.
1 The
empty string is obtained as
V
a∈Σ
a=0 and the empty set as ¬ true.
7
For simplicity, we often denote the regular languages in B’s transition function by regular expressions.
A run of B on a tree t is a labeling λ : Dom(t) → Q such that, for every
v ∈ Dom(t) with n children, λ(v1) · · · λ(vn) ∈ δ(λ(v), labt (v)). Note that, when
v has no children, the criterion reduces to ε ∈ δ(λ(v), labt (v)). A run is accepting
if the root is labeled with an accepting state, that is, λ(ε) ∈ F . A tree is accepted
if there is an accepting run. The set of all accepted trees is denoted by L(B)
and is called a regular tree language.
A tree automaton is bottom-up deterministic if, for all q, q ′ ∈ Q with q 6= q ′
and a ∈ Σ, δ(q, a) ∩ δ(q ′ , a) = ∅. We denote the set of bottom-up deterministic
NTAs by DTA.
Example 3.5. We give a bottom-up deterministic tree automaton B = (Q, Σ,
δ, F ) which accepts the parse trees of well-formed Boolean expressions that are
true. Here, the alphabet Σ is {∧, ∨, ¬, true, false}. The states set Q contains
the states qtrue and qfalse , and the accepting state set F is the singleton {qtrue }.
The transition function of B is defined as follows:
• δ(qtrue , true) = ε. We assign the state qtrue to leafs with label “true”.
• δ(qfalse , false) = ε. We assigns the state qfalse to leafs with label “false”.
∗
• δ(qtrue , ∧) = qtrue qtrue
.
• δ(qfalse , ∧) = (qtrue + qfalse )∗ qfalse (qtrue + qfalse )∗ .
• δ(qtrue , ∨) = (qtrue + qfalse )∗ qtrue (qtrue + qfalse )∗ .
∗
.
• δ(qfalse , ∨) = qfalse qfalse
• δ(qtrue , ¬) = qfalse .
• δ(qfalse , ¬) = qtrue .
Consider the tree t depicted in Figure 3(a). The unique accepting run r of
B on t can be graphically represented as shown in Figure 3(b). Formally, the
run of B on t is the function λ : Dom(t) → Q : u 7→ labr (u). Note that B is a
DTA.
As for DTDs, we parameterize NTAs by the formalism used to represent the
regular languages in the transition functions δ(q, a). So, for a class of representations of regular languages M, we denote by NTA(M) the class of NTAs
where all transition functions are represented
by elements of M. The size of
P
an automaton B then is |Q| + |Σ| + q∈Q,a∈Σ |δ(q, a)|. Here, by |δ(q, a)|, we
denote the size of the automaton accepting δ(q, a). Unless explicitly specified
otherwise, δ(q, a) is always represented by an NFA.
In our proofs, we will use reductions from the following decision problems
for string automata:
Emptiness: Given an automaton A, is L(A) = ∅?
8
∧
∧
∨
false
¬
false
true
∨
false
false
true
false
(a) The tree t.
qtrue
qtrue
qfalse
qtrue
qtrue
qfalse
qtrue
qtrue
qfalse
qfalse
qtrue
qfalse
(b) Graphical representation of the run r of B on t.
Figure 3: Illustrations for Example 3.5.
Universality: Given an automaton A, is L(A) = Σ∗ ?
Intersection emptiness: Given the automata A1 , . . . , An , is L(A1 ) ∩ · · · ∩
L(An ) = ∅?
The corresponding decision problems for tree automata are defined analogously.
In the Appendix, we show that the following statements hold over the alphabet {0, 1} (Corollary .2):
1. Intersection emptiness of an arbitrary number of DFAs is pspace-hard.
2. Universality of NFAs is pspace-hard.
Over the alphabet {0, 1, 0′, 1′ }, the following statement holds:
(3) Intersection emptiness of an arbitrary number of TDBTAs is exptimehard.
3.3
Transducers
We adhere to transducers as a formal model for simple transformations corresponding to structural recursion [8] and a fragment of top-down XSLT. As
in [25], the abstraction focuses on structure rather than on content. We next
define the tree transducers used in this paper. To simplify notation, we restrict
9
ourselves to one alphabet. That is, we consider transducers mapping Σ-trees to
Σ-trees.2
For a set Q, denote by HΣ (Q) (respectively TΣ (Q)) the set of Σ-hedges
(respectively trees) where leaf nodes are labeled with elements from Σ ∪ Q
instead of only Σ.
Definition 3.6. A tree transducer is a 4-tuple T = (Q, Σ, q 0 , R), where Q is
a finite set of states, Σ is the input and output alphabet, q 0 ∈ Q is the initial
state, and R is a finite set of rules of the form (q, a) → h, where a ∈ Σ, q ∈ Q,
and h ∈ HΣ (Q). When q = q 0 , h is restricted to be either empty, or consist of
only one tree with a Σ-symbol as its root label.
The restriction on rules with the initial state ensures that the output is
always a tree rather than a hedge. Transducers are required to be deterministic:
for every pair (q, a), there is at most one rule in R.
The translation defined by a tree transducer T = (Q, Σ, q 0 , R) on a tree t
in state q, denoted by T q (t), is inductively defined as follows: if t = ε then
T q (t) := ε; if t = a(t1 · · · tn ) and there is a rule (q, a) → h ∈ R then T q (t)
is obtained from h by replacing every node u in h labeled with state p by the
hedge T p (t1 ) · · · T p (tn ). Note that such nodes u can only occur at leaves. So, h
is only extended downwards. If there is no rule (q, a) → h ∈ R then T q (t) := ε.
0
Finally, the transformation of t by T , denoted by T (t), is defined as T q (t),
interpreted as a tree.
For a ∈ Σ, q ∈ Q and (q, a) → h ∈ R, we denote h by rhs(q, a). If q and
a
are
not important, we say that h is an rhs. The size of T is |Q| + |Σ| +
P
q∈Q,a∈Σ |rhs(q, a)|, where |rhs(q, a)| denotes the number of nodes in rhs(q, a).
In the sequel, we always use p, p1 , p2 , . . . and q, q1 , q2 , . . . to denote states.
Let q be a state of tree transducer T and a ∈ Σ. We then define qT [a] :=
top(T q (a)). For a string w = a1 · · · an , we define qT [w] := qT [a1 ] · · · qT [an ]. In
the sequel, we leave T implicit whenever T is clear from the context.
We give an example of a tree transducer:
Example 3.7. Let T = (Q, Σ, p, R) where Q = {p, q}, Σ = {a, b, c, d, e}, and R
contains the rules
(p, a) → d(e)
(q, a) → c p
(p, b) → d(q)
(q, b) → c(p q)
Note that the right-hand side of (q, a) → c p is a hedge consisting of two trees,
while the other right-hand sides consist of only one tree.
Our tree transducers can be implemented as XSLT programs in a straightforward way. For instance, the XSLT program equivalent to the above transducer
is given in Figure 4 (we assume the program is started in mode p).
2 In general, of course, one can define transducers where the input alphabet differs from
the output alphabet.
10
<xsl:template match="a" mode ="p">
<d>
<e/>
</d>
</xsl:template>
<xsl:template match="b" mode ="p">
<d>
<xsl:apply-templates mode="q"/>
</d>
</xsl:template>
<xsl:template match="a" mode ="q">
<c/>
<xsl:apply-templates mode="p"/>
</xsl:template>
<xsl:template match="b" mode ="q">
<c>
<xsl:apply-templates mode="p"/>
<xsl:apply-templates mode="q"/>
</c>
</xsl:template>
Figure 4: The XSLT program equivalent to the transducer of Example 3.7.
Example 3.8. Consider the tree t shown in Figure 5(a). In Figure 5(b) we
give the translation of t by the transducer of Example 3.7. In order to keep
the example simple, we did not list T q (ε) and T p (ε) explicitly in the process of
translation.
We discuss two important features of tree transducers: copying and deletion.
In Example 3.7, the rule (q, b) → c(p q) copies the children of the current node
in the input tree twice: one copy is processed in state p and the other in state
q. The symbol c is the parent node of the two copies. So, one could say that
the current node is translated in the new parent node labeled c. The rule
(q, a) → c p copies the children of the current node only once. However, no
parent node is given for this copy. So, there is no node in the output tree that
can be interpreted as the translation of the current node in the input tree. We
therefore say that it is deleted. For instance, T q (a(b)) = c d where d corresponds
to b and not to a.
We define some relevant classes of transducers. A transducer is non-deleting
if no states occur at the top-level of any rhs. We denote by Tnd the class of
non-deleting transducers and by Td the class of transducers where we allow
deletion. Furthermore, a transducer T has copying width k if there are at most
k occurrences of states in every sequence of siblings in an rhs. For instance,
the transducer in Example 3.7 has copying width 2. Given a natural number
k, which we will leave implicit, we denote by Tbc the class of transducers of
copying width k. The abbreviation “bc” stands for bounded copying. We denote
intersections of these classes by combining the indexes. For instance, Tnd,bc is
11
T p (t)
↓
d
T q (b)
T q (b(ab))
↓
d
c
c
T p (a)
T q (a(b))
T p (b)
c
T p (b)
T q (a)
T q (b)
↓
d
c
b
b
b
a
b
d
a
b
(a) The tree t of Example 3.8.
c
d
c
c
d
c
e
(b) The translation of t by the transducer T of Example 3.7.
Figure 5: A tree and its translation.
the class of non-deleting transducers with bounded copying. When we want to
emphasize that we also allow unbounded copying in a certain application, we
write, for instance, Tnd,uc instead of Tnd .
3.4
The Typechecking Problem
Definition 3.9. A tree transducer T typechecks with respect to to an input tree
language Sin and an output tree language Sout , if T (t) ∈ Sout for every t ∈ Sin .
We now define the problem central to this paper.
Definition 3.10. Given Sin , Sout , and T , the typechecking problem consists in
verifying whether T typechecks with respect to Sin and Sout .
We parameterize the typechecking problem by the kind of tree transducers
and tree languages we allow. Let T be a class of transducers and S be a representation of a class of tree languages. Then TC[T , S] denotes the typechecking
problem where T ∈ T and Sin , Sout ∈ S. Examples of classes of tree languages
are those defined by tree automata or DTDs. Classes of transducers are discussed in the previous section. The complexity of the problem is measured in
terms of the sum of the sizes of the input and output schemas Sin and Sout and
the transducer T .
Table 1 summarizes the results obtained in [23]. Unless specified otherwise,
all problems are complete for the mentioned complexity classes. In the set12
d,uc
nd,uc
NTA
exptime
exptime
nd,bc
exptime
DTA
exptime
exptime
in exptime
pspace-hard
DTD(NFA)
exptime
pspace
DTD(DFA)
exptime
pspace
DTD(SL)
exptime
conp
pspace
ptime
conp
Table 1: Results of [23] (upper and lower bounds). The top row shows the
representation of the input and output schemas and the left column shows the
class of tree transducer: “d”, “nd”, “uc”, and “bc” stand for “deleting”, “nondeleting”, “unbounded copying”, and “bounded copying” respectively.
ting of [23], typechecking is only tractable when restricting to non-deleting and
bounded copying transducers in the presence of DTDs with DFAs.
Recall that, in this article, we are interested in variants of the typechecking
problem where the input and/or output schema is fixed. We therefore introduce
some notations that are central to the paper. We denote the typechecking
problem where the input schema, the output schema, or both are fixed by
TCi [T , S], TCo [T , S], and TCio [T , S], respectively. The complexity of these
subproblems is measured in terms of the sum of the sizes of the input and
output schemas Sin and Sout , and the transducer T , minus the size of the fixed
schema(s).
4
Main Results
fixed
in, out,
in+out
in
out
in+out
TT
d,uc
d,bc
nd,uc
nd,bc
nd,uc
nd,bc
nd,uc
nd,bc
NTA
DTA
DTD(NFA)
DTD(DFA)
DTD(SL)
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
EXPTIME
PSPACE
PSPACE
EXPTIME
EXPTIME
PSPACE
NL
EXPTIME
EXPTIME
PSPACE
PSPACE
EXPTIME
EXPTIME
PTIME
ptime
conp
conp
EXPTIME
EXPTIME
NL
NL
NL
EXPTIME
EXPTIME
NL
NL
NL
in
in
PTIME
PTIME
Table 2: Complexities of the typechecking problem in the new setting (upper and
lower bounds). The top row shows the representation of the input and output
schemas, the leftmost column shows which schemas are fixed, and the second
column to the left shows the class of tree transducer: “d”, “nd”, “uc”, and
“bc” stand for “deleting”, “non-deleting”, “unbounded copying”, and “bounded
copying” respectively. In the case of deleting transformations, the different
possibilities are grouped as all complexities coincide.
13
As argued in the Introduction, it makes sense to consider the input and/or
output schema not as part of the input for some scenarios. From a complexity
theory point of view, it is important to note here that the input and/or output
alphabet then also becomes fixed. In this article, we revisit the results of [23]
from that perspective.
The results are summarized in Table 2. As some results already follow from
proofs in [23], we printed the results requiring a new proof in bold. The entries
where the complexity was lowered (assuming that the complexity classes in
question are different) are underlined. Again, all problems are complete for the
mentioned complexity classes unless specified otherwise.
We discuss the obtained results: for non-deleting transformations, we get
three new tractable cases: (i) fixed input schema, unbounded copying, and
DTD(SL)s; (ii) fixed output schema, bounded copying and DTD(NFA)s; and,
(iii) fixed input and output, unbounded copying and all DTDs. It is striking,
however, that in the presence of deletion or tree automata (even deterministic
ones) typechecking remains exptime-hard for all scenarios.
Mostly, we only needed to strengthen the lower bound proofs of [23].
4.1
Deletion: Fixed Input Schema, Fixed Output Schema,
and Fixed Input and Output Schema
The exptime upper bound for typechecking already follows from [23]. Therefore, it remains to show the lower bounds for TCio [Td,bc ,DTD(DFA)] and TCio [Td,bc ,DTD(SL)],
which we do in Theorem 4.1. In fact, if follows from the proof that the lower
bounds already hold for transducers with copying width 2.
We require the notion of top-down deterministic binary tree automata in the
proof of Theorem 4.1. A binary tree automaton (BTA) is a non-deterministic
tree automaton B = (Q, Σ, δ, F ) operating on binary trees. These are trees
where every node has zero, one, or two children. We assume that the alphabet
is partitioned in internal labels and leaf labels. When a label a is an internal
label, the regular language δ(q, a) only contains strings of length one or two.
When a is a leaf label, the regular language δ(q, a) only contains the empty
string. A binary tree automaton is top-down deterministic if (i) F is a singleton
and, (ii) for every q, q ′ ∈ Q with q 6= q ′ and a ∈ Σ, δ(q, a) contains at most
one string. We abbreviate “top-down deterministic binary tree automaton” by
TDBTA.
Theorem 4.1.
1. T C io [Td,bc , DT D(DF A)] is exptime-complete; and
2. T C io [Td,bc , DT D(SL)] is exptime-complete.
Proof. The exptime upper bound follows from Theorem 11 in [23]. We proceed
by proving the lower bounds.
We give a logspace reduction from the intersection emptiness problem of
an arbitrary number of top-down deterministic binary tree automata (TDBTAs) over the alphabet Σ = {0, 1, 0′, 1′ }. The intersection emptiness problem of
14
s
#
#
..
.
#
t
Figure 6: Structure of the trees defined by the input schema in the proof of
Theorem 4.1.
TDBTAs over alphabet {0, 1, 0′, 1′ } is known to be exptime-hard (cfr. Corollary .2(3) in the Appendix).
For i = 1, . . . , n, let Ai = (Qi , Σ, δi , {starti }) be a TDBTA, with Σ =
{0, 1, 0′, 1′ }. Without loss of generality, we can assume that the state sets Qi are
pairwise disjoint. We call 0 and 1 internal labels and 0′ and 1′ leaf labels. In our
proof, we use the markers ‘ℓ’ and ‘r’ to denote that a certain node is a left or a
right child. Formally, define Σℓ := {aℓ | a ∈ Σ} and Σr := {ar | a ∈ Σ}. We use
symbols from Σℓ and Σr for the left and right children of nodes, respectively.
Tn
We now define a transducer T and two DTDs din and dout such that i=1 L(Ai ) =
∅ if and only if T typechecks with respect to din and dout . In the construction,
we exploit the copying power of transducers to make n copies of the input tree:
one for each Ai . By using deleting states, we can execute each Ai on its copy
of the input tree without producing output. When an Ai does not accept, we
output an error symbol under the root of the output tree. The output DTD
should then only check that an error symbol always appears. A bit of care
needs to be taken, as a bounded copying transducer can not make an arbitrary
number of copies of the input tree in the same rule. The transducer therefore
goes through an initial copying phase where it repeatedly copies part of the
input tree twice, until there are (at least) n copies. The transducer remains in
the copying phase as long as it processes special symbols “#”. The input trees
are therefore of the form as depicted in Figure 6. In addition, the transducer
should verify that the number of #-symbols in the input equals ⌈log n⌉.
The input DTD (din , s), which we will describe next, uses the alphabet
Σℓ ∪ Σr ∪ {s, #}, and defines all trees of the form as described in Figure 6,
where s and # are alphabet symbols, and every internal node of t (which is
depicted in Figure 6) has one or two children. When a node is an only child, it
is labeled with an element of Σℓ . Otherwise, it is labeled with an element of Σℓ
or an element of Σr if it is a left child or a right child, respectively. In this way,
15
the transducer knows whether a node is a left or a right child by examining the
label. The root symbol of t is labeled with a symbol from Σℓ . Furthermore, all
internal nodes of t are labeled with labels in {0ℓ , 0r , 1ℓ , 1r } and all leaf nodes
are labeled with labels in {0′ℓ , 0′r , 1′ℓ , 1′r }. As explained above, we will use the
sequence of #-symbols to make a sufficient number of copies of t.
The input DTD (din , s) is defined as follows:
• din (s) = # + 0ℓ + 1ℓ ;
• din (#) = # + 0ℓ + 1ℓ ;
• for each a ∈ {0ℓ, 1ℓ , 0r , 1r },
din (a) = (0ℓ + 1ℓ + 0′ℓ + 1′ℓ ) + (0ℓ + 1ℓ + 0′ℓ + 1′ℓ )(0r + 1r + 0′r + 1′r ); and,
• for each a ∈ {0′ℓ, 1′ℓ , 0′r , 1′r }, din (a) = ε.
Obviously, (din , s) can be expressed as a DTD(DFA). It can also be expressed
as a DTD(SL), as follows
=1
′ =1
din (a) = (ϕ[0=1
] ∨ ϕ[(1′ℓ )=1 ])
ℓ ] ∨ ϕ[1ℓ ] ∨ ϕ[(0ℓ )
=1
′ =1
] ∨ ϕ[(1′ℓ )=1 ])
⊕ (ϕ[0=1
ℓ ] ∨ ϕ[1ℓ ] ∨ ϕ[(0ℓ )
∧
(ϕ[0=1
r ]
∨
ϕ[1=1
r ]
∨
ϕ[(0′r )=1 ]
∧ s=0 ∧ #=0
∨
ϕ[(1′r )=1 ])
for every a ∈ {0ℓ , 1ℓ , 0r , 1r }, where
• ⊕ denotes the “exclusive or”;
• for every i ∈ {ℓ, r} and x ∈ {0i , 1i , 0′i , 1′i }, ϕ[x=1 ] denotes the conjunction
^
y =0 ).
(x=1 ∧
y∈{0i ,1i ,0′i ,1′i }\{x}
Notice that the size of the SL-formula expressing din (a) is constant.
ε
We construct a tree transducer T = (QT , ΣT , qcopy
, RT ). The alphabet of T
is ΣT = Σℓ ∪ Σr ∪ {s, #, error, ok}. The state set QT is defined to be the set
{q ℓ , q r | q ∈ Qi , i ∈ {1, . . . , n}}. The transducer will use ⌈log n⌉ special copying
j
states qcopy
to make at least n copies of the input tree. To define QT formally,
we first introduce the notation D(k), for k = 0, . . . , ⌈log n⌉. Intuitively, D(k)
corresponds to the set of nodes of a complete binary tree of depth k + 1. For
example, D(1) = {ε, 0, 1} and D(2) = {ε, 0, 1, 00, 01, 10, 11}. The idea is that,
if i ∈ D(k) \ D(k − 1), for k > 0, then i represents the binary encoding of a
number in {0, . . . , 2kS− 1}. Formally, if k = 0, then D(k) = {ε}; otherwise,
D(k) = D(k − 1) ∪ j=0,1 {ij | i ∈ D(k − 1)}. The state set QT is then the
union of the sets Qℓ = {q ℓ | q ∈ Qj , 1 ≤ j ≤ n}, Qr = {q r | q ∈ Qj , 1 ≤ j ≤ n},
16
j
the set {qcopy
| j ∈ D(⌈log n⌉)} and the set {startℓj | n + 1 ≤ j ≤ 2⌈log n⌉ }. Note
that the last set can be empty. It only contains dummy states translating any
input to the empty string.
We next describe the action of the tree transducer T . Roughly, the operation
of T on the input s(#(#(· · · #(t)))) can be divided in two parts: (i) copying
the tree t a sufficient number of times while reading the #-symbols; and, (ii)
simulating one of the TDBTAs on each copy of t. The tree transducer outputs
the symbol “error” when one of the TDBTAs rejects t, or when the number
of #-symbols in its input is not equal to ⌈log n⌉. Apart from copying the root
symbol s to the output tree, T only writes the symbol “error” to the output.
Hence, the output tree always has a root labeled s which has zero or more
children labeled “error”. The output DTD, which we define later, should then
verify whether the root has always one “error”-labeled child.
Formally, the transition rules in RT are defined as follows:
ε
0
1
• (qcopy
, s) → s(qcopy
qcopy
). This rule puts s as the root symbol of the
output tree.
i
i0
i1
• (qcopy
, #) → qcopy
qcopy
for i ∈ D(⌈log n⌉ − 1) − {ε}. These rules copy
the tree t in the input at least n times, provided that there are enough
#-symbols.
i
• (qcopy
, #) → startℓk , where i ∈ D(⌈log n⌉) − D(⌈log n⌉ − 1), and i is the
binary representation of k. This rule starts the in-parallel simulation of the
Ai ’s. For i = n + 1, . . . , 2⌈log n⌉ , startℓi is just a dummy state transforming
everything to the empty tree.
i
, a) → error for a ∈ Σ and i ∈ D(⌈log n⌉). This rule makes sure that
• (qcopy
the output of T is accepted by the output tree automaton if there are not
enough #-symbols in the input.
• (startℓk , #) → error for all k = 1, . . . , 2⌈log n⌉ . This rule makes sure that
the output of T is accepted by the output tree automaton if there are too
much #-symbols in the input.
• (q ℓ , ar ) → ε and (q r , aℓ ) → ε for all q ∈ Qj , j = 1, . . . , n. This rule ensures
that tree automata states intended for left (respectively right) children are
not applied to right (respectively left) children.
• (q ℓ , aℓ ) → q1ℓ q2r and (q r , ar ) → q1ℓ q2r , for every q ∈ Qi , i = 1, . . . , n, such
that δi (q, a) = q1 q2 , and a is an internal symbol. This rule does the actual
simulation of the tree automata Ai , i = 1, . . . , n.
• (q ℓ , aℓ ) → q1ℓ and (q r , ar ) → q1ℓ , for every q ∈ Qi , i = 1, . . . , n, such
that δi (q, a) = q1 and a is an internal symbol. This rule does the actual
simulation of the tree automata Ai , i = 1, . . . , n.
• (q ℓ , aℓ ) → ε and (q r , ar ) → ε for every q ∈ Qi , i = 1, . . . , n, such that
δi (q, a) = ε and a is a leaf symbol. This rule simulates accepting computations of the Ai ’s.
17
• (q ℓ , aℓ ) → error and (q r , ar ) → error for every q ∈ Qi , i = 1, . . . , n, such
that δi (q, a) is undefined. This rule simulates rejecting computations of
the Ai ’s.
It is straightforward to verify that, on input s(#(#(· · · #(t)))), T outputs the
tree s if and only if there are ⌈log n⌉ #-symbols in the input and t ∈ L(A1 ) ∩
· · · ∩ L(An ).
Finally, dout (s) = error error∗ , which can easily be defined as a DTD(DFA)
and as a DTD(SL).
It is easy to see that the reduction can be carried out in deterministic logarithmic space, that T has copying width 2, and that din and dout do not depend
on A1 , . . . , An .
4.2
Non-deleting: Fixed Input Schema
We turn to the typechecking problem in which we consider the input schema as
fixed. We start by showing that typechecking is in ptime in the case where we
use DTDs with SL-expressions and the tree transducer is non-deleting (Theorem 4.3). To this end, we recall a lemma and introduce some necessary notions
that are needed for the proof of Theorem 4.3.
For an SL-formula φ, we say that two strings w1 and w2 are φ-equivalent
(denoted w1 ≡φ w2 ) if w1 |= φ if and only if w2 |= φ.
For a ∈ Σ and w ∈ Σ∗ , we denote by #a (w) the number of a’s occurring in
w. We recall Lemma 17 from [23]:
Lemma 4.2. Let φ be an SL-formula and let k be the largest integer occurring
in φ. For every w, w′ ∈ Σ∗ , for every a ∈ Σ,
• if #a (w′ ) > k when #a (w) > k, and
• #a (w′ ) = #a (w), otherwise,
then w ≡φ w′ .
For a hedge h and a DTD d, we say that h partly satisfies d if for every
u ∈ Dom(h), labh (u1) · · · labh (un) ∈ L(d(labh (u))) where u has n children.
Note that there is no requirement on the root nodes of the trees in h. Hence,
the term “partly”.
We are now ready to show the first ptime result:
Theorem 4.3. T C i [Tnd,uc , DTD(SL)] is in ptime.
Proof. Denote the tree transformation by T = (QT , Σ, qT0 , RT ) and the input
and output DTDs by (din , sin ) and (dout , sout ), respectively. As din is fixed, we
can assume that din is reduced.
Intuitively, the typechecking algorithm is successful when T does not typecheck with respect to din and dout . The outline of the typechecking algorithm
is as follows:
18
(q0T , sin )
T
b
(q, a)
c
∈ din (a)
w
6∈ dout(c)
T
Figure 7: Illustration of the typechecking algorithm in the proof of Theorem 4.3.
1. Compute the set RP of “reachable pairs” (q, a) for which there exists a
tree t ∈ L(din ) and a node u ∈ Dom(t) such that labt (u) = a and T visits
u in state q. That is, we compute all pairs (q, a) such that either
• q = qT0 and a = sin ; or
• (q ′ , a′ ) ∈ RP , there is a q-labeled node in rhs(q ′ , a′ ), and there exists
a string w1 aw2 ∈ din (a′ ) for w1 , w2 ∈ Σ∗ .
2. For each such pair (q, a) and for each node v ∈ Dom(rhs(q, a)), test
whether there exists a string w ∈ din (a) such that T q (a(w)) does not
partly satisfy dout . We call w a counterexample.
The algorithm is successful, if and only if there exists a counterexample.
We illustrate the general operation of the typechecking algorithm in Figure 7.
In this figure, T visits the a-labeled node on the left in state q. Consequently,
T outputs the hedge rhs(q, a), which is illustraded by dotted lines on the right.
The typechecking algorithm searches for a node u in rhs(q, a) (which is labeled
by c in the figure), such that the string of children of u is not in L(dout (c)).
Notice that the typechecking algorithm does not assume that dout is reduced
(recall the definition of a reduced DTD from Section 3.2). We need to show that
the algorithm is correct, that is, there exists a counterexample if and only if T
does not typecheck with respect to din and dout . Clearly, when the algorithm
does not find a counterexample, T typechecks with respect to din and dout . Conversely, suppose that the algorithm finds a pair (q, a) and a string w such that
T q (a(w)) does not partly satisfy dout . So, since we assumed that din is reduced,
19
there exists a tree t ∈ L(din ) and a node u ∈ Dom(t) such that labt (u) = a and
u is visited by T in state q. Also, there exists a node v in T q (a(w)), such that
the label of u is c and the string of children of u is not in dout (c). We argue
that T (t) 6∈ L(dout ). There are two cases:
(i) if L(dout ) contains a tree with a c-labeled node, then T (t) 6∈ dout since
T q (a(w)) does not partly satisfy dout ; and
(ii) if L(dout ) does not contain a tree with a c-labeled node, then T (t) 6∈ dout
since T (t) contains a c-labeled node.
We proceed by showing that the algorithm can be carried out in polynomial
time. As the input schema is fixed, step (1) of the algorithm is in polynomial
time. Indeed, we can compute the set RP of reachable pairs (q, a) in a top-down
manner by a straightforward reachability algorithm.
To show that step (2) of the typechecking algorithm is in polynomial time, fix
a tuple (q, a) that was computed in step (1) and a node u in rhs(q, a) with label b.
Let z0 q1 z1 · · · qn zn be the concatenation of u’s children, where all z0 , . . . , zn ∈ Σ∗
and q1 , . . . , qn ∈ QT . We now search for a string w ∈ Σ∗ for which w |= din (a),
but for which z0 q1 [w]z1 · · · qn [w]zn 6|= dout (b). Recall from Section 3.3 that q[w]
is the homomorphic extension of q[a] for a ∈ Σ, which is top(rhs(q, a))) in the
case of non-deleting tree transducers.
Denote din (a) by φ. Let {a1 , . . . , as } be the different symbols occurring in
φ and let k be the largest integer occurring in φ. According to Lemma 4.2,
ms
1
every Σ-string is φ-equivalent to a string of the form w = am
with
1 · · · as
s
0 ≤ mi ≤ k + 1 for each i = 1, . . . , s. Note that there are (k + 1) such strings,
which is a constant number, as it only depends on the input schema. For the
following, the algorithm considers each such string w.
Fix such a string w such that w |= φ. For each symbol c in dout (b), the
number #c (z0 q1 [w]z1 · · · qn [w]zn ) is equal to the linear sum
c
× #aℓ+1 (w) + ksc × #as (w) + k c ,
k1c × #a1 (w) + · · · + kℓc × #aℓ (w) + kℓ+1
where k c = #c (z0 · · · zn ) and for each i = 1, . . . , s, kic = #c (q1 [ai ] · · · qn [ai ]). We
now must test if there exists a string w′ ≡φ w such that z0 q1 [w′ ]z1 · · · qn [w′ ]zn 6|=
dout (b). Let a1 , . . . , aℓ be the symbols that occur at least k + 1 times in w
and aℓ+1 , . . . , as be the symbols that occur at most k times in w, respectively.
Then, deciding whether w′ exists is equivalent to finding an integer solution to
the variables xa1 , . . . , xas for the boolean combination of linear (in)equalities
Φ = Φ1 ∧ ¬Φ2 , where
• Φ1 states that w′ ≡φ w, that is,
Φ1 =
ℓ
^
(xai > k) ∧
i=1
s
^
j=ℓ+1
and
20
xaj = #aj (w) ;
• Φ2 states that z0 q1 (w′ )z1 · · · qn (w′ )zn |= dout (b), that is, Φ2 is defined by
replacing every occurrence of c=i or c≥i in dout (b) by the equation
s
X
(kjc × xaj ) + k c = i
j=1
or by
s
X
(kjc × xaj ) + k c ≥ i,
j=1
respectively.
In the above (in)equalities, xai , 1 ≤ i ≤ s, represents the number of occurrences
of ai in w′ .
Finding a solution for Φ now consists of finding integer values for xa1 , . . . , xas
so that Φ evaluates to true. Corollary .7 in the Appendix shows that we can
decide in ptime whether such a solution for Φ exists.
Theorem 4.4. T C i [Tnd,bc , DTD(DFA)] is nlogspace-complete.
Proof. In Theorem 4.9(2), we prove that the problem is nlogspace-hard, even
if both the input and output schemas are fixed. Hence, it remains to show that
the problem is in nlogspace.
Let us denote the tree transformation by T = (QT , Σ, qT0 , RT ) and the input
and output DTDs by (din , r) and dout , respectively. We can assume that din is
reduced.3
Then, the typechecking algorithm can be summarized as follows:
1. Guess a sequence of pairs (q0 , a0 ), (q1 , a1 ), . . . , (qn , an ) in QT × Σin , such
that
• (q0 , a0 ) = (qT0 , r); and
• for every pair (qi , ai ), qi+1 occurs in rhs(qi , ai ) and ai+1 occurs in
some string in L(din (ai )).
We only need to remember (qn , an ) as a result of this step.
2. Guess a node u in rhs(qn , an ) — say that u is labeled with b — and test
whether there exists a string w ∈ din (an ) such that T q (an (w)) does not
partly satisfy dout .
The algorithm is successful if and only if w exists and, hence, the problem does
not typecheck.
The first step is a straightforward reachability algorithm, which is in nlogspace.
It remains to show that the second step is in nlogspace.
Let (q, a) be the pair (qn , an ) computed in step two. Let dout (b) = (Qout , Σ, δout ,
{pI }, {pF }) be a DFA and let k be the copying bound of T . Let z0 q1 z1 · · · qℓ zℓ
3 Reducing
din would be ptime-complete otherwise, see Corollary .9 in the Appendix.
21
be the concatenation of u’s children, where ℓ ≤ k. So we want to check whether
there exists a string w such that z0 q1 [w]z1 · · · qℓ [w]zℓ is not accepted by dout (b).
We guess w one symbol at a time and simulate in parallel ℓ copies of dout (b)
and one copy of din (a).
By δ̂ we denote the canonical extension of δ to strings in Σ∗ . We start by
guessing states p1 , . . . , pℓ of dout (b), where p1 = δ̂out (pI , z0 ), and keep a copy of
these on tape, to which we refer as p′1 , . . . , p′ℓ . Next, we keep on guessing symbols
c of w, whereafter we replace each pi by δ̂out (pi , qi (c)). The input automaton
obviously starts in its initial state and is simulated in the straightforward way.
The machine non-deterministically stops guessing, and checks whether, for
each i = 1, . . . , ℓ − 1, δ̂out (pi , zi ) = p′i+1 and δ̂out (pℓ , zℓ ) = pF . For the input
automaton, it simply checks whether the current state is the final state. If the
latter tests are positive, then the algorithm accepts, otherwise, it rejects.
We only keep 2ℓ + 1 states on tape, which is a constant number, so the
algorithm runs in nlogspace.
Theorem 4.5.
1. T C i [Tnd,uc , DTD(DFA)] is pspace-complete; and
2. T C i [Tnd,bc , DTD(NFA)] is pspace-complete.
Proof. In [23], it was shown that both problems are in pspace. We proceed by
showing that they are also pspace-hard.
(1) We reduce the intersection emptiness problem of an arbitrary number of
deterministic finite automata with alphabet {0, 1} to the typechecking problem.
This problem is known to be pspace-hard, as shown in Corollary .2(1) in the
Appendix. Our reduction only requires logarithmic space. We define a transducer T = (QT , {0, 1, #0, . . . , #n }, qT0 , RT ) and two DTDs dTin and dout such
n
that T typechecks with respect to din and dout if and only if i=1 L(Mi ) = ∅.
The DTD (din , s) defines trees of depth two, where the string formed by
the children of the root is an arbitrary string in {0, 1}∗, so din (s) = (0 + 1)∗ .
The transducer makes n copies of this string, separated by the delimiters #i :
QT = {q, qT0 } and RT contains the rules (qT0 , s) → s(#0 q#1 q . . . #n−1 q#n ) and
(q, a) → a, for every a ∈ Σ. Finally, (dout , s) defines a tree of depth two as
follows:
dout (s) = {#0 w1 #1 w2 #2 · · · #n−1 wn #n |
∃j ∈ {1, . . . , n} such that Mj does not accept wj }.
Clearly, dout (s) can be represented by a DFA whose size is polynomial in the sizes
of the Mi ’s. Indeed, the DFA just simulates every Mi on the string following
#i−1 , until it encounters #i . It then verifies that at least one Mi rejects.
It is easy to see that this reduction can be carried out by a deterministic
logspace algorithm.
(2) This is an easy reduction from the universality problem of an NFA N with
alphabet {0, 1}. The latter problem is pspace-hard, as shown in Corollary .2(2)
in the Appendix. Again, the input DTD (din , s) defines a tree of depth two
where din (s) = (0 + 1)∗ . The tree transducer is the identity transformation.
22
The output DTD dout has as start symbol s and dout (s) = L(N ). Hence, this
instance typechecks if and only if {0, 1}∗ ⊆ L(N ).
This reduction can be carried out by a deterministic logspace algorithm.
4.3
Non-deleting: Fixed Output Schema
Again, upper bounds carry over from [23]. Also, when the output DTD is
a DTD(NFA), we can convert it into an equivalent DTD(DFA) in constant
time. As the ptime typechecking algorithm for TC[Tnd,bc ,DTD(DFA)] in [23]
also works when the input DTD is a DTD(NFA), we have that the problem TCo [Tnd,bc ,DTD(NFA)] is in ptime. As the ptime-hardness proof for
TC[Tnd,bc ,DTD(DFA)] in [23] uses a fixed output schema, we immediately obtain the following.
Theorem 4.6. T C o [Tnc,bc , DT D(N F A)] is ptime-complete.
The lower bound in the presence of tree automata will be discussed in Section 4.4. The case requiring some real work is T C o [Tnd,uc , DTD(DFA)].
Theorem 4.7. T C o [Tnd,uc , DTD(DFA)] is pspace-complete.
Proof. In [23], it was shown that the problem is in pspace. We proceed by
showing pspace-hardness.
We use a logspace reduction from the corridor tiling problem [9]. Let (T, V,
H, ϑ̄, β̄) be a tiling system, where T = {ϑ1 , . . . , ϑk } is the set of tiles, V ⊆ T × T
and H ⊆ T × T are the sets of vertical and horizontal constraints respectively,
and ϑ̄ and β̄ are the top and bottom row, respectively. Let n be the width of
ϑ̄ and β̄. The tiling system has a solution if there is an m ∈ N such that the
space m × n (m rows and n columns) can be correctly tiled with the additional
requirement that the bottom and top row are β̄ and ϑ̄, respectively.
We define the input DTD din over the alphabet Σ := {(i, ϑj ) | j ∈ {1, . . . , k}, i ∈
{1, . . . , n}} ∪ {r}; r is the start symbol. Define
∗
din (r) = #β̄# Σ1 · Σ2 · · · Σn # ϑ̄#,
where we denote by Σi the set {(i, ϑj ) | j ∈ {1, . . . , k}}. Here, # functions as
a row separator. For all other alphabet symbols a ∈ Σ, din (a) = ε. So, din
encodes all possible tilings that start and end with the bottom row β̄ and the
top row ϑ̄, respectively.
0
We now construct a tree transducer B = (QB , Σ, qB
, RB ) and an output
DTD dout such that T has no correct corridor tiling if and only if B typechecks
with respect to din and dout . Intuitively, the transducer and the output DTD
have to work together to determine errors in input tilings. There can only
be two types of error: two tiles do not match horizontally or two tiles do not
match vertically. The main difficulty is that the output DTD is fixed and can,
therefore, not depend on the tiling system. The transducer is constructed in such
23
a way that it prepares in parallel the verification for all horizontal and vertical
constraints by the output schema. In particular, the transducer outputs specific
symbols from a fixed set independent of the tiling system allowing the output
schema to determine whether an error occurred.
The state set QB is partitioned into two sets, Qhor and Qver :
• Qhor is for the horizontal constraints: for every i ∈ {1, . . . , n − 1} and
ϑ ∈ T , qi,ϑ ∈ Qhor transforms the rows in the tiling such that it is possible
to check that when position i carries a ϑ, position i + 1 carries a ϑ′ such
that (ϑ, ϑ′ ) ∈ H; and,
• Qver is for the vertical constraints: for every i ∈ {1, . . . , n} and ϑ ∈ T ,
pi,ϑ ∈ Qver transforms the rows in the tiling such that it is possible to check
that when position i carries a ϑ, the next row carries a ϑ′ on position i
such that (ϑ, ϑ′ ) ∈ V .
The tree transducer B always starts its transformation with the rule
0
, r) → r(w),
(qB
where w is the concatenation of all of the above states, separated by the delimiter
$. The other rules are of the following form:
• Horizontal constraints: for all
where qi,ϑ ∈ Qhor and
trigger
other
ok
α=
error
other
(j, ϑ) ∈ Σ add the rule (qi,ϑ , (j, ϑ′ )) → α
if
if
if
if
if
j
j
j
j
j
= i and ϑ = ϑ′
= i and ϑ 6= ϑ′
= i + 1 and (ϑ, ϑ′ ) ∈ H
= i + 1 and (ϑ, ϑ′ ) 6∈ H
6= i and j 6= i + 1
Finally, (qi,ϑ , #) → hor.
The intuition is as follows: if the i-th position in a row is labeled with ϑ,
then this position is transformed into trigger. Position i + 1 is transformed to ok when it carries a tile that matches ϑ horizontally. Otherwise,
it is transformed to error. All other symbols are transformed into an
other.
On a row, delimited by two hor-symbols, the output DFA rejects if and
only if there is a trigger immediately followed by an error. When there
is no trigger, then position i was not labeled with ϑ. So, the label
trigger acts as a trigger for the output automaton.
• Vertical constraints: for all
where pi,ϑ ∈ Qver and
trigger1
trigger2
ok
α=
error
other
(j, ϑ) ∈ Σ, add the rule (pi,ϑ , (j, ϑ′ )) → α
if
if
if
if
if
(j, ϑ′ ) = (i, ϑ) and (ϑ, ϑ) ∈ V
(j, ϑ′ ) = (i, ϑ) and (ϑ, ϑ) 6∈ V
j = i, ϑ 6= ϑ′ , and (ϑ, ϑ′ ) ∈ V
j = i, ϑ 6= ϑ′ , and (ϑ, ϑ′ ) 6∈ V
j 6= i
24
Finally, (pi,ϑ , #) → ver.
The intuition is as follows: if the i-th position in a row is labeled with ϑ,
then this position is transformed into trigger1 when (ϑ, ϑ) ∈ V and to
trigger2 when (ϑ, ϑ) 6∈ V . Here, both trigger1 and trigger2 act as a
trigger for the output automaton: they mean that position i was labeled
with ϑ. But no trigger1 and trigger2 can occur in the same transformed
row as either (θ, θ) ∈ V or (θ, θ) 6∈ V . When position i is labeled with
ϑ′ 6= ϑ, then we transform this position into ok when (ϑ, ϑ′ ) ∈ V , and in
error when (ϑ, ϑ′ ) 6∈ V . All other positions are transformed into other.
The output DFA then works as follows. If a position is labeled trigger1
then it rejects if there is a trigger2 or a error occurring after the next
ver. If a position is labeled trigger2, then it rejects if there is a trigger2
or a error occurring after the next ver. Otherwise, it accepts that row.
By making use of the delimiters ver and hor, both above described automata
can be combined into one automaton, taking care of the vertical and the horizontal constraints. This automaton resets to its initial state whenever it reads
the delimiter symbol $. Note that the output automaton is defined over the fixed
alphabet {trigger, trigger1, trigger2, error, ok, other, hor, ver, $}.
Although the results in [23] were formulated in the context of variable
schemas, the proofs for bounded copying, non-deleting tree transducers with
DTD(SL) and with DTD(DFA) schemas actually used a fixed output schema.
We can therefore sharpen these results as follows.
Theorem 4.8.
1. T C o [Tnd,bc , DTD(SL)] is conp-complete;
2. T C o [Tnd,bc , DTD(DFA)] is ptime-complete.
4.4
Non-deleting: Fixed Input and Output Schema
We turn to the case where both input and output schemas are fixed. The
following two theorems give us several new tractable cases.
Theorem 4.9.
1. T C io [Tnd,bc , DTD(SL)] is nlogspace-complete.
2. T C io [Tnd,bc , DTD(DFA)] is nlogspace-complete.
Proof. For both problems, membership in nlogspace follows from Theorem 4.10.
Indeed, every DTD(SL) can be rewritten into an equivalent DTD(NFA) in constant time as the input and output schemas are fixed.
We proceed by showing nlogspace-hardness. We say that an NFA N =
(QN , Σ, δN , IN , FN ) has degree of nondeterminism 2 if (i) IN has at most two
elements and (ii) for every q ∈ QN and a ∈ Σ, the set δN (q, a) has at most
two elements. We give a logspace reduction from the emptiness problem of an
NFA with alphabet {0, 1} and a degree of nondeterminism 2 to the typechecking
problem. According to Lemma .3 in the Appendix, this problem is nlogspacehard. Intuitively, the input DTD will define all possible strings over alphabet
25
{0, 1}. The tree transducer simulates the NFA and outputs “accept” if a computation branch accepts, and “error” if a computation branch rejects. The output
DTD defines trees where all leaves are labeled with “error”.
0
More concretely, let N = (QN , {0, 1}, δN , {qN
}, FN ) be an NFA with degree
of nondeterminism 2. The input DTD (din , r) defines all unary trees, where
the unique leaf is labeled with a special marker #. That is, din (r) = din (0) =
din (1) = (0 + 1 + #) and din (#) = ε. Note that these languages can be defined
by SL-formulas or DFAs which are sufficiently small for our purpose.
Given a tree t = r(a1 (· · · (an (#)) · · · )), the tree transducer will simulate
every computation of N on the string a1 · · · an . The tree transducer T =
(QT , {r, #, 0, 1, error, accept}, qT0 , RT ) simulates N ’s nondeterminism by copying the remainder of the input twice in every step. Formally, QT is the union
of {qT0 } and QN , and RT contains the following rules:
0
• (qT0 , r) → r(qN
). This rule puts r as the root symbol of the output tree
and starts the simulation of N .
1
2
• (qN , a) → a(qN
, qN
), where qN ∈ QN , a ∈ {0, 1} and δN (qN , a) =
1
2
{qN , qN }. This rule does the actual simulation of N . By continuing in
1
2
both states qN
and qN
, we simulate all possible computations of N .
• (qN , a) → error if δN (qN , a) = ∅. If N rejects, we output the symbol
“error”.
• (qN , #) → error for qN 6∈ FN ; and
• (qN , #) → accept for qN ∈ FN . These last two rules verify whether N is
in an accepting state after reading the entire input string.
Notice that T outputs the symbol “error” (respectively “accept”) if and only
if a computation branch of N rejects (respectively accepts).
The output of T is always a tree in which only the symbols “error” and
“accept” occur at the leaves. The output DTD then needs to verify that only
the symbol “error” occurs at the leaves. Formally, dout (r) = dout (0) = dout (1) =
{0, 1, error}+ and dout (error) = ε. Again, these languages can be defined by
sufficiently small SL-formulas or DFAs.
It is easy to see that the reduction only requires logarithmic space.
Theorem 4.10. T C io [Tnd,uc , DTD(NFA)] is nlogspace-complete.
Proof. The nlogspace-hardness of the problem follows from Theorem 4.9(b),
where it shown that the problem is already nlogspace-hard when DTD(DFA)s
are used as input and output schema.
We show that the problem is also in nlogspace. Thereto, let T = (QT , Σ, qT0 ,
RT ) be the tree transducer, and let (din , r) and dout be the input and output
DTDs, respectively. As both din and dout are fixed, we can assume without loss
of generality that they are reduced.4 For the same reason, we can also assume
that the NFAs in din and dout are determinized.
4 In
general, reducing a DTD(NFA) is ptime-complete (Section 3.2).
26
We guess a sequence of state-label pairs (p0 , a0 ), (p1 , a1 ), . . . , (pn , an ) where
n < |QT ||Σ| such that
• (p0 , a0 ) = (qT0 , r); and
• for every pair (pi , ai ), pi+1 occurs in rhs(pi , ai ) and ai+1 occurs in some
string in L(din (ai )).
Each time we guess a new pair in this sequence, we forget the previous one,
so that we only keep a state, an alphabet symbol, a counter, and the binary
representation of |QT ||Σ| on tape.
For simplicity, we write (pn , an ) as (p, a) in the remainder of the proof. We
guess a node u ∈ Dom(rhs(p, a)). Let b = labrhs(p,a) (u) and let z0 q1 z1 · · · qk zk
be the concatenation of u’s children, where every z0 , . . . , zk ∈ Σ∗ and every
q1 , . . . , qk ∈ QT , then we want to check whether there exists a string w ∈ din (a)
such that z0 q1 [w]z1 · · · qk [w]zk is not accepted by dout (b). Recall from Section 3.3
that, for a state q ∈ QT , we denote by q[w] the homomorphic extension of q[c]
for c ∈ Σ, which is top(rhs(q, c))) in the case of non-deleting tree transducers.
We could do this by guessing w one symbol at a time and simulating k copies
of dout (b) and one copy of din (a) in parallel, like in the proof of Theorem 4.4.
However, as k is not fixed, the algorithm would use superlogarithmic space.
0
So, we need a different approach. To this end, let A = (Qin , Σ, δin , qin
, Fin )
0
and B = (Qout , Σ, δout , qout , Fout ) be the DFAs accepting din (a) and dout (b),
respectively. To every q ∈ QT , we associate a function
fq : Qout × Σ → Qout : (p′ , c) 7→ δ̂out (p′ , q[c]),
where δ̂out denotes the canonical extension of δout to strings in Σ∗ . Note that
there are maximally |Qout ||Qout ||Σ| such functions. Let K be the cardinality of
the set {fq | q ∈ QT }. Hence, K is bounded from above by |Qout ||Qout ||Σ| ,
which is a constant (with respect to the input). Let f1 , . . . , fK an arbitrary
enumeration of {fq | q ∈ QT }.
The typechecking algorithm continues as follows. We start by writing the
0
′
′
(1+K·|Qout |)-tuple (qin
, q1′ , . . . , q|Q
, . . . , q1′ , . . . , q|Q
) on tape, where Qout =
out |
out |
′
′
{q1 , . . . , q|Qout | }. We will refer to this tuple as the tuple p̄ := (p′0 , . . . , p′K·|Qout | ).
We explain how we update p̄ when guessing w symbol by symbol. Every time
when we guess the next symbol c of w, we overwrite the tuple p̄ by
δin (p′0 , c), f1 (p′1 , c), . . . , f1 (p′|Qout | , c), . . .
. . . , fK (p′(K−1)·|Qout |+1 , c), . . . , fK (p′K·|Qout | , c) .
Notice that there are at most |Qin | · K · |Qout |2 different (K · |Qout | + 1)-tuples of
this form. We nondeterministically determine when we stop guessing symbols
of w.
It now remains to verify whether w was indeed a string such that w ∈ din (a)
and z0 q1 [w]z1 · · · qk [w]zk 6∈ dout (b). The former condition is easy to test: we
simply have to test whether p′0 ∈ Fin . To test the latter condition, we read the
27
string z0 q1 z1 · · · qk zk from left to right while performing the following tests. We
keep a state of dout (b) in memory and refer to it as the “current state”.
0
1. The initial current state is qout
.
2. If the current state is p′ and we read zj , then we change the current state
to δ̂out (p′ , zj ).
3. If the current state is p′ and we read qj , then we change the current
state to p′i in p̄, where for i, the following condition holds. Let ℓ, m =
1, . . . , K · |Qout | be the smallest integers such that
• p′ = qℓ′ in Qout , and
• fqj = fm .
Then i = (m − 1)K + ℓ.
Note that deciding whether p′ = qℓ′ and fqj = fm can be done deterministically in logarithmic space, as the output schema is fixed. Consequently,
i can also be computed in constant time and space.
4. We stop and accept if the current state is a non-accepting state after
reading zk .
Theorem 4.11. T C io [Tnd,bc , DTA(DFA)] is exptime-complete.
Proof. The proof is quite analogous to the proof of Theorem 4.1. As deletion is
now disallowed, whereas it was allowed in Theorem 4.1, we need to define the
ε
rules of the transducer T = (QT , ΣT , qcopy
, RT ) differently.
The language defined by the input schema is exaclty the same as in Theorem 4.1. The transition rules in RT are defined as follows:
ε
0
1
• (qcopy
, s) → s(qcopy
qcopy
);
i
i0
i1
• (qcopy
, #) → #(qcopy
qcopy
) for i ∈ D(⌈log n⌉ − 1) − {ε};
i
• (qcopy
, #) → #(startℓk ), where i ∈ D(⌈log n⌉) − D(⌈log n⌉ − 1), and i is
the binary representation of k;
i
• (qcopy
, a) → error for a ∈ Σ and i ∈ D(⌈log n⌉);
• (startℓk , #) → error for all k = 1, . . . , 2⌈log n⌉ ;
• (q ℓ , ar ) → ε and (q r , aℓ ) → ε for all q ∈ Qj , j = 1, . . . , n;
• (q ℓ , aℓ ) → aℓ (q1ℓ q2r ) and (q r , ar ) → ar (q1ℓ q2r ), for every q ∈ Qi , i = 1, . . . , n,
such that δi (q, a) = q1 q2 , and a is an internal symbol;
• (q ℓ , aℓ ) → aℓ (q1ℓ ) and (q r , ar ) → ar (q1ℓ ), for every q ∈ Qi , i = 1, . . . , n,
such that δi (q, a) = q1 and a is an internal symbol;
28
• (q ℓ , aℓ ) → ε and (q r , ar ) → ε for every q ∈ Qi , i = 1, . . . , n, such that
δi (q, a) = ε and a is a leaf symbol; and
• (q ℓ , aℓ ) → error and (q r , ar ) → error for every q ∈ Qi , i = 1, . . . , n, such
that δi (q, a) is undefined.
It is straightforward to verify that, on input s(#(#(· · · #(t)))), T performs the
identity transformation if and only if there are ⌈log n⌉ #-symbols in the input
and t ∈ L(A1 ) ∩ · · · ∩ L(An ). All other outputs contain at least one leaf labeled
“error”.
Finally, the output tree automaton accepts all trees with at least one leaf
that is labeled “error”. So the only counterexamples for typechecking are those
trees that are accepted by all automata.
It is easy to see that the reduction can be carried out in deterministic logarithmic space, that T has copying width 2, and that the input and output
schemas do not depend on A1 , . . . , An .
5
Conclusion
We considered the complexity of typechecking in the presence of fixed input
and/or output schemas. We have settled an open question in [23], namely that
TC[Tnd,bc , DT A] is exptime-complete.
In comparison with the results in [23], fixing input and/or output schemas
only lowers the complexity in the presence of DTDs and when deletion is disallowed. Here, we see that the complexity is lowered when
1. the input schema is fixed, in the case of DTD(SL)s;
2. the input schema is fixed, in the case of DTD(DFA)s;
3. the output schema is fixed, in the case of DTD(NFA)s; and
4. both input and output schema are fixed, in all cases.
In all of these cases, the complexity of the typechecking problem is in polynomial
time.
It is striking, however, that in many cases, the complexity of typechecking
does not decrease significantly by fixing the input and/or output schema, and
most cases remain intractable. We have to leave the precise complexity (that
is, the ptime-hardness) of T C i [Tnd,uc , DTD(SL)] as an open problem.
Acknowledgments
We thank Giorgio Ghelli for raising the question about the complexity of typechecking in the setting of a fixed output schema. We also thank Joos Heintz for
providing us with a useful reference to facilitate the proof of Proposition .6.
29
Appendix: Definitions and Basic Results
The purpose of this Appendix is to prove some lemmas that we use in the
body of the paper. We first introduce some notations and definitions needed for
the propositions and proofs further on in this Appendix. We also survey some
complexity bounds on decisions problems concerning automata that are used
throughout the paper.
We show that the complexities of the classical decision problems of string and
tree automata are preserved when the automata operate over fixed alphabets.
We will consider the following decision problems for string automata:
Emptiness: Given an automaton A, is L(A) = ∅?
Universality: Given an automaton A, is L(A) = Σ∗ ?
Intersection emptiness: Given the automata A1 , . . . , An , is L(A1 ) ∩ · · · ∩
L(An ) = ∅?
The corresponding decision problems for tree automata are defined analogously.
We associate to each label a ∈ Σ a unique binary string enc(a) ∈ {0, 1}∗ of
length ⌈log |Σ|⌉. For a string s = a1 · · · an , enc(s) = enc(a1 ) · · · enc(an ). This
encoding can be extended to string languages in the obvious way.
We show how to extend the encoding “enc” to trees over alphabet {0, 1, 0′, 1′ }.
Here, 0 and 1 are internal labels, while 0′ and 1′ are leaf labels. Let enc(a) =
b1 · · · bk for a ∈ Σ. Then we denote by tree-enc(a) the unary tree b1 (b2 (· · · (bk )),
if a is an internal label, and the unary tree b1 (b2 (· · · (b′k )), otherwise. Then, the
enc-fuction can be extended to trees as follows: for t = a(t1 · · · tn ),
enc(t) = tree-enc(a)(enc(t1 ) · · · enc(tn )).
Note that we abuse notation here. The hedge enc(t1 ) · · · enc(tn ) is intended to
be the child of the leaf in tree-enc(a). The encoding can be extended to tree
languages in the obvious way.
Proposition .1. Let B be a TDBTA. Then there is a TDBTA B ′ over the
alphabet {0, 1, 0′, 1′ } such that L(B ′ ) = enc(L(B)). Moreover, B ′ can be constructed from B in logspace.
Proof. Let B = (QB , ΣB , δB , FB ) be a TDBTA. Let k := ⌈log |ΣB |⌉. We define
B ′ = (QB ′ , {0, 1, 0′, 1′ }, δB ′ , FB ′ ). Set QB ′ = {qx | q ∈ QB and x is a prefix of
enc(a), where a ∈ ΣB } and FB ′ = {qε | q ∈ FB }. To define the transition function, we introduce some notation. For each a ∈ Σ and i, j = 1, . . . , ⌈log |ΣB |⌉,
denote by a[i : j] the substring of enc(a) from position i to position j (we abbreviate a[i : i] by a[i]). For each transition δB (q, a) = q 1 q 2 , add the transitions δB ′ (qε , a[1]) = qa[1] , δB ′ (qa[1] , a[2]) = qa[1:2] , . . . , δB ′ (qa[1:k−1] , a[k]) = qε1 qε2 .
Other transitions are defined analogously. Clearly, B ′ is a TDBTA, L(B ′ ) =
enc(L(B)), and B ′ can be constructed from B in logspace.
30
It is straightforward to show that Proposition .1 also holds for NFAs and
DFAs (the proofs are analogous). It is immediate from Proposition .1, that lower
bounds of decision problems for automata over arbitrary alphabets [16, 29, 31]
carry over to automata working over fixed alphabets. Hence, we obtain the
following corollary to Proposition .1:
Corollary .2. Over the alphabet {0, 1}, the following statements hold:
1. Intersection emptiness of an arbitrary number of DFAs is pspace-hard.
2. Universality of NFAs is pspace-hard.
Over the alphabet {0, 1, 0′, 1′ }, the following statement holds:
(3) Intersection emptiness of an arbitrary number of TDBTAs is exptimehard.
Lemma .3 now immediately follows from nlogspace-hardness of the reachability problem on graphs with out-degree 2 [15].
Lemma .3. The emptiness problem for an NFA with alphabet {0, 1} degree of
nondeterminism 2 is nlogspace-hard.
We now aim at proving Proposition .6, which states that we can find integer
solutions to arbitrary Boolean combinations of linear (in)equalities in polynomial time, when the number of variables is fixed. To this end, we revisit a
lemma that is due to Ferrante and Rackoff [13].
First, we need some definitions. We define logical formulas with variables
x1 , x2 , . . . and linear equations with factors in Q. A term is an expression of the
form a1 /b1 , a1 /b1 x1 + · · · + an /bn xn , or a1 /b1 x1 + · · · + an−1 /bn−1 xn−1 + an /bn
where ai , bi ∈ N for i = 1, . . . , n. An atomic formula is either the string “true”,
the string “false”, or a formula of the form ϑ1 = ϑ2 , ϑ1 < ϑ2 , or ϑ1 > ϑ2 .
A formula is built up from atomic formulas using conjunction, disjunction,
negation, and the symbol ∃ in the usual manner. Formulas are interpreted in
the obvious manner over Q. For instance, the formula ¬∃x1 , x2 (x1 < x2 ) ∧
¬ ∃x3 (x1 < x3 ∧ x3 < x2 ) states that for every two different rational numbers,
there exists a third rational number that lies strictly between them.
The size of a formula Φ is the sum of the number of brackets, Boolean
connectives, the sizes of the variables, and the sizes of all rational constants
occurring in Φ. Here, we assume that all rational constants are written as a/b,
where a and b are integers, written in binary notation. We assume that variables
are written as xi , where i is written in binary notation.
Lemma .4 (Lemma 1 in [13]). Let Φ(x1 , . . . , xn ) be a quantifier-free formula.
Then there exists a ptime procedure for obtaining another quantifier-free formula, Φ′ (x1 , . . . , xn−1 ), such that
Φ′ (x1 , . . . , xn−1 ) is equivalent to ∃xn Φ(x1 , . . . , xn ).
Proof. Let Φ(x1 , . . . , xn ) be a quantifier-free formula.
31
Step 1: Solve for xn in each atomic formula of Φ. That is, obtain a quantifierfree formula, Ψn (x1 , . . . , xn ), such that every atomic formula of Ψn either
does not involve xn or is of the form (i) xn < ϑ, (ii) xn > ϑ, or (iii) xn = ϑ,
where ϑ is a term not involving xn .
Step 2: We now make the following definitions:
Given Ψn (x1 , . . . , xn ), to get Ψn−∞ (x1 , . . . , xn−1 ), respectively, Ψn∞ (x1 , . . . , xn−1 ),
replace
xn < ϑn in Ψ by “true” (respectively, “false”);
xn > ϑn in Ψ by “false” (respectively, “true”); and,
xn = ϑn in Ψ by “false” (respectively, “false”).
The intuition is that, for any rational numbers r1 , . . . , rn−1 , if r is a sufficiently small rational number, then Ψn (r1 , . . . , rn−1 , r) and Ψn−∞ (r1 , . . . , rn−1 )
are equivalent. A similar statement can be made for Ψn∞ for r sufficiently
large.
Step 3: We will now eliminate the quantifier from ∃xn Ψ(x1 , . . . , xn ). Let U
be the set of all terms ϑ (not involving xn ) such that xn > ϑ, xn < ϑ,
or xn = ϑ is an atomic formula of Ψ. Lemma 1.1 in [13] then shows that
∃xn Ψ(x1 , . . . , xn ) is equivalent to the quantifier-free formula Φ′ (x1 , . . . , xn−1 )
defined to be
_
′
Ψn−∞ ∨ Ψn∞ ∨
Ψn,ϑ,ϑ ,
ϑ,ϑ′ ∈U
′
′
.
where Ψn,ϑ,ϑ = Ψn x1 , . . . , xn−1 , ϑ+ϑ
2
The following proposition is implicit in the work by Ferrante and Rackoff [13],
we prove it for completeness:
Proposition .5. Let Φ(x1 , . . . , xn ) be a quantifier-free formula. If n is fixed,
then satisfiability of Φ over Q can be decided in ptime. Moreover, if Φ is
satisfiable, we can find (v1 , . . . , vn ) ∈ Qn such that Φ(v1 , . . . , vn ) is true in
polynomial time.
Proof. We first show that satisfiability can be decided in ptime. To this end,
we simply iterate over the three steps in the proof of Lemma .4 until we obtain a
formula without variables. Hence, in each iteration, one variable xi is eliminated
from Φ. For every i = 1, . . . , n, let Φi be the formula obtained after eliminating
variable xi .
Notice that, in each iteration of the algorithm, the number of atomic formulas
grows quadratically when going from Φi to Φi−1 . However, as there are only a
constant number of iterations, the number of atomic formulas in the resulting
formula Φ1 is still polynomial. Moreover, Ferrante and Rackoff show that the
absolute value of every integer constant occurring in any rational constant in Φi
32
n
is at most (s0 )14 , where s0 is the largest absolute value of any integer constant
occurring in any rational constant in Φ (cfr. page 73 in [13]). As n is a fixed
number, we can decide whether Φ1 is satisfiable in polynomial time.
Suppose that Φ is satisfiable. We now show how we can construct (v1 , . . . , vn ) ∈
Qn in polynomial time such that Φ(v1 , . . . , vn ) is true. For a term ϑ using variables x1 , . . . , xi , we denote by ϑ(v1 , . . . , vi−1 ) the rational number obtained by
replacing the variables x1 , . . . , xi in ϑ by v1 , . . . , vi−1 and evaluating the resulting expression.
′
For every i = 1, . . . , n, we construct vi from Ψi , Ψi−∞ , Ψi∞ , and Ψi,ϑ,ϑ
(which are defined in the proof of Lemma .4) as follows:
′
(1) If Ψi,ϑ,ϑ is satisfiable, then vi =
ϑ(v1 ,...,vi−1 )+ϑ′ (v1 ,...,vi−1 )
.
2
(2) Otherwise, if Ψi∞ is satisfiable, then vi = max{ϑ(v1 , . . . , vi−1 ) | xi < ϑ or
xi > ϑ or xi = ϑ is an atomic formula in Ψi } + 1.
(3) Otherwise, if Ψi−∞ is satisfiable, then vi = min{ϑ(v1 , . . . , vi−1 ) | xi < ϑ or
xi > ϑ or xi = ϑ is an atomic formula in Ψi } − 1.
It remains to show that we can represent every vi in a polynomial manner.
In the proof of Lemma 2 in [13], Ferrante and Rackoff show that, if wi is the
maximum absolute value of any integer occurring in the definition of v1 , . . . , vi ,
then we have the recurrence
cn
wi+1 ≤ (s0 )2
· (wi )i ,
for a constant c and s0 defined as before. Let c′ = 2cn . Hence, the maximum
′
number
of bits needed to represent the largest integer in vi+1 is log (s0 )c ·
(wi )i = c′ log(s0 ) · i log(wi ), which is polynomially larger than log(wi ), the
number of bits needed to represent the largest integer in vi .
As we only have a constant number of iterations, the number of bits needed to
represent the largest integer occurring in the definition of vn is also polynomial.
The following proposition is a generalization of a well-known theorem by
Lenstra which states that there exists a polynomial time algorithm to find an
integer solution for a conjunction of linear (in)equalities with rational factors
and a fixed number of variables [18].
Proposition .6. There exists a ptime algorithm that decides whether a Boolean
combination of linear (in)equalities with rational factors and a fixed number of
variables has an integer solution.
Proof. Note that we cannot simply put the Boolean combination into disjunctive
normal form, as this would lead to an exponential increase of its size.
Let Φ(x1 , . . . , xn ) be a Boolean combination of formulas ϕ1 , . . . , ϕm with
variables x1 , . . . , xn that range over Z. Here, n is a constant integer greater
33
than zero. Without loss of generality, we can assume that every ϕi is of the
form
ki,1 × x1 + · · · + ki,n × xn + ki ≥ 0,
where ki , ki,1 , . . . , ki,n ∈ Q.
We describe a ptime procedure for finding a solution for x1 , . . . , xn , that is,
for finding values v1 , . . . , vn ∈ Z such that Φ(v1 , . . . , vn ) evaluates to true.
First, we introduce some notation and terminology. For every i = 1, . . . , m,
we denote by ϕ′i the formula ki,1 × x1 + · · · + ki,n × xn + ki = 0. In the
following, we freely identify ϕ′i with the hyperplane it defines in Rn . For an
n-tuple y = (y1 , . . . , yn ) ∈ Qn , we denote by ϕ′i (y) the rational number ki,1 ×
y1 + · · · + ki,n × yn + ki .
Given a set of hyperplanes H in Rn , we say that C ⊆ Rn is a cell of H when
(i) for every hyperplane ϕ′i in H, and for every pair of points y, z ∈ C, we
have that ϕ′i (y) θ 0 if and only if ϕ′i (z) θ 0, where, θ denotes “<”, “>”, or
“=”; and
(ii) there exists no C ′ ) C with property (i).
Let H be the set of hyperplanes {ϕ′i | 1 ≤ i ≤ m}.
We now describe the ptime algorithm. The algorithm iterates over the
following steps:
(1) Compute (v1′ , . . . , vn′ ) ∈ Qn such that Φ(v1′ , . . . , vn′ ) is true.5 If no such
(v1′ , . . . , vn′ ) exists, the algorithm rejects.
(2) For every ϕ′i ∈ H, let θi ∈ {<, >, =} be the relation such that
ki,1 × v1′ + · · · + ki,n × vn′ + ki θi 0.
For every i = 1, . . . , m, let ϕ′′i = ki,1 × x1 + · · · + ki,n × xn + ki θi 0. So, for
every i = 1, . . . , m, ϕ′′i defines the half-space or hyperplane that contains
the point (v1′ , . . . , vn′ ).
Let Φ′ (x1 , . . . , xn ) be the conjunction
^
ϕ′′i .
1≤i≤n
Notice that the points satisfying Φ′ (x1 , . . . , xn ) are precisely the points in
the cell C of H that contains (v1′ , . . . , vn′ ).
(3) Solve the integer programming problem for Φ′ (x1 , . . . , xn ). That is, find a
(v1 , . . . , vn ) ∈ Zn such that Φ′ (v1 , . . . , vn ) evaluates to true.
(4) If (v1 , . . . , vn ) ∈ Zn exists, then write (v1 , . . . , vn ) to the output and accept.
5 Note
that we abuse notation here, as the variables in Φ range over Z and not Q.
34
(5) If (v1 , . . . , vn ) ∈ Zn does not exist, then overwrite Φ(x1 , . . . , xn ) with
Φ′′ (x1 , . . . , xn ) = Φ(x1 , . . . , xn ) ∧ ¬Φ′ (x1 , . . . , xn )
and go back to step (1).
We show that the algorithm is correct. Clearly, if the algorithm accepts, Φ
has a solution. Conversely, suppose that Φ has a solution. Hence, the algorithm
computes a value (v1′ , . . . , vn′ ) ∈ Qn in step (1) of its first iteration. It follows
from the following two observations that the algorithm accepts:
(i) If the algorithm computes (v1′ , . . . , vn′ ) ∈ Qn in step (1), and the cell C of
H containing (v1′ , . . . , vn′ ) also contains a point in Zn , then step (3) finds
a solution (v1 , . . . , vn ) ∈ Zn ; and,
(ii) If the algorithm computes (v1′ , . . . , vn′ ) ∈ Qn in step (1), and the cell C
of H containing (v1′ , . . . , vn′ ) does not contain a point in Zn , then step (3)
does not find a solution. By construction of Φ′′ in step (5), the solutions
to the formula Φ′′ are the solutions of Φ, minus the points in C. As C
did not contain a solution, we have that Φ has a solution if and only if
Φ′′ has a solution. Moreover, there exists no (v1′′ , . . . , vn′′ ) ∈ C such that
Φ′′ (v1′′ , . . . , vn′′ ) evaluates to true.
To show that the algorithm can be implemented to run in polynomial time,
we first argue that there are at most a polynomial number of iterations. This
follows from the observation in step (2) that the points satisfying Φ′ (x1 , . . . , xn )
are precisely all the points in a cell C of H. Indeed, when we do not find a
solution to the problem in step (3), we adapt Φ to exclude all the points in cell
C in step (5). Hence, in the following iteration, step (1) cannot find a solution
in cell C anymore. It follows that the number of iterations is bounded by the
number of cells in H, which is Θ(mn ) (see, e.g. [7], or Theorem 1.3 in [11] for a
more recent reference).
Finally, we argue that every step of the algorithm can be computed in ptime.
Step (1) can be solved by the quantifier elimination method of Ferrante
and Rackoff (Lemma .4). Proposition .5 states that we can find (v1′ , . . . , vn′ ) in
polynomial time.
Step (2) is easily to be seen to be in ptime: we only have to evaluate every
ϕ′i once on (v1′ , . . . , vn′ ).
Step (3) can be executed in ptime by Lenstra’s algorithm for integer linear
programming with a fixed number of variables [18].
Step (4) is in ptime (trivial).
Step (5) replaces Φ(x1 , . . . , xn ) by the formula Φ(x1 , . . . , xn )∧¬Φ′ (x1 , . . . , xn ).
As the size of Φ′ (x1 , . . . , xn ) is bounded by n plus the sum of the sizes of ϕ′′i for
i = 1, . . . , n, the formula Φ only grows by a linear term in each iteration. As
the number of iterations is bounded by a polynomial, the maximum size of Φ is
also bounded by a polynomial.
It follows that the algoritm is correct, and can be implemented to run in
polynomial time.
35
Corollary .7. There exists a ptime algorithm that decides whether a Boolean
combination of linear (in)equalities with rational factors and a fixed number of
variables has a solution of positive integers.
Proof. Given a Boolean combination Φ(x1 , . . . , xn ) of linear (in)equalities with
rational factors, we simply apply the algorithm of Proposition .6 to the formula
^
Φ′ (x1 , . . . , xn ) = Φ(x1 , . . . , xn ) ∧
xi ≥ 0.
1≤i≤n
In the following proposition, we treat the emptiness problem for DTDs: given
a DTD d, is L(d) = ∅? Note that L(d) can be empty even when d is not. For
instance, the trivial grammar a → a generates no finite trees.
Proposition .8. The emptiness problem is (1) ptime-complete for DTD(NFA)
and DTD(DFA), and (2) conp-complete for DTD(SL).
Proof. (1) The upper bound follows from a reduction to the emptiness problem
for NTA(NFA)s, which is in ptime (cf. Theorem 19(1) in [23])
For the lower bound, we reduce from path systems [10], which is known to
be ptime-complete. path systems is the decision problem defined as follows:
given a finite set of propositions P , a set A ⊆ P of axioms, a set R ⊆ P × P × P
of inference rules and some p ∈ P , is p provable from A using R? Here, (i) every
proposition in A is provable from A using R and, (ii) if (p1 , p2 , p3 ) ∈ R and if
p1 and p2 are provable from A using R, then p3 is also provable from A using
R.
In our reduction, we construct a DTD (d, p) such that (d, p) is not empty if
and only if p is provable. Concretely, for every (a, b, c) ∈ R, we add the string
ab to d(c); for every a ∈ A, d(a) = {ε}. Clearly, (d, p) satisfies the requirements.
(2) We provide an np algorithm to check whether a DTD(SL) (d, r) defines
a non-empty language. Intuitively, the algorithm computes the set S = {a ∈
Σ | L((d, a)) 6= ∅} in an iterative manner and accepts when r ∈ S.
Let k be the largest integer occurring in any SL-formula in d. Initially, S is
empty.
The iterative step is as follows. Guess a sequence of different symbols
b1 , . . . , bm in S. Then guess a vector (v1 , . . . , vm ) ∈ {0, . . . , k + 1}m , where
k is the largest integer occurring in any SL-formula in d. Intuitively, the vector
(v1 , . . . , vm ) represents the string bv11 · · · bvmm . From Lemma 4.2 it follows that any
SL-formula in d is satisfiable if and only if it is satisfiable by a string of the form
au1 1 · · · aunn , where Σ = {a1 , . . . , an }, and for all i = 1, . . . , n, ui ∈ {0, . . . , k + 1}.
Now add to S each a ∈ Σ for which bv11 · · · bvmm |= d(a). Note that this condition
can be checked in ptime. Repeat the iterative step at most |Σ| times and accept
when r ∈ S.
The conp-lowerbound follows from an easy reduction of non-satisfiability.
Let φ be a propositional formula with variables x1 , . . . , xn . Let Σ be the set
{a1 , . . . , an }. Let (d, r) be the DTD where d(r) = φ′ , where φ′ is the formula
36
φ in which every xi is replaced by a=1
i . Hence, (d, r) defines the empty tree
language if and only if φ is unsatisfiable.
Reducing a grammar is the act of finding an equivalent reduced grammar.
Corollary .9. Reducing a DTD(NFA) is ptime-complete; and reducing a DTD(SL)
is np-complete.
Proof. We first show the upper bounds. Let (d, s) be a DTD(NFA) or DTD(SL)
over alphabet Σ. In both cases, the algorithm performs the following steps for
each a ∈ Σ:
(i) Test whether a is reachable from s. That is, test whether there is a sequence
of Σ-symbols a1 , . . . , an such that
• a = s and an = a; and
• for every i = 2, . . . , n, there exists a string w1 ai w2 ∈ d(ai−1 ), for
w1 , w2 ∈ Σ∗ .
(ii) Test whether L((d, a)) 6= ∅.
Symbols that do not pass test (i) and (ii) are deleted from the alphabet of the
DTD. Let c be such a deleted symbol. In the case of SL, every atom c≥i and
c=i is replaced by true when i = 0 and false otherwise. Further, in the case of
NFAs, every transition mentioning c is removed.
In the case of a DTD(NFA), step (i) is in nlogspace and step (ii) is in
ptime. In the case of a DTD(SL), both tests (i) and (ii) are in np.
For the lower bound, we argue that
(1) if there exists an nlogspace-algorithm for reducing a DTD(NFA), then
emptiness of a DTD(NFA) is in nlogspace; and,
(2) if there exists a ptime-algorithm for reducing a DTD(SL), then emptiness
of a DTD(SL) is in ptime.
Statements (1) and (2) are easy to show: one only has to observe that an
emptiness test of a DTD can be obtained by reducing the DTD and verifying
whether the alphabet of the DTD still contains the start symbol.
References
[1] N. Alon, T. Milo, F. Neven, D. Suciu, and V. Vianu. Typechecking XML
views of relational databases. ACM Transactions on Computational Logic,
4(3):315–354, 2003.
[2] N. Alon, T. Milo, F. Neven, D. Suciu, and V. Vianu. XML with data
values: Typechecking revisited. Journal of Computer and System Sciences,
66(4):688–727, 2003.
37
[3] A. Balmin, Y. Papakonstantinou, and V. Vianu. Incremental validation of
XML documents. ACM Transactions on Database Systems, 29(4):710–751,
2004.
[4] V. Benzaken, A. Frisch, and G. Castagna. CDuce: an XML-centric generalpurpose language. In Proceedings of the 8th ACM SIGPLAN International
Conference on Functional Programming (ICFP 2003), pages 51–63. ACM
Press, 2003.
[5] G. J. Bex, S. Maneth, and F. Neven. A formal model for an expressive
fragment of XSLT. Information Systems, 27(1):21–39, 2002.
[6] A. Brüggemann-Klein, M. Murata, and D. Wood. Regular tree and regular
hedge languages over unranked alphabets: Version 1, april 3, 2001. Technical Report HKUST-TCSC-2001-0, The Hongkong University of Science
and Technology, 2001.
[7] R. Buck. Partition of space. American Mathematical Monthly, 50(9):541–
544, 1943.
[8] P. Buneman, M. Fernandez, and D. Suciu. UnQl: a query language and
algebra for semistructured data based on structural recursion. The VLDB
Journal, 9(1):76–110, 2000.
[9] B. S. Chlebus. Domino-tiling games. Journal of Computer and System
Sciences, 32(3):374–392, 1986.
[10] S. A. Cook. An observation on time-storage trade-off. Journal of Computer
and System Sciences, 9(3):308–316, 1974.
[11] H. Edelsbrunner. Algorithms in Combinatorial Geometry. Springer-Verlag,
1987.
[12] J. Engelfriet and S. Maneth. A comparison of pebble tree transducers with
macro tree transducers. Acta Informatica, 39:613–698, 2003.
[13] J. Ferrante and C. Rackoff. A decision procedure for the first order theory
of real addition with order. SIAM Journal on Computing, 4(1):69–76, 1975.
[14] H. Hosoya and B. C. Pierce. XDuce: A statically typed XML processing
language. ACM Transactions on Internet Technology (TOIT), 3(2):117–
148, 2003.
[15] D. S. Johnson. A catalog of complexity classes. In J. van Leeuwen, A. R.
Meyer, N. Nivat, M. S. Paterson, and D. Perrin, editors, Handbook of
Theoretical Computer Science, volume A, chapter 2, pages 67–161. NorthHolland, 1990.
[16] D. Kozen. Lower bounds for natural proof systems. In Proceedings 18th
Annual Symposium on Foundations of Computer Science (FOCS 1977),
pages 254–266. IEEE, 1977.
38
[17] D. Lee, M. Mani, and M. Murata. Reasoning about XML schema languages
using formal language theory. Technical report, IBM Almaden Research
Center, 2000. Log# 95071.
[18] H. W. Lenstra. Integer programming with a fixed number of variables.
Mathematics of Operations Research, 8:538–548, 1983.
[19] S. Maneth and F. Neven. Structured document transformations based on
XSL. In R. Connor and A. Mendelzon, editors, Research Issues in Structured and Semistructured Database Programming (DBPL 1999), volume
1949 of Lecture Notes in Computer Science, pages 79–96. Springer, 2000.
[20] S. Maneth, T. Perst, A. Berlea, and H. Seidl. XML type checking with
macro tree transducers. In Proceedings of the 24th Symposium on Principles
of Database Systems (PODS 2005), pages 283–294. ACM Press, 2005.
[21] W. Martens and F. Neven.
Frontiers of tractability for typechecking simple XML transformations.
Journal of Computer and
System Sciences.
To Appear. Preliminary version available at
http://alpha.uhasselt.be/wim.martens.
[22] W. Martens and F. Neven. Frontiers of tractability for typechecking simple
XML transformations. In Proceedings of the 23d Symposium on Principles
of Database Systems (PODS 2004), pages 23–34. ACM Press, 2004.
[23] W. Martens and F. Neven. On the complexity of typechecking top-down
XML transformations. Theoretical Computer Science, 336(1):153–180,
2005.
[24] T. Milo and D. Suciu. Type inference for queries on semistructured data.
In Proceedings of the Eighteenth Symposium on Principles of Database Systems (PODS 1999), pages 215–226. ACM Press, 1999.
[25] T. Milo, D. Suciu, and V. Vianu. Typechecking for XML transformers.
Journal of Computer and System Sciences, 66(1):66–97, 2003.
[26] F. Neven. Automata theory for XML researchers.
31(3):39–46, 2002.
SIGMOD Record,
[27] F. Neven and T. Schwentick. XML schemas without order. Unpublished
manuscript, 1999.
[28] Y. Papakonstantinou and V. Vianu. DTD inference for views of XML data.
In Proceedings of the 19th Symposium on Principles of Database Systems
(PODS 2000), pages 35–46, New York, 2000. ACM Press.
[29] H. Seidl. Haskell overloading is DEXPTIME-complete. Information Processing Letters, 52(2):57–60, 1994.
[30] C.M. Sperberg-McQueen and H. Thompson.
http://www.w3.org/XML/Schema, 2005.
39
XML
Schema.
[31] L.J. Stockmeyer and A.R. Meyer. Word problems requiring exponential
time: Preliminary report. In Conference Record of Fifth Annual ACM
Symposium on Theory of Computing (STOC 1973), pages 1–9, New York,
1973. ACM Press.
[32] D. Suciu. Typechecking for semistructured data. In Proceedings of the
8th Workshop on Data Bases and Programming Languages (DBPL 2001),
pages 1–20, Berlin, 2001. Springer.
[33] D. Suciu. The XML typechecking problem. SIGMOD Record, 31(1):89–96,
2002.
[34] A. Tozawa. Towards static type checking for XSLT. In Proceedings of
the ACM Symposium on Document Engineering (DOCENG 2001), pages
18–27, 2001.
40
| 2cs.AI
|
C++ programming language for an abstract
massively parallel SIMD architecture
arXiv:cs/0005023v1 [cs.PL] 19 May 2000
Alessandro Lonardo1, Emanuele Panizzi2,1, and Benedetto Proietti3
1
Istituto Nazionale di Fisica Nucleare (INFN), P.le Aldo Moro, 2, 00185 Roma, Italy
2
Università degli Studi dell’Aquila, Dipartimento di Ingegneria Elettrica,
Monteluco di Roio, 67100 L’Aquila, Italy
3
Università degli Studi “La Sapienza”, Dipartimento di Matematica
P.le Aldo Moro, 2, 00185 Roma, Italy
{alessandro.lonardo, emanuele.panizzi, benedetto.proietti}@roma1.infn.it
Abstract. The aim of this work is to define and implement an extended
C++ language to support the SIMD programming paradigm. The C++
programming language has been extended to express all the potentiality
of an abstract SIMD machine consisting of a central Control Processor and a N-dimensional toroidal array of Numeric Processors. Very few
extensions have been added to the standard C++ with the goal of minimising the effort for the programmer in learning a new language and
to keep very high the performance of the compiled code. The proposed
language has been implemented as a porting of the GNU C++ Compiler
on a SIMD supercomputer.
1
Introduction
The aim of this work is to define and implement an extended C++ [1] language
to support the SIMD [2] programming paradigm. Our goal is to add minimal
extensions to the standard C++ language [3] in order to minimise the syntactical differences when porting standard C++ applications or writing new codes.
Our decision to be always as close as possible to the standard lead to the definition of an extended C++ language with very few constructs to learn for C++
programmers, and relatively easy to use.
Using our language, the SIMD parallelism is efficiently achieved with traditional sequential programming plus a couple of new constructs (used to perform
memory mapped internode communication and to inhibit execution of code on
some processing nodes) and some knowledge of the native data types and their
allocation. The programmer can thus focus on the realization of the algorithm
and on the data distribution, which are the key points to exploit the parallel
architecture.
The proposed language has been implemented as a porting of the GNU C++
Compiler 1 [4] for the APEmille parallel supercomputer [5,6]. Some modifications
of the GNU C++ Compiler have been introduced, as well as the complete redefinition of the back-end for the target machine [4]. APEmille is a parallel SIMD
1
Release 2.95.1
computer developed at INFN (Italian National Institute for Nuclear Physics)
capable of peak performance of 1 Teraflop in a configuration with 2048 processing
nodes.
The simplicity and low number of extensions to the standard language helped
reaching the goal of efficiency of the executable parallel codes, main goal for any
number crunching application running on a massively parallel supercomputer.
In this paper, we describe the proposed SIMD C++ language, and especially
those aspects which extend the standard C++ syntax or semantics. Section 3 is
devoted to this description. Section 2 explains the abstract SIMD architecture
which we refer to, while section 4 reports on those works related to our either
for the language used (extensions of C/C++) or for a similar target architecture
or parallel paradigm. Section 5 contains the conclusions.
2
The Abstract SIMD machine
SIMD machines consists of synchronized processing elements with an associated
unique control processor. The Control Processor (later, CP) broadcasts the same
instruction stream to all processing elements. All processing elements execute the
same instruction at each clock cycle on their own data. In the proposed architecture the processing elements are specialized processors for numeric applications:
we will call them Numeric Processors or NPs. The NPs form an N-dimensional
toroidal array. Each NP consists of an ALU, an own register file, a local memory and a local memory mass storage. There is no shared memory across the
whole machine: communication among NPs is achieved through a memory mapping mechanism that allows each NP to access memory of its neighbours. The
machine is SIMD and guarantees no conflicts in memory accesses.
2.1
Control Processor
The CP handles integer data types, executes branches, function calls, and generates memory address. Every instruction is sent by the CP to each NP in the
machine, at the same clock cycle.
Global addresses are broadcasted to all NPs. They will use them to address
their own memory.
2.2
Numeric Processor
NPs are specialized in numeric instructions, in fact they natively support floating
point data types – scalar (both single and double precision), vector/complex
(couple of single precision) and the integer data type.
NPs, in order to perform conditional execution, can test local conditions,
and, when they are not met, can disable the effect of the following numeric
instructions. We call the conditional test a where instruction.
CP can also make all NPs test their local conditions. Then it can perform a
global branch if any, all or none of the NPs has met the condition.
NPs can address their own memory using the global address generated by
the CP and, eventually, adding a local offset.
3
Proposed SIMD C++ language
In this section we describe our extended C++ analyzing all the aspects more related to parallelism. Subsection 3.9 contains a resume of the main characteristics
and discusses general topics.
3.1
Types, Declarations and Allocation
Basic Data Types With basic data types or basic types we refer to the types
natively supported by the language, as int or float. The C++ language proposed in this paper includes all the basic data types supported by the illustrated
abstract SIMD machine. Namely:
int, float, double, complex, vector, localint
The localint data type are integer variables allocated in the numeric processors
(see later), while the types vector and complex represent a pair of float and are
treated as native types by our abstract machine. Pointers, arrays and function
pointers are supported for every type and every level of indirection.
These types are all signed. All the ”standard” C++ types (e.g.: long long,
long double, char, etc. and all unsigned types) could be supported, performing
software emulation for those not supported by the physical machine.
Declaring variables is absolutely identical to standard C++. It is also possible to declare new types with the typedef keyword just as in C++, with the
standard rules and no limitation.
Allocation The variables declared are allocated in the Control Processor (CP)
or in the Numeric Processor (NP) depending on their type. We divide the basic types into two groups: the Control Processor types (int, pointers) and the
Numeric Processor types (float, double, vector, complex, localint).
CP variables are allocated in the (unique) Control Processor, so there is just
one instance of them; NP variables are allocated in each NP data memory, and,
most important, at the same location in every one: so there are multiple instances
of numeric variables. This is the essence of the SIMD programming paradigm,
and implies a very important fact: memory images of NPs are all identical; each
allocation, both static and dynamic, is the same for every NP.
Arrays of any type are allocated in the same processor of the base type. For
example, the array
double a[100000];
is allocated in the NPs’ memories. However the base of the array, which is known
at compile time, is a pointer, and so it is handled by the CP.
The allocation mechanism is automatic and controlled by the compiler: there
is no way and no reason for the programmer to alter it. On the other hand data
distribution is left to the programmer. We will discuss this topic in subsection
3.8.
3.2
Expressions
Expressions within the same type Handling expressions among variables
of the same type is not ambiguous, because they are allocated in the same kind
of processor so code for ”that” processor will be generated to handle them. For
example:
CP allocation
int i,j,k;
double a,b,c;
NP allocation
CP code
i = j+1;
a = 1.0;
NP code
b = a*c-b;
NP code
CP code
k++;
Promotions >> int CP ptr NP ptr float double vector complex localint
int
CP pointer
NP pointer
float
double
vector
complex
localint
yes
yes
yes
no
no
no
no
no
yes
yes
yes
no
no
no
no
no
yes
yes
yes
no
no
no
no
yes
yes
no
no
yes
yes
no
no
yes
yes
no
no
yes
yes
no
no
yes
yes
no
no
yes
yes
yes
no
yes
yes
no
no
yes
yes
no
yes
yes
yes
no
no
yes
yes
no
no
yes
Table 1. Allowed Promotions
Mixed-types expressions They are handled by promoting types, or by explicit
cast by the programmer. There are specific rules about cast and promotions.
– cast/promotions from CP to NP types are ALWAYS allowed.
– cast/promotions from NP to CP types are NEVER allowed.
– cast/promotions between two types of the same group are allowed depending
on the specific types.
Casts >>
int
CP pointer
NP pointer
float
double
vector
complex
localint
int CP ptr NP ptr float double vector complex localint
yes no
no
yes
yes
yes
yes
yes
no yes
no
no
no
no
no
no
no yes
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
yes
Table 2. Allowed Casts
It is obvious that a cast/promotion from a CP type to NP generates multiple
instances of one value.
3.3
Multiple Addressing
As stated before, the abstract SIMD machine includes the ability to add a local
offset when accessing local NP memory, so every NP could access a different location in memory. This is realized with the localint variables, that are integers
allocated in the NPs. These values can be used to add a local displacement when
accessing local memory.
A pseudo function localoffset() can be called with a localint argument
to set the local offset for the following memory access.
int i;
localint li;
float r, a[100];
// ...
localoffset(li);
r = a[i];
In the example above, access in array a is at index i + li.
3.4
Type Constructors: Structs, Classes and Unions
It is possible to declare a new type using struct, class or union as in standard
C++. Structs and classes can contain data fields of any other data type (both
CP types and NP types), while unions must contain fields associated to the same
kind of processor (only CP or only NP), due to allocation reasons, as explained
before.
class Mixed {
int a;
float x;
public:
Mixed (int aa, float xx) : a(aa), x(xx) {};
};
Each field is allocated in the respective processor so that multiple instances
of numeric field exist. The effort to address them and to keep pointers consistent is made by the compiler. Fields must be accessed directly with pointers:
incrementing and decrementing pointers to ”navigate” through a struct or class
could generate unpredictable results because the object is allocated on different
memories. The space allocated is compacted, so that only the necessary size is
allocated in each kind of processor.
3.5
Object Oriented Features:
Encapsulation, Inheritance, Polymorphism
Encapsulation is handled as in standard C++ with no other extension nor limitation. Field allocation follows what stated in 3.4. Methods are called passing
them the invocation object as an hidden argument.The same method is executed
by each NP.
Also Inheritance and Polymorphism have no extensions nor limitations. Nonvirtual base class members are inserted in the CP or NP instance layout of the
object after their type class. Virtual base class members and virtual classes
information are inserted into the CP instance of the object: in fact they are
pointers.
3.6
Communication
Communication among Numeric Processors is achieved through memory mapping. The proposed C++ language allows to address an array element in a remote
NP by summing a constant to the array index or to the pointer that would be
used for local access. Different pre-defined constants are associated to neighbour
NPs. These constants specify the relative position of the NP to be accessed with
respect to the current NP. The constants are generated and handled on the CP,
so they are the same for all NPs. The following example shows communication
between NPs:
float r, v[100000];
r = v[3+XPLUS_NP]; // each NP accesses the 3rd element of
// the nearest neighbour on the x axis
// in the positive direction
The constant XPLUS_NP is machine dependent.
It is possible to use a remote object as parameter or invocation object of a
method. In this way, a code like:
class C
{public:
float x;
void f(float y) { x = y; }
};
int main()
{ float a;
C v[10];
// ...
v[0+XPLUS_NP].f(a);
// ...
}
assigns to the x field of the v[0] object on each node the value of a on the
adjacent (XMINUS_NP) node.
3.7
Local Conditions
The instruction flow being unique, it is possible to branch only when global
conditions (conditions on the CP) are met. Conditions on local variables (on
NPs) can be handled with the where-elsewhere keywords2. Conditioned code
will be executed only by those NPs that met the condition. All the other NPs
will execute NOPs. CP instructions inside a where block, on the other hand, are
always executed. where-elsewhere are used just like if-else, as shown in the
following example.
int i;
double x,y;
// ...
where (x != 0.0)
{ y = 1/x;
}
elsewhere
{ y = 0;
}
3.8
Examples of implementation of SIMD programs
When writing a program for a SIMD machine using the proposed C++ extensions, the “SI” part of the SIMD paradigm is realized using a single instruction
stream, as the C++ language naturally does; the “MD” part of SIMD is achieved
allocating multiple instances of the numeric variables.
2
In our implementation, these are not keywords but function names
The initial loading of different data in each Numeric Processor data memory
is made by the operating system, while the slicing of a big array into the NPs
must be determined by the programmer.
For example, suppose that the problem needs to handle an array of 10000 ×
10000 elements, and that the target machine has a 2D square topology with
10 × 10 = 100 Numeric Processors. The programmer will declare a 1000 × 1000
array (and every NP will have its own instance of this array, that represents a
slice of the big array). A very trivial example of code that implements the sum
of two 10000 × 10000 elements arrays can be useful to explain the parallelization
mechanism and the data distribution:
int main()
{
const int dimx = 1000;
const int dimy = 1000;
const int size_per_node = dimx*dimy;
float m1[dimx][dimy], m2[dimx][dimy], m3[dimx][dimy];
const char *filename = "myfile.data";
// ...
distributed_load(m1, filename, size_per_node);
distributed_load(m2, filename, size_per_node);
for (int i=0; i<dimx; i++)
for (int j=0; j<dimy; j++)
m3[i][j] = m1[i][j] + m2[i][j];
distributed_store(m3, filename, size_per_node);
// ...
}
Numeric instructions will be executed in parallel by the Numeric Processors
on their local data, while Control Processor will execute flow control, integer instructions and will generate memory addresses. The distributed_load() and
distributed_store() functions perform machine dependent system calls supposed to load and store data in the appropriate way.
3.9
General remarks on the language proposed
The proposed C++ is very similar to the standard C++, is easy to learn, to use
and to debug, produces highly efficient executable codes, and can be used in a
professional environment.
The most important aspect of the proposed language is that it is a minimal
extension to the C++ standard. This is a key feature as we want the programmer
to concentrate on the application development rather than paying attention to
implementation aspects.
Our language strictly conforms to the machine architectural characteristics
in order to fully exploit the simplification that the SIMD synchronous structure
and the memory mapped internode communications introduce in the task of
writing the parallel algorithm.
As a result, there is no need to develop multi-threaded programs nor to use
any special communication library.
Finally, object distribution is obtained by the simple allocation model described above. All objects are replicated on each processing element and invocation of a method is executed on all NPs (or on the subset of them satisfying an
eventual WHERE condition).
4
Related work
In this section we compare our language to a couple of parallel C/C++ extensions and to High Performance Fortran, which is the standard for data parallel
applications.
HPC++ [7,8] is a set of class libraries and tools that extend the C++ language. It also has a set of runtime systems that are required for remote access. It
refers to a very general architectural model, so it can be used on a variety of machine architectures. There are two main execution modes for programs written
in HPC++:
1. multi-thread shared memory mode, suitable for coarse-grained applications
with some particular collective operations for thread synchronization.
2. Single Program Multiple Data (SPMD) mode, in which n copies of the same
program run on the n processing nodes. This mode is similar to using C/C++
with MPI or PVM. The programmer must manage the data distribution and
the synchronization of processes.
HPC++ has thus a totally different approach compared to our one, and this
approach is not applicable to our language architecture which is focused on
single threaded programs.
pC++ [9] is a C++ extension that provides a thread-based programming
model and a simple way to encapsulate SPMD code in it, together with a mechanism for data distribution similar to the one adopted in HPF (see below). The
key concept of this extension is the collection of objects. It is possible to invoke
a method on an entire collection or on a part of it. The compilation of pC++
code is achieved as a translation into standard C++ by a preprocessor.
The HPF [10] programming model has the following key points:
1. single threaded control;
2. global namespace, low-level of data distribution and remote communication
details hidden to the programmer;
3. loosely synchronous: synchronization of program execution on different nodes
is accomplished only at special points (e.g. the completion of a loop) and
not instruction by instruction;
4. parallel operations: operations on array elements executed at the same time
over all nodes.
HPF extends the Fortran language adding compiler directives, libraries and
new language constructs.
The most relevant compiler directives to our purposes are those related to
data distribution. This is accomplished in three steps: in the first step an alignment is defined for arrays, in the second step aligned data are mapped on a
abstract set of processors and finally this set is mapped onto physical processors. This is quite different from our approach as we force the programmer to
take care of allocation of large matrices as described above. In particular we
rely on operating system calls to perform something similar to the BLOCK and
DEGENERATE distribution types.
A similarity between HPF and our language is the specification of locally conditioned code execution via the WHERE statement, although the HPF version
accepts logical-array arguments while ours accepts any logical condition between
NP data. Other parallel constructs, such as FORALL, are not provided by our
language.
HPF is the parallel extension to a standard language that, compared with
the other two, best matches with our approach.
5
Conclusions
The next topics to be analyzed will include exception handling, RTTI (Run Time
Type Information) and namespaces.
Our implementation of the compiler based on a porting of the GNU CC
compiler on the APEmille architecture is currently under test. We plan to discuss
our implementation in a further paper.
References
1. Stroustrup, B.: The C++ Programming Language, third ed. Addison-Wesley 1997
2. Hwang, K., Briggs, F.A.: Computer Architecture and Parallel Processing. McGrawHill 1985
3. Koenig, A.E. 1996. Working paper for Draft Proposed International Standard for Information Systems - Programming Language C++. Document Number: X3J16/960225 WG21/N1043. URL: http://www.maths.warwick.ac.uk/c++/pub/wp/
4. Stallman, R.: Using and Porting GNU CC, for Version 2.95 Free Software Foundation ISBN: 1-882114-38-8
5. Aglietti, F. et al.: The teraflop supercomputer APEmille: architecture, software and
project status report. Computer Physics Communications 110 (1998) 216–219
6. Tripiccione, R.: APEmille. Parallel Computing 25 (1999) 1297-1309
7. The HPC++ working group. HPC++ white papers. Technical Report 95633, Center
for Research on Parallel Computation, Dec. 1995
8. http://www.extreme.indiana.edu/hpc++
9. http://www.extreme.indiana.edu/sage
10. Koelbel, C.H. et al.: The High Performance Fortran HandBook. The MIT Press,
1991
11. High
Performance
Fortran
Forum
Home
Page:
http://www.crpc.rice.edu/HPFF/home.html
| 6cs.PL
|
arXiv:1509.05205v2 [math.GT] 10 Jul 2016
A FINITE GENERATING SET FOR THE LEVEL 2 TWIST
SUBGROUP OF THE MAPPING CLASS GROUP OF A CLOSED
NON-ORIENTABLE SURFACE
RYOMA KOBAYASHI AND GENKI OMORI
Abstract. We obtain a finite generating set for the level 2 twist subgroup of
the mapping class group of a closed non-orientable surface. The generating
set consists of crosscap pushing maps along non-separating two-sided simple
loops and squares of Dehn twists along non-separating two-sided simple closed
curves. We also prove that the level 2 twist subgroup is normally generated
in the mapping class group by a crosscap pushing map along a non-separating
two-sided simple loop for genus g ≥ 5 and g = 3. As an application, we
calculate the first homology group of the level 2 twist subgroup for genus
g ≥ 5 and g = 3.
1. Introduction
Let Ng,n be a compact connected non-orientable surface of genus g ≥ 1 with
n ≥ 0 boundary components. The surface Ng = Ng,0 is a connected sum of g
real projective planes. The mapping class group M(Ng,n ) of Ng,n is the group
of isotopy classes of self-diffeomorphisms on Ng,n fixing the boundary pointwise
and the twist subgroup T (Ng,n ) of M(Ng,n ) is the subgroup of M(Ng,n ) generated by all Dehn twists along two-sided simple closed curves. Lickorish [7] proved
that T (Ng ) is an index 2 subgroup of M(Ng ) and the non-trivial element of
M(Ng )/T (Ng ) ∼
= Z/2Z =: Z2 is represented by a “Y-homeomorphism”. We define
a Y-homeomorphism in Section 2. Chillingworth [1] gave an explicit finite generating set for T (Ng ) and showed that T (N2 ) ∼
= Z2 . The first homology group H1 (G)
of a group G is isomorphic to the abelianization Gab of G. The group H1 (T (Ng ))
is trivial for g ≥ 7, H1 (T (N3 )) ∼
= Z12 , H1 (T (N4 )) ∼
= Z2 ⊕ Z and H1 (T (Ng )) ∼
= Z2
for g = 5, 6. These results were shown by Korkmaz [5] for g ≥ 7 and by Stukow [10]
for the other cases.
Let Σg,n be a compact connected orientable surface of genus g ≥ 0 with n ≥ 0
boundary components. The mapping class group M(Σg,n ) of Σg,n is the group
of isotopy classes of orientation preserving self-diffeomorphisms on Σg,n fixing the
boundary pointwise. Let S be either Ng,n or Σg,n . For n = 0 or 1, we denote
by Γ2 (S) the subgroup of M(S) which consists of elements acting trivially on
H1 (S; Z2 ). Γ2 (S) is called the level 2 mapping class group of S. For a group G,
a normal subgroup H of G and a subset X of H, H is normally generated in G
by X if H is the normal closure of X in G. In particular, for X = {x1 , . . . , xn },
if H is the normal closure of X in G, we also say that H is normally generated
in G by x1 , . . . , xn . In the case of orientable surfaces, Humphries [3] proved that
Date: November 13, 2017.
2010 Mathematics Subject Classification. 57M05, 57M07, 57M20, 57M60.
1
2
R. KOBAYASHI AND G. OMORI
Γ2 (Σg,n ) is normally generated in M(Σg,n ) by the square of the Dehn twist along
a non-separating simple closed curve for g ≥ 1 and n = 0 or 1. In the case of nonorientable surfaces, Szepietowski [11] proved that Γ2 (Ng ) is normally generated in
M(Ng ) by a Y-homeomorphism for g ≥ 2. Szepietowski [12] also gave an explicit
finite generating set for Γ2 (Ng ). This generating set is minimal for g = 3, 4. Hirose
and Sato [2] gave a minimal generating set for Γ2 (Ng ) when g ≥ 5 and showed that
g
g
∼ Z(3)+(2) .
H1 (Γ2 (Ng )) =
2
We denote by T2 (Ng ) the subgroup of T (Ng ) which consists of elements acting
trivially on H1 (Ng ; Z2 ) and we call T2 (Ng ) the level 2 twist subgroup of M(Ng ).
Recall that T (N2 ) ∼
= Z2 and Chillingworth [1] proved that T (N2 ) is generated by
the Dehn twist along a non-separating two-sided simple closed curve. T2 (N2 ) is
a trivial group because Dehn twists along non-separating two-sided simple closed
curves induce nontrivial actions on H1 (Ng ; Z2 ). Let Aut(H1 (Ng ; Z2 ), ·) be the group
of automorphisms on H1 (Ng ; Z2 ) preserving the intersection form · on H1 (Ng ; Z2 ).
Since the action of M(Ng ) on H1 (Ng ; Z2 ) preserves the intersection form ·, there
is the natural homomorphism from M(Ng ) to Aut(H1 (Ng ; Z2 ), ·). McCarthy and
Pinkall [8] showed that the restriction of the homomorphism to T (Ng ) is surjective.
Thus T2 (Ng ) is finitely generated.
In this paper, we give an explicit finite generating set for T2 (Ng ) (Theorem 3.1).
The generating set consists of “crosscap pushing maps” along non-separating twosided simple loops and squares of Dehn twists along non-separating two-sided simple
closed curves. We review the crosscap pushing map in Section 2. We can see the
generating set for T2 (Ng ) in Theorem 3.1 is minimal for g = 3 by Theorem 1.2. We
prove Theorem 3.1 in Section 3. In the last part of Subsection 3.2, we also give the
smaller finite generating set for T2 (Ng ) (Theorem 3.14). However, the generating
set consists of crosscap pushing maps along non-separating two-sided simple loops,
squares of Dehn twists along non-separating two-sided simple closed curves and
squares of Y-homeomorphisms.
By using the finite generating set for T2 (Ng ) in Theorem 3.1, we prove the
following theorem in Section 4.
Theorem 1.1. For g = 3 and g ≥ 5, T2 (Ng ) is normally generated in M(Ng ) by a
crosscap pushing map along a non-separating two-sided simple loop (See Figure 1).
T2 (N4 ) is normally generated in M(N4 ) by a crosscap pushing map along a
non-separating two-sided simple loop and the square of the Dehn twist along a nonseparating two-sided simple closed curve whose complement is a connected orientable
surface (See Figure 2).
The x-marks as in Figure 1 and Figure 2 mean Möbius bands attached to boundary components in this paper and we call the Möbius band the crosscap. The group
which is normally generated in M(Ng ) by the square of the Dehn twist along a
non-separating two-sided simple closed curve is a subgroup of T2 (Ng ) clearly. The
authors do not know whether T2 (Ng ) is generated by squares of Dehn twists along
non-separating two-sided simple closed curves or not.
As an application of Theorem 1.1, we calculate H1 (T2 (Ng )) for g ≥ 5 in Section 5
and we obtain the following theorem.
GENERATING LEVEL 2 TWIST SUBGROUP
3
Figure 1. A crosscap pushing map along a non-separating twosided simple loop is described by a product of Dehn twists along
non-separating two-sided simple closed curves as in the figure.
Figure 2. A non-separating two-sided simple closed curve on N4
whose complement is a connected orientable surface.
Theorem 1.2. For g = 3 and g ≥ 5, the first homology group of T2 (Ng ) is as
follows:
( 2
Z ⊕ Z2
if g = 3,
∼
H1 (T2 (Ng )) =
(g3)+(g2)−1
if g ≥ 5.
Z2
In this proof, we use the five term exact sequence for an extension of a group for
g ≥ 5. The authors do not know the first homology group of T2 (N4 ).
2. Preliminaries
2.1. Crosscap pushing map. Let S be a compact surface and let e : D′ ֒→ intS
be a smooth embedding of the unit disk D′ ⊂ C. Put D := e(D′ ). Let S ′ be the
surface obtained from S − intD by the identification of antipodal points of ∂D. We
call the manipulation that gives S ′ from S the blowup of S on D. Note that the
image M of the regular neighborhood of ∂D in S − intD by the blowup of S on
D is a crosscap, where a crosscap is a Möbius band in the interior of a surface.
Conversely, the blowdown of S ′ on M is the following manipulation that gives S
from S ′ . We paste a disk on the boundary obtained by cutting S along the center
line µ of M . The blowdown of S ′ on M is the inverse manipulation of the blowup
of S on D.
Let x0 be a point of Ng−1 and let e : D′ ֒→ Ng−1 be a smooth embedding of a
unit disk D′ ⊂ C to Ng−1 such that the interior of D := e(D′ ) contains x0 . Let
M(Ng−1 , x0 ) be the group of isotopy classes of self-diffeomorphisms on Ng−1 fixing
the point x0 , where isotopies also fix x0 . Then we have the blowup homomorphism
ϕ : M(Ng−1 , x0 ) → M(Ng )
4
R. KOBAYASHI AND G. OMORI
that is defined as follows. For h ∈ M(Ng−1 , x0 ), we take a representative h′ of h
which satisfies either of the following conditions: (a) h′ |D is the identity map on D,
(b) h′ (x) = e(e−1 (x)) for x ∈ D. Such h′ is compatible with the blowup of Ng−1
on D, thus ϕ(h) ∈ M(Ng ) is induced and well defined (c.f. [11, Subsection 2.3]).
The point pushing map
j : π1 (Ng−1 , x0 ) → M(Ng−1 , x0 )
is a homomorphism that is defined as follows. For γ ∈ π1 (Ng−1 , x0 ), j(γ) ∈
M(Ng−1 , x0 ) is described as the result of pushing the point x0 once along γ. Note
that for x, y ∈ π1 (Ng−1 ), yx means yx(t) = x(2t) for 0 ≤ t ≤ 21 and yx(t) = y(2t−1)
for 12 ≤ t ≤ 1, and for elements [f ], [g] of the mapping class group, [f ][g] means
[f ◦ g].
We define the crosscap pushing map as the composition of homomorphisms:
ψ := ϕ ◦ j : π1 (Ng−1 , x0 ) → M(Ng ).
For γ ∈ π1 (Ng−1 , x0 ), we also call ψ(γ) the crosscap pushing map along γ. Remark
that for γ, γ ′ ∈ π1 (Ng−1 , x0 ), ψ(γ)ψ(γ ′ ) = ψ(γγ ′ ). The next two lemmas follow
from the description of the point pushing map (See [6, Lemma 2.2, Lemma 2.3]).
Lemma 2.1. For a two-sided simple loop γ on Ng−1 based at x0 , suppose that γ1 ,
γ2 are two-sided simple closed curves on Ng−1 such that γ1 ⊔ γ2 is the boundary
of the regular neighborhood N of γ in Ng−1 whose interior contains D. Then for
some orientation of N , we have
t−1
,
ψ(γ) = ϕ(tγ1 t−1
f
γ 2 ) = tγ
1 γ
f
2
where γe1 , γe2 are images of γ1 , γ2 to Ng by blowups respectively (See Figure 3).
Let µ be a one-sided simple closed curve and let α be a two-sided simple closed
curve on Ng such that µ and α intersect transversely at one point. For these simple
closed curves µ and α, we denote by Yµ,α a self-diffeomorphism on Ng which is
described as the result of pushing the regular neighborhood of µ once along α.
We call Yµ,α a Y-homeomorphism (or crosscap slide). By Lemma 3.6 in [11], Yhomeomorphisms are in Γ2 (Ng ).
Lemma 2.2. Suppose that γ is a one-sided simple loop on Ng−1 based at x0 such
that γ and ∂D intersect at antipodal points of ∂D. Then we have
ψ(γ) = Yµ,eγ ,
where γ
e is a image of γ to Ng by a blowup and µ is a center line of the crosscap
obtained from the regular neighborhood of ∂D in Ng−1 by the blowup of Ng−1 on D
(See Figure 4).
Remark that the image of a crosscap pushing map is contained in Γ2 (Ng ). By
Lemma 2.1, if γ is a two-sided simple loop on Ng , then ψ(γ) is an element of T2 (Ng ).
We remark that Y-homeomorphisms are not in T (Ng ) (See [7]).
2.2. Notation of the surface Ng . Let ei : Di′ ֒→ Σ0 for i = 1, 2, . . . , g be smooth
embeddings of unit disks Di′ ⊂ C to a 2-sphere Σ0 such that Di := ei (Di′ ) and
Dj are disjoint for distinct 1 ≤ i, j ≤ g, and let xi ∈ Σ0 for i = 1, 2, . . . , g be g
points of Σ0 such that xi is contained in the interior of Di as the left-hand side
of Figure 5. Then Ng is diffeomorphic to the surface obtained from Σ0 by the
blowups on D1 , . . . , Dg . We describe the identification of ∂Di by the x-mark as
GENERATING LEVEL 2 TWIST SUBGROUP
5
Figure 3. A crosscap pushing map along two-sided simple loop γ.
Figure 4. A crosscap pushing map along one-sided simple loop γ
(Y-homeomorphism Yµ,eγ ).
the right-hand side of Figure 5. We call the crosscap which is obtained from the
regular neighborhood of ∂Di in Σ0 by the blowup of Σ0 on Di the i-th crosscap.
(k)
We denote by Ng−1 the surface obtained from Σ0 by the blowups on Di for every
(k)
i 6= k. Ng−1 is diffeomorphic to Ng−1 . Let xk;i be a simple loop on Ng based at
(k)
(k)
xk for i 6= k as Figure 6. Then the fundamental group π1 (Ng−1 ) = π1 (Ng−1 , xk ) of
(k)
Ng−1 has the following presentation.
(k)
π1 (Ng−1 ) = xk;1 , . . . , xk;k−1 , xk;k+1 , . . . , xk;g | x2k;1 . . . x2k;k−1 x2k;k+1 . . . x2k;g = 1 .
Figure 5. The embedded disks D1 , D2 , . . . , Dg on Σ0 and the
surface Ng .
(k)
2.3. Notations of mapping classes. Let ψk : π1 (Ng−1 ) → M(Ng ) be the cross(k)
(k)
cap pushing map obtained from the blowup of Ng−1 on Dk and let π1 (Ng−1 )+ be
6
R. KOBAYASHI AND G. OMORI
Figure 6. The simple loop xk;i for 1 ≤ i ≤ k − 1 and xk;j for
(k)
k + 1 ≤ j ≤ g on Ng−1 based at xk .
(k)
(k)
the subgroup of π1 (Ng−1 ) generated by two-sided simple loops on Ng−1 based at
(k)
xk . By Lemma 2.1, we have ψk (π1 (Ng−1 )+ ) ⊂ T2 (Ng ). We define non-separating
(k)
two-sided simple loops αk;i,j and βk;i,j on Ng−1 based at xk as in Figure 7 for distinct 1 ≤ i < j ≤ g and 1 ≤ k ≤ g. We also define αk;j,i := αk;i,j and βk;j,i := βk;i,j
for distinct 1 ≤ i < j ≤ g and 1 ≤ k ≤ g. We have the following equations:
αk;i,j = xk;i xk;j
for i < j < k or j < k < i or k < i < j,
βk;i,j = xk;j xk;i
for i < j < k or j < k < i or k < i < j.
Denote the crosscap pushing maps ak;i,j := ψk (αk;i,j ) and bk;i,j := ψk (βk;i,j ). Remark that ak;i,j and bk;i,j are contained in the image of ψk |π1 (N (k) )+ . Let η be the
g−1
self-diffeomorphism on Ng which is the rotation of Ng such that η sends the i-th
crosscap to the (i + 1)-st crosscap for 1 ≤ i ≤ g − 1 and the g-th crosscap to the
1-st crosscap as Figure 8. Then we have ak;i,j = η k−1 a1;i−k+1,j−k+1 η −(k−1) and
bk;i,j = η k−1 b1;i−k+1,j−k+1 η −(k−1) for each distinct 1 ≤ i, j, k ≤ g.
(k)
Figure 7. Two-sided simple loops αk;i,j and βk;i,j on Ng−1 based
at xk .
For distinct i1 , i2 , . . . , in ∈ {1, 2, . . . , g}, we define a simple closed curve αi1 ,i2 ,...,in
on Ng as in Figure 9. The arrow on the side of the simple closed curve αi1 ,i2 ,...,in
GENERATING LEVEL 2 TWIST SUBGROUP
7
Figure 8. The self-diffeomorphism η on Ng .
in Figure 9 indicates the direction of the Dehn twist tαi1 ,i2 ,...,in along αi1 ,i2 ,...,in if
n is even. We set the notations of Dehn twists and Y-homeomorphisms as follows:
for 1 ≤ i < j ≤ g,
Ti,j
:=
tαi,j
Ti,j,k,l
:=
tαi,j,k,l
Yi,j
:=
Yαi ,αi,j = ψi (xi;j )
for g ≥ 4 and 1 ≤ i < j < k < l ≤ g,
for distinct 1 ≤ i, j ≤ g.
2
2
Note that Ti,j
and Ti,j,k,l
are elements of T2 (Ng ), Yi,j is an element of Γ2 (Ng ) but
2
Yi,j is not an element of T2 (Ng ). We remark that ak;i,j = b−1
k;i,j = Ti,j for any
distinct i, j, k ∈ {1, 2, 3} when g = 3.
Figure 9. The simple closed curve αi1 ,i2 ,...,in on Ng .
3. Finite generating set for T2 (Ng )
In this section, we prove the main theorem in this paper. The main theorem is
as follows:
Theorem 3.1. For g ≥ 3, T2 (Ng ) is generated by the following elements:
(i) ak;i,i+1 , bk;i,i+1 , ak;k−1,k+1 , bk;k−1,k+1 for 1 ≤ k ≤ g, 1 ≤ i ≤ g and
i 6= k − 1, k,
(ii) a1;2,4 , bk;1,4 , al;1,3 for k = 2, 3 and 4 ≤ l ≤ g when g is odd,
2
(iii) T1,j,k,l
for 2 ≤ j < k < l ≤ g when g ≥ 4,
where the indices are considered modulo g.
We remark that the number of generators in Theorem 3.1 is 16 (g 3 + 6g 2 + 5g − 6)
for g ≥ 4 odd, 16 (g 3 + 6g 2 − g − 6) for g ≥ 4 even and 3 for g = 3.
8
R. KOBAYASHI AND G. OMORI
(k)
3.1. Finite generating set for π1 (Ng−1 )+ . First, we have the following lemma:
(k)
(k)
Lemma 3.2. For g ≥ 2, π1 (Ng−1 )+ is an index 2 subgroup of π1 (Ng−1 ).
(k)
Proof. Note that π1 (Ng−1 ) is generated by xk;1 , . . . , xk;k−1 , xk;k+1 , . . . , xk;g . If g =
(k)
2, π1 (Ng−1 ) is isomorphic to Z2 which is generated by a one-sided simple loop.
(k)
Hence π1 (Ng−1 )+ is trivial and we obtain this lemma when g = 2.
We assume that g ≥ 3. For i 6= k, we have
xk;i = x−1
k;k−1 · xk;k−1 xk;i .
(k)
Since xk;k−1 xk;i = βk;i,k−1 ∈ π1 (Ng−1 )+ , the equivalence classes of xk;i and x−1
k;k−1
(k)
(k)
in π1 (Ng−1 )/π1 (Ng−1 )+ are the same. We also have
2
xk;k−1 = x−1
k;k−1 · xk;k−1 .
(k)
Since x2k;k−1 ∈ π1 (Ng−1 )+ , the equivalence classes of xk;k−1 and x−1
k;k−1 in
(k)
(k)
(k)
(k)
π1 (Ng−1 )/π1 (Ng−1 )+ are the same. Thus π1 (Ng−1 )/π1 (Ng−1 )+ is generated by
the equivalence class [xk;k−1 ] whose order is 2 and we have completed the proof of
Lemma 3.2.
(k)
Ng−1 is diffeomorphic to the surface on the left-hand side (resp. right-hand
side) of Figure 10 when g − 1 = 2h + 1 (resp. g − 1 = 2h + 2). We take a
diffeomorphism which sends xk;i for i 6= k and xk as in Figure 6 to xk;i for i 6= k
(k)
and xk as in Figure 10 and identify Ng−1 with the surface in Figure 10 by the
]
(k)
(k)
diffeomorphism. Denote by pk : Ng−1 ։ Ng−1 the orientation double covering
]
(k)
(k)
of Ng−1 as in Figure 11. Then Hk := (pk )∗ (π1 (Ng−1 )) is an index 2 subgroup
]
(k)
(k)
of π1 (Ng−1 ). Note that when g − 1 = 2h + 1, π1 (Ng−1 ) is generated by yk;i for
]
(k)
1 ≤ i ≤ 4h, and when g−1 = 2h+2, π1 (Ng−1 ) is generated by yk;i for 1 ≤ i ≤ 4h+2,
]
(k)
where yk;i is two-sided simple loops on Ng−1 based at the lift x
fk of xk as in Figure 11.
We have the following Lemma.
Lemma 3.3. For g − 1 ≥ 1 and 1 ≤ k ≤ g,
(k)
Hk = π1 (Ng−1 )+ .
(k)
(k)
Proof. Note that π1 (Ng−1 )+ is an index 2 subgroup of π1 (Ng−1 ) by Lemma 3.2. It
(k)
is sufficient for proof of Lemma 3.3 to prove Hk ⊂ π1 (Ng−1 )+ because the index of
Hk in
(k)
π1 (Ng−1 )
is
(k)
(k)
(k)
(k)
2 = [π1 (Ng−1 ) : Hk ] = [π1 (Ng−1 ) : π1 (Ng−1 )+ ][π1 (Ng−1 )+ : Hk ]
(k)
= 2 · [π1 (Ng−1 )+ : Hk ]
(k)
if Hk ⊂ π1 (Ng−1 )+ .
(k)
We define subsets of π1 (Ng−1 )+ as follows:
A :=
{xk;j+1 xk;j , xk;k+1 xk;k−1 | 1 ≤ j ≤ g − 1, j 6= k − 1, k},
B
{xk;j xk;j+1 , xk;k−1 xk;k+1 | 1 ≤ j ≤ g − 1, j 6= k − 1, k},
:=
GENERATING LEVEL 2 TWIST SUBGROUP
9
(k)
Figure 10. Ng−1 is diffeomorphic to the surface on the left-hand
side (resp. right-hand side) of the figure when g − 1 = 2h + 1 (resp.
g−1 = 2h+2). We regard the above surface on the left-hand side as
the surface identified antipodal points of the boundary component,
and the above surface on the right-hand side as the surface attached
their boundary components along the orientation of the boundary.
C
:=
{x2k;1 }
{x2k;2 }
if k 6= 1,
if k = 1.
]
(k)
π1 (Ng−1 ) is generated by yk;i . For i ≤ 2h when g − 1 = 2h + 1 (resp. i ≤ 2h + 1
when g − 1 = 2h + 2), we can check that
x
x
if 2 ≤ i ≤ 2h and g − 1 = 2h + 1,
k;ρ(i+1) k;ρ(i)
xk;g xk;g−1
if i = 1 and g − 1 = 2h + 1,
(pk )∗ (yk;i ) =
x
x
if 2 ≤ i ≤ 2h + 1 and g − 1 = 2h + 2,
k;ρ(i+1) k;ρ(i)
xk;g xk;g−1
if i = 1 and g − 1 = 2h + 2,
and (pk )∗ (yk;i ) is an element of A, where ρ is the order reversing bijection from
{1, 2, . . . , 2h} (resp. {1, 2, . . . , 2h+1}) to {1, 2, . . . , g−1}−{k}. Since if g−1 = 2h+1
and g − 1 = 2h′ + 2, we have
2
xk;1 if k 6= 1,
(pk )∗ (yk;2h+1 ) =
x2k;2 if k = 1,
2
xk;1 if k 6= 1,
(pk )∗ (yk;2h′ +2 ) =
x2k;2 if k = 1,
(pk )∗ (yk;2h+1 ) and (pk )∗ (yk;2h′ +2 ) are elements of C respectively (See Figure 12).
Finally, for i ≥ 2h + 2 when g − 1 = 2h + 1 (resp. i ≥ 2h + 3 when g = 2h + 2), we
can also check that
x ′ x ′
if 2h + 2 ≤ i ≤ 4h − 1 and g − 1 = 2h + 1,
k;ρ (i) k;ρ (i+1)
xk;g−1 xk;g
if i = 4h and g − 1 = 2h + 1,
(pk )∗ (yk;i ) =
′ (i) xk;ρ′ (i+1)
x
if 2h + 3 ≤ i ≤ 4h + 1 and g − 1 = 2h + 2,
k;ρ
xk;g−1 xk;g
if i = 4h + 2 and g − 1 = 2h + 2,
and (pk )∗ (yk;i ) = xk;ρ′ (i) xk;ρ′ (i+1) is an element of B, where ρ′ is the order preserving bijection from {2h + 2, 2h + 3, . . . , 4h} (resp. {2h + 3, 2h + 4, . . . , 4h + 2}) to
{1, 2, . . . , g − 1} − {k} (See Figure 13). We obtain this lemma.
10
R. KOBAYASHI AND G. OMORI
]
(k)
Figure 11. The total space Ng−1 of the orientation double cov]
(k)
(k)
ering pk of Ng−1 and two-sided simple loops yk;i on Ng−1 based at
x
fk .
Figure 12. The representative of x2k;1 when k 6= 1 or x2k;2 when
k = 1.
By the proof of Lemma 3.3, we have the following proposition.
(k)
Proposition 3.4. For g ≥ 2, π1 (Ng−1 )+ is generated by the following elements:
GENERATING LEVEL 2 TWIST SUBGROUP
11
Figure 13. The representative of xk;j xk;j+1 and xk;k−1 xk;k+1 for
j 6= k − 1, k.
(1) xk;i+1 xk;i , xk;i xk;i+1 , xk;k+1 xk;k−1 , xk;k−1 xk;k+1 for 1 ≤ i ≤ g − 1 and
i 6= k − 1, k,
(2) x2k;2 when k = 1,
(3) x2k;1 when 2 ≤ k ≤ g.
We remark that xk;i+1 xk;i = βk;i,i+1 , xk;i xk;i+1 = αk;i,i+1 , xk;k+1 xk;k−1 =
2
2
αk;k−1,k+1 , xk;k−1 xk;k+1 = βk;k−1,k+1 and Yi,j
= Yj,i
. Let G be the subgroup of
(k)
T2 (Ng ) generated by ∪gk=1 ψk (π1 (Ng−1 )+ ). The next corollary follows from Proposition 3.4 immediately.
Corollary 3.5. For g ≥ 2, G is generated by the following elements:
(i) ak;i,i+1 , bk;i,i+1 , ak;k−1,k+1 , bk;k−1,k+1 for 1 ≤ k ≤ g, 1 ≤ i ≤ g − 1 and
i 6= k − 1, k,
2
(ii) Y1,j
when 2 ≤ j ≤ g,
where the indices are considered modulo g.
The simple loop x2k;1 and x2k;2 are separating loops. By the next proposition,
(k)
π1 (Ng−1 )+ is generated by finitely many two-sided non-separating simple loops.
(k)
Proposition 3.6. For g ≥ 2, π1 (Ng−1 )+ is generated by the following elements:
(1) xk;i+1 xk;i , xk;i xk;i+1 , xk;k+1 xk;k−1 , xk;k−1 xk;k+1 for 1 ≤ i ≤ g and i 6=
k − 1, k,
(2) xk;2 xk;4 when k = 1 and g − 1 is even,
(3) xk;1 xk;4 when k = 2, 3 and g − 1 is even,
(4) xk;1 xk;3 when 4 ≤ k ≤ g and g − 1 is even,
where the indices are considered modulo g.
Proof. When g − 1 is odd, since we have
−1
−1
−1
x21;2 = x1;2 x1;3 · x−1
1;3 x1;4 · x1;4 x1;5 · · · · · x1;g−1 x1;g · x1;g x1;2
12
R. KOBAYASHI AND G. OMORI
and
−1
−1
−1
x2k;1 = xk;1 xk;2 · x−1
k;2 xk;3 · xk;3 xk;4 · · · · · xk;g−1 xk;g · xk;g xk;1
for 2 ≤ k ≤ g, this proposition is clear.
When g − 1 is even, we use the relation
x2k;1 . . . x2k;k−1 x2k;k+1 . . . x2k;g = 1.
For k = 1, we have
x21;2 = x1;2 x1;4 · x1;4 x1;5 · · · · · x1;g x1;2 · x1;2 x1;3 · x1;3 x1;2 .
By a similar argument, we also have following equations:
x2 = x2;1 x2;4 · x2;4 x2;5 · · · · · x2;g x2;1 · x2;1 x2;3 · x2;3 x2;1 if k = 2,
22;1
x3;1 = x3;1 x3;4 · x3;4 x3;5 · · · · · x3;g x3;1 · x3;1 x3;2 · x3;2 x3;1 if k = 3,
x2k;1 =
x2k;1 = xk;1 xk;3 · xk;3 xk;4 · · · · · xk;k−1 xk;k+1 · · · ·
·xk;g xk;1 · xk;1 xk;2 · xk;2 xk;1
if 4 ≤ k ≤ g.
We obtain this proposition.
We remark that xk;1 xk;g = βk;1,g , xk;g xk;1 = αk;1,g , x1;2 x1;4 = α1;2,4 , xk;1 xk;4 =
βk;1,4 for k = 2, 3 and xk;1 xk;3 = αk;1,3 for 4 ≤ k ≤ g. By the above remarks, we
have the following corollary.
Corollary 3.7. For g ≥ 2, G is generated by the following elements:
(i) ak;i,i+1 , bk;i,i+1 , ak;k−1,k+1 , bk;k−1,k+1 for 1 ≤ k ≤ g, 1 ≤ i ≤ g and
i 6= k − 1, k,
(ii) a1;2,4 , bk;1,4 , al;1,3 for k = 2, 3 and 4 ≤ l ≤ g when g is odd,
where the indices are considered modulo g.
3.2. Proof of Main-Theorem. First, we obtain a finite generating set for T2 (Ng )
by the Reidemeister-Schreier method. We use the following minimal generating set
for Γ2 (Ng ) given by Hirose and Sato [2] when g ≥ 5 and Szepietowski [12] when
g = 3, 4 to apply the Reidemeister-Schreier method. See for instance [4] to recall
the Reidemeister-Schreier method.
Theorem 3.8. [2, 12] For g ≥ 3, Γ2 (Ng ) is generated by the following elements:
(1) Yi,j for 1 ≤ i ≤ g − 1, 1 ≤ j ≤ g and i 6= j,
2
(2) T1,j,k,l
for 2 ≤ j < k < l ≤ g when g ≥ 4.
Proposition 3.9. For g ≥ 3, T2 (Ng ) is generated by the following elements:
2
(1) Yi,j Y1,2 , Yi,j
for 1 ≤ i ≤ g − 1, 1 ≤ j ≤ g and i 6= j,
−1 2
2
for 2 ≤ j < k < l ≤ g when g ≥ 4.
(2) Y1,2 T1,j,k,l Y1,2 , T1,j,k,l
Proof. Note that T2 (Ng ) is the intersection of Γ2 (Ng ) and T (Ng ). Hence we have
the isomorphisms
Γ2 (Ng )/(Γ2 (Ng ) ∩ T (Ng )) ∼
= Z2 [Y1,2 ].
= (Γ2 (Ng )T (Ng ))/T (Ng ) ∼
We remark that Γ2 (Ng )T (Ng ) = M(Ng ) and the last isomorphism is given by
Lickorish [7]. Thus T2 (Ng ) is an index 2 subgroup of Γ2 (Ng ).
Set U := {Y1,2 , 1} and X as the generating set for Γ2 (Ng ) in Theorem 3.8, where
1 means the identity element. Then U is a Schreier transversal for T2 (Ng ) in Γ2 (Ng ).
GENERATING LEVEL 2 TWIST SUBGROUP
13
For x ∈ Γ2 (Ng ), define x as the element of U such that [x] = [x] in Γ2 (Ng )/T2 (Ng ).
By the Reidemeister-Schreier method, for g ≥ 4, T2 (Ng ) is generated by
B ={wu−1 wu | w ∈ X ± , u ∈ U, wu 6∈ U }
±1
Y1,2
={Yi,j
−1
±1
±1
Yi,j
Y1,2 , Yi,j
±2
Y1,2
∪ {T1,j,k,l
−1
−1
±1
Yi,j
| 1 ≤ i ≤ g − 1, 1 ≤ j ≤ g, i 6= j}
±2
±2
T1,j,k,l
Y1,2 , T1,j,k,l
−1
±2
T1,j,k,l
| 2 ≤ j < k < l ≤ g}
±1
−1 ±1
={Yi,j
Y1,2 , Y1,2
Yi,j | 1 ≤ i ≤ g − 1, 1 ≤ j ≤ g, i 6= j}
−1 ±2
±2
∪ {Y1,2
T1,j,k,l Y1,2 , T1,j,k,l
| 2 ≤ j < k < l ≤ g},
where X ± := X ∪ {x−1 | x ∈ X} and note that equivalence classes of Y−1 ±1
∓1
Y1,2 )−1
homeomorphisms in Γ2 (Ng )/T2 (Ng ) is nontrivial. Since Y1,2
Yi,j = (Yi,j
−1
−2
and Yi,j Y1,2 = Yi,j · Yi,j Y1,2 , we have the following generating set for T2 (Ng ):
2
B ′ ={Yi,j Y1,2 , Yi,j
| 1 ≤ i ≤ g − 1, 1 ≤ j ≤ g, i 6= j}
−1 2
2
∪ {Y1,2
T1,j,k,l Y1,2 , T1,j,k,l
| 2 ≤ j < k < l ≤ g}.
By a similar discussion, T2 (N3 ) is generated by
2
B ′ ={Yi,j Y1,2 , Yi,j
| 1 ≤ i ≤ 2, 1 ≤ j ≤ 3, i 6= j}.
We obtain this proposition.
Let G be the group generated by the elements of type (i), (ii) and (iii) in Theorem 3.1. Then G is a subgroup of T2 (Ng ) clearly and it is sufficient for the proof
of Theorem 3.1 to prove B ′ ⊂ G, where B ′ is the generating set for T2 (Ng ) in the
(k)
proof of Proposition 3.9. By Corollary 3.7, we have ψk (π1 (Ng−1 )+ ) ⊂ G for any
(i)
2
1 ≤ k ≤ g. Thus Yi,j
= ψi (x2i;j ) ∈ ψi (π1 (Ng−1 )+ ) ⊂ G. We complete the proof of
−1 2
Theorem 3.1 if Yi,j Y1,2 and Y1,2
T1,j,k,l Y1,2 are in G.
−1 2
Lemma 3.10. For g ≥ 4, Y1,2
T1,j,k,l Y1,2 ∈ G.
−1 2
−1 2
Proof. Since Y1,2
T1,j,k,l Y1,2 is a Dehn twist along
T1,j,k,l Y1,2 = tY −1 (α1,j,k,l ) , Y1,2
1,2
the two-sided simple closed curve as in Figure 14. Then we have a1;k,l (α1,2,k,l ) =
−1
−2
−1
Y1,2
(α1,2,k,l ) and Y1,2
a1;2,j a1;k,l (α1,j,k,l ) = Y1,2
(α1,j,k,l ) for 3 ≤ j ≤ g
and the local orientation of the regular neighborhood of a1;k,l (α1,2,k,l ) (resp.
−2
−1
−1
Y1,2
a1;2,j a1;k,l (α1,j,k,l )) and Y1,2
(α1,2,k,l ) (resp. Y1,2
(α1,j,k,l )) are different. Therefore we have
−1 2
−2
Y1,2
T1,2,k,l Y1,2 =a1;k,l T1,2,k,l
a−1
1;k,l ,
−1 2
−2
−2
−1
2
Y1,2
T1,j,k,l Y1,2 =Y1,2
a1;2,j a1;k,l T1,j,k,l
a−1
1;k,l a1;2,j Y1,2 for 3 ≤ j ≤ g.
(1)
By Corollary 3.7, a1;k,l , a1;2,j ∈ ψ1 (π1 (Ng−1 )+ ) ⊂ G. We obtain this lemma.
Szepietowski [11, Lemma 3.1] showed that for any non-separating two-sided simple closed curve γ, t2γ is a product of two Y-homeomorphisms. In particular, we
have the following lemma.
Lemma 3.11 ([11]). For distinct 1 ≤ i, j ≤ g,
2
Ti,j for i < j,
−1
−1
Yj,i
Yi,j = Yj,i Yi,j
=
−2
Ti,j
for j < i.
14
R. KOBAYASHI AND G. OMORI
Figure 14. The upper side of the figure is the simple closed curve
−1
Y1,2
(α1,2,k,l ) on Ng and the lower side of the figure is the simple
−1
closed curve Y1,2
(α1,j,k,l ) on Ng for 3 ≤ j ≤ g.
2
Lemma 3.12. For distinct 1 ≤ i, j ≤ g, Ti,j
∈ G.
Proof. We discuss by a similar argument in proof of Lemma 3.5 in [12]. Let γi be
(i)
the two-sided simple loop on Ng−1 for i = 3, . . . , g as in Figure 15. Then we have
(i)
2
T1,2
= ψg (γg ) · · · ψ4 (γ4 )ψ3 (γ3 ) (see Figure 15). Since γi ∈ π1 (Ng−1 )+ , each ψi (γi )
2
is an element of G by Corollary 3.7. Hence we have T1,2
∈ G.
We denote by σi,j the self-diffeomorphism on Ng which is obtained by the transposition of the i-th crosscap and the j-th crosscap as in Figure 16. σi,j is called the
crosscap transposition (c.f. [9]). For 1 ≤ i < j ≤ g, put fi,j ∈ M(Ng ) as follows:
f1,2 := 1,
f1,j := σj−1,j · · · σ3,4 σ2,3 for 3 ≤ j ≤ g,
fi,j := σi−1.i · · · σ2,3 σ1,2 f1,j for 2 ≤ i < j ≤ g.
−1
−1
−1
−1
2
2
Then Ti,j
= fi,j T1,2
fi,j
= fi,j ψg (γg )fi,j
· · · fi,j ψ4 (γ4 )fi,j
·fi,j ψ3 (γ3 )fi,j
. Since the
−1
action of σi,j on Ng preserves the set of i-th crosscaps for 1 ≤ i ≤ g, fi,j ψk (γk )fi,j
is
(k′ )
−1
∈
an element of ψk′ (π1 (Ng−1 )) for some k ′ . By Corollary 3.7, we have fi,j ψk (γk )fi,j
G and we obtain this lemma.
Finally, by the following proposition, we complete the proof of Theorem 3.1.
Proposition 3.13. For distinct 1 ≤ i, j, k, l ≤ g, Yk,l Yi,j ∈ G.
Proof. Yk,l Yi,j is the following product of elements of G.
(a) case (k, l) = (i, j):
2
Yk,l Yi,j = Yi,j
.
By Corollary 3.7, the right-hand side is an element of G.
(b) case (k, l) = (j, i):
−1
2
Yj,i Yi,j = Yj,i
· Yj,i
Yi,j
Lem. 3.11
=
±2
(a) · Ti,j
.
By Lemma 3.12, the right-hand side is an element of G.
(c) case k = i and l 6= j:
Yi,l Yi,j = ψi (xi;l )ψi (xi;j ) = ψi (αi;j,l ) = ai;j,l
Cor. 3.7
∈
G.
GENERATING LEVEL 2 TWIST SUBGROUP
15
2
Figure 15. T1,2
is a product of crosscap pushing maps along γ3 , γ4 , . . . , γg .
Figure 16. The crosscap transposition σi,j .
(d) case k 6= i and l = j:
−1 −1
Yk,j Yi,j = Yk,j Yk,i · Yk,i
Yi,k · Yi,k Yi,j = (c) · (b) · (c) ∈ G.
(e) case k = j and l 6= i:
−1
Yj,l Yi,j = Yj,l Yj,i · Yj,i
Yi,j
Lem. 3.11
±2
(c) · Ti,j
Lem. 3.12
Lem. 3.11
±2
Ti,k
· (c)
Lem. 3.12
=
∈
G.
(f) case k 6= j and l = i:
−1
Yk,i Yi,j = Yk,i Yi,k
· Yi,k Yi,j
=
∈
G.
(g) case {k, l} ∩ {i, j} is empty:
−1 −1
Yk,l Yi,j = Yk,l Yk,j · Yk,j
Yk,j · Yk,j Yi,j = (c) · (a) · (d) ∈ G.
We have completed this proposition.
By a similar discussion in Subsection 3.2 and Corollary 3.5, we obtain the following theorem.
Theorem 3.14. For g ≥ 3, T2 (Ng ) is generated by following elements:
(i) ak;i,i+1 , bk;i,i+1 , ak;k−1,k+1 , bk;k−1,k+1 for 1 ≤ k ≤ g, 1 ≤ i ≤ g − 1 and
i 6= k − 1, k,
16
R. KOBAYASHI AND G. OMORI
2
(ii) Y1,j
for 2 ≤ j ≤ g,
2
(iii) T1,j,k,l for 2 ≤ j < k < l ≤ g when g ≥ 4,
where the indices are considered modulo g.
Since the number of generators in Theorem 3.14 is 61 (g 3 + 6g 2 − 7g − 12) for
g ≥ 4 and 3 for g = 3, the number of generators in Theorem 3.14 is smaller than
the number of generators in Theorem 3.1. On the other hand, by Theorem
1.2,
the dimension of the first homology group H1 (T2 (Ng )) of T2 (Ng ) is g3 + g2 − 1 =
1 3
2
6 (g − g − 6) for g ≥ 4. The difference of them is g − g − 1. The authors do not
know the minimal number of generators for T2 (Ng ) when g ≥ 4.
4. Normal generating set for T2 (Ng )
The next lemma is a generalization of the argument in the proof of Lemma 3.5
in [12].
Lemma 4.1. Let γ be a non-separating two-sided simple closed curve on Ng such
that Ng − γ is a non-orientable surface. Then t2γ is a product of crosscap pushing
maps along two-sided non-separating simple loops such that their crosscap pushing
maps are conjugate to a1;2,3 in M(Ng ).
Proof of Theorem 1.1. By Theorem 3.1, T2 (Ng ) is generated by (I) crosscap push2
ing maps along non-separating two-sided simple loops and (II) T1,j,k,l
for 2 ≤ j <
2
2
2
2
k < l ≤ g. When g = 3, T2 (Ng ) is generated by T1,2 , T1,3 , T2,3 . Recall Ti,j
= a−1
k;i,j
when g = 3. Since Ng − αi,j is non-orientable for g ≥ 3, ak;i,j is conjugate to ak′ ;i′ ,j ′
in M(Ng ). Hence Theorem 1.1 is clear when g = 3.
(k)
Assume g ≥ 4. For a non-separating two-sided simple loop c on Ng−1 based at
xk , by Lemma 2.1, there exist non-separating two-sided simple closed curves c1 and
c2 such that ψ(c) = tc1 t−1
c2 , where c1 and c2 are images of boundary components of
(k)
regular neighborhood of c in Ng−1 to Ng by a blowup. Then the surface obtained
by cutting Ng along c1 and c2 is diffeomorphic to a disjoint sum of Ng−3,2 and
N1,2 . Thus mapping classes of type (I) is conjugate to a1;2,3 in M(Ng ). We obtain
Theorem 1.1 for g = 4.
Assume g ≥ 5. Simple closed curves αi,j,k,l satisfy the condition of Lemma 4.1.
2
Therefore T1,j,k,l
is a product of crosscap pushing maps along non-separating twosided simple loops and such crosscap pushing maps are conjugate to a1;2,3 in
M(Ng ). We have completed the proof of Theorem 1.1.
5. First homology group of T2 (Ng )
By the argument in the proof of Proposition 3.9, for g ≥ 2, we have the following
exact sequence:
(5.1)
1 −→ T2 (Ng ) −→ Γ2 (Ng ) −→ Z2 −→ 0,
where Z2 is generated by the equivalence class of a Y-homeomorphism.
The level 2 principal congruence subgroup Γ2 (n) of GL(n, Z) is the kernel of the
natural surjection GL(n, Z) ։ GL(n, Z2 ). Szepietowski [12, Corollary 4.2] showed
that there exists an isomorphism θ : Γ2 (N3 ) → Γ2 (2) which is induced by the action
of Γ2 (N3 ) on the free part of H1 (N3 ; Z). Since the determinant of the action of a
GENERATING LEVEL 2 TWIST SUBGROUP
17
Dehn twist on the free part of H1 (N3 ; Z) is 1, we have the following commutative
diagram of exact sequences:
/ Γ2 (N3 )
/ T2 (N3 )
1
(5.2)
θ|T2 (N3 )
/1
/ Z2
/ 1,
θ
/ SL(2, Z)[2]
1
/ Z2
/ Γ2 (2)
det
where SL(n, Z)[2] := Γ2 (n) ∩ SL(n, Z) is the level 2 principal congruence subgroup
of the integral special linear group SL(n, Z). By the commutative diagram (5.2),
T2 (N3 ) is isomorphic to SL(2, Z)[2].
Proof of Theorem 1.2. For g = 3, the first homology group H1 (T2 (N3 )) is isomorphic to H1 (SL(2, Z)[2]) by the commutative diagram (5.2). The restriction of the
natural surjection from SL(2, Z) to the projective special linear group P SL(2, Z)
to SL(n, Z)[2] gives the following commutative diagram of exact sequences:
(5.3)
1
/ Z2 [−E]
/ SL(2, Z)
/ P SL(2, Z)
/1
/ SL(2, Z)[2]
/ P SL(2, Z)[2]
/ 1,
id
1
/ Z2 [−E]
where E is the identity matrix and P SL(n, Z)[2] := SL(n, Z)/{±E} is the level 2
principal congruence subgroup of P SL(2, Z). Since P SL(2, Z)[2] is isomorphic to
the free group F2 of rank 2 and −E commutes with all matrices, the exact sequence
in the lower row of Diagram (5.3) is split and SL(2, Z)[2] is isomorphic to F2 ⊕ Z2 .
Thus H1 (T2 (N3 )) is isomorphic to Z2 ⊕ Z2 .
For g ≥ 2, the exact sequence (5.1) induces the five term exact sequence between
these groups:
H2 (Γ2 (Ng )) −→ H2 (Z2 ) −→ H1 (T2 (Ng ))Z2 −→ H1 (Γ2 (Ng )) −→ H1 (Z2 ) −→ 0,
where
H1 (T2 (Ng ))Z2 := H1 (T2 (Ng ))/ f m − m | m ∈ H1 (T2 (Ng )), f ∈ Z2 .
−1
For m ∈ H1 (T2 (Ng )) and f ∈ Z2 , f m := [f ′ m′ f ′ ] ∈ H1 (T2 (Ng )) for some
representative m′ ∈ T2 (Ng ) and f ′ ∈ Γ2 (Ng ). Since H2 (Z2 ) ∼
= H2 (RP ∞ ) = 0 and
∼
H1 (Z2 ) = Z2 , we have the short exact sequence:
0 −→ H1 (T2 (Ng ))Z2 −→ H1 (Γ2 (Ng )) −→ Z2 −→ 0.
(g)+(g)
Since Hirose and Sato [2] showed that H1 (Γ2 (Ng )) ∼
= Z23 2 , it is sufficient for the
proof of Theorem 1.2 when g ≥ 5 to prove that the action of Z2 ∼
= Γ2 (Ng )/T2 (Ng )
on the set of the first homology classes of generators for T2 (Ng ) is trivial.
By Theorem 1.1, T2 (Ng ) is generated by crosscap pushing maps along nonseparating two-sided simple loops for g ≥ 5. Let ψ(γ) = tγ1 t−1
γ2 be a crosscap
pushing map along a non-separating two-sided simple loop γ, where γ1 and γ2 are
images of boundary components of the regular neighborhood of γ in Ng−1 to Ng by
a blowup. The surface S obtained by cutting Ng along γ1 and γ2 is diffeomorphic
to a disjoint sum of Ng−3,2 and N1,2 . Since g − 3 ≥ 5 − 3 = 2, we can define
a Y-homeomorphism Y on the component of S. The Y-homeomorphism is not
a product of Dehn twists. Hence [Y ] is the nontrivial element in Z2 and clearly
18
R. KOBAYASHI AND G. OMORI
Y ψ(γ)Y −1 = ψ(γ) in Γ2 (Ng ), i.e. [Y ψ(γ)Y −1 ] = [ψ(γ)] in H1 (T2 (Ng )). Therefore
the action of Z2 on H1 (T2 (Ng )) is trivial and we have completed the proof of
Theorem 1.2.
Remark 5.1. When g = 3, we have the exact sequence
0 −→ H1 (T2 (N3 ))Z2 −→ H1 (Γ2 (N3 )) −→ Z2 −→ 0
by the argument in the proof of Theorem 1.2. Since H1 (Γ2 (N3 )) ∼
= H1 (Γ2 (2)) ∼
= Z42 ,
∼
we showed that the action of Z2 = Γ2 (N3 )/T2 (N3 ) on H1 (T2 (N3 )) is not trivial by
Theorem 1.2 when g = 3.
Acknowledgement:
The authors would like to express their gratitude to
Hisaaki Endo and Susumu Hirose, for his encouragement and helpful advices. The
authors also wish to thank Masatoshi Sato for his comments and helpful advices.
The second author was supported by JSPS KAKENHI Grant number 15J10066.
References
[1] D. R. J. Chillingworth, A finite set of generators for the homeotopy group of a non-orientable
surface, Math. Proc. Camb. Philos. Soc. 65 (1969), 409–430.
[2] S. Hirose, M. Sato, A minimal generating set of the level 2 mapping class group of a nonorientable surface, Math. Proc. Camb. Philos. Soc. 157 (2014), no. 2, 345–355.
[3] S. P. Humphries, Normal closures of powers of Dehn twists in mapping class groups, Glasgow
Math. J. 34 (1992), 313–317.
[4] D. L. Johnson, Presentations of Groups, London Math. Soc. Stud. Texts 15 (1990).
[5] M. Korkmaz, First homology group of mapping class group of nonorientable surfaces, Math.
Proc. Camb. Phil. Soc. 123 (1998), 487–499.
[6] M. Korkmaz, Mapping class groups of nonorientable surfaces, Geom. Dedic. 89 (2002), 109–
133.
[7] W. B. R. Lickorish, On the homeomorphisms of a non-orientable surface, Math. Proc. Camb.
Philos. Soc . 61 (1965), 61–64.
[8] J. D. McCarthy and U. Pinkall. Representing homology automorphisms of nonorientable
surfaces, Max Planck Inst. Preprint MPI/SFB 85-11, revised version written on 26 Feb 2004,
available from http://www.math.msu.edu/˜mccarthy/publications/selected.papers.html.
[9] L. Paris and B. Szepietowski. A presentation for the mapping class group of a nonorientable
surface, arXiv:1308.5856v1 [math.GT], 2013.
[10] M. Stukow, The twist subgroup of the mapping class group of a nonorientable surface, Osaka
J. Math. 46 (2009), 717–738.
[11] B. Szepietowski. Crosscap slides and the level 2 mapping class group of a nonorientable
surface, Geom. Dedicata 160 (2012), 169–183.
[12] B. Szepietowski. A finite generating set for the level 2 mapping class group of a nonorientable
surface. Kodai Math. J. 36 (2013), 1–14.
(Ryoma Kobayashi) Department of General Education, Ishikawa National College
of Technology, Tsubata, Ishikawa, 929-0392, Japan
E-mail address: kobayashi [email protected]
(Genki Omori) Department of Mathematics, Tokyo Institute of Technology, Ohokayama, Meguro, Tokyo 152-8551, Japan
E-mail address: [email protected]
| 4math.GR
|
Santa Fe Institute Working Paper 14-XX-XXX
arxiv.org:14XX.XXXX [cond-mat.stat-mech]
Understanding and Designing Complex Systems:
Response to
“A framework for optimal high-level descriptions
in science and engineering—preliminary report”
arXiv:1412.8520v1 [cond-mat.stat-mech] 30 Dec 2014
James P. Crutchfield,1, ∗ Ryan G. James,1, † Sarah Marzen,2, ‡ and Dowman P. Varn1, §
1
Complexity Sciences Center and Department of Physics,
University of California at Davis, One Shields Avenue, Davis, CA 95616
2
Department of Physics and Redwood Center for Theoretical Neuroscience
University of California at Berkeley, Berkeley, CA 94720-5800
(Dated: December 31, 2014)
We recount recent history behind building compact models of nonlinear, complex processes and
identifying their relevant macroscopic patterns or “macrostates”. We give a synopsis of computational mechanics, predictive rate-distortion theory, and the role of information measures in monitoring model complexity and predictive performance. Computational mechanics provides a method
to extract the optimal minimal predictive model for a given process. Rate-distortion theory provides methods for systematically approximating such models. We end by commenting on future
prospects for developing a general framework that automatically discovers optimal compact models.
As a response to the manuscript cited in the title above, this brief commentary corrects potentially
misleading claims about its state space compression method and places it in a broader historical
setting.
Keywords: information theory, rate-distortion theory, computational mechanics, information bottleneck, macrostates, microstates, statistical physics, coarse-graining, dimension reduction, minimum description length
PACS numbers: 02.50.-r 89.70.+c 05.45.Tp 02.50.Ey 02.50.Ga
I.
INTRODUCTION
Building compact models of nonlinear processes goes to
the heart of our understanding the complex world around
us—a world replete with unanticipated, emergent patterns. Via discovery mechanisms that we do not yet understand well, we eventually do come to know many of
these patterns, even if we have never seen them before.
Such discoveries can be substantial. At a minimum, compact models that capture such emergent “macrostates”
are essential tools in harnessing complex processes to useful ends. Most ambitiously, one would hope to automate
the discovery process itself, providing an especially useful
tool for the era of Big Data.
One key problem in the larger endeavor of pattern discovery is dimension reduction: reduce the high-dimensional
state space of a stochastic dynamical system into smaller,
more manageable models that nonetheless still capture
the relevant dynamics. The study of complex systems
always requires this. For better or worse, it is frequently
∗
†
‡
§
[email protected]
[email protected]
[email protected]
[email protected]
accomplished in an ad hoc fashion [1, 2]. Indeed, it is
desirable to have an overarching framework for this kind
of analysis that can be applied across the many manifestations of complex systems, but to date such a broad
theory has not been forthcoming. Thus, the need for this
kind of research remains and is more timely than ever [3].
Not surprisingly, it has a long and active history.
This is the setting into which steps a recent arxiv.org
preprint “A framework for optimal high-level descriptions
in science and engineering—preliminary report” [4]. As
a solution to the problem of dimension reduction, it advocates for state space compression (SSC): Form a compressed variable Yt that predicts a target statistic Ωt of
a system’s behavior . . . Xt−1 Xt Xt+1 . . .. When viewed
in a historical context, it is unclear if SSC is more than
an alternative notation and vocabulary for extant approaches to dimension reduction. Here, we explain this
question. We are concerned about several instances in
Ref. [4] where statements are made about research we either participated in or are quite familiar with that do not
accurately reflect that work. The following comments air
our concerns, providing several constructive suggestions.
Our response to Ref. [4] is organized as follows. We
recall the history over the last half century driving interest and research on reconstructing optimal minimal
2
models, specifically as it bears on nonlinear complex systems. We briefly recount the approach of computational
mechanics, which defined what optimal predictive models are and gave the general solution to finding them.
We then draw connections to predictive rate-distortion
theory that systematically approximates those optimal
models. We also comment on information-theoretic ways
to quantify model complexity and predictive performance
and how they trade-off against each other. Our goal is to
respond directly and briefly to Ref. [4], but not to review
the broad and extensive literature on the topic of optimal descriptions of complex systems. As such, citations
are intentionally narrowed to support a single narrative
thread.
II.
RECONSTRUCTING LOW-DIMENSIONAL
MODELS OF COMPLEX PROCESSES
The research program to automate theory building for
complex systems has its origins in fluid turbulence—a
high-dimensional system, if there ever was one—studied
for many decades, of great practical import, and at the
time, according to Heisenberg [5], one of the premier
problems in nonlinear physics. Cracking this problem
relied on developing a connection between the abstract
theory of dynamical systems and hydrodynamic experiment. This came in the attractor reconstruction method
that extracted “Geometry from a Time Series” [6] using
one or a few signals from a high-dimensional complex system. Attractor reconstruction eventually led to demonstrating that deterministic chaos is the mechanism generating weak turbulence [7], verifying a long-standing conjecture [8] in dynamical systems theory that over-threw
the decades-old quasi-periodic theory of turbulence due
to Landau and Lifschitz [9].
The reconstruction method, though, only produced a
reduced-dimension space of effective states of the infinitedimensional fluid dynamics, ignoring the dynamical
mechanism that generated the turbulent motion. Generalizing attractor reconstruction, Refs. [10, 11] introduced methods to infer the effective theory or equations
of motion from time-series generated by chaotic dynamical systems. Reference [10], in particular, also provided a critique of the general reconstruction approach,
highlighting its subjectivity—one must choose a class of
model representation. Such ad hoc choices preclude an
objective measure of a system’s complexity. Reference
[12] solved the subjectivity problem. It was the first to
state the problem of optimal predictive, minimal models of complex processes and provided its solution—the
ǫ-machine and its causal states. Using this foundation,
it was able to define the statistical complexity—the first
consistent, constructive measure of structural complexity. And, it introduced intrinsic computation—the idea
that all physical systems store and process information.
Influenced by the goals of artificial intelligence at the
time, it also challenged future researchers to develop an
artificial science—automating the construction of minimal causal models of complex systems. Using these
methods, which came to be called computational mechanics, Ref. [13] went on to give the first quantitative definition of the emergence of macrostate organization in
terms of increasing structural complexity and intrinsic
computation. It argued that there is a natural hierarchy
of causal organization and introduced a renormalizationgroup method for moving up the hierarchy—a method of
genuine pattern discovery. A recent review of this history
is found in Ref. [14].
III.
CAUSAL STATES AND MACROSTATES
Identifying emergent organization, especially if pursued
as a problem in theoretical physics, is often couched in
terms of finding system macrostates. The metaphorical
intuition behind this framing is roughly that emergent organization is the analog of the macroscopic, measurable
properties of a thermodynamic system. A macrostate—
say, given by particular values of temperature and
pressure—is the set of the many microscopic molecular
configurations or microstates that lead to the same measurable properties. In this framing, macrostates emerge
from the microstates under the action of the dynamics
on the microscopic scale as it relaxes to equilibrium [15].
The practicing statistical physicist, more specifically, often begins the analysis of a system’s properties by searching for an “order parameter” or for insightful “coarsegrainings”, which are analogous concepts at our general
level of discussion here.
A system’s causal states, being groupings of microscopic
trajectories that capture a system’s emergent behaviors
and organization, play a role very analogous to the system’s macrostates and not its microstates [16]. Reference
[4] offers a view that is substantially at variance with this.
SSC addresses this in terms of the descriptions to which
given behavior is compressed. Notably, Ref. [4, page 6]
starts by restricting the microscopic dynamics:
The microstate ... evolves in time according
to a (usually) Markovian process . . .
And then, on page 41 it notes that the observations of
the stochastic processes analyzed in computational me-
3
chanics:
. . . do not evolve according to a first-order
Markov process, and so cannot be identified
with the fine-grained values Xt of SSC. On
the other hand, . . . , evolution over the space
of causal sets is first-order Markov. This suggests that we identify the causal states with
SSC’s fine-grained space, not its compressed
space.
In the page 6 quote, however, SSC assumed firstorder Markov dynamics on the underlying process (microstates). In the context described above, in which the
target Ωt is the future Xt Xt+1 . . . and Yt compresses the
past . . . Xt−2 Xt−1 , the causal states are interpreted as
coarse-grainings only of Xt . Thus, in SSC the dynamics
on the causal states and the observed process are both
first-order Markov. These restrictions are unnecessary.
is called predictive rate distortion (PRD) [17]. SSC
largely ignores this important connection to prior work.
This is strange since Shannon introduced rate-distortion
theory explicitly to solve dimension reduction problems
[18, 19]: find a compact “encoding” of a data source subject to a set of constraints, such as transmission rate,
accuracy, processing, delay, and the like. Optimal modeling can be framed in just this way, for example, as found
in Rissanen’s minimum description length approach [20].
The physics metaphor for building models extends to
this setting, too: data are the microstates, compressed
model variables are macrostates, and coding constraints
are physical boundary conditions.
A telling consequence of SSC’s misidentification of how
computational mechanics is used is that the assumption of a first-order Markov process for the microstates
simplifies the dimension-reduction problem to the point
that the computational complexity of identifying causal
states would fall in P and no longer be NP-hard, as it is
more generally. Overall, the simplification to first-order
Markov obscures the relationship between computational
mechanics and SSC. Computational mechanics can be
applied to any stochastic process, including the Markovian one assumed to govern SSC’s microscopic variables
. . . Xt−1 Xt Xt+1 . . ..
Take SSC’s target Ωt to be the future of Xt , and find
some Yt that compresses the past of Xt to retain expected accuracy in predicting Ωt without unnecessarily
increasing asymptotic coding cost of transmitting Yt to
another observer. When accuracy is heavily prioritized
over coding cost, the causal states are recovered, as shown
by Refs. [21, 22] and as discussed in Ref. [23]. In fact, the
causal states are essentially the answer to the question:
Among all possible compressions with minimal accuracy
cost, which has the minimal coding cost? Both issues
of accuracy and coding cost can be extended directly to
the case in which the target Ωt is some coarse-graining
of the future of Xt , so that one searches for compressed
predictive and perceptual features or macrostates. Recent PRD work shows, in fact, how to extract just such
coarse-grained macrostates, given a process’s ǫ-machine
[17], noting that the latter can be calculated theoretically
[24, 25] or estimated empirically [26–28].
Apparently, despite being defined as a coarse-graining,
SSC (incorrectly) associates causal states with its microstates Xt for no reason other than their Markovian
nature. This is a confusing association for two reasons:
First, the Xt values were introduced as being (usually)
Markovian, not necessarily Markovian; and second, by
construction computational mechanics’ causal states are
a coarse-graining of trajectories and so by definition are
a form of state-space compression. SSC’s association is
based on reasoning that is fundamentally flawed, reflecting a basic misunderstanding of causal states and how to
apply computational mechanics.
However, a major conceptual difference between PRD
and the SSC framework is that SSC explicitly restricts
Xt to be first-order Markov; whereas, PRD can address
arbitrary order and infinite-order Markov processes. The
first-order Markov assumption is simply not realistic. Microstates are rarely experimentally accessible; e.g., the
spike trains from neurophysiological studies are a very
coarse-grained observable of underlying membrane voltage fluctuations [29] and, as such, their dynamics is often infinite-order Markov [30]. This remains true even
if the underlying membrane voltage dynamics are firstorder Markov.
IV.
STATE SPACE COMPRESSION VERSUS
RATE DISTORTION THEORY
Now, let’s turn to consider how SSC defines its
“macrostates” via coarse-graining microstates. In this,
we find a basic connection between SSC and Shannon’s rate-distortion theory applied to prediction—what
PRD and computational mechanics actually do state
space compression. They find compressed predictive representations of a time series. Moreover, computational
mechanics finds a hidden Markov model (HMM) for the
given (non-Markovian) time series. Indeed, that’s the
point of the ǫ-machine and, for lossy versions, the point
of causal rate-distortion theory [17] and the recursive information bottleneck method [23].
4
V.
CODING COST
This all leads to the central question of how to quantify the organization captured by these macrostates.
PRD and computational mechanics use the statistical complexity—the Shannon information in the causal
states or in their coarse-grainings. SSC takes issue with
the use of information theory in these approaches [4, page
40]:
Statistical complexity plays a role in the objective function of [77] that is loosely analogous to the role played by accuracy cost in
the SSC objective function.
[Emphasis ours.] However, SSC’s coding cost H[Y0 ] is the
information contained in its version of the causal states.
This is a simplified version of the statistical complexity
Cµ , once one restricts to first-order Markov processes and
predictive mappings from pasts . . . Xt−2 Xt−1 to Yt . That
is, Cµ is a coding cost, not an accuracy cost.
However, there is perhaps another underlying difficulty
with using mutual information as an accuracy cost: It is
used alternatively as either the coding cost or an inverse
accuracy cost in the information bottleneck method [36].
And, there, mutual information as an ‘inverse accuracy
cost’ really amounts to an accuracy cost that employs the
KL-divergence between Pr(Ωt |Yt ) and Pr(Ωt |Xt ) in the
more general PRD framework.
Generally, though, the rate-distortion theorem [37] justifies the use of mutual information as a coding cost regardless of distortion measure. And, helpfully, this has
been extended to the nonasymptotically large block coding limit [38, 39], demonstrating that one should not be
glib about introducing new coding costs.
VII.
GENERALITY
SSC is offered as an improvement on the current literature for being a principled and constructive method of
generating macrostates. Reference [4]’s abstract states:
Moreover, on Ref. [4, page 40], we read:
. . . the authors consider a ratio of costs rather
than a linear combination of them, like we
consider here.
However, PRD objective functions—e.g., as described by
Refs. [23, 31]—are linear combinations of mutual informations.
VI.
MUTUAL INFORMATION AS ACCURACY
COST
Reference [4, page 46] criticizes the use of general multivariate mutual information in noting a difficulty with
systems governed by a time-varying dynamic:
The underlying difficulty is inherent in the
very notion of using mutual information to
define accuracy cost, and is intrinsic to consideration of the relation between Y and X
at more than two moments in time.
Using mutual information at more than two moments
in time is simply not a problem for ergodic stationary
processes [32]. Moreover, it’s not necessarily a problem
for nonergodic or nonstationary processes either. For
these, several examples in Ref. [33], as well as those in
Refs. [34, 35], calculate the past-future mutual information (excess entropy) that involves all moments in time.
This State Space Compression framework
makes it possible to solve for the optimal
high-level description of a given dynamical
system . . ..
This brings up two concerns, one explicit in the quote
and one implicit.
First, what is provided is not a constructive framework;
it does not provide methods to solve for the macrostates.
Moreover, each of the provided macrostate examples is
constructed in an ad hoc manner.
Second, the burden of proof lies with SSC’s advocates.
Since alternative frameworks exist and are contrasted
against, at this date progress behooves the authors to
provide examples where their framework succeeds and
others fail.
SCC’s lack in these regards should be compared with how
alternatives had to develop new calculational methods for
the challenging problems that are entailed in modeling
complex systems. For example, computational mechanics extended methods from holomorphic functional calculus that now give closed-form expressions for a process’s
information measures [40].
We emphasize that SSC’s H(Y0 ) is exactly the computational cost-based objective function used to identify
causal states, in which we constrain ourselves to deterministic mappings Yt = f (Xt ) such that Pr(Ωt |Yt0 ) =
Pr(Ωt |f (Xt0 )). Indeed, more broadly interpreted, causal
states are the unique and minimal macrostate choice for
5
SSC in the limit of high premium on accuracy and minimal concern about computational cost. Adapting the
proof of Thm. 1 in Ref. [22] will be helpful here. In any
case, the causal states are the minimal sufficient statistic
for prediction [41]. In other words, any process statistic
can be calculated from them and this raises the bar quite
high for SSC’s proposed alternative macrostates.
In terms of implementations, PRD and computational
mechanics are constructive.
In rate-distortion theory generally there’s the Blahut-Arimoto algorithm
and its generalizations for calculating coarse-grained
macrostates [37, 42]. And, there are several statistical
inference algorithms that estimate ǫ-machines from various kinds of data [26, 27, 43, 44].
The general problem of dimension reduction for complex
systems is an important one, and we encourage efforts
along these lines. We appreciate that the synopses of
computational mechanics, PRD, and information measures above address but a small part SSC’s stated goals
and the goals earlier researchers have set. It is important, though, that SSC start with correct assumptions
and an understanding of its antecedents. In any case, we
hope that our comments remedy, in a constructive way,
misleading impressions that Ref. [4] gives of the state of
the art of computational mechanics and predictive rate
distortion.
ments. Despite this, we are still optimists. Those wishing
to contribute, however, should pick at least one application area and drill all the way down to show their alternative discovers a particular new scientific phenomenon.
This is a necessary calibrating exercise for any attempt
at generality.
In contrast to SSC, its antecedents have done their due
diligence. Rate-distortion theory, developed for over a
decade into the information bottleneck method [36], has
been applied to test how close sensory spike trains are
to stimuli predictive information functions [45]. For our
part, computational mechanics led to a new theory of
structure in disordered materials [28] and to measuring
novel information processing in neural spike trains [30].
Looking forward to future engineering of complex systems, the structural understanding developed in these
applications moves us in a direction to design novel semiconducting materials for a new generation of computing
technology and the next generation of nanoscale spiketrain probes that will scale to monitoring the activity of
thousands of neurons [46].
ACKNOWLEDGMENTS
Finally, let’s end on a practical note. We are advocates
for a broad, even pluralistic approach to automated nonlinear model building—i.e., for artificial science. However, our repeated experience is that general “frameworks” seriously stub their toes on application to experi-
The authors thank the Santa Fe Institute for its hospitality during visits and Chris Ellison, Dave Feldman,
and John Mahoney for helpful comments. JPC is an
SFI External Faculty member. This material is based
upon work supported by, or in part by, the U.S. Army
Research Laboratory and the U. S. Army Research Office under contracts W911NF-13-1-0390, W911NF-13-10340, and W911NF-12-1-0288. S.M. is a National Science Foundation Graduate Student Research Fellow and
a U.C. Berkeley Chancellor’s Fellow.
[1] A. N. Gorban, N. Kazantzis, I. G. Kevrekidis, H. C. Ottinger, and C. Theodoropoulos, editors. Model Reduction
and Coarse-Graining Approaches for Multiscale Phenomena. Series in Complexity. Springer, 2006.
[2] W. H. A. Schilders, H. A. van der Vorst, and J. Rommes.
Model Order Reduction: Theory, Research Aspects and
Applications, volume 13 of Mathematics in Industry.
Springer, 2008.
[3] J. P. Crutchfield. The dreams of theory. WIRES Comp.
Stat., 6(March/April):75–79, 2014.
[4] D. H. Wolpert, J. A. Grochow, E. Libby, and S. DeDeo. A
framework for optimal high-level descriptions in science
and engineering—Preliminary report, 2014. arxiv.org:
1409.7403.
[5] W. Heisenberg. Nonlinear problems in physics. Physics
Today, 20:23–33, 1967.
[6] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S.
Shaw. Geometry from a time series. Phys. Rev. Let.,
45:712, 1980.
[7] A. Brandstater, J. Swift, Harry L. Swinney, A. Wolf,
J. D. Farmer, E. Jen, and J. P. Crutchfield. Lowdimensional chaos in a hydrodynamic system. Phys. Rev.
Lett., 51:1442, 1983.
[8] D. Ruelle and F. Takens. On the nature of turbulence.
Comm. Math. Phys., 20:167–192, 1971.
[9] L. D. Landau and E. M. Lifshitz, editors. Fluid Mechanics. Pergamon Press, Oxford, United Kingdom, 1959.
[10] J. P. Crutchfield and B. S. McNamara. Equations of
motion from a data series. Complex Systems, 1:417 –
452, 1987.
[11] J. D. Farmer and J. Sidorowitch. Predicting chaotic time
series. Phys. Rev. Lett., 59:366, 1987.
VIII.
RUBBER, MEET ROAD
6
[12] J. P. Crutchfield and K. Young. Inferring statistical complexity. Phys. Rev. Let., 63:105–108, 1989.
[13] J. P. Crutchfield. The calculi of emergence: Computation, dynamics, and induction. Physica D, 75:11–54,
1994.
[14] J. P. Crutchfield. Between order and chaos. Nature
Physics, 8(January):17–24, 2012.
[15] M. C. Cross and P. C. Hohenberg. Pattern formation
outside of equilibrium. Rev. Mod. Phys., 65(3):851–1112,
1993.
[16] We describe them as mesoscopic states—a structural
level intermediate between raw system configurations
(microstates) and the macrostates of thermodynamics.
[17] S. Marzen and J. P. Crutchfield. Circumventing the curse
of dimensionality in prediction: Causal rate-distortion for
infinite-order markov processes. 2014. SFI Working Paper 14-12-047; arxiv.org:1412.2859 [cond-mat.stat-mech].
[18] C. E. Shannon. A mathematical theory of communication. Bell Sys. Tech. J., 27:379–423, 623–656, 1948.
[19] C. E. Shannon. Coding theorems for a discrete source
with a fidelity criterion. IRE National Convention
Record, Part 4, 7:142–163, 623–656, 1959.
[20] J. Rissanen. Stochastic complexity and modeling. Ann.
Statistics, 14:1080, 1986.
[21] S. Still and J. P. Crutchfield. Structure or noise? 2007.
Santa Fe Institute Working Paper 2007-08-020; arxiv.org
physics.gen-ph/0708.0654.
[22] S. Still, J. P. Crutchfield, and C. J. Ellison. Optimal
causal inference: Estimating stored information and approximating causal architecture. CHAOS, 20(3):037111,
2010.
[23] S. Still. Information bottleneck approach to predictive
inference. Entropy, 16:968–989, 2014.
[24] J. P. Crutchfield and D. P. Feldman. Statistical complexity of simple one-dimensional spin systems. Phys. Rev.
E, 55(2):R1239–R1243, 1997.
[25] D. P. Feldman and J. P. Crutchfield. Discovering noncritical organization: Statistical mechanical, information
theoretic, and computational views of patterns in simple
one-dimensional spin systems. 1998. Santa Fe Institute
Working Paper 98-04-026.
[26] C. R. Shalizi, K. L. Shalizi, and J. P. Crutchfield. Pattern
discovery in time series, Part I: Theory, algorithm, analysis, and convergence. 2002. Santa Fe Institute Working
Paper 02-10-060; arXiv.org/abs/cs.LG/0210025.
[27] C. C. Strelioff and J. P. Crutchfield. Bayesian structural inference for hidden processes. Phys. Rev. E,
89:042119, 2014. Santa Fe Institute Working Paper 1309-027, arXiv:1309.1392 [stat.ML].
[28] D. P. Varn and J. P. Crutchfield. Chaotic crystallography: How the physics of information reveals structural order in materials. Curr. Opin. Chem. Eng., 7:47–
56, 2015. Santa Fe Institute Working Paper 14-09-036;
arxiv.org:1409.5930 [cond-mat].
[29] F. Rieke, D. Warland, R. de Ruyter van Steveninck, and
W. Bialek. Spikes: Exploring the Neural Code. Bradford
Book, New York, 1999.
[30] S. Marzen and J. P. Crutchfield. Time resolution dependence of continuous-time information measures. in
preparation, 2014.
[31] C. R. Shalizi and J. P. Crutchfield. Information bottlenecks, causal states, and statistical relevance bases: How
to represent relevant information in memoryless transduction. Adv. Compl. Sys., 5(1):91–95, 2002.
[32] R. G. James, C. J. Ellison, and J. P. Crutchfield.
Anatomy of a bit: Information in a time series observation. CHAOS, 21(3):037109, 2011.
[33] W. Bialek, I. Nemenman, and N. Tishby. Complexity
through nonextensivity. Physica A, 302:89–99, 2001.
[34] N. Travers and J. P. Crutchfield. Infinite excess entropy processes with countable-state generators. Entropy, 16:1396–1413, 2014. SFI Working Paper 11-11-052;
arxiv.org:1111.3393 [math.IT].
[35] M. Tchernookov and I. Nemenman. Predictive information in a nonequilibrium critical model. J. Stat. Phys.,
153:442–459, 2013.
[36] W. Bialek, I. Nemenman, and N. Tishby. Predictability,
complexity, and learning. Neural Computation, 13:2409–
2463, 2001.
[37] T. M. Cover and J. A. Thomas. Elements of Information
Theory. Wiley-Interscience, New York, second edition,
2006.
[38] V. Kostina and S. Verdu. Fixed-length lossy compression
in the finite blocklength regime. IEEE Trans. Info. Th.,
58(26):3309–3338, 2012.
[39] V. Kostina and S. Verdu. Lossy joint source-channel coding in the finite blocklength regime. IEEE Trans. Info.
Th., 59(5):2545–2574, 2013.
[40] J. P. Crutchfield, P. Riechers, and C. J. Ellison. Exact
complexity: Spectral decomposition of intrinsic computation. submitted. Santa Fe Institute Working Paper
13-09-028; arXiv:1309.3792 [cond-mat.stat-mech].
[41] C. R. Shalizi and J. P. Crutchfield. Computational mechanics: Pattern and prediction, structure and simplicity.
J. Stat. Phys., 104:817–879, 2001.
[42] K. Rose. A mapping approach to rate-distortion computation and analysis. IEEE Trans. Info. Th., 40(6):1939–
1952, 1994.
[43] D. Pfau, N. Bartlett, and F. Wood. Probabilistic deterministic infinite automata. In Adv. Neural Info. Proc.
Sys., pages 1930–1938. MIT Press, 2011.
[44] D. P. Varn, G. S. Canright, and J. P. Crutchfield. ǫMachine spectral reconstruction theory: A direct method
for inferring planar disorder and structure from X-ray
diffraction studies. Acta. Cryst. Sec. A, 69(2):197–206,
2013.
[45] S. E. Palmer, O. Marre, M. J. Berry II, and W. Bialek.
Predictive information in a sensory population. 2013.
arXiv:1307.0225.
[46] A. P. Alivisatos, M. Chun, G. M. Church, R. J.
Greenspan, M. L. Roukes, and R. Yuste. The brain activity map project and the challenge of functional connectomics. Neuron, 74:970–974, 2012.
| 5cs.CE
|
The impact of Entropy and Solution Density on
selected SAT heuristics
Dor Cohen and Ofer Strichman
arXiv:1706.05637v1 [cs.AI] 18 Jun 2017
Information systems engineering, IE, Technion, Haifa, Israel
Abstract. In a recent article [4], Oh examined the impact of various key
heuristics (e.g., deletion strategy, restart policy, decay factor, database
reduction) in competitive SAT solvers. His key findings are that their
expected success depends on whether the input formula is satisfiable or
not. To further investigate these findings, we focused on two properties
of satisfiable formulas: the entropy of the formula, which approximates
the freedom we have in assigning the variables, and the solution density,
which is the number of solutions divided by the search space. We found
that both predict better the effect of these heuristics, and that satisfiable
formulas with small entropy ‘behave’ similarly to unsatisfiable formulas.
1
Introduction
In a recent article [4], Oh examined the impact of various key heuristics in
competitive SAT solvers. His key findings are that the average success of those
heuristics depends on whether the input formula is satisfiable or not. In particular the effect of the deletion strategy, restart policy, decay factor, and database
reduction is different, on average, between satisfiable and unsatisfiable formulas. This observation can be used for designing solvers that specialize in one of
them, and for designing a hybrid solver that alternates between SAT / UNSAT
‘modes’. Indeed certain variants of COMiniSatPS [4] work this way.
We do not see an a priory reason to believe that the SAT/UNSAT divide—
corresponding to the distinction between zero or more solutions—explains best
the differences in the effect of the various heuristics.1 In this work we investigate
further his findings, and show empirically that there are more refined measures
(properties) than the satisfiability of the formula, that predict better the effectiveness of these heuristics. In particular, we checked how it correlates with two
measures of satisfiable formulas: the entropy of the formula (to be defined below), which approximates the freedom we have in assigning the variables, and the
solution density (henceforth density), which is the number of solutions divided
by the search space. Our experiments show that both are strongly correlated to
the effectiveness of the heuristics, but the entropy measure seems to be a better predictor. Generally our findings confirm Oh’s observations regarding which
heuristic works better with satisfiable formulas. But we also found that satisfiable formulas with small entropy ‘behave’ similarly to unsatisfiable formulas.
1
While proving Unsat and Sat belong to separate complexity classes, there is no
known connection of this fact to effectiveness of heuristics.
Fig. 1. (left) Depicting the entropy function (1), for a satisfiable formula with 11
solutions. (right) The distribution of e(v) of a formula with 100 variables.
2
Entropy
Let ϕ be a propositional CNF formula, var(ϕ) its set of variables and lit(ϕ) its set
of literals. In the following we will use v, v̄ to denote the literals corresponding to
a variable v when the distinction between variables and literals is clear from the
context. If ϕ is satisfiable, we denote by r(l), for l ∈ lit(ϕ), the ratio of solutions
to ϕ that satisfy l. Hence for all v ∈ var(ϕ), it holds that r(v) + r(v̄) = 1. We
now define:
Definition 1 (variable entropy). For a satisfiable formula ϕ, the entropy of
a variable v ∈ var(ϕ) is defined by
.
e(v) = −r(v) log2 r(v) − r(v̄) log2 r(v̄) .
(1)
where 0 · log2 0 is taken as being equal to 0.
This definition is inspired by Shannon’s definition of entropy in the context of
information theory [6]. Figure 1 (left) depicts (1).
Intuitively, entropy reflects how ‘balanced’ a variable is with respect to the
solution space of the formula. In particular e(v) = 0 when r(v) = 0 or r(v) = 1,
which means that ϕ ⇒ v̄ or ϕ ⇒ v, respectively. In other words, e(v) = 0 implies
that v is a backbone variable, since its value is implied by the formula. The other
extreme is e(v) = 1; this happens when r(v) = r(v̄) = 0.5, which means that v
and v̄ appear an equal number of times in the solution space.
Definition 2 (formula entropy). The entropy of a satisfiable formula is the
average entropy of its variables.
As an example, Fig. 1 (right) is a histogram of e(v) for a particular formula
ϕ, where for 24 out of the 100 variables r(v) = 0.
Entropy is hard to compute : Let #SAT (ϕ) denote the number of solutions
a formula ϕ has. Then it is easy to see that
r(v) =
#(ϕ ∧ v)
#ϕ
and
r(v̄) = 1 − r(v) .
(2)
Hence computing e(v) amounts to two calls to a model counter. But since the
denominator #ϕ is fixed for ϕ, computing e(ϕ) amounts to |var(ϕ)| + 1 calls to
a model counter. Since model counting is a #P problem, we can only compute
this value for small formulas.
The benchmark set : Using the model-counter Cachet [5], we computed
the precise entropy of 5000 3-SAT random formulas with 100 variables and 400
clauses. These are formulas taken from SAT-lib, in which the number of backbone
variables is known. Specifically, there is an equal number of formulas in this set
with 10,30,50,70 and 90 backbone variables (i.e., a 1000 formulas of each number
of backbone variables), which gave us a near-uniform distribution of entropy
among the formulas.
3
A preliminary: standardized linear regression
We assume the reader is somewhat familiar with linear regression. It is a standard
technique for building a linear model ŷ = β̂0 + β̂1 x, where ŷ in our case is a
predictor of the number of conflicts, and x is either the entropy or the density
of the formula. We will focus on two results of linear regression: the value of
β̂1 and the p-value. The latter is computed with respect to a null hypothesis,
denoted H0 , that β̂1 = 0, and an alternative hypothesis H1 . H1 can be either the
complement of H0 (β̂1 6= 0) or a ‘one-sided hypothesis’, e.g., H1 : β̂1 > 0. In the
former case, p = 2P r(Z ≤ z | H0 ), where Z ∼ N (0, 1) and z = β̂1 −0 . The ‘0’
std(β̂1 )
in the numerator comes from the specific value in H0 . In other words, assuming
H0 is correct, the p-value indicates the probability that a random value from a
standard normal distribution N (0, 1), is less than z, the standardized value of
β̂1 . In the latter case p = P r(Z ≤ z | H0 ).
We list below several important points about the analysis that we applied.
.
– Standardization of the data: given data points X = x1 , . . . , xn , their stan.
dardization X 0 = x01 , . . . , x0n is defied for 1 ≤ i ≤ n by
xi − x̄
,
σ
where x̄ is the average value of X and σ is its standard deviation. Now X 0
has no units, and hence two standardized sets of data are comparable even
if they originated from different types of measures (in our case, entropy and
density). All the data in our experiments was standardized.
– Bootstrapping: Bootstrapping, parameterized by a value k, is a well-known
technique for improving the precision of various statistics, such as the confidence interval. Technically, bootstrap is applied as follows: Given the original
n samples, uniformly sample it n times with replacement (i.e., without taking
the sampled points out, which implies that the same point can be selected
more than once); repeat this process k times. Hence we now have n · k data
points. For our experiments we took k = 1000, which is a rather standard
value when using this technique. Hence, we have 5 · 106 data points.
x0i =
– Two regression tests: The entropy and density data consists of pairs of the
form hentropy, conf licts[i]i, and hdensity, conf licts[i]i, respectively, where
i ∈ {1, 2} is the index of the heuristic. Hence the corresponding data is four
series of points (e1 , c1 [i]), . . . , (en , cn [i]), and (d1 , c1 [i]), . . . , (dn , cn [i]), where
i ∈ {1, 2}. In order to compare the predictive power of entropy, density and
Oh’s criterion of SAT/UNSAT, we performed two statistical tests (recall that
the data is standardized, and hence comparable):
• The ∆ test: A linear regression test over the series (e1 , c1 [1] − c1 [2]) . . .
(en , cn [1] − cn [2]), and the series (d1 , c1 [1] − c1 [2]) . . . (dn , cn [1] − cn [2]).
• The ∆β̂1 test: A linear regression test over the series (e1 , c1 [1]) . . . (en , cn [1])
and (e1 , c1 [2]) . . . (en , cn [2]), and similarly for density (i.e., four tests all
together). We then checked the significance of β̂1 for each of these 4 tests
(in all such tests the significance was clear). In addition, we checked the
hypothesis H0 : β̂1 [1] − β̂1 [2] = 0 for each of the measures. The result of
this last test is what we will list in the results table in Appendix B.
Intuitively, the two models tell us slightly different things: the first tells us
whether the gap between the two heuristics is correlated with the measure,
and the second tells us whether there is a significant difference in the value
of β̂1 (the slope of the linear model) between the two heuristics. As we will
see in the results, the p-value obtained by these models can be very different.
– Plots: The plots are based on the original (non-standardized) data. To reduce the clutter (from 5000 points), we rounded all values to 2 decimal points
and then aggregated them. Aggregation means that points (x, y1 ) . . . (x, yn )
(i.e., n points with an equal xvalue) are replaced with a single point (x,
avg(y1 . . . yn )). However the trend-lines in the various plots are depicted according to the original data, before rounding and aggregation. The statistical
significance of these trend-lines appears in Appendix B.
4
Entropy and density predict hardness
We checked the correlation between hardness, as measured by the number of
conflicts, and the two measures described above, namely entropy and density. We
use the number of conflicts as a proxy of the run-time, because these are all easy
formulas for SAT, and hence the differences in run-time are rather meaningless.
The two plots in Fig. 2 depict this data based on our experiments with the solver
MiniSat-HACK-999ED. It is apparent that higher entropy and higher density
imply a smaller number of conflicts. A detailed regression analysis appears in
Appendix A, for seven solvers.
We also checked the correlation between the two measures themselves: perhaps formulas with higher entropy also have a higher density (each variable v
with high entropy, e.g., e(v) = 1, nearly doubles the number of solutions). It
turns out that in our benchmarks these two measures are not correlated: the
confidence-interval for β̂1 is [0.144–0.156] with a p-value which is practically 0.
Fig. 2. Entropy (left) and density (right) as predictors of the number of conflicts (based
on MiniSat-HACK-999ED). It is apparent that higher entropy and higher density imply
a smaller number of conflicts.
5
Empirical findings
In this section we describe each of the experiments of Oh [4], and our own version
of the experiment based on entropy and density, when applied to the benchmarks
mentioned above. We omit the details of one experiment, in which Oh examined
the effect of canceling database reduction, the reason being that this heuristic
is only activated after 2000 conflicts, and most of our benchmarks are solved
before that point.2 Raw data as well as charts and regression analysis of our full
set of experiments can be found online in [1].
1. Deletion strategy: Different solvers use different criteria for selecting the
learned clauses for deletion. It was shown in [4] that for SAT instances learned
clauses with low Literal Block Distance (LBD) [2] value can help, whereas others
have no apparent effect. In one of the experiments, whose results are copied
here at the top part of Fig. 3, Oh compared the criterion of ‘core LBD-cut’3
5 and clause size 12. In other words, either save (i.e., do not delete) clauses
with an LBD-cut of 5 and lower, or clauses with size 12 or lower. It shows that
for UNSAT instances the former is better, whereas the opposite conclusion is
reached for the SAT instances. The results of our own experiments are depicted
at the bottom of the figure. They show that the latter is indeed slightly better
with our benchmarks (all satisfiable, recall). But what is more important, is
that the difference becomes smaller with lower entropy—hence the decline of
the trend-line (recall that the trend-lines are based on the raw data, whereas the
diagram itself is computed after rounding and aggregation to improve visibility).
Hence it is evident that formulas with small entropy ‘behave’ more similar to
unsat formulas. The ascending trend-line in the right figure shows, surprisingly,
an opposite effect of density.
2. Deletion with different LBD-cut value Related to the previous heuristic, in [4] it was found that deletion based on larger LBD-cut values, up to a
2
3
Our attempt to use an approximate model-counter with larger formulas failed: the
inaccuracies were large enough to make the analysis show results that are senseless.
An LBD-cut is the lowest value of LBD a learned clause had so far, assuming this
value is recalculated periodically.
Fig. 3. The effect of the deletion criterion. The results of [4] appear in the table at the
top of the figure (the numbers indicate the solved instances). It shows that for SAT
instances keeping everything with clause size 12 is better than keeping everything with
LBD 5, whereas the result is opposite for the UNSAT instances. Our own experiments
(bottom left) show that within SAT instances, the clause size criterion becomes better
with higher entropy, but not with higher density. Note that the y-axis corresponds to
the difference in the # of conflicts. On average on these instances the # of conflicts
across both methods was ≈ 290.
point, improve the performance of the solver with unsat formulas, but not with
SAT ones. Fig. 4 (top) is an excerpt from his results for various LBD-cut values.
We repeated his experiment with LBD-cut 1 and LBD-cut 5. The plots show that
lower values of entropy and (independently) lower values of density yield a bigger advantage to LBD-cut 5, which again demonstrates that satisfiable formulas
with these values ‘behave’ similarly to unsat formulas.
3. Restarts policy: The Luby restart strategy [3] is based on a fixed sequence of time intervals, whereas the Glucose restarts are more rapid and
dynamic. It initiates a restart when the solver identifies that learned clauses
have higher LBD than average. According to the competitions’ results this is
generally better in unsat instances. Oh confirmed the hypothesis that this is related to the restart strategy: indeed his results show that for satisfiable instances
Luby restart is better.
Our own results can be seen in Fig. 5 and in Appendix B. The fact that
the gap in the number of conflicts between Luby and Glucose-style restarts is
negative, implies that the former is generally better, which is consistent with Oh’s
results for satisfiable formulas. Observe that the trend-line slightly declines with
entropy (β̂1 = −15), which implies that Glucose restarts are slightly better with
low entropy. So again we observe that low entropy formulas ‘behave’ more similar
to UNSAT formulas than those that have high entropy. The table in Appendix B
shows that this result has a relatively high p-value. We speculate that with
high-entropy instances, the solver hits more branches that can be extended to
a solution, hence Glucose’s rapid restarts can be detrimental. Density seems to
have an opposite effect, although again only with low statistical confidence.
Fig. 4. The results of [4] (top) show that unsat formulas are solved faster with high
LBD-cut. Our results (bottom) show that low-entropy and low-density formulas behave
more similarly to unsat formulas.
4. The variable decay factor: The well-known VSIDS branching heuristic
is based on an activity score of literals, which decay over time, hence giving higher
priority to literals that appear in recently-learned clauses. In the solver MiniSat HACK 999ED, there is a different decay factor for each of the two restart
phases: this solver alternates between a Glucose-style (G) restart policy phase
and a no-restart (NR) phase (these two phases correspond to good heuristics
for SAT and UNSAT formulas, respectively). In [4] Oh compares different decay
factors for each of these restart phases, on top of MiniSat HACK 999ED. His
results show that for UNSAT instances slower decay gives better performance,
while for SAT instances it is unclear. His results appear at the top of Fig. 6. We
experimented with the two extreme decay factors in that table: 0.95 and 0.6.
Note that since our benchmarks are relatively easy, the solver never reaches the
NR phase. The plot at the bottom of the figure shows the gap in the number of
conflicts between these two values. A higher value means that with strong decay
(0.6) the results are worse. We can see that the results are worse with strong
decay when the entropy is low, which demonstrates again that the effect of the
variable decay factor is similar for unsat formulas and satisfiable formulas with
low entropy. A similar phenomenon happens with small density.
Conclusions : We defined the entropy property of satisfiable formulas, and used
it, together with solution density, to further investigate the results achieved by
Oh in [4]. We showed that both are strongly correlated with the difficulty of
solving the formula (as measured by the number of conflicts). Furthermore, we
showed that they predict better the effect of various SAT heuristics than Oh’s
sat/unsat divide, and that satisfiable formulas with small entropy ‘behave’ similarly to unsatisfiable formulas. Since both measures are hard to compute we do
not expect these results to be applied directly (e.g., in a portfolio), but perhaps
future research will find ways to cheaply approximate them. For example, a high
Fig. 5. The effect of the restart strategy, comparing Luby and Glucose-style restarts.
The results of [4] (top) show that the Glucose strategy (rapid restarts) has an advantage
in unsat formulas. Our results (bottom) show that the same phenomenon is apparent
in formulas with low entropy. Indeed observe that the number of conflicts with Glucose
becomes smaller than it is with Luby (hence the negative gap), in satisfiable formulas
with low entropy.
Fig. 6. The effect of variable decay: the results of [4] (top) generally show that unsat
formulas are better solved with a high decay factor. The restart policy in his solver
is hybrid: it alternates between a ‘no-restart’ (NR) phase and a ‘Glucose’ (G) phase.
The ‘NR’ and ‘G’ columns hold the decay factor during these phases. The plots at the
bottom show the gap in the number of conflicts between G = 0.6 and G = 0.95. It
shows that with low entropy, strong decay (i.e., G = 0.6) is worse, similar to the effect
that it has on unsat formulas. With low density (right) a similar effect is visible.
backbone count (variables with a value at decision level 0) may be correlated to
low entropy, because such variables contribute 0 to the formula’s entropy.
Acknowledgement We thank Dr. David Azriel for his guidance regarding
statistical techniques.
References
1. Full experimental results. http://ie.technion.ac.il/~ofers/entropy/supp.zip.
2. G. Audemard and L. Simon. Predicting learnt clauses quality in modern SAT
solvers. In C. Boutilier, editor, IJCAI 2009, Proceedings of the 21st International
Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17,
2009, pages 399–404, 2009.
3. M. Luby, A. Sinclair, and D. Zuckerman. Optimal speedup of las vegas algorithms.
Inf. Process. Lett., 47(4):173–180, 1993.
4. C. Oh. Between SAT and UNSAT: the fundamental difference in CDCL SAT. In
SAT, volume 9340 of LNCS, pages 307–323, 2015.
5. T. Sang, F. Bacchus, P. Beame, H. A. Kautz, and T. Pitassi. Combining component
caching and clause learning for effective model counting. In SAT, 2004.
6. C. E. Shannon. A mathematical theory of communication. Bell System Technical
Journal, 27(3):379423, 1948.
A
Predicting hardness: a regression analysis
Denote by β̂1E and β̂1S the β̂1 -value of the linear models for entropy vs. conflicts
and density vs. conflicts, respectively. The table below shows strong correlation
between both measures to the number of conflicts (the p-value in both cases, for
all engines, is practically 0). The Last two columns show the gap β̂1E − β̂1S and the
corresponding p-value for H0 : β̂1E − β̂1S = 0, H1 : β̂1E − β̂1S 6= 0, when measured
across the k = 1000 iterations of the bootstrap method that was described in
Sec. 3. For engines with high p-value we cannot reject H0 with confidence.
Solver
β̂1E
β̂1S
β̂1E − β̂1S
p-value
MiniSat-HACK-999ED (-84.29, -72.58 ) (-84.93, -73.56 ) ( 5.37, 16.96 ) 0.716
MiniSat-HACK-999ED
(modified to luby) (-86.31, -75.36 ) (-82.97, -72.64 ) (-7.51, 1.44 ) 0.200
MiniSat-HACK-999ED
(modified for 2 phases) (-72.84, -63.61 ) (-72.31, -62.91 ) (-4.80, 3.57 ) 0.738
SWDiA5BY
(-91.61, -79.17 ) (-90.97, -78.77 ) (-5.95, 4.92 ) 0.84
COMiniSatPS
(-74.68, -64.58 ) (-75.41, -65.43 ) (-3.79, 5.37 ) 0.76
lingeling-ayv
(-76.19, -66.61 ) (-71.70, -61.76 ) (-8.99, -0.35 ) 0.029
Glucose
(-91.24, -79.34 ) (-90.56, -78.88 ) (-6.00, 4.85 ) 0.845
Table 1. For each solver, we list the 95% confidence interval of its β̂1E (entropy) and β̂1S
(solutions). For all engines the corresponding p-value is practically 0 (i.e.,. ≤ 10−100 ).
The last two columns refer to the gap between these measures.
B
Regression-tests results
The table below lists the confidence interval and corresponding p-value, for the
two regression tests ∆ and ∆β̂1 (in the latter we also list the results for β̂0 ) that
were explained in Sec. 3, and the four experiments described in Sec. 5. H1 is
one-sided.
Exp. Measure Conf. interval p-val Conf. interval p-val Conf. interval p-val
(∆)
(∆β̂1 )
(∆β̂0 )
1
2
3
4
Entropy
Density
Entropy
Density
Entropy
Density
Entropy
Density
(-2.76, 2.64 )
(-0.81, 4.35 )
(-3.72, 0.25 )
(-3.40, 0.59 )
(-8.31, 3.52 )
(-4.41, 7.36 )
(-15.1, -10.6 )
(-3.92, 0.60 )
0.48 (-2.75, 2.46)
0.09 (-0.83, 4.43)
0.04 (-3.78, 0.25 )
0.09 (-3.34, 0.69 )
0.22 (-8.01, 3.67 )
0.30 (-4.34, 7.50 )
0
(-15.1, -10.7 )
0 (-13.60, -8.86 )
0.05
0.39
0.39
0.47
0.001
0.05
0.125
0.475
(-12.06, -6.60)
0
(-12.06, -6.64)
0
(0.48, 4.61 ) 0.01
(0.47, 4.56 ) 0.01
(-36.12, -23.78 ) 0
(-35.99, -23.90 ) 0
(20.99, 25.44 ) 0
(20.96, 25.47 ) 0
Table 2. Regression-tests results for the four experiments in Sec. 5. p-value ≤ 10−10
are rounded to 0.
| 2cs.AI
|
arXiv:0906.3485v4 [math.AC] 10 Dec 2013
The Uniformization of Certain Algebraic
Hypergeometric Functions
Robert S. Maier
Depts. of Mathematics and Physics, University of Arizona, Tucson, AZ 85721, USA
Abstract
The hypergeometric functions n Fn−1 are higher transcendental functions, but
for certain parameter values they become algebraic, because the monodromy
of the defining hypergeometric differential equation becomes finite. It is shown
that many algebraic n Fn−1 ’s, for which the finite monodromy is irreducible but
imprimitive, can be represented as combinations of certain explicitly algebraic
functions of a single variable; namely, the roots of trinomials. This generalizes a result of Birkeland, and is derived as a corollary of a family of binomial
coefficient identities that is of independent interest. Any tuple of roots of a
trinomial traces out a projective algebraic curve, and it is also determined when
this so-called Schwarz curve is of genus zero and can be rationally parametrized.
Any such parametrization yields a hypergeometric identity that explicitly uniformizes a family of algebraic n Fn−1 ’s. Many examples of such uniformizations
are worked out explicitly. Even when the governing Schwarz curve is of positive
genus, it is shown how it is sometimes possible to construct explicit single-valued
or multivalued parametrizations of individual algebraic n Fn−1 ’s, by parametrizing a quotiented Schwarz curve. The parametrization requires computations in
rings of symmetric polynomials.
Key words: hypergeometric function, algebraic function, imprimitive
monodromy, trinomial equation, binomial coefficient identity, algebraic curve,
Belyi cover, uniformization, symmetric polynomial
2000 MSC: 33C20, 33C80, 14H20, 14H45, 05E05
1. Introduction
The hypergeometric functions n Fn−1 (ζ), n > 1, are parametrized special
functions of fundamental importance. Each n Fn−1 (ζ) is a function of a single complex variable, and in general is a higher transcendental function. It
is parametrized by complex numbers a1 , . . . , an ; b1 , . . . , bn−1 , and is written as
Email address: [email protected] (Robert S. Maier)
URL: http://math.arizona.edu/~rsm (Robert S. Maier)
Preprint submitted to Elsevier
December 18, 2017
n Fn−1 (a1 , . . . , an ; b1 , . . . , bn−1 ; ζ).
n Fn−1
a1 , . . . , an
b1 , . . . , bn−1
It is analytic on |ζ| < 1, with definition
ζ
=
∞
X
k=0
(a1 )k · · · (an )k
ζk.
(b1 )k · · · (bn−1 )k (1)k
(1.1)
Here (c)k := c(c + 1) · · · (c + k − 1), and the lower parameters b1 , . . . , bn−1 may
not be non-positive integers. The n = 1 function 1 F0 (a1 ; –; ζ) equals (1 − ζ)−a1 ,
and the n = 2 function 2 F1 (a1 , a2 ; b1 ; ζ) is the Gauss hypergeometric function.
If its parameters are suitably chosen, n Fn−1 (ζ) will become an algebraic
function of ζ. Equivalently, if one regards n Fn−1 as a single-valued function on
a certain Riemann surface, defined by continuation from the disk |ζ| < 1, then
the surface will become compact. If n = 1, this occurs when a1 ∈ Q. In the first
nontrivial case n = 2, the characterization of the triples (a1 , a2 ; b1 ) for which
2 F1 (a1 , a2 ; b1 ; ζ) is algebraic is a classical result of Schwarz. There is a finite list
of possible normalized triples (the famous ‘Schwarz list’), and 2 F1 is algebraic
iff (a1 , a2 ; b1 ) is a denormalized version of a triple on the list. Denormalization
involves integer displacements of the parameters. For specifics, see [10, § 2.7.2].
More recently, Beukers and Heckman [3] treated n > 3, and obtained a
complete characterization of the parameters (a1 , . . . , an ; b1 , . . . , bn−1 ) that yield
algebraicity. Like the n = 2 result of Schwarz, their result was based on the fact
that the function n Fn−1 (a1 , . . . , an ; b1 , . . . , bn−1 ; ζ) satisfies an order-n differential equation on the Riemann sphere P1ζ , called En (a1 , . . . , an ; b1 , . . . , bn−1 , 1)ζ
below. In modern language, En specifies a flat connection on an n-dimensional
vector bundle over P1ζ . It has singular points at ζ = 0, 1, ∞, and its (projective) monodromy group is generated by loops about these three points. The
monodromy, resp. projective monodromy group is a subgroup of GLn (C), resp.
PGLn (C), and algebraicity occurs iff the monodromy is finite. Schwarz exploited
the classification of the finite subgroups of PGL2 (C). In a tour de force, Beukers
and Heckman handled the n > 3 case by exploiting the Shephard–Todd classification of the finite subgroups of GLn (C) generated by complex reflections. Their
characterization result, however, is non-constructive: it supplies an algorithm
for determining whether a given function n Fn−1 (a1 , . . . , an ; b1 , . . . , bn−1 ; ζ) is
algebraic in ζ, but it does not yield a polynomial equation (with coefficients
polynomial in ζ) satisfied by the function.
In this paper, a new approach is taken to the problem of constructing equations satisfied by algebraic n Fn−1 ’s. Several classes of hypergeometric function,
known to be algebraic by Beukers–Heckman, are made explicit by being uniformized. Recall that an algebraic function F may be of genus zero, i.e., may
have a ‘uniformization,’ or parametrization, by rational functions. That is, one
may have F (R1 (t)) = R2 (t) for certain rational functions R1 , R2 , so that formally, F = R2 ◦ R1−1 . In this case the Riemann surface of F = F (ζ), on which ζ
and F are single-valued meromorphic functions, is isomorphic to the Riemann
sphere P1t . A sample result of this nature, obtained below, is the following. Let
2
n = p + 2, where p > 1 is odd. Then
n Fn−1
!
4nn t2 (1 − t2 )2p [(1 + t)p + (1 − t)p ]p
n
pp
[(1 + t)n + (1 − t)n ]
(1 + t)n + (1 − t)n a
. (1.2)
= 12 (1 + t)−2a + (1 − t)−2a
(1 + t)p + (1 − t)p
a+(n−1)
a
n, . . . ,
n
a+p 1
a+1
p ,..., p , 2
Equation (1.2) holds as an equality for all t in a neighborhood of t = 0, or equivalently if one expands each side about t = 0, as an equality between power series.
It is a hypergeometric identity, which is more general than a uniformization of
a single F . This is because a ∈ C is arbitrary. (If any lower hypergeometric
parameter is a non-positive integer, the identity must be interpreted in a lim-
a+p 1
iting sense.) If a ∈ Q, it follows that n Fn−1 na , . . . , a+(n−1)
; a+1
n
p , . . . , p , 2; ζ
is algebraic in ζ. If moreover a ∈ Z, in which case (1.2) is a uniformization in
the strict sense, this algebraic n Fn−1 must be of genus zero.
Most of the algebraic n Fn−1 ’s parametrized below, or more specifically uniformized, are of the following type. The n Fn−1 in (1.2) is the q = 2 case of
!
a+(n−1)
a
,
.
.
.
,
n
n
ζ ,
(1.3)
n Fn−1
a+p 1
q−1
a+1
p ,..., p , q,..., q
where n = p + q. The order-n differential equation En on P1ζ associated to (1.3),
En
a+(n−1) a+1
q−1
1
a
; p , . . . , a+p
n, . . . ,
n
p , q,..., q ,1 ζ
,
(1.4)
plays a role in the Beukers–Heckman analysis of algebraicity (see § 2). The
treatment of (1.3) and (1.4) relies on the following fact. If gcd(p, q) = 1, the
solution space of (1.4) is spanned by xγ1 (ζ), . . . , xγn (ζ), where γ := −qa and the
n algebraic functions x1 (ζ), . . . , xn (ζ) are the roots of the trinomial equation
xn − g xp − β = 0.
(1.5)
Here g 6= 0 is arbitrary but fixed, and β is determined implicitly by
ζ = (−)q
nn β q
.
pp q q g n
(1.6)
(This statement assumes γ ∈
/ Z.) As defined by (1.6), ζ is a projectivized and
normalized version of the discriminant of (1.5): if β 6= 0, the trinomial has
coincident roots if and only if ζ = 1.
Formulas relating such hypergeometric functions as (1.3) to trinomial roots
can be traced to the 1758 work of Lambert, who expressed such roots as finite
sums of hypergeometric series. (See [2, p. 72].) Once trinomial roots have been
expressed in terms of hypergeometric series, typically by Lagrange inversion,
formulas for hypergeometric functions in terms of trinomial roots can be derived.
Two especially useful reversed formulas of this sort were obtained in the 1920s
3
by Birkeland [4]. (See [1, 27] for general reviews; the classical literature is cited
in [18], and some explicit series for the roots of (1.5) are given in [9, 13].) For
example, Birkeland expressed the algebraic function (1.3) as a combination of
only q of the n functions xγ1 (ζ), . . . , xγn (ζ).
The first result of this paper is the extension of Birkeland’s two formulas to
the case when the n Fn−1 being P
represented
terms of trinomial roots has a
Pin
n−1
n
so-called parametric excess S = i=1 bi − i=1 ai that does not equal 21 , as it
does in (1.3), but rather is an arbitrary half-odd-integer. There are in fact two
families of formulas indexed by ℓ ∈ Z, with S = 21 − ℓ, that subsume Birkeland’s
‘ℓ = 0’ formulas. The new families are shown to follow from a family of binomial
coefficient identities, which is derived first and is of independent interest: it
subsumes several combinatorial identities that are listed individually in [16].
The derivation of the family of binomial coefficient identities resembles that of
Chu [6]. The technique used is not Lagrange inversion (explicitly) but rather
the Vandermonde convolution transform of Gould [15].
The heart of this paper is a study of the trinomial equation (1.5), focused
on basic algebraic geometry rather than on controlling or bounding its roots.
(For the latter, see [11] and papers cited therein.) The key concept is that of a
Schwarz curve. For each coprime p, q > 1 with n := p+q, this is defined following
(n)
Kato and Noumi [21] to be a complex projective curve Cp,q ⊂ Pn−1 comprising
n−1
all [x1 : . . . : xn ] ∈ P
such that x1 , . . . , xn are the roots of some equation of
(n)
the form (1.5). That is, Cp,q is the common zero set of the n − 2 elementary
symmetric polynomials σ1 , . . . , σq−1 and σq+1 , . . . , σn−1 in x1 , . . . , xn . Clearly
g = (−)q−1 σq and β = (−)n−1 σn , giving a formula for ζ; so there is a degree-n!
(n)
(n)
covering Cp,q → P1ζ . The curve Cp,q is irreducible [21, Cor. 4.7].
(k)
For each k, n − 1 > k > 2, a subsidiary Schwarz curve Cp,q ⊂ Pk−1 , also
irreducible, is defined to include all [x1 : . . . : xk ] ∈ Pk−1 such that x1 , . . . , xk
(2)
are k of the n roots of some equation of the form (1.5). The curve Cp,q is P1
(1)
itself, and one can introduce a final curve Cp,q , also isomorphic to P1 . These
(k)
(k)
(k−1)
(k)
curves are joined by maps φp,q : Cp,q → Cp,q with deg φp,q = n − k + 1, i.e.,
(n)
φp,q
(n−1)
φp,q
φ(3)
p,q
φ(2)
p,q
φ(1)
p,q
1
(1)
(n−1)
−→ · · · −→ C(2)
C(n)
p,q −→ Cp,q −→ Pζ .
p,q −→ Cp,q
(1.7)
(k)
Each covering Cp,q → P1ζ is a Belyı̆ map, i.e., is ramified only over ζ = 0, 1, ∞.
(2)
(1)
The Schwarz curves are not in general of genus zero, though Cp,q and Cp,q
(2)
(1)
are. One can write Cp,q ∼
= P1t and Cp,q ∼
= P1s , where t and s are rational
parameters, and make the concrete choice t = (x1 + x2 )/(x1 − x2 ). It is a
(2)
(1)
consequence of the rational parametrizations of Cp,q and Cp,q , substituted into
the appropriate Birkeland-style representation formulas, that when q = 2 or
q = 1, the function n Fn−1 of (1.3) can be parametrized by t or s, with its
argument ζ being respectively a degree-[n(n − 1)] rational function of t, and a
degree-n rational function of s. This turns out to be the origin of the sample
identity (1.2). Its right side is the sum of two terms, proportional to (1 + t)−2a ,
(1 − t)−2a ; and these come from the exponentiated roots xγ1 , xγ2 .
4
The fundamental fact is that if a hypergeometric function n Fn−1 (ζ) can be
represented in terms of a k-tuple of roots of the trinomial equation (1.5) with
(k)
n = p + q, any parametrization of the curve Cp,q will yield a parametrization
of n Fn−1 (ζ); and if this curve is of genus zero, with a rational parameter, the
parametrization of n Fn−1 (ζ) will be by rational functions of the parameter.
Many interesting examples of this are worked out below, going well beyond the
(n)
sample identity (1.2). The curves employed include ‘top’ Schwarz curves Cp,q ,
(k)
as well as subsidiary curves Cp,q with k = 1, 2, 3.
To derive parametrizations of certain algebraic n Fn−1 (ζ)’s, a more sophisticated technique is useful. If there is a representation of an n Fn−1 (ζ) in terms of
k trinomial roots that is symmetric in the roots, the parametrization is really
(k)
by a quotiented Schwarz curve, obtained from Cp,q by quotienting out the sym(k)
metric group Sk . Even if Cp,q is of positive genus, the quotiented curve may
be of genus zero. Quotienting out the cyclic group Ck or the dihedral group Dk
may also be useful. Several examples of explicit parametrizations of algebraic
n Fn−1 ’s that can be obtained by quotienting are worked out, by computation
in rings of symmetric polynomials.
The paper is organized as follows. The first half focuses on hypergeometric equations and series, and series identities. In § 2 the equations En are introduced, and their monodromy is discussed. In § 3, the family of binomial
coefficient identities is derived with the aid of the Vandermonde convolution
transform. In § 4, families of formulas of Birkeland type for various hypergeometric functions, including algebraic ones, are derived from these identities.
The main results are Thm. 4.4, which extends Birkeland’s two representations
and is parametrized by ℓ ∈ Z, and Thm. 4.6. The former expresses n Fn−1 ’s of
a type that generalizes (1.3) in terms of trinomial roots. The latter expresses
certain non-algebraic n+1 Fn ’s in terms of 2 F1 ’s and 3 F2 ’s.
The second half is explicitly algebraic-geometric. In §§ 5 and 6, the Schwarz
curves are introduced and studied. Their ramification structures and genera
are computed, and the curves of genus zero are determined. (Theorem 6.3,
the genus formula, generalizes a theorem of [21].) In § 7, many hypergeometric
identities with a free parameter a ∈ C are derived, including uniformizations of
algebraic n Fn−1 ’s of genus zero. A typical result is Thm. 7.3, of which Eq. (1.2)
(2)
above is a special case; it comes from Cp,2 . In § 8, the quotiented Schwarz curves
are defined and in some cases parametrized, yielding parametrizations in which
a ∈ Q is fixed. Some of these are multivalued parametrizations with radicals.
2. Monodromy and degeneracy
The parametrized functions n Fn−1 (a1 , . . . , an ; b1 , . . . , bn−1 ; ζ) are locally defined by Eq. (1.1). Two sorts of degeneracy are of interest, for the functions
themselves, and for the equations that they satisfy. The first is when an upper
and a lower parameter equal each other. If so, they may be ‘cancelled’, reducing
the n Fn−1 to an n−1 Fn−2 . The second arises when one of the lower parameters is taken to an integer −m, and one of the upper ones to an integer −m′ ,
5
′
with −m 6 −m′ 6 0. Let [ζ >m ]F denote the sum of the terms proportional
to ζ k , k > m′ , in the series for F ; and let (a) and (b) denote (a1 , . . . , an−1 )
and (b1 , . . . , bn−2 ). The following auxiliary lemma permits such hypergeometric
identities as (1.2) to be interpreted in a limiting sense; it will be used in § 8.
Lemma 2.1. Let an → −m′ , bn−1 → −m, with (an + m′ )/(bn−1 + m) → α.
Then
′
(a), an
ζ −→
[ζ >m ] n Fn−1
(b), bn−1
−1
(a)m+1
(a) + m + 1, m − m′ + 1
m−m′ m
m+1
α (−)
ζ .
ζ
n Fn−1
(b) + m + 1, m + 2
m′
(b)m+1 (1)m+1
If m = m′ , this limit simplifies to
α · [ζ
>m′
] n−1 Fn−2
(a)
(b)
ζ .
Proof. Take the term-by-term limit of the series in (1.1).
It is well known that the function n Fn−1 (a1 , . . . , an ; b1 , . . . bn−1 ; ζ), defined
on the disk |ζ| < 1, satisfies the bn = 1 case of the order-n equation Dn F = 0,
where (with Dζ := d/dζ)
Dn = Dn (a1 , . . . , an ; b1 , . . . , bn )
(2.1)
= (ζDζ + b1 − 1) · · · (ζDζ + bn − 1) − ζ (ζDζ + a1 ) · · · (ζDζ + an ).
The equation Dn F = 0 is denoted by En (a1 , . . . , an ; b1 , . . . , bn ) here; a subscript ζ will be added as appropriate to indicate the independent variable. This
equation has three regular singular points on P1ζ , namely ζ = 0, 1, ∞, with respective characteristic exponents 1 − b1 , . . . , 1 − bn ; and 0, 1, . . . , n − 2, S; and
a1 , . . . , an . In this,
!
!
n
n
X
X
(2.2)
ai − 1
bi −
S=
i=1
i=1
Pn−1
P
is the ‘parametric excess,’ which reduces to i=1
bi − ni=1 an if bn = 1. The
canonical solution n Fn−1 , defined if none of b1 , . . . , bn−1 is a non-positive integer,
is analytic at ζ = 0 and belongs to the exponent 1 − bn = 0. Allowing bn to
differ from unity in En is not a major matter, since adding δ to each parameter
of En merely multiplies all solutions by ζ −δ . The following is a standard fact.
/ Z, ∀j, j ′ , so that the singular point ζ = 0 is nonLemma 2.2. If bj − bj ′ ∈
logarithmic, the solution space of En at ζ = 0 is spanned by the functions
a1 − bj + 1, . . . , an − bj + 1
ζ
, j = 1, . . . , n.
ζ 1−bj n Fn−1
\
b1 − bj + 1, . . . , bj −
bj + 1, . . . , bn − bj + 1
6
The differential equation En on P1ζ is the canonical order-n one with three
regular singular points, the monodromy at one of them being special. This
is seen as follows. Associated to any En is a monodromy representation of
π1 (X, ζ0 ), the fundamental group of the triply punctured sphere X = P1ζ \
{0, 1, ∞}. (The base point ζ0 plays no role, up to a similarity transformation.) The image H of π1 (X, ζ0 ) in GLn (C) is the monodromy group of the
specified En . It is generated by h0 , h1 , h∞ , the monodromy matrices around
ζ = 0, 1, ∞, which satisfy h∞ h1 h0 = I. Their eigenvalues are exponentiated
characteristic exponents, so they have characteristic polynomials
det(λI − h∞ ) =
n
Y
det(λI − h−1
0 )=
(λ − αj ),
n
Y
j=1
j=1
n−1
det(λI − h1 ) = (λ − 1)
(λ − e
2πiS
),
(λ − βj ),
(2.3)
/ Z then h1 is diagonalizable [3],
where αj = e2πiaj , βj = e2πibj . Moreover, if S ∈
i.e., the singular point ζ = 1 is not logarithmic. Hence if S ∈
/ Z, h1 will act on Cn
as multiplication by e2πiS on a 1-dimensional subspace, and as the identity on
a complementary (n − 1)-dimensional subspace. It will be a complex reflection.
The monodromy representation is irreducible iff no upper parameter aj and
lower one bj ′ differ by an integer, i.e., αj 6= βj ′ , ∀j, j ′ ; and it is unaffected
(up to similarity transformations) by integer displacements of parameters [3]. If
exactly one upper and one lower parameter of n Fn−1 (a1 , . . . , an ; b1 , . . . , bn−1 ; ζ)
are equal, permitting cancellation, then the corresponding En will have reducible
monodromy. One can show that in this case, the differential operator Dn will
have an order-(n − 1) right factor, which comes from the n−1 Fn−2 to which the
n Fn−1 reduces. In the theory of the Gauss function 2 F1 , the reducible case,
when a1 or a2 differs by an integer from b1 or b2 = 1, is often called the ‘degenerate’ case [10, § 2.2]. Reducibility of E2 facilitates the explicit construction of
solutions [22].
A monodromy group H < GLn (C) of an equation En , if irreducible, is said
to be imprimitive if there is a direct sum decomposition V1 ⊕ · · · ⊕ Vd of Cn with
d > 2 and dim Vj > 1, such that each element of H permutes the spaces Vj . The
following is a key theorem of Beukers and Heckman. It refers to the complex
reflection subgroup Hr of H, which is generated by the elements hk∞ · h1 · h−k
∞ ,
k ∈ Z. A ‘scalar shift’ by δ ∈ C means the addition of δ to each parameter.
Theorem 2.3 (cf. [3], Thm. 5.8). Suppose that H, Hr , computed from En , are
irreducible in GLn (C). Then H will be imprimitive if and only if there exist
relatively prime p, q > 1 with p + q = n, and a ∈ C, such that En takes on one
of the two forms
!
!
a+(n−1)
−a+(p−1) a
a+(q−1)
a
−a
,
.
.
.
,
,
.
.
.
,
;
,
.
.
.
,
n
n
p
p
q
q
En
, En
,
a+p 1
q
a+1
1
n
p ,..., p ; q,..., q
n, . . . , n
up to integer displacements of parameters, and up to a scalar shift by some
δ ∈ C. (For irreducibility, one must have that qa ∈
/ Z, resp. na ∈
/ Z; this
7
/ Z,
condition ensures that the upper and lower parameters satisfy aj − bj ′ ∈
∀j, j ′ .) Moreover, if a ∈ Q then H will be finite.
This theorem yields a class of algebraic n Fn−1 ’s, including (1.3). One must
choose the shift δ so that one of the lower parameters equals 1; e.g., δ = 0.
Beukers and Heckman do not compute the monodromy group H of the two En ’s
in the theorem, but it follows readily from their proof that if, e.g., a = −1/mq,
resp. a = −1/mn, with m > 2, then H is of order mn−1 n!, and is isomorphic
to the ‘symmetric’ index-m subgroup of the wreath product Cm ≀ Sn , where Cm
and Sn are the usual cyclic and symmetric groups. So is the corresponding
projective monodromy group H < PGLn (C) (see [21, Cor. 4.6]).
If a ∈ Z then Thm. 2.3 does not apply, due to reducibility: in each En
an upper and a lower parameter will differ by an integer. But, one has the
following.
Theorem 2.4 (cf. [3], Prop. 5.9). If there are equal upper and lower parameters in either En of Theorem 2.3, and a ∈ Z (e.g., if a = ±1), then the
En−1 obtained by cancelling them will have monodromy group H < GLn−1 (C)
isomorphic to Sn .
This yields a class of algebraic n−1 Fn−2 ’s. The hypergeometric functions
appearing in the following sections will include algebraic n Fn−1 ’s of the type
arising from Thm. 2.3, but for certain choices of parameter, they will reduce to
n−1 Fn−2 ’s. On the power series level, Lemma 2.1 will perform some of these
reductions.
3. Sequence transformations and coefficient identities
In this section and § 4, it is shown that the solutions of hypergeometric
equations En of the types introduced in Thm. 2.3, with monodromy that is irreducible but imprimitive, include algebraic functions that are solutions of trinomial equations. Explicit expressions for the associated n Fn−1 ’s as combinations
of algebraic functions will be derived, which extend those of Birkeland [4].
The chief result of this section is Theorem 3.4, which provides a family of
binomial coefficient identities, the family being indexed by ℓ ∈ Z. They are
derived from the standardized trinomial equation y − 1 − zy B = 0, and will be
used in § 4. Theorem 3.7 provides two more identities, which are more sophisticated in the sense that they have an additional free parameter. Some of these
identities were found by Lambert, Ramanujan, Pólya, and Gould, and the family indexed by ℓ ∈ Z was explored by Chu [6] in a little-noticed paper. The tool
employed in deriving the identities of Theorems 3.4 and 3.7 is the Vandermonde
convolution transform of Gould [14, 15], which is useful in deriving identities
relating coefficients of related series, such as binomial coefficient identities.
In what follows,
a
:= (a − r + 1)r /r!
(3.1)
r
8
extends the binomial coefficient to the case of arbitrary upper parameter, and
(
(a) · · · (a + r − 1),
r > 0;
(a)r :=
(3.2)
−1
[(a − s) · · · (a − 1)] , r = −s 6 0,
−1
extends the usual rising factorial, so that (a)r = [(a + r)−r ] for all r ∈ Z. The
standard forward finite difference operator ∆A,B , which acts on functions of A,
is defined by ∆A,B [h](A) = h(A + B) − h(A). Its n’th power ∆nA,B satisfies
∆nA,B [h](A) =
n
X
(−)n−k
k=0
n
h(A + kB).
k
(3.3)
Theorem 3.1 ([15]). Let A, B ∈ C be given, with B 6= 0, 1. Let f (k), k > 0,
be an infinite sequence of complex numbers, and let the sequence fˆ(n), n > 0,
be defined by
n
X
A + Bk
k n
ˆ
f (k).
(3.4)
f (n) =
(−)
k
n
k=0
In a neighborhood of z = 0, let y near 1 be defined implicitly by the standardized
trinomial equation y − 1 − z y B = 0. Then
n
∞
∞
X
X
1−y
A + Bk
fˆ(n)
f (k) z k = y A
(3.5)
k
y
n=0
k=0
holds as an equality between power series in z; and hence, if either has a positive
radius of convergence, as an equality in a neighborhood of z = 0.
If the theorem applies, the sequence fˆ(n), n > 0, will be called the (A, B)Vandermonde convolution transform of the sequence f (k), k > 0. If f (0) = 1,
then fˆ(0) = 1 also. For the inverse transform, see [15].
Definition 3.2. For each ℓ ∈ Z, the sequence fℓ (A, B; k), k > 0, is defined by
fℓ (A, B; k) :=
(A + ℓ)1−ℓ
(A + Bk + 1)ℓ−1
=
.
(A + 1)ℓ−1
(A + Bk + ℓ)1−ℓ
(3.6)
It is normalized so that fℓ (A, B; 0) = 1. Note that f1 (A, B; k) ≡ 1.
It is an immediate consequence of (3.3) that the (A, B)-Vandermonde convolution transform of fℓ (A, B; k), defined by Eq. (3.4), has the representation
(−)n
A
∆nA,B
(A + 1)ℓ−1
fˆℓ (A, B; n) =
(A + 1)ℓ−1
n
(3.7)
(−)n
=
∆nA,B [(A − n + 1)n+ℓ−1 ] .
n! (A + 1)ℓ−1
Here and in subsequent equations and identities, parameters such as A, B are
constrained, sometimes tacitly, so that division by zero occurs on neither side.
9
Theorem 3.3. (i) For each ℓ = −m 6 0, fˆℓ (A, B; n) is nonzero only if n =
0, . . . , m, and is rational in A, B for each n. (ii) For each ℓ > 1, fˆℓ (A, B; n)
equals a certain degree-(ℓ − 1) polynomial in n with coefficients polynomial in
A, B, multiplied by the function (n + 1)ℓ−1 (−B)n /(A + 1)ℓ−1 of n.
Proof. Part (i) follows from the fact that if n > m + 1, then (A − n + 1)n−m−1 ,
appearing in (3.7), is a polynomial in A of degree n − m − 1; hence its n’th finite
difference must equal zero. Part (ii) comes from a finite difference computation
(cf. [19, § 76]). Explicitly, fˆℓ (A, B; n) equals (−B)n /(A + 1)ℓ−1 times
ℓ−1 X
i
X
s(n + ℓ − 1, n + i) S(n + i, n + j) (n + 1)j B i
i=0 j=0
(A + ℓ − 1)/B
, (3.8)
j
where s, S are the Stirling numbers of the first and second kinds. But s(ν, ν −ρ),
S(ν, ν − ρ) are degree-2ρ polynomials in ν that are divisible by (ν − ρ + 1)ρ , so
each summand is of degree at most 2(ℓ−1) in n and is divisible by (n+1)ℓ−1 .
Remark 3.3.1. The formula given in the preceding proof, valid when ℓ > 1, can
be written as
ℓ−1
X
(A + ℓ − 1)/B
(−)n
,
C(n + ℓ − 1, n + j; B) (n + 1)j
fˆℓ (A, B; n) =
j
(A + 1)ℓ−1 j=0
(3.9)
P
where the coefficients C(n, k; B) := ni=k s(n, i)B i S(i, k) are ‘C-numbers’ [5].
Closed-form expressions for C(n, k; B) when B = −1, 12 , 2 are known [7, p. 158].
However, these values of B play no special role in the present analysis.
Examples of transformed sequences fˆℓ (A, B; n), n > 0, include
fˆ−1 (A, B; n) = δn,0 +
AB
δn,1 ,
A+B−1
(3.10a)
fˆ0 (A, B; n) = δn,0 ,
fˆ1 (A, B; n) = (−B)n ,
(n + 1)(−B)n
fˆ2 (A, B; n) = (A + 1) + 12 (B − 1)n
.
A+1
(3.10b)
(3.10c)
(3.10d)
These follow readily from the representation (3.7), or with more effort from (3.9).
For later reference, note that
n
(−B)
fˆ2 (A, B; n) + B fˆ2 (A, B; n − 1) = [A + 1 + (B − 1)n]
.
A+1
(3.11)
Theorem 3.4. For each ℓ ∈ Z, there is a certain rational function of y,
Fℓ (A, B; y), with coefficients polynomial in A, B, such that
A
y Fℓ (A, B; y) = 1 +
∞
X
(A + Bk + 1)ℓ−1 A + Bk
k=1
(A + 1)ℓ−1
10
k
zk
(3.12)
Table 1: The first few rational functions Fℓ , with S :=
ℓ
−1
0
1
2
S
3
2
1
2
1
−2
− 23
1
2
− ℓ.
Fℓ (A, B; y)
AB−(A−1)(B−1)y
(A+B−1) y
1
y
(1−B)y+B
2
y [(B−1)(B−A−1)y−B(B−A−2)]
3
(A+1)[(1−B)y+B]
n
o
2(B−1)(B−A−1)y−B(2B−2A−3)
= y 2 Dy
2(A+1)(B−1)[(1−B)y+B]2
holds in a neighborhood of z = 0, if y near 1 is the solution of the standardized
trinomial equation y − 1 − z y B = 0. (By assumption, B 6= 0, 1.) Specifically,
Fℓ (A, B; y) =
∞
X
fˆℓ (A, B; n)
n=0
1−y
y
n
,
(3.13)
where the sequence fˆℓ was defined in (3.7) and characterized in Theorem 3.3.
The rational function Fℓ has at most one pole on P1y . If ℓ = −m < 0, the
pole is at y = 0 and is of order m; and if ℓ > 0, it is at y = B/(B − 1) and is
of order 2ℓ − 1. If ℓ > 0, Fℓ can be written as (y 2 Dy )ℓ−1 F̃ℓ , where Dy = d/dy
and F̃ℓ is rational with a pole at y = B/(B − 1) of order ℓ.
Proof. The identity (3.12) is the specialization of (3.5) to the case f = fℓ , fˆ = fˆℓ .
By Thm. 3.3, Fℓ is a degree-m polynomial in v := (1 − y)/y if ℓ = −m 6 0, and
if ℓ > 0, it is rational on P1v with a unique pole (of order 2ℓ − 1) at v = (−B)−1 ,
i.e., at y = B/(B − 1). Since (n + 1)ℓ−1 | fˆℓ , the function Fℓ on P1v is the
(ℓ − 1)’th derivative of some rational function with a unique finite-v pole (of
order ℓ) at v = (−B)−1 . The final claim thus follows from Dv = −y 2 Dy .
As examples of the use of formula (3.13), four of the rational functions Fℓ
are listed in Table 1. They come from the sequences fˆℓ given in (3.10).
Theorem 3.4, with the aid of Table 1, yields the following binomial coefficient
identities, which are indexed by ℓ = −1, 0, 1, 2, respectively. They hold near
11
z = 0, if y near 1 is defined by y − 1 − z y B = 0.
y
A
∞
X
AB − (A − 1)(B − 1)y
A−1
A
A + Bk k
=1+
z ,
(A + B − 1) y
A + Bk − 1 A + Bk
k
k=1
(3.14a)
∞
X
y
A
A + Bk k
yA = 1 +
z ,
A + Bk
k
k=1
∞
X
A + Bk k
y
=1+
z ,
yA
(1 − B)y + B
k
k=1
(
)
y 2 [(B − 1)(B − A − 1)y − B(B − A − 2)]
A
(A + 1) [(1 − B)y + B]
3
= 1+
(3.14b)
(3.14c)
(3.14d)
∞
X
A + Bk + 1 A + Bk k
z .
A+1
k
k=1
The ℓ = 0 identity (3.14b) and the ℓ = 1 identity (3.14c) are well known. The
former was derived by Lambert in 1758 (and by Ramanujan [2, p. 72]), and the
latter by Pólya [28]. They can be proved by Lagrange inversion [25, Thm. 2.1],
and imply each other [29, Lem. 1]. But the identities (3.14a),(3.14d) are less
familiar. The ℓ = 3, 4, . . . and ℓ = −2, −3, . . . identities can also be worked out.
Each identity in this family indexed by ℓ ∈ Z can be obtained by differentiating the preceding. This is formalized in the following recurrence, which is
equivalent to one obtained by Chu [6]. By starting with F0 ≡ 1 and iterating,
one can generate all Fℓ , ℓ > 0. By doing the reverse, i.e., by integrating (and
exploiting the fact that Fℓ (1) = 1, ∀ℓ ∈ Z, which determines each constant of
integration), one can generate all Fℓ , ℓ < 0.
Theorem 3.5. The rational functions Fℓ (A, B; y), ℓ ∈ Z, satisfy
Fℓ+1 (A, B; y) =
(A − B + 2)ℓ−1 y −A+B+1
Dy y A−B+1 Fℓ (A − B + 1, B; y) .
(A + 1)ℓ
(1 − B)y + B
Proof. Apply Dz to both sides of (3.12), using Dz =
follows from y − 1 − z y B = 0.
y 2B
y B −B(y−1)y B−1 Dy ,
which
In the binomial coefficient identities of the form (3.12) with ℓ = 3, 4, . . .
and ℓ = −2, −3, . . . , not given here, the left-hand functions Fℓ (A, B; y) become
increasingly complicated. But for all ℓ ∈ Z, the right-hand power series in z has
radius of convergence (B − 1)B−1 /B B if B 6= 1, and unity if B = 1. This is
consistent with the presence (when ℓ > 0, at least) of a pole at y = B/(B − 1)
on the left-hand side, since y = B/(B − 1) corresponds to z = (B − 1)B−1 /B B .
As B → 1, the pole moves to y = ∞, i.e., to z = 1.
The following theorem reveals when the evaluation of the rational function
Fℓ (A, B; y), in any identity in this family, may lead to a division by zero.
12
Theorem 3.6. The rational function Fℓ (A, B; y) in (3.12) is of the form
Π̃m (A, B; y)
,
(A + B − m)m (A + 2B − m)m−1 · · · (A + mB − m)1 y m
ℓ = −m 6 0; (3.15a)
y ℓ Πℓ−1 (A, B; y)
,
(A + 1)ℓ−1 [(1 − B)y + B]2ℓ−1
ℓ > 1.
(3.15b)
where Π̃m , Πℓ−1 are polynomials in y, the subscripts indicating their degrees.
Proof. Iterate Thm. 3.5 toward negative and positive ℓ.
It should be mentioned that in the identities indexed by ℓ > 1, the poles
in A that are evident in (3.15b), located at A = −1, . . . , −ℓ + 1, are removable.
The denominator factor (A + 1)ℓ−1 is present on the right as well as the left
side of (3.12), and can simply be cancelled. But the poles in A present in Fℓ ,
ℓ 6 −1, which are evident in (3.15a), are less easily removed.
This family of identities can be generalized by modifying the initial sequence
fℓ (A, B, k) of (3.6) to include one or more free parameters. A pair of such
generalizations, involving the functions 2 F1 and 3 F2 , will prove useful. Let
gℓ (A, B, C; k) :=
A+C +ℓ
fℓ+1 (A, B; k),
A + C + Bk + ℓ
k > 0,
(3.16)
so that gℓ is ‘interpolating’: it reduces to fℓ if C = 0 and to fℓ+1 as C → ∞. It
follows from (3.3), much as in (3.7), that the (A, B)-transforms of g0 , g1 are
A
1
n
n
,
(3.17a)
ĝ0 (A, B, C; n) = (−) (A + C) ∆A,B
n
A+C
A
A+1
A+C +1
∆nA,B
. (3.17b)
ĝ1 (A, B, C; n) = (−)n
A+1
A+C +1
n
By a finite difference computation or an expansion in partial fractions (cf. [14,
§ 6]), these transformed sequences satisfy
ĝ0 (A, B, C; n) = (−)n
n
ĝ1 (A, B, C; n) + B ĝ1 (A, B, C; n − 1) = (−)
(C)n
,
+1 n
(3.18a)
A+C
B
(C)n
A+C+1
B
A+1
B−1
+1
+1
n
n
,
A+1
B−1 n
(3.18b)
for each n > 0, resp. n > 1, showing by comparison with (3.10) and (3.11) that
ĝ0 (A, B, C; n) reduces to fˆ0 (A, B; n) = (−B)n if C = 0 and to fˆ1 (A, B; n) = δn,0
as C → ∞; and that ĝ1 (A, B, C; n) reduces to fˆ1 (A, B; n) = 1 if C = 0, and to
fˆ2 (A, B; n), given in (3.10d), as C → ∞. To avoid division by zero in (3.18),
one may assume that B 6= 0, 1 and that (A + C)/B is not a negative integer.
The right sides of (3.18a),(3.18b) have the form of coefficients of 2 F1 and 3 F2
series. Applying Thm. 3.1 therefore yields a pair of hypergeometric identities.
13
Theorem 3.7. Under the preceding assumptions, one has the interpolating
identities
∞
X
A+C
A + Bk k
y−1
C, 1
A
=1+
z ,
y 2 F1
A+C
y
A + C + Bk
k
B +1
k=1
A+1
y
C, B−1
+ 1, 1
y−1
A
y
3 F2
A+C+1
A+1
+ 1, B−1
(1 − B)y + B
y
B
∞
X
A+C +1
A + Bk + 1 A + Bk k
=1+
z ,
A + C + Bk + 1
A+1
k
k=1
which are valid near z = 0, if y near 1 is defined by y − 1 − z y B = 0.
Of these two identities the first was found by Gould [14, § 6], but the more
complicated second one is new. The parametrized functions multiplied by y A
on their left sides will be denoted by Gℓ (A, B, C; y), ℓ = 0, 1, respectively.
The first identity reduces if C = 0 to (3.14b), the ℓ = 0 identity of Lambert and Ramanujan, and as C → ∞ to (3.14c), the ℓ = 1 identity of Pólya.
The function G0 (A, B, C; y) reduces to F0 (A, B; y) ≡ 1 and F1 (A, B; y) =
y/[(1 − B)y + B], respectively. Similarly, the second identity reduces if C = 0
to Pólya’s identity, and as C → ∞ to the ℓ = 2 identity (3.14d). The function
G1 (A, B, C; y) reduces respectively to F1 (A, B; y) and F2 (A, B; y).
4. Algebraic and hypergeometric functions
In this section certain algebraic functions, namely the solutions of trinomial
equations, are expressed in terms of n Fn−1 ’s that are solutions of En ’s with
monodromy that is irreducible but imprimitive. In Thm. 4.4, which is the
main result of this section, the n Fn−1 ’s are expressed in terms of algebraic
functions by ‘inverting’ the just-mentioned representations. If the parameter
ℓ ∈ Z in Thm. 4.4 is set to zero, the expressions reduce to those of Birkeland [4].
Theorem 4.6 is a partial extension of Thm. 4.4 that has an additional free
parameter, and is less algebraic.
The chief result of § 3 was Theorem 3.4, which provides a family of binomial
coefficient identities indexed by ℓ. It was based on the standardized trinomial
equation y − 1 − z y B = 0, and the solution y that is near 1 for z near 0. The
theorem is now restated in terms of the general trinomial equation
xn − g xp − β = 0.
(4.1)
Here n = p + q, for integers p, q > 1, and g, β ∈ C with at most one of g, β equal
to zero. The condition gcd(p, q) = 1 will be added later.
Let the n solutions of (4.1), with multiplicity, be denoted x1 , . . . , xn . In the
limit β → 0 with fixed g > 0, one may choose
(
ε−(j−1)
g 1/q ,
j = 1, . . . , q,
q
xj =
(4.2)
0,
j = q + 1, . . . , n.
14
In the limit g → 0 with fixed β > 0, one may choose
xj = ε−(j−1)
β 1/n ,
n
j = 1, . . . , n.
(4.3)
Here and below, εr := exp(2πi/r) signifies a primitive r’th root of unity.
To reduce the first case of (4.1), i.e., that of β near zero, to y − 1 − z y B = 0,
let
y = g −1 xq ,
z = ε(j−1)n
g −n/q β,
q
B = −p/q.
(4.4)
To reduce the second case, i.e., that of g near zero, let
y = β −1 xn ,
z = ε(j−1)q
β −q/n g,
n
B = p/n.
(4.5)
By undoing these two reductions, and letting A = γ/q, resp. A = γ/n, one
obtains the following ‘de-standardized’ version of Thm. 3.4, in which the rational
functions Fℓ (A, B; y), ℓ ∈ Z, were defined.
Theorem 4.1. The following hold for ℓ ∈ Z and γ ∈ C.
(i) In a neighborhood of β = 0 with fixed g > 0, and with the trinomial root xj
−(j−1) 1/q
near εq
g
defined by (4.1),(4.2),
[εq(j−1) g −1/q xj ]γ Fℓ (γ/q, −p/q; g −1 xqj )
=1+
∞
X
(γ/q − pk/q + 1)ℓ−1 γ/q − pk/q h (j−1)n −n/q ik
εq
g
β ,
(γ/q + 1)ℓ−1
k
k=1
for j = 1, . . . , q.
(ii) In a neighborhood of g = 0 with fixed β > 0, and with the trinomial root xj
−(j−1) 1/n
near εn
β
defined by (4.1),(4.3),
[εn(j−1) β −1/n xj ]γ Fℓ (γ/n, p/n; β −1 xnj )
∞
X
(γ/n + pk/n + 1)ℓ−1 γ/n + pk/n h (j−1)q −q/n ik
=1+
εn
β
g ,
(γ/n + 1)ℓ−1
k
k=1
for j = 1, . . . , n.
In (i) and (ii), γ ∈ C is constrained so that no division by zero occurs in the
evaluation of either the rational function Fℓ or the first factor of the summand.
The connection to hypergeometric equations En with imprimitive monodromy,
and their canonical solutions n Fn−1 , can now be made. Henceforth, let a Riemann sphere P1ζ be parametrized by
ζ := (−)q
nn β q
.
pp q q g n
(4.6)
The following formulas extend those of Birkeland [4, § 3], who considered only
the case ℓ = 0. Since F0 ≡ 1, his formulas contain no left-hand Fℓ factor.
15
Theorem 4.2. The following hold for ℓ ∈ Z and a ∈ C, with ζ defined by
(4.6).
(i) In a neighborhood of β = 0 with fixed g > 0, and with the trinomial root xj
−(j−1) 1/q
near εq
g
defined by (4.1),(4.2),
[ε(j−1)
g −1/q xj ]−qa Fℓ (−a, −p/q; g −1 xqj )
q
=
q−1
X
κ=0
εn
q
(j−1)κ
n Fn−1
(−)κ (a)1−ℓ
(a + nκ/q)1−ℓ−κ (1)κ
(−)q pp q q
·ζ
nn
a
+ κq , . . . , a+(n−1)
+ κq
n
n
a−ℓ+1
+ κq , . . . , a−ℓ+p
+ κq ; 1q
p
p
κ/q
\
+ κq , . . . , q−κ
+ κq , . . . ,
q
q
q
+
!
,
!
,
ζ
κ
q
for j = 1, . . . , q.
(ii) In a neighborhood of g = 0 with fixed β > 0, and with the trinomial root xj
−(j−1) 1/n
near εn
β
defined by (4.1),(4.3),
[ε(j−1)
β −1/n xj ]−na Fℓ (−a, p/n; β −1 xn
n
j)
=
n−1
X
(εqn )(j−1)κ
κ=0
n Fn−1
(−)κ (a)1−ℓ
(a + qκ/n)1−ℓ−κ (1)κ
(−)q pp q q
·ζ
nn
−a+ℓ
+ nκ , . . . , −a+ℓ+(p−1)
+ nκ ; aq
p
p
n
1
\
+ nκ , . . . , n−κ
+ nκ , . . . , n
+ nκ
n
n
−κ/n
+
κ
, . . . , a+(q−1)
n
q
+
κ
n
ζ
−1
for j = 1, . . . , n.
In (i) and (ii), a ∈ C is constrained so that no division by zero occurs on either
side, and (in (i)) so that no lower parameter of any n Fn−1 is a non-positive
integer.
Remark 4.2.1. The quantity in square brackets on the right sides of (i),(ii) equals
g −n β q . It is raised to the power κ/q, resp. power −κ/n. The branch should be
chosen so that the result equals g −κn/q β κ , resp. β −κq/n g κ .
Proof. Partition the series of Thm. 4.1 into residue classes of k mod q, resp.
k mod n. That is, let k = κ + i · q, resp. k = κ + i · n, and for each κ, sum over
i = 0, 1, . . . , using the identity
. . . t+r−1
.
(4.7)
(t)i·r = rr rt i t+1
r
r
i
i
Also, let γ = −qa, resp. γ = −na. (That is, A = −a.)
Corollary 4.3. The following hold for ℓ ∈ Z and a ∈ C.
−(j−1) 1/q
g
defined by
(i) Near ζ = 0 on P1ζ , with the trinomial root xj near εq
(4.1),(4.2), the coefficient g > 0 being fixed and arbitrary and the coefficient β
being defined in terms of ζ by (4.6), the quantities
x−qa
Fℓ (−a, −p/q; g −1 xqj ),
j
16
j = 1, . . . , q,
regarded as functions of ζ, lie in the solution space of the differential equation
!
a
, . . . , a+(n−1)
n
n
En
.
a−ℓ+1
, . . . , a−ℓ+p
; 1q , . . . , qq
p
p
ζ
−(j−1)
(ii) Near ζ = ∞ on P1ζ , with the trinomial root xj near εn
β 1/n defined by
(4.1),(4.3), the coefficient β > 0 being fixed and arbitrary and the coefficient g
being defined in terms of ζ by (4.6), the quantities
xj−na Fℓ (−a, p/n; β −1 xnj ),
j = 1, . . . , n,
regarded as functions of the reciprocal variable ζ̃ := ζ −1 , lie in the solution space
near ζ̃ = 0 of the differential equation
!
−a+ℓ
, . . . , −a+ℓ+(p−1)
; aq , . . . , a+(q−1)
p
p
q
En
.
n
1
n, . . . , n
ζ̃
In (i), it is assumed that a ∈ C is such that no pair of lower parameters differs
by an integer; i.e., that the singular point ζ = 0 of the En is non-logarithmic.
Proof. Immediate, by Lemma 2.2 applied to Thm. 4.2.
If in either En of Corollary 4.3, a is chosen so that no upper parameter differs
by an integer from a lower one, then the monodromy group of the En will be
irreducible; and moreover, if gcd(p, q) = 1 then the En will be of the imprimitive
irreducible type characterized in Thm. 2.3. The latter fact is obvious if ℓ = 0,
though if ℓ ∈ Z \ {0}, integer displacements of parameters are needed.
The representations of Theorem 4.2 can now be inverted, to express the
n Fn−1 ’s satisfying these En ’s in terms of the trinomial roots x1 , . . . , xn . In
part (i) of the theorem below, only x1 , . . . , xq are needed; but in part (ii), all
are needed.
Theorem 4.4. If gcd(p, q) = 1, the following hold for each ℓ ∈ Z and a ∈ C.
(i) Near ζ = 0 on P1ζ , for arbitrary fixed g > 0 and for κ = 0, . . . , q − 1,
ζ
κ/q
n Fn−1
a+(n−1)
a
κ
+ κq
n + q,...,
n
a−ℓ+1
+ κq , . . . , a−ℓ+p
+ κq ; 1q
p
p
q
\
κ
+ κq , . . . , q−κ
q + q,..., q +
n κ/q
κ
q
ζ
!
(−)κ (1)κ (a + nκ/q)1−ℓ−κ (−)q n
=
(a)1−ℓ
pp q q
q
X
(j−1)κ (j−1) −1/q
[εq
g
xj ]−qa Fℓ (−a, −p/q; g −1 xqj ),
ε−n
× q −1
q
j=1
−(j−1)
in which the trinomial root xj near εq
g 1/q , j = 1, . . . , q, is defined by
(4.1),(4.2), with the coefficient β determined by ζ according to (4.6).
17
(ii) Near ζ = ∞ on P1ζ , for arbitrary fixed β > 0 and for κ = 0, . . . , n − 1,
ζ
−a+ℓ
+ nκ , . . . , −a+ℓ+(p−1)
+ nκ ; aq + nκ , . . . , a+(q−1)
+ nκ
p
p
q
F
ζ −1
n n−1
κ
κ
n
κ
n−κ
1
\
n + n, . . . , n + n, . . . , n + n
−κ/n
κ
(−) (1)κ (a + qκ/n)1−ℓ−κ (−)q nn
=
(a)1−ℓ
pp q q
n
X
(j−1)κ (j−1) −1/n −na
[εn
β
xj ]
Fℓ (−a, p/n; β −1 xnj ),
ε−q
× n−1
n
j=1
−κ/n
!
−(j−1)
in which the trinomial root xj near εn
β 1/n , j = 1, . . . , n, is defined by
(4.1),(4.3), with the coefficient g determined by ζ according to (4.6).
In (i) and (ii), a ∈ C is constrained so that no division by zero occurs, and
(in (i)) so that no lower parameter of any n Fn−1 , for any κ, is a non-positive
integer.
Remark 4.4.1. A convention for choosing branches of fractional powers must be
adhered to, in interpreting these formulas. In (i), there are q choices for each of
ζ
1/q
,
(−)q nn
pp q q
1/q
,
β.
(The last is evident from (4.6).) Any choices will work, so long as the formal
q’th root of Eq. (4.6), i.e.,
ζ 1/q =
(−)q nn
pp q q
1/q
· g −n/q · β,
(4.8)
is satisfied. This removes one degree of freedom. Part (ii) is similarly interpreted.
Proof. Each of the two formulas in Thm. 4.2 has the form P
of a linear transformation expressed as a matrix–vector product, i.e., uj = κ Mjκ vκ . Here,
M = (Mjκ ) = (ε(j−1)κ ) is a q × q, resp. n × n matrix, with ε := εnq , resp. εqn . If
gcd(p, q) = 1, or equivalently gcd(q, n) = 1, then the root of unity ε will be a
primitive q’th, resp. n’th, root, and M will be nonsingular. The vκ are expressed
in terms of the uj by inverting M. But M−1 = q −1 M∗ , resp. M−1 = n−1 M∗ .
Elementwise multiplication by M∗ yields the claimed summation formulas.
Corollary 4.5. If gcd(p, q) = 1, then the following holds for a ∈ C with
−(j−1) 1/n
na ∈
/ Z. Near ζ = ∞ on P1ζ , if xj near εn
β , j = 1, . . . , n, is defined by
(4.1),(4.3), with the coefficient g determined by ζ according to (4.6), then the
exponentiated roots
x−na
,
j = 1, . . . , n,
j
18
regarded as functions of ζ̃ := ζ −1 , span the solution space near ζ̃ = 0 of the
differential equation
!
−a+(p−1) a
a+(q−1)
−a
,
.
.
.
,
;
,
.
.
.
,
p
p
q
q
En
.
n
1
n, . . . , n
ζ̃
Proof. Immediate by Lemma 2.2, applied to the ℓ = 0 case of Thm. 4.4(ii). (The
condition na ∈
/ Z ensures linear independence of the exponentiated roots.)
The representations of certain n Fn−1 ’s given in Thm. 4.4, in terms of the
solutions of trinomial equations, are the point of departure for the rest of this
paper. In each n Fn−1 , the parametric excess S (the sum of the lower parameters, minus the sum of the upper ones) equals 12 − ℓ, by examination. As was
mentioned, the ℓ = 0 case of the theorem, when the Fℓ factor on each right-hand
side degenerates to unity, was previously obtained by Birkeland.
Many classical hypergeometric identities, such as the cubic transformations
of 3 F2 found by Bailey, are restricted to n Fn−1 ’s with S = 12 . Sometimes
there are ‘companion’ identities that are satisfied by hypergeometric functions
with S = − 21 . (For Bailey’s identities, see [12, (5.3,5.4) and (5.6,5.7)].) It is
remarkable that in the present context, the cases S = 21 and − 12 , i.e., ℓ = 0 and 1,
are merely the most easily treated steps on an infinite ladder of possibilities.
The preceding theorems relating hypergeometric and algebraic functions,
including Thm. 4.4, came ultimately from Thm. 3.4, which applied the Vandermonde convolution transform to the sequences fℓ given in Defn. 3.2. One
could start instead with Thm. 3.7, i.e., with the ‘interpolating’ sequences g0 , g1
parametrized by C, which reduce to f0 , f1 if C = 0 and to f1 , f2 as C → ∞.
Deriving the following theorem, which interpolates between the ℓ = 0, 1 and 1, 2
cases of Thm. 4.4, is straightforward. (In it, c stands for −C.)
Theorem 4.6. If gcd(p, q) = 1, the following hold for ℓ = 0, 1 and a ∈ C, with
y−1
C, 1
,
G0 (A, B, C; y) := 2 F1
A+C
y
B +1
A+1
y
y−1
C, B−1
+ 1, 1
G1 (A, B, C; y) :=
.
3 F2
A+1
A+C+1
+ 1, B−1
(1 − B)y + B
y
B
(i) Near ζ = 0 on P1ζ , for arbitrary fixed g > 0 and for κ = 0, . . . , q − 1,
ζ
κ/q
n+1 Fn
=
a
+ κq , . . . , a+(n−1)
+ κq ;
n
n
a−ℓ+(p−1)
a−ℓ
+ κq , . . . ,
+ κq ; 1q
p
p
+
\
κ
, . . . , q−κ
+ κq , . . . , qq
q
q
(−)κ (1)κ (a + nκ/q)−ℓ−κ (a + c − ℓ + pκ/q)
(a)−ℓ (a + c − ℓ)
× q −1
q
X
j=1
ε−n
q
(j−1)κ
(j−1) −1/q
[εq
g
19
(−)q nn
pp q q
+
κ
;
q
a+c−ℓ
+ κq
p
a+c−ℓ+p
+ κq
p
κ/q
xj ]−qa Gℓ (−a, −p/q, −c; g −1 xqj ),
ζ
−(j−1)
g 1/q , j = 1, . . . , q, is defined by
in which the trinomial root xj near εq
(4.1),(4.2), with the coefficient β determined by ζ according to (4.6).
(ii) Near ζ = ∞ on P1ζ , for arbitrary fixed β > 0 and for κ = 0, . . . , n − 1,
ζ
−κ/n
n+1 Fn
=
−a+ℓ+1
κ
κ a
+n
, . . . , −a+ℓ+p
+n
; q
p
p
\
n−κ
1
κ
κ
n
κ
+ n,..., n + n,..., n + n
;
n
+
(−)κ (1)κ (a + qκ/n)−ℓ−κ (a + c − ℓ − pκ/n)
(a)−ℓ (a + c − ℓ)
× n−1
n
X
j=1
ε−q
n
(j−1)κ
(j−1) −1/n
[εn
β
κ
, . . . , a+(q−1)
n
q
(−)q nn
pp q q
+
κ
;
n
−a−c+ℓ
κ
+n
p
−a−c+ℓ+p
κ
+n
p
−κ/n
xj ]−na Gℓ (−a, p/n, −c; β −1 xn
j ),
−(j−1)
β 1/n , j = 1, . . . , n, is defined by
in which the trinomial root xj near εn
(4.1),(4.3), with the coefficient g determined by ζ according to (4.6).
In (i) and (ii), a, c ∈ C are constrained so that no division by zero occurs, and
so that no lower parameter of any n+1 Fn , for any κ, is a non-positive integer.
The ℓ = 0, 1 representations of n+1 Fn ’s given in Thm. 4.6 reduce if c = 0 to
the ℓ = 0, 1 representations of n Fn−1 ’s given in Thm. 4.4, by cancelling equal
upper and lower parameters; and in the limit c → ∞, to the ℓ = 1, 2 ones.
It should be noted that unlike Thm. 4.4, which involves rational right-hand
functions Fℓ , ℓ ∈ Z, Thm. 4.6 is essentially non-algebraic: the right-hand functions Gℓ , ℓ = 0, 1, which depend on the additional parameter c, are hypergeometric and are generically transcendental.
5. Schwarz curves: Generalities
This section introduces what will be called Schwarz curves, which are projective algebraic curves that parametrize ordered k-tuples of roots (with multiplicity) of the general degree-n trinomial equation
xn − g xp − β = 0.
(5.1)
As in § 4, it is assumed that n = p + q for relatively prime integers p, q > 1, and
that g, β ∈ C with at most one of g, β equaling zero. For each k, any ordered
k-tuple of roots will trace out a Schwarz curve as
ζ = (−)q
nn β q
pp q q g n
(5.2)
varies. Any Schwarz curve is a projective curve, since multiplying each root by
any c 6= 0 will multiply g, β by cq , cn, but leave ζ invariant. The dependence
on ζ, which is really a projectivized (and normalized) version of the discriminant
of (5.1), will be interpreted as specifying a covering of P1ζ by the Schwarz curve,
the genus of which will be calculated. Determining the dependence of a point on
the curve on the base point ζ will be of value when uniformizing the solutions of
En ’s with imprimitive monodromy, since the formulas of Thm. 4.4(i),(ii) express
many solutions in terms of q-tuples, resp. n-tuples of roots of (5.1).
20
ζ
−1
Let σl = σl (x1 , . . . , xn ) denote the l’th P
elementary symmetric polynomial in
n
the roots x1 , . . . , xn , so that σ0 = 1, σ1 = i=1 xi , etc. From (5.1),
β = (−)n−1 σn ,
g = (−)q−1 σq ;
(5.3)
and σl = 0 for l = 1, . . . , q − 1 and q + 1, . . . , n − 1. Parametrizing tuples of
roots, given ζ ∈ P1 , means parametrizing them given the ratio [σn q : σq n ].
(n)
Definition 5.1. The top Schwarz curve Cp,q ⊂ Pn−1 is the algebraic curve
comprising all points [x1 : . . . : xn ] ∈ Pn−1 such that x1 , . . . , xn satisfy the n − 2
homogeneous equations
σl (x1 , . . . , xn ) = 0,
l = 1, . . . , q − 1 and q + 1, . . . , n − 1.
(5.4)
(n)
That is, Cp,q comprises all [x1 : . . . : xn ] ∈ Pn−1 such that x1 , . . . , xn are the
n roots (with multiplicity) of some trinomial equation of the form (5.1). The
(n)
curve Cp,q is stable under the action of the symmetric group Sn on [x1 : . . . : xn ].
(n)
(n)
An associated degree-n! covering map πp,q : Cp,q → P1ζ is defined by
ζ = (−)n
nn σn q
.
pp q q σq n
(5.5)
(n)
Proposition 5.2 ([21], Cor. 4.7). The curve Cp,q is irreducible.
Definition 5.3. For each k with n > k > 2, the subsidiary Schwarz curve
(k)
(n)
Cp,q ⊂ Pk−1 , also irreducible, is the image of Cp,q under the map [x1 : . . . : xn ] 7→
(n)
(k)
[x1 : . . . : xk ], which on Cp,q is (n− k)!-to-1. A more concrete definition of Cp,q is
the following. It is the closure in Pk−1 of the set of all points [x1 : . . . : xk ] ∈ Pk−1
such that x1 , . . . , xk are k of the n roots (with multiplicity) of some trinomial
(k)
equation of the form (5.1). The curve Cp,q is stable under the action of Sk on
[x1 : . . . : xk ].
(k−1)
(k)
(k)
For each k with n > k > 2, let a projection φp,q : Cp,q → Cp,q be defined by
(n)
(k+1)
(k)
φp,q ([x1 : . . . : xk ]) = [x1 : . . . : xk−1 ], so that φp,q ◦ · · · ◦ φp,q is the projection
(k)
(n)
(k)
(n)
Cp,q → Cp,q . Since Cp,q → Cp,q is (n − k)!-to-1 for each k, it follows that for
(k)
each k, φp,q is (n − k + 1)-to-1, i.e., is a degree-(n − k + 1) map.
(k)
Remark 5.3.1. Each subsidiary curve Cp,q ⊂ Pk−1 , n > k > 2, is the solution set
of a system of k−2 homogeneous equations in x1 , . . . , xk , obtained by eliminating
xk+1 , . . . , xn from the system (5.4). The details of this will be given shortly.
(k)
Remark 5.3.2. As will be explained in § 6.1, defining Cp,q as a closure in Pk−1
appends at most a finite number of points to it. Specifically, if k 6 p, it appends
each point [x1 : . . . : xk ] ∈ Pk−1 in which x1 , . . . , xk are distinct p’th roots of
unity. There are (p − k + 1)k−1 = (p − k + 1)k /p such points in Pk−1 , none of
which comes directly (via a k-tuple of roots) from a trinomial equation of the
form (5.1). Rather, each is a limit in Pk−1 of points that do.
21
(k)
(k)
(n)
Remark 5.3.3. The projections Cp,q → Cp,q and φp,q are really only partial
maps, being undefined at a finite number of singular points, as is typical of
maps between algebraic curves. (The symbol 99K could be used instead of →.)
(k)
Specifically, if x1 = . . . = xk−1 = 0 then φp,q ([x1 : . . . : xk ]) is undefined. The
problems with zero tuples, associated to trinomials with β = 0, will be dealt
(k)
(k)
with in § 6, where each curve Cp,q will be lifted to a desingularized curve C̃p,q .
The reason for the term ‘Schwarz curve’ is this. The ratios of any n independent solutions y1 , . . . , yn of an order-n differential equation on P1ζ define
a (multivalued) Schwarz map from P1ζ to Pn−1 . Its image is a curve in Pn−1 ,
and in some cases the inverse map from the image is single-valued, i.e., supplies
a covering of P1ζ by the curve. Studying Schwarz maps is a standard way of
computing the monodromy groups of differential equations [20, 21], and goes
back to Schwarz’s classical work on E2 , the Gauss hypergeometric equation. In
the present paper, solutions of En ’s with imprimitive monodromy have been
expressed in terms of tuples of solutions of trinomial equations, which are algebraic in ζ; so a purely algebraic use of the Schwarz map and curve concepts
(n)
seems warranted. The top Schwarz curve Cp,q was introduced and used by Kato
and Noumi [21], though not under that name; the subsidiary Schwarz curves
seem not to have been treated or exploited before.
(n)
(n−1)
(2)
It is evident that as defined, Cp,q ∼
= Cp,q and Cp,q ∼
= P1t , where birational
equivalence is meant. Here, t is any homogeneous degree-1 rational function
of x1 , x2 . Henceforth the choice t = (x1 + x2 )/(x1 − x2 ) will be made, so
(2)
[x1 : x2 ] = [t + 1 : t − 1] will be a uniformization of Cp,q by the rational
(1)
1
parameter t. It will also prove useful to define Cp,q := P . This will make it
(2)
(2)
(1)
(1)
(2)
possible to define φp,q : Cp,q → Cp,q and φp,q : Cp,q → P1ζ in such a way that
(n)
(n)
πp,q
= φ(1)
p,q ◦ · · · ◦ φp,q ,
(5.6)
in the sense of being an equality on a cofinite domain. (See (5.18) and (5.21)
below.) For each k with n > k > 1,
(k)
(k)
πp,q
= φ(1)
p,q ◦ · · · ◦ φp,q
(k)
(5.7)
(k)
will then define a subsidiary (partial) map πp,q : Cp,q → P1ζ of degree (n−k+1)k .
(k)
One can derive a system of polynomial equations for each Cp,q , n > k > 2,
in terms of x1 , . . . , xk , by using resultants to eliminate xk+1 , . . . , xn . But one can
also eliminate them by hand in the following way, which will incidentally indicate
(1)
how best to define the curve Cp,q . For any k, 0 < k < n, let σ̄m , σ̂m denote the
m’th elementary symmetric polynomial in x1 , . . . , xk , resp. xk+1 , . . . , xn . Then
σl =
l
X
σ̄m σ̂l−m ,
0 6 l 6 n,
(5.8)
m=0
it being understood that σ̄m = 0 if m ∈
/ {0, . . . , k}, and similarly, that σ̂m = 0
if m ∈
/ {0, . . . , n − k}, with σ̄0 = σ̂0 = 1. The defining equations of the top
22
(n)
curve Cp,q , by Defn. 5.1, are
Pn
1 = σ0 = m=0 σ̄m σ̂0−m ;
Pn
Pn
. . . , 0 = σq−1 = m=0 σ̄m σ̂(q−1)−m ;
1 =
m=0 σ̄m σ̂1−m ,
0 = σP
n
σq = m=0 σ̄m σ̂q−m ;
P
P
0 = σq+1 = n σ̄m σ̂(q+1)−m , . . . , 0 = σn−1 = n σ̄m σ̂(n−1)−m ;
m=0
m=0
P
σn = nm=0 σ̄m σ̂n−m .
(5.9)
n−k
Lemma 5.4. {σ̂m }m=0
can be expressed in terms of x1 , . . . , xk (and σn ) by
X
mk
1
(−)m
xm
1 · · · xk ,
m1 +···+mk =m
(∀j) 06mj 6m
m = 0, . . . , min(q − 1, n − k);
σ̂m =
n−k−m
X
(−)
σn
mk
1
xm
1 · · · xk ,
(x1 · · · xk )n−k−m+1
m1 +···+mk =(k−1)(n−k−m)
(∀j) 06mj 6n−k−m
m = max(q − k + 1, 0), . . . , n − k.
Note that 0 6 m 6 min(q − 1, n − k) and max(q − k + 1, 0) 6 m 6 n − k, the
m-ranges of validity of these two formulas, may overlap.
Proof. The first formula is proved by induction. The idea is that one solves the
zeroth equation in (5.9) for σ̂0 , then the first equation for σ̂1 , etc.; stopping with
the min(q − 1, n − k)’th equation, since the next one may involve σq , which is
not known. But since the l’th equation, for 1 6 l 6 min(q − 1, n − k), says that
σ̂l = −(σ̄1 σ̂l−1 + · · · + σ̄k σ̂l−k ),
(5.10)
the inductive step amounts to verifying that
P (m) = σ̄1 P (m − 1) − σ̄2 P (m − 2) + · · · + (−)k−1 σ̄k P (m − k),
X
mk
1
P (m) :=
xm
1 · · · xk .
(5.11)
m1 +···+mk =m
(∀j) 06mj 6m
This is a well-known identity.
The second formula in the lemma is dual to the first and is proved by a similar
induction, ‘downward.’ One solves the n’th equation in (5.9) for σ̂n−k , then the
(n − 1)’st equation for σ̂n−k−1 , etc.; stopping with the max(q − k + 1, 0)’th
equation, since the next one may involve σq , which is not known. As with the
first formula, the inductive step reduces to a verification of Eq. (5.11).
The formulas of the lemma are accompanied by constraints. First, there are
the implicit constraints coming from the equivalence between the two formulas
for σ̂m , for each m in the doubly covered range
max(q − k + 1, 0) 6 m 6 min(q − 1, n − k).
23
(5.12)
Second, there are the equations σl = 0 in (5.9), for each l in the ranges
min(n − k + 1, q) 6 l 6 q − 1,
q + 1 6 l 6 max(k − 1, q),
(5.13)
which were not exploited in the proof of the lemma. In all, there are k − 1
constraint equations. They serve (if k > 2) to do two things: (i) they yield an
expression for σn as a rational function of σ̂1 , . . . , σ̂n−k , and hence as a rational
function of x1 , . . . , xk , homogeneous of degree n; and, (ii) they impose k − 2
homogeneous conditions on x1 , . . . , xk , which are the desired defining equations
(k)
of the curve Cp,q ⊂ Pk−1 .
The formula for σq in (5.9), as yet unused, yields an expression for σq as
a rational function of x1 , . . . , xk , homogeneous of degree q. Therefore, ζ ∝
(k)
σn q /σq n is homogeneous of degree zero and is in the function field of Cp,q , and
1
Pk−1 ⊃ C(k)
p,q ∋ [x1 : . . . : xk ] 7→ ζ ∈ P
(5.14)
(k)
(k)
yields a formula for the degree-(n − k + 1)k (partial) map πp,q : Cp,q → P1ζ , i.e.,
an explicit formula ζ = ζ(x1 , . . . , xk ).
The just-sketched elimination procedure can be carried out by hand for the
cases k = 2 (see Lemma 5.5 below) and k = 3 (see Thm. 6.5). When k increases
further, carrying it out by hand becomes increasingly difficult.
Lemma 5.5. Applying the elimination procedure to the case k = 2 yields the
formulas
xn − xn2
xq1 − xq2
σq = (−)q−1 1p
,
p
p,
x1 − x2
x1 − xp2
nn
(xp − xp2 )p (xq1 − xq2 )q
nn σ q
,
ζ = (−)n p q nn = p q (x1 x2 )pq 1
p q σq
p q
(xn1 − xn2 )n
σn = (−)n (x1 x2 )p
(2)
(2)
the last of which defines a degree-[n(n − 1)] covering map πp,q : Cp,q → P1ζ .
Proof. By elementary algebra.
(k)
(k)
Remark 5.5.1. If one simply defines πp,q : Cp,q → P1ζ as the composition of
(2)
the two rational maps [x1 : . . . : xk ] 7→ [x1 : x2 ] and πp,q , then it will follow
(k)
immediately from Lemma 5.5 that ζ is in the function field of Cp,q , without
actually employing the just-sketched elimination procedure to derive an explicit
rational formula ζ = ζ(x1 , . . . , xk ).
The case k = 1 obviously requires special treatment, since one cannot express
σn , σq in terms of x1 alone. If k = 1, the system (5.9) reduces to
0 = σ1 = x1 + σ̂1 ,
. . . , 0 = σq−1 = x1 σ̂q−2 + σ̂q−1 ;
σ = x σ̂
q
1 q−1 + σ̂q ;
(5.15)
0
=
σ
q+1 = x1 σ̂q + σ̂q+1 , . . . , 0 = σn−1 = x1 σ̂n−2 + σ̂n−1 ;
σn = x1 σ̂n−1 ,
24
the solution of which can be written as
σn = (−)n−q−1 x1n−q σ̂q ,
σq = σ̂q + (−)q−1 xq1 .
(5.16)
(n)
This suggests focusing on s, the element of the function field of Cp,q defined by
1 − s := (−)q−1 σq /xq1
= g/xq1
s := (−)n−1 σn /xn1 ,
= β/xn1 ,
(5.17)
(the two definitions being equivalent), and defining the special k = 1 Schwarz
(1)
curve Cp,q to be P1s .
(2)
(2)
(1)
There is then a degree-(n − 1) map φp,q : Cp,q → Cp,q given by the map t =
(x1 + x2 )/(x1 − x2 ) 7→ s, since both σn , σq are rational functions of x1 , x2 . By
comparing the expressions for σn , σq in Lemma 5.5 with those for σn /xn1 , σq /xq1
(2)
in (5.17), one finds that this degree-(n − 1) map φp,q is given by
xp2 xq1 − xq2
1 xn − xn2
= 1 − q 1p
q
p
p
x1 x1 − x2
x1 x1 − xp2
(5.18)
(t − 1)p (t + 1)q − (t − 1)q
(t + 1)n − (t − 1)n
1
=−
=1−
.
(t + 1)q (t + 1)p − (t − 1)p
(t + 1)q (t + 1)p − (t − 1)p
s=−
(2)
(n)
(n)
(1)
The degree-(n − 1)! composition φp,q ◦ · · · ◦ φp,q : Cp,q → Cp,q ∼
= P1s is made
explicit by the ratios zj = xj /x1 , j = 1, . . . , n, being the solutions z of
z n − (1 − s)z p − s = 0,
(5.19)
so that z2 , . . . , zn are the inverse images (with multiplicity) of s under the corresponding degree-(n − 1) rational function s = s(z), which is
s=−
z p (1 + z + · · · + z q−1 )
z p (1 − z q )
=
−
.
1 − zp
1 + z + · · · + z p−1
(5.20)
(1)
(1)
There is a final degree-n map φp,q : Cp,q ∼
= P1s → P1ζ , given by
ζ = (−)n
sq
nn
nn σn q
= (−)q p q
.
p
q
n
p q σq
p q (1 − s)n
(5.21)
(n)
It completes the sequence of maps leading from the top curve Cp,q down to P1ζ .
Pulling everything together yields the following theorem, which summarizes
the results of this section.
Theorem 5.6. For any pair of relatively prime integers p, q > 1 with n = p+ q,
there is a sequence of algebraic curves and (partial ) maps
(n)
φp,q
(n−1)
φp,q
φ(3)
p,q
φ(2)
p,q
φ(1)
p,q
(n−1)
(1)
1
C(n)
−→ · · · −→ C(2)
p,q −→ Cp,q
p,q −→ Cp,q −→ Pζ ,
(k)
(5.22)
(k)
in which deg φp,q = n − k + 1. For k = n, n − 1, . . . , 2, the curve Cp,q ⊂ Pk−1 ,
prior to closure, comprises all [x1 : . . . : xk ] in which x1 , . . . , xk is a nonzero
25
ordered k-tuple of roots (with multiplicity) of some trinomial equation of the
(k)
form (5.1). Each partial map φp,q , n > k > 2, takes [x1 : . . . : xk ], where at
(k)
least one of x1 , . . . , xk−1 is nonzero, to [x1 : . . . : xk−1 ]. One writes πp,q for
(1)
(k)
φp,q ◦ · · · ◦ φp,q , which is of degree (n − k + 1)k . Any [x1 : . . . : xk ] in the domain
(k)
of πp,q is taken by it to ζ, computed from the associated trinomial equation by
Eq. (5.2).
(k)
(2)
The final two curves Cp,q , k = 2, 1, are of genus zero; i.e., Cp,q ∼
= P1t and
(1) ∼ 1
Cp,q = Ps , where t := (x1 + x2 )/(x1 − x2 ) and s are rational parameters. The
(2)
(1)
final two maps φp,q , φp,q are given by
(t − 1)p (t + 1)q − (t − 1)q
,
(t + 1)q (t + 1)p − (t − 1)p
n
sq
q n
,
ζ = φ(1)
p,q (s) = (−)
p
q
p q (1 − s)n
s = φ(2)
p,q (t) = −
(2)
(1)
(2)
and their composition πp,q = φp,q ◦ φp,q by
p
p p
q
q q
nn 2
(2)
(2)
pq [(t + 1) − (t − 1) ] [(t + 1) − (t − 1) ]
ζ = πp,q
(t) = φ(1)
.
p,q φp,q (t) = p q (t −1)
n
p q
[(t + 1)n − (t − 1)n ]
6. Schwarz curves: Ramifications and genera
(k)
(q)
(n)
The Schwarz curves Cp,q introduced in § 5, in particular Cp,q , Cp,q , will be
used in § 7 to parametrize the solutions of hypergeometric differential equations
(En ’s) with imprimitive monodromy. The explicit formulas of Thm. 5.6 will be
(1)
(2)
especially useful. They exploit the parametrizations of Cp,q , Cp,q by respective
(2)
(1)
rational (i.e., P1 -valued) parameters t, s, which exist because Cp,q , Cp,q are of
genus zero.
(k)
The question arises whether ‘higher’ Cp,q , such as the family of projective
(3)
plane curves Cp,q ⊂ P2 , are ever of genus zero. In principle, the genus of each
(k)
Cp,q ⊂ Pk−1 , n > k > 2, can be calculated from the Hurwitz formula applied
(k)
(k)
(k)
to the map πp,q : Cp,q → P1ζ . But some care is needed, since in general, Cp,q is
a singular projective curve, and not being smooth, is not a Riemann surface.
By resolving singularities, one must first construct a smooth desingularization
(k)
(k)
(k)
C̃p,q → Cp,q . (Up to birational equivalence, C̃p,q is unique.) The Hurwitz
(k)
(k)
formula can then be applied to the lifted map π̃p,q : C̃p,q → P1ζ , which is a
degree-(n − k + 1)k holomorphic map of Riemann surfaces.
(k)
Section 6.1 counts the singular points of Cp,q and determines their multiplic(k)
ities. (See Table 2.) Section 6.2 determines the ramification structure of π̃p,q .
(3)
Section 6.3 explicitly parametrizes several plane curves Cp,q ⊂ P2 , which ac(k)
cording to the resulting formula for the genus of Cp,q (Thm. 6.3) are of genus
zero.
26
6.1. Desingularization
(k)
Recall that Cp,q ⊂ Pk−1 , n > k > 2, is an algebraic curve including each of
the points [x1 : . . . : xk ] ∈ Pk−1 in which x1 , . . . , xk are k of the n roots (with
multiplicity) of some trinomial equation
xn − g xp − β = 0
(6.1)
(with n = p + q, gcd(p, q) = 1, and at most one of g, β equaling zero). Cremona
inversion in Pk−1, i.e., the substitution xj = 1/x′j , j = 1, . . . , k, induces an
(k)
(k)
(k)
equivalence Cp,q ∼
= Cq,p . It was noted (see Remark 5.3.2) that if k 6 p, Cp,q must
k−1
really be defined as a closure in P
, and the taking of the closure appends
a finite number of limit points, which do not come directly from any trinomial
equation of the form (6.1). This will be elucidated below.
It is a standard fact (see § 6.2) that if β 6= 0, at most two roots of (6.1) can
coincide; and coincidence occurs if and only if
ζ := (−)q
nn β q
pp q q g n
(6.2)
(k)
equals unity. Also, if a point [x1 : . . . : xk ] ∈ Cp,q comes directly from an
equation of the form (6.1), the coefficients g, β are determined by the point, up
to the scaling (g, β) 7→ (cq g, cn β) for some c 6= 0. (To see this, consider the
system xn1 − gxp1 − β = 0, xn2 − gxp2 − β = 0, where by the preceding, x1 6= x2
can be assumed. Since gcd(n, p) = 1, if this system has a solution (g, β), it
has a unique solution; but of course (x1 , x2 ) can be multiplied by c.) The
(k)
(k)
formula (6.2), which is unaffected by scaling, defines the map πp,q : Cp,q → P1ζ .
(k)
(k)
The points on Cp,q of special interest here are those taken by πp,q to ζ =
0, 1, ∞. Among points coming from trinomials, the “ζ = ∞” points are ones
with g = 0, and the “ζ = 0” points are ones with β = 0. (It will be seen that
each of the limit points present if k ≤ p is also a ζ = 0 point, by continuity.)
(k)
(k)
A ‘generic’ point on Cp,q is one that is taken by πp,q to some ζ 6∈ {0, 1, ∞}.
(k)
k−1
Which points of Cp,q ⊂ P
are non-smooth will now be determined. First,
(k)
consider any generic point [a] = [a1 : . . . : ak ] ∈ Cp,q . That is, consider
a1 , . . . , ak , distinct and nonzero, which are k of the roots (with multiplicity)
of a unique trinomial xn − g0 xp − β0 with g0 , β0 6= 0. Near [a], treat β as a local
(k)
(k)
parameter on Cp,q . That is, for β near β0 , let [a(β)] = [a1 (β) : . . . : ak (β)] ∈ Cp,q
n
p
come from the roots of x − g0 x − β. For each j,
d n
daj
=
(x − g0 xp )
dβ
dx
x=aj
−1
−1
= najn−1 − pg0 ap−1
,
j
(6.3)
which (by distinctness of roots) is the reciprocal of a nonzero quantity, for β
near β0 . Without loss of generality, assume [a] is in the affine chart on Pk−1 with
(k)
coordinate (z1 , . . . , zk−1 ), where zj = xj /xk . To show that Cp,q is smooth at [a],
27
it suffices to show that (d/dβ)(z1 , . . . , zk−1 )|β=β0 6= (0, . . . , 0), with zj = aj /ak .
Suppose not; in fact, suppose merely that
d
daj
dak
−2
(aj /ak ) = ak
ak
− aj
(6.4)
dβ
dβ
dβ
is zero at β = β0 for j equal to some j ∗ , 1 6 j ∗ 6 k − 1. This is equiv∗
alent to a−1
j (d/dβ)aj |β=β0 taking equal values when j = j , k, or (by (6.3))
p
n
naj − pg0 aj taking equal values. But, equal values are also taken when j = j ∗, k
by anj − g0 apj = β0 . Hence anj∗ = ank and apj∗ = apk ; which since gcd(n, p) = 1,
contradicts aj∗ 6= ak .
(k)
Points [a] ∈ Cp,q coming from trinomials with g = g0 = 0 (i.e., “ζ = ∞”
points) can similarly be shown to be smooth, by using g rather than β as a
1/n π
local parameter. Any point [a] coming from xn − β0 = 0 has aj = β0 εnj,
where π1 , . . . , πk are distinct elements of {0, 1, . . . , n − 1}. For g near g0 = 0,
(k)
let [a(g)] = [a1 (g) : . . . : ak (g)] ∈ Cp,q come from the roots of xn − gxp − β0 , and
use the affine coordinate (z1 , . . . , zk−1 ) on Pk−1 , where zj = xj /xk . Then
−1
d q
−p
= (nβ0 )−1 ap+1
,
(6.5)
(x − β0 x ) x=a
=
j
g=g0 =0
j
dx
daj
dak
d
= (nβ0 )−1 (aj /ak )(apj − apk ).
(aj /ak ) g=g0 =0 = a−2
ak
− aj
k
g=g0 =0
dg
dg
dg
(6.6)
daj
dg
The latter is zero for j = 1, . . . , k − 1 only if ap1 , . . . , apk are equal, but they are
(k)
not. So, (d/dg)(z1 , . . . , zk−1 )|g=g0 =0 6= (0, . . . , 0), and Cp,q is smooth at [a].
(k)
The curve Cp,q ⊂ Pk−1 can also be shown to be smooth at any “ζ = 1”
point, i.e., any point coming from a trinomial with β 6= 0 that has a pair of
(k)
coincident roots. Let [a] be a point on Cp,q with aj1 = aj2 , coming from a
trinomial xn − g0 xp − β0 with β0 6= 0. On a neighborhood of [a], a parameter
for the curve can be chosen to be t = (β − β0 )1/2 , so that aj1 (t), aj2 (t) are
aj1 (0) ± Ct + O(t2 ) for some C 6= 0, and aj (t) = aj (0) + Cj t2 + O(t3 ) for each
j 6= j1 , j2 . The derivative with respect to t of any affine coordinate (zj )j6=j ∗ =
(xj /xj ∗ )j6=j ∗ on Pk−1 , where j ∗ is chosen so that j ∗ 6= j1 , j2 , is defined and
(k)
nonzero at t = 0. Hence, Cp,q is smooth at [a].
(k)
In summary, a point in Cp,q ⊂ Pk−1 coming from a trinomial xn −gxp −β can
be non-smooth only if g 6= 0 and β = 0. That is, it must be one of the “ζ = 0”
(k)
points, which may display high-order coincidences of roots. Limit points on Cp,q
coming not directly but indirectly from trinomials, via the taking of the closure
in Pk−1 (and in fact, from a β → 0 limit), will be considered later.
(k)
For ease of understanding, the singular points of Cp,q will first be determined
k−1
in the ‘top’ case k = n, in which the closure in P
need not be taken. There
(n)
n−1
are n!/p!q points p = [x1 : . . . : xn ] in Cp,q ⊂ P
that come from trinomials
28
xn − gxp − β with g 6= 0 and β = 0. One of them (cf. (4.2)) is p0 , defined by
(
εq−(j−1) g 1/q ,
j = 1, . . . , q,
xj =
(6.7)
0,
j = q + 1, . . . , n.
Here g 1/q signifies one of the q’th roots of g, chosen arbitrarily. Each point p
(n)
at which Cp,q can be non-smooth is obtained from p0 by a permutation of
x1 , . . . , xn .
It will now be shown that p0 is a ordinary (p − 1)!-fold multiple point of the
(n)
curve Cp,q : there are exactly (p − 1)! branches (local irreducible components)
(n)
of Cp,q at p0 , so that the curve is non-smooth at p0 if and only p > 2. Moreover,
in a neighborhood of p0 , each of the (p − 1)! branches at p0 is holomorphically
parametrized by ξ := (−β)1/p . This is a consequence of the fact that near
β = β0 = 0 (g 6= 0 being held fixed), each of the n roots of xn − gxp − β has a
convergent Puiseux expansion in (−β)1/p .
Assume the roots are ordered so that [x1 : . . . : xn ] → p0 as β → 0, i.e., so
that for each j, the limit of xj agrees with (6.7). (This assumption fixes the
ordering of x1 , . . . , xq , but leaves unspecified that of xq+1 , . . . , xn .) Then the
Puiseux expansions of x1 , . . . , xn in ξ will be of the form
i
h
X∞
ε−(j−1)
j = 1, . . . , q,
cji ξ pi ,
g 1/q 1 +
q
i=1
i
h
(6.8)
xj (ξ) =
X∞
qi
j
(εpᾱj ξ) g −1/p 1 +
,
j
=
q
+
1,
.
.
.
,
n,
ξ)
c̄i (εᾱ
p
i=1
where (ᾱq+1 , . . . , ᾱn ) is an unspecified permutation of (0, 1, . . . , p − 1). Of these
two types of expansion, the first is obtained from xn − gxp − β = 0 by applying
the implicit function theorem to xn − gxp + ξ p = 0, and the second by applying
the theorem to ξ q y n − gy p + 1 = 0 (where y := x/ξ), which is equivalent.
(n)
The branch of Cp,q at p0 that results from (6.8), locally parametrized by ξ, is
affected by the choice of the permutation (ᾱq+1 , . . . , ᾱn ). But cyclically shifting
(ᾱq+1 , . . . , ᾱn ) by (1, . . . , 1) is equivalent to multiplying ξ by εp , which does not
alter the branch. Hence when counting branches, one must quotient out the
group Cp of cyclic permutations: there are not p! but (p − 1)! distinct branches,
with distinct tangent lines.
(n)
Like p0 , each of the n!/p!q points p ∈ Cp,q coming from a trinomial with β = 0
(n)
is an ordinary (p − 1)!-fold multiple point; therefore the curve Cp,q ⊂ Pn−1 is
singular if and only if p > 2. Each such point p, including p0 , lifts to (p − 1)!
(n)
(n)
(n)
points on the desingularization C̃p,q → Cp,q , and each of the other points on Cp,q
(n)
lifts to a single point. Each point p̃ ∈ C̃p,q lifted from one of the n!/p!q points
(n)
(n)
p ∈ Cp,q has a neighborhood in C̃p,q that is biholomorphic to a neighborhood of
the point 0 on the complex ξ-line. That is, near each of these (n!/p!q)×(p−1)! =
(n)
n!/pq lifted points p̃, ξ can be used as a local parameter on C̃p,q .
(n)
(n)
(n)
(n)
Composing C̃p,q → Cp,q with the map πp,q : Cp,q → P1ζ , yields the holomor(n)
(n)
phic lifted map π̃p,q : C̃p,q → P1ζ , which near any of these points p̃ is effectively
29
(k)
Table 2: Branching data for the ordinary multiple points on the algebraic curve Cp,q ⊂ Pk−1,
k > 2. They are partitioned into types indexed by ν, with max(0, k − p) 6 ν 6 min(q, k). If
P (k, ν) points of type ν.
nonzero, ν counts the j, 1 6 j 6 k, for which xj 6= 0. There are Np,q
ν=0
0<ν<k
ν=k
P
Np,q
(k, ν)
(p− k + 1)k−1
k
(q − ν + 1)ν−1
ν
(q − k + 1)k−1
T
Np,q
(k, ν)
1
(p − k + ν + 1)k−ν−1
1
Mp,q (k, ν)
p
pq
q
a holomorphic function ζ = ζ(ξ), defined on a neighborhood of ξ = 0. Since ζ
(n)
as defined by (6.2) is proportional to β q and hence to ξ pq , the composition π̃p,q
takes each of these points p̃ to ζ = 0 with multiplicity pq.
(k)
The desingularization of any subsidiary Schwarz curve Cp,q , n > k > 2, is
(n)
similar to that of the top curve Cp,q . With g 6= 0 fixed, consider any k-tuple
of roots x1 , . . . , xk of the trinomial (6.1), and let ν denote the number of these
k roots that tend to nonzero values as β → 0. Necessarily, 0 6 ν 6 k with ν 6 q
and k − ν 6 p. These k roots can be expanded as power series in ξ := (−β)1/p ,
and up to a permutation of x1 , . . . , xk , the expansions will be of the form
h
i
X∞
j 1/q
εα
cji ξ pi ,
1+
j = 1, . . . , ν,
q g
i
h i=1
(6.9)
xj (ξ) =
X∞
qi
−1/p
j
j
(εᾱ
,
j = ν + 1, . . . , k,
c̄i (εᾱ
1+
p ξ)
p ξ) g
i=1
for certain distinct αj ∈ {0, . . . , q − 1} and certain distinct ᾱj ∈ {0, . . . , p − 1}.
In (6.9), it is the initial ν components of (x1 , . . . , xk ) that do not tend to zero
as β → 0, and the final k − ν components that do.
The case ν = 0 occurs only when k 6 p and is clearly special, since the limit
of (x1 , . . . , xk ) as β → 0 in this case is (0, . . . , 0); but the limit of [x1 : . . . : xk ] as
ᾱk
1
β → 0 exists in Pk−1 as the point [εᾱ
p : . . . : εp ]. This explains Remark 5.3.2,
(k)
on the need to take the closure when defining Cp,q ⊂ Pk−1 , if k 6 p. On the fibre
(k)
of Cp,q over ζ = 0, there are (p − k + 1)k−1 = (p − k + 1)k /p distinct limit points
[x1 : . . . : xk ] of this ν = 0 type, in which x1 , . . . , xk are distinct p’th roots of
unity. The case ν = k occurs only when k 6 q, and is also rather special. On
the fibre over ζ = 0, there are (q − k + 1)k−1 = (q − k + 1)k /q distinct points of
the ν = k type, in which x1 , . . . , xk are distinct q’th roots of unity.
A straightforward extension of the preceding treatment of the top curve
(n)
Cp,q , which in retrospect was a treatment of the case (k, ν) = (n, q), yields the
P
(k, ν) is the number of ordinary multiple points
branching data in Table 2. Np,q
(k)
p ∈ Cp,q ⊂ Pk−1 of the type indexed by ν, coming from a trinomial equation with
T
β = 0 (or in the ν = 0 case, from a β → 0 limit). For each such point, Np,q
(k, ν)
is the multiplicity: the number of distinct branches at p, or equivalently the
(k)
(k)
number of points p̃ to which p lifts on the desingularization C̃p,q → Cp,q . For
T
P
example, Np,q
(n, q) = n!/p!q and Np,q
(n, q) = (p − 1)!; these are the previously
30
derived top-curve values.
P
The formula kν (q − ν + 1)ν−1 , given in the table for Np,q
(k, ν) if 0 < ν < k,
comes from (i) choosing ν of the roots x1 , . . . , xk to be the ones with nonzero
limits as β → 0, and (ii) choosing the ν associated exponents αj (i.e., powers
of εq ) to be distinct elements of {0, 1, . . . , q − 1}, with the group Cq of cyclic
(k)
permutations quotiented out. At any p ∈ Cp,q specified in this way, there are
(k)
T
Np,q
(k, ν) = (p − k + ν + 1)k−ν−1 branches of Cp,q . This is because a branch
is specified by a choice of k − ν distinct exponents ᾱj (i.e., powers of εp ) from
{0, 1, . . . , p − 1}, with the group Cp of cyclic permutations quotiented out.
P
T
For each ν with 0 < ν < k, each of the Np,q
(k, ν) × Np,q
(k, ν) lifted points
(k)
p̃ ∈ C̃p,q has a neighborhood that is biholomorphic to a neighborhood of ξ = 0.
The cases ν = 0, resp. k are special, since the appropriate local parameter for
(k)
[x1 : . . . : xk ] ∈ Cp,q is not ξ but rather ξ ′ := ξ q , resp. ξ ′′ := ξ p , as is evident
T
from (6.9). In both these cases there is only one branch, i.e., Np,q
(k, 0) = 1
T
(each appended limit point is a smooth point), resp. Np,q (k, k) = 1.
The final quantity Mp,q (k, ν) in the table is the multiplicity with which each
(k)
(k)
lifted point p̃ ∈ C̃p,q is taken to the point ζ = 0 in P1ζ by the lifted map π̃p,q ,
′
which is locally a holomorphic function ζ = ζ(ξ), or ζ = ζ(ξ ) resp. ζ = ζ(ξ ′′ ).
This multiplicity is usually pq, as previously seen in the top case, but in the
special case ν = 0, resp. k, it is p, resp. q, because ζ ∝ ξ pq = (ξ ′ )p = (ξ ′′ )q .
Note that
min(q,k)
X
P
T
Np,q
(k, ν) · Np,q
(k, ν) · Mp,q (k, ν) = (n − k + 1)k ,
(6.10)
ν=max(0,k−p)
since the left side equals the number of points (with multiplicity) on the fibre
(k)
(k)
of π̃p,q above ζ = 0, i.e., deg π̃p,q = (n − k + 1)k . By substituting the data of
Table 2 into (6.10), one obtains a familiar binomial coefficient identity.
6.2. Genera
It is now possible to determine, for each k > 2, the ramifications of the
(k)
(k)
degree-(n − k + 1)k holomorphic map of Riemann surfaces π̃p,q : C̃p,q → P1ζ ,
(k)
(k)
(k)
where C̃p,q → Cp,q is the desingularization. Any p̃ ∈ C̃p,q can be mapped with
nontrivial multiplicity to P1ζ only if (i) its image, as in § 6.1, is some point
(k)
(k)
p ∈ Cp,q on the fibre of πp,q over ζ = 0; or, (ii) its image p is taken with
(k)
nontrivial multiplicity by πp,q to some point ζ 6= 0. Case (ii) can occur only if
(k)
(k)
(k)
ζ = πp,q (p) is a critical value of πp,q , i.e., only if πp,q −1 ζ comprises fewer than
(n)
(k)
(n − k + 1)k points on Cp,q ; i.e., only if πp,q −1 ζ comprises fewer than n! points
(n)
on Cp,q .
Case (ii) can accordingly occur only if two of the n roots (with multiplicity)
of the trinomial xn − gxp − β giving rise to p coincide; or, if the roots are distinct
n−1
but are proportional to {εm
n }m=0 , the n’th roots of unity, so that there are only
(n − 1)! distinct ordered n-tuples [x1 : . . . : xn ] ∈ Pn−1 . By evaluating the
31
compact and elegant formula for the discriminant of the trinomial [17, 23], one
finds that when β 6= 0, the first of these two subcases occurs only if
(pg/n)n − (−pβ/q)q = 0,
(6.11)
which by (6.2) is equivalent to ζ = 1. In this subcase, exactly two of the n roots
coincide. (For a description of the monodromy of the roots around ζ = 1, see [26,
§ 2].) The second subcase occurs only if g = 0, i.e., ζ = ∞. One concludes that
(k)
(k)
π̃p,q : C̃p,q → P1ζ can be ramified only over the three points ζ = 0, 1, ∞. It is a
so-called Belyı̆ cover.
If ζ = 1, i.e., Eq. (6.11) holds, without loss of generality one can take
g = n/p, β = −q/p, so that the trinomial is proportional to pxn −nxp +q, with a
(n)
single doubled root at x = 1. There are n!/2 distinct points [x1 : . . . : xn ] ∈ Cp,q
on the fibre over ζ = 1: namely, points in which x1 , . . . , xn are permutations of
the roots of pxn − nxp + q, including the single doubled root.
Definition 6.1. For each p, q > 1 with gcd(p, q) = 1, a degree-(p + q − 2)
Pp+q−2
polynomial with simple roots Tp,q (x) = j=0 tj xj , satisfying
p xp+q − (p + q)xp + q = (x − 1)2 Tp,q ,
is defined by
tj =
(
(j + 1)q,
(p + q − 1 − j)p,
0 6 j 6 p − 1,
p − 1 6 j 6 p + q − 2.
Its roots will be denoted x∗p,q;α , 1 6 α 6 p + q − 2.
Theorem 6.2. For each k, n > k > 2, the degree-(n − k + 1)k map of Riemann
(k)
(k)
surfaces π̃p,q : C̃p,q → P1ζ is a Belyı̆ cover: it is ramified only over ζ = 0, 1, ∞.
1. The fibre over ζ = 0 contains (p − k + 1)k−1 , resp. (q − k + 1)k−1 points
of multiplicity p, resp. q, all other points being of multiplicity pq.
2. The fibre over ζ = 1 contains (n − k − 1)k points of unit multiplicity, all
other points being of multiplicity 2.
3. The fibre over ζ = ∞ contains (n − k + 1)k−1 points of multiplicity n.
Remark 6.2.1. The case k = n of this theorem, dealing with the top Schwarz
(n)
curve Cp,q , was proved by Kato and Noumi [21]. Note that the fibre over ζ = 0
contains no points of multiplicity p if k > p, and none of multiplicity q if k > q;
and that the fibre over ζ = 1 contains no points of unit multiplicity if k > n − 1.
Proof. The facts about the fibre over ζ = 0 can be read off from Table 2.
(k)
Above ζ = 1, the points of Cp,q that occur with unit multiplicity are the
∗
∗
points [xp,q;χ(1) : . . . : xp,q;χ(k) ], where the roots x∗p,q;α , 1 6 α 6 n − 2 were
defined above, and χ(1), . . . , χ(k) are distinct integers selected from 1, . . . , n− 2.
(k)
The points of Cp,q that occur with multiplicity 2 are those in which, instead,
one or two of the xj ’s are equal to the double root 1.
32
(k)
As was already remarked, the points of Cp,q above ζ = ∞ are the points
[x1 : . . . : xk ] ∈ Pk−1 in which x1 , . . . , xk are distinct n’th roots of unity. There
are exactly (n − k + 1)k−1 = (n − k + 1)k /n such points.
(k)
Theorem 6.3. For each n > k > 1, the genus of Cp,q ⊂ Pk−1 as an algebraic
(k)
curve, and the topological genus of the Riemann surface C̃p,q , are equal to
1+
(k − 1)(2n − k − 2)
n
p−1
q−1
−
(p−k+1)k−1 −
(q−k+1)k−1 .
(n−k+1)k−1 −
4(n − 1)
2pq
2q
2p
Proof. Apply the Hurwitz genus formula to the data given in Thm. 6.2 (which
(k)
(k)
trivially extends to k = 1). The genus is stable under p ↔ q, since Cp,q ∼
= Cq,p ,
even though the singular points (i.e., their number and type) are not.
Corollary 6.4. (i) The following Schwarz curves, and only these, are rational,
i.e., of genus zero.
(1)
(2)
◦ The trivial curves Cp,q ∼
= P1s and Cp,q ∼
= P1t , for all coprime p, q > 1.
(3)
◦ Cp,q for {p, q} = {1, 2} and {1, 3}.
(4)
◦ Cp,q for {p, q} = {1, 3}.
(q)
In consequence, Cp,q can be rationally parametrized only if q = 1, q = 2, or
(n)
(p, q) = (1, 3); and the top curve Cp,q only if {p, q} = {1, 1}, {1, 2}, or {1, 3}.
(ii) The following Schwarz curves, and only these, are elliptic, i.e., of genus 1.
(3)
◦ Cp,q for {p, q} = {1, 4}.
Proof. By Thm. 6.3, these curves are of genus 0 and genus 1 as claimed. The
‘only these’ statements remain to be proved.
(k)
(k−1)
It follows from Thm. 6.2 that for each k with n > k > 2, C̃p,q → C̃p,q is
(k−1)
a covering with nontrivial branching. By the Hurwitz formula, if g(C̃p,q ) > 0
(k)
(k−1)
then g(C̃p,q ) must be strictly greater than g(C̃p,q ). Therefore, to determine
which Schwarz curves with k > 3 are of genus 0 or genus 1, one should focus on
the case k = 3. Substituting k = 3 and n = p + q into Thm. 6.3 yields
2
2
g(C̃(3)
(6.12)
p,q ) = (p + 4 pq + q ) − 9(p + q) + 14 / 2.
By examination, this equals 0 only if {p, q} = {1, 2} or {1, 3}, and equals 1 only
(4)
(4)
if {p, q} = {1, 4}. Moreover, by Thm. 6.3, g(C̃1,3 ) = 0 and g(C̃1,4 ) > 1; so the
‘only these’ statements are proved.
6.3. Projective plane curves
(3)
The curves Cp,q are projective plane curves and include the first Schwarz
(3)
curves of positive genus. The following theorem makes each Cp,q ⊂ P2 and the
(3)
(3)
associated covering map πp,q : Cp,q → P1ζ quite concrete.
33
Theorem 6.5. For all coprime p, q > 1 with at least one of p, q greater than 1
(3)
and n := p + q, the curve Cp,q ⊂ P2 has defining equation
xp1 (xn2 − xn3 ) + xp2 (xn3 − xn1 ) + xp3 (xn1 − xn2 )
= 0,
(x1 − x2 )(x2 − x3 )(x3 − x1 )
(6.13)
where the left side is a symmetric homogeneous polynomial in x1 , x2 , x3 that is
(3)
of degree n + p − 3 and is of degree n − 2 in any single variable. The curve Cp,q
2
goes through the fundamental points of P (i.e., [1 : 0 : 0], [0 : 1 : 0], [0 : 0 : 1])
if and only if p > 2, and is singular if and only if p > 3, in which case each
fundamental point is an ordinary (p − 1)-fold multiple point, and there are no
(3)
(3)
other singular points. The (partial ) covering map πp,q : Cp,q → P1ζ is given by
#
"
q+1
q+1
x
(x
−
x
)
+
cycl.
1
p
p
p
2
3
(−)n−1 x1 x2 x3
,
p > 1,
x1 (xp2 xp+1
− xp+1
xp3 ) + cycl.
3
2
"
#
σn =
xq1 (x2 − x3 ) + cycl.
p
p
p
n−1
x1 x2 x3
,
q > 1;
(−)
(xp2 − xp3 ) + cycl.
xp+1
1
#
"
p n+1
n+1 p
x
(x
x
−
x
x
)
+
cycl.
1
2 3
2
3
,
p > 1,
(−)q−1
x1 (xp2 xp+1
− xp+1
xp3 ) + cycl.
3
2
"
#
σq =
xp+1
(xn2 − xn3 ) + cycl.
q−1
1
(−)
,
q > 1;
xp+1
(xp2 − xp3 ) + cycl.
1
nn σn q
ζ = (−)n p q n ;
p q σq
and is of degree n(n − 1)(n − 2).
Proof. To get (6.13), substitute the formulas for β = (−)n−1 σn , g = (−)q−1 σq
in Lemma 5.5 into the trinomial equation (6.1), and relabel x as x3 . To get the
other formulas, apply the elimination procedure sketched in § 5. (Cf. Lemma 5.5
itself, which results from applying the procedure when k = 2; the details when
k = 3 are long and are omitted.) The singular points (0 or 3 in number) come
from Table 2. If p > 3, there are singular points on the fibre over ζ = 0: each
fundamental point is a “ν = 1” one, but there are none of types ν = 0, 2, 3.
(3)
(3)
Remark 6.5.1. The curves Cp,q , Cq,p ⊂ P2 are transforms of each other: the
(3)
defining equation (6.13) of Cp,q is taken by a Cremona inversion in P2 (i.e.,
xi = 1/x′i ) to itself, altered by p ↔ q; as is the formula for ζ = ζ(x1 , x2 , x3 ).
(3)
The genus of Cp,q was given in (6.12), but comes equally well by substituting
the data of Thm. 6.5 into the standard genus–degree formula
X
ri
N −1
.
(6.14)
g=
−
2
2
i
(3)
Here N = n + p − 3 = 2p + q − 3 is the degree of Cp,q , and the sum is over its
three singular points, each (if p > 3) of multiplicity r = p − 1.
34
(3)
Parametrizing the plane curve Cp,q is a key step leading to a hypergeometric
identity; especially, if either q or n = p + q equals 3. The following examples of
parametrizations are illustrative, and will be exploited in § 7.
(3)
(3)
Example 6.6. {p, q} = {1, 2}, the simplest case. The top curves C1,2 , C2,1 ⊂ P2
are of genus zero (see Cor. 6.4). The defining equations are x1 + x2 + x3 = 0
and x1 x2 + x2 x3 + x3 x1 = 0. They are respectively a line, and a conic through
the fundamental points of P2 ; and are related by a Cremona inversion.
(n)
(n)
(n−1)
(n)
In general, Cp,q ∼
= Cp,q , which is reflected in the fact that φp,q : Cp,q →
(n−1)
is of degree 1. Hence t := (x1 + x2 )/(x1 − x2 ), used here to uniformize
Cp,q
(3)
(2)
(3)
any Cp,q , can be used as a rational parameter for C1,2 , C2,1 , yielding
(
[t + 1 : t − 1 : −2t],
(p, q) = (1, 2);
[x1 : x2 : x3 ] =
(6.15)
2
[−2t(t + 1) : −2t(t − 1) : t − 1],
(p, q) = (2, 1).
But to respect the S3 symmetry it is better to use an alternative parameter t̃
(related to t by a Möbius transformation), thus:
[x1 : x2 : x3 ] =
(
[ω(1 + t̃) : ω̄(1 + ω t̃) : (1 + ω̄ t̃)],
(p, q) = (1, 2);
[ω(1 + ω t̃)(1 + ω̄ t̃) : ω̄(1 + t̃)(1 + ω t̃) : (1 + ω̄ t̃)(1 + t̃)],
(p, q) = (2, 1),
(6.16)
where ω := ε3 and ω̄ := ε3 2 .
(3)
(3)
(3)
(3)
According to Thm. 6.5, the covering π1,2 : C1,2 → P1ζ resp. π2,1 : C2,1 → P1ζ
is performed by a function ζ = ζ(x1 , x2 , x3 ), namely
(x1 x2 x3 )2
27
,
(p, q) = (1, 2);
−
4 (x1 x2 + x2 x3 + x3 x1 )3
(6.17)
ζ=
27
x1 x2 x3
−
,
(p,
q)
=
(2,
1).
4 (x1 + x2 + x3 )3
Substituting (6.16) yields a degree-6 rational map t̃ 7→ ζ, i.e., the Belyı̆ map
1 − t̃ + t̃2
27 s2
(p, q) = (1, 2);
◦
3 2
3
4 (1 − s)
(1 + t̃)2
(1 + t̃ )
ζ=
=
(6.18)
4t̃3
27
s
(1 + t̃)2
−
◦
(p,
q)
=
(2,
1),
4 (1 − s)3 1 − t̃ + t̃2
(2)
(2)
(3)
(3)
(3)
(3)
as the covering π1,2 : C1,2 ∼
= C2,1 → P1ζ . Equa= C1,2 → P1ζ , resp. π2,1 : C2,1 ∼
(2)
(1)
(2)
(2)
(1)
(2)
tion (6.18) exhibits the compositions π1,2 = φ1,2 ◦ φ1,2 and π2,1 = φ2,1 ◦ φ2,1 .
(2)
(2)
(1)
(1)
The maps φ1,2 resp. φ2,1 and φ1,2 resp. φ2,1 , i.e., t̃ 7→ s and s 7→ ζ, are of
degrees 2 and 3. They come from Thm. 5.6, if one takes account of the Möbius
transformation relating t and t̃.
(3)
(3)
Example 6.7. {p, q} = {1, 3}. The curves C1,3 , C3,1 are of genus zero (see
(3)
Cor. 6.4). First consider (p, q) = (1, 3). The defining equation of C1,3 ⊂ P2 is
x21 + x22 + x23 + x1 x2 + x2 x3 + x3 x1 = 0,
35
(6.19)
which follows from Thm. 6.5, or more directly by eliminating x4 from the equations σ1 = 0, σ2 = 0. This is a conic that does not go through the fundamental
points of P2 . It can be parametrized by inspection, with parameter u ∈ P1 , as
[x1 : x2 : x3 ] = [ω(1 − ω̄ u)(1 + 2ω̄ u) : ω̄(1 − ω u)(1 + 2ω u) : (1 − u)(1 + 2u)],
(6.20)
where ω := ε3 . Substituting (6.20) into the function ζ = ζ(x1 , x2 , x3 ) of
Thm. 6.5 yields a degree-24 rational map u 7→ ζ, i.e., the Belyı̆ map
ζ=−
256
(x1 x2 x3 )3 (x1 + x2 + x3 )3
27 (x1 + x2 )4 (x2 + x3 )4 (x3 + x1 )4
(6.21a)
u3 (1 − u3 )3 (1 + 8u3 )3
(1 + 8u6 )2 (1 + 88u3 − 8u6 )2
=1−
3
6
4
(1 − 20u − 8u )
(1 − 20u3 − 8u6 )4
3
2
s
(1 − t)(1 + 3t )
ω − ω̄ 1 − 2u − 2u2
256
,
◦
◦
=−
27 (1 − s)4
(1 + t)3
3
1 + 2u2
= −256
(6.21b)
(6.21c)
(3)
(3)
as the covering π1,3 : C1,3 ∼
= P1u → P1ζ . Equation (6.21c) exhibits the composition
(3)
(1)
(2)
(3)
(3)
(2)
(1)
π1,3 = φ1,3 ◦ φ1,3 ◦ φ1,3 . The maps φ1,3 , φ1,3 , φ1,3 , i.e., u 7→ t, t 7→ s, s 7→ ζ, are of
degrees 2, 3, 4. They come from Theorem 5.6 and the fact that the parameter t
(2)
on any curve Cp,q equals (x1 + x2 )/(x1 − x2 ).
(4)
(3)
The top Schwarz curve C1,3 ⊂ P3 is birationally equivalent to C1,3 , since the
(4)
(3)
map φ(4) : C1,3 → C1,3 is of degree 1; so it too can be parametrized by u. Solving
the equation σ1 = 0 for x4 = x4 (u) yields x4 = −3u, hence
[x1 : x2 : x3 : x4 ]
= [ω(1 − ω̄ u)(1 + 2ω̄ u) : ω̄(1 − ω u)(1 + 2ω u) : (1 − u)(1 + 2u) : −3u] (6.22)
(4)
is an (asymmetric) rational parametrization of C1,3 . Equation (6.21b) can be
(4)
(4)
viewed as defining the degree-24 covering π1,3 : C1,3 ∼
= P1u → P1ζ .
The treatment of the case (p, q) = (3, 1) is similar. The genus-zero curve
(3)
(3)
C3,1 ⊂ P2 is obtained from C1,3 ⊂ P2 by Cremona inversion (xi = 1/x′i ), i.e., by
the standard quadratic transformation (xi = x′j x′k ). It has defining equation
x21 x22 + x22 x23 + x23 x21 + x21 x2 x3 + x22 x3 x1 + x23 x1 x2 = 0
(6.23)
and is a trinodal quartic; it goes through the fundamental points [1 : 0 : 0],
[0 : 1 : 0], [0 : 0 : 1], and each is an ordinary double point (‘node’). The smooth
points [1 : ω : ω̄], [1 : ω̄ : ω] are notable for being appended limit points in the
sense of § 6.1: they come indirectly rather than directly from trinomial roots.
By undoing the quadratic transformation, a rational parametrization by u
(3)
(4)
of C3,1 can be obtained from (6.20), and one of C3,1 from (6.22). Thus the
(3)
(3) ∼
degree-24 rational map u 7→ ζ given in (6.21b) can be used as π : C
=
3,1
(4)
(4)
P1u → P1ζ and π3,1 : C3,1 ∼
= P1u → P1ζ .
36
3,1
Example 6.8. {p, q} = {1, 4}. A discussion of the case (p, q) = (1, 4) will
(3)
suffice. The defining equation of C1,4 ⊂ P2 is
x31 + x32 + x33 + x1 x22 + x1 x23 + x2 x23 + x2 x21 + x3 x21 + x3 x22 + x1 x2 x3 = 0, (6.24)
which follows from Thm. 6.5, or more directly by eliminating x4 , x5 from the
equations σ1 = 0, σ2 = 0, σ3 = 0. This is a smooth cubic that does not
go through the fundamental points of P2 . It is elliptic, i.e., of genus 1, with
Klein–Weber invariant j = −52 /2. It can therefore be uniformized with the aid
of elliptic functions, but it is easier to construct a multivalued parametrization
with radicals. As usual, let t = (x1 +x2 )/(x1 −x2 ), so that [x1 : x2 ] = [t+1 : t−1],
and notice that as Thm. 6.5 predicts, Eq. (6.24) is of degree n − 2 = 3 in x3 . By
symmetry, x3 , x4 , x5 are the three roots, and they are computable in terms of
(3)
(4)
(5)
radicals from t by Cardano’s formula. It follows that each of C1,4 , C1,4 , C1,4 has a
multivalued parametrization with radicals in terms of t. These parametrizations
are respectively 3, 6, and 6-valued.
The technique of the last example immediately yields the following theorem.
Theorem 6.9. For all coprime p > 1, q > 2 with n := p + q 6 6, one can
construct multivalued parametrizations with radicals for the subsidiary curve
(n)
(q)
Cp,q ⊂ Pq−1 and the top curve Cp,q ⊂ Pn−1 , respectively (p + 1)q−2 -valued and
(n − 2)!-valued.
7. Identities with free parameters
The results of §§ 5 and 6 will now be put to use, by deriving an interesting
collection of parametrized hypergeometric identities. The source of many is
Thm. 4.4, which expressed certain n Fn−1 ’s in terms of algebraic functions. The
expressions in parts (i) and (ii) of that theorem involve the roots x1 , . . . , xq ,
resp. x1 , . . . , xn , of the trinomial equation
xn − g xp − β = 0,
(7.1)
with n := p + q and gcd(p, q) = 1. It follows that any parametrization of the
(q)
(n)
Schwarz curve Cp,q , resp. Cp,q , will yield a hypergeometric identity. The curves
(q)
(n)
Cp,q , Cp,q of genus zero were classified in Cor. 6.4, and the resulting identities
are given in §§ 7.1, 7.2, 7.3 below, the respective parameters used being s, t, u.
(1)
(2)
These are respectively the rational parameters for any Cp,q , any Cp,q , and (by
(3)
(3)
Example 6.7) the plane curves C1,3 , C3,1 .
Each identity derived from Thm. 4.4 in this section is really a family of
identities: the represented n Fn−1 depends on a discrete parameter ℓ ∈ Z, and
the identity involves a rational function Fℓ = Fℓ (A, B; y). The functions Fℓ
were defined in § 3 (see Table 1 and Thm. 3.5). The reader will recall that in
particular, F0 ≡ 1 and F1 (A, B; y) = y/[(1 − B)y + B].
37
Each of the n Fn−1 ’s also depends on a parameter a ∈ C. If a is chosen so
that no upper parameter of the corresponding differential equation En differs
by an integer from a lower one, the monodromy group of the En will be of
the imprimitive irreducible type characterized in Thm. 2.3; and if a ∈ Q, the
group will be finite. (The case of equal upper and lower parameters, permitting
‘cancellation,’ is possible only for a finite number of choices of a, such as a = ±1
when ℓ = 0; it was mentioned in Thm. 2.4.) One must treat with care the
possibility that one of the lower parameters may be a non-positive integer, in
which case n Fn−1 is undefined (though it may still be possible to interpret the
identity in a limiting sense; cf. Lemma 2.1). It is assumed for simplicity in this
section that a ∈ C is chosen so that this does not occur, and so that no division
by zero occurs.
Several of the identities below are rationally parametrized formulas for n+1 Fn ’s
rather than n Fn−1 ’s. They come from Thm. 4.6 rather than Thm. 4.4, and instead of the rational functions Fℓ , ℓ ∈ Z, they involve interpolating functions
G0 , G1 that were defined in Thm. 4.6 in terms of 2 F1 , 3 F2 . Each of these identities has an additional free parameter c ∈ C, and a similar caveat applies.
7.1. Parametrizations by s
(q)
(1)
The rational parametrization of the curve Cp,q = Cp,1 by s ∈ P1 given in Eqs.
(5.17) and (5.21), when substituted into part (i) of each of Thms. 4.4 and 4.6,
yields the following.
Theorem 7.1. For each p > 1, with n := p + 1, define a degree-n Belyı̆ map
P1s → P1ζ by
nn
s
ζ = ζp,1 (s) := − p
.
p (1 − s)n
Then for all a ∈ C and ℓ ∈ Z, one has that near s = 0,
!
a
, . . . , a+(n−1)
n
n
ζp,1 (s) = (1 − s)a Fℓ −a, −p; (1 − s)−1 .
n Fn−1
a−ℓ+p
a−ℓ+1
,..., p
p
Moreover, for all a ∈ C, c ∈ C and ℓ = 0, 1, one has that near s = 0,
n+1 Fn
a+(n−1)
a
;
n, . . . ,
n
a−ℓ+(p−1)
a−ℓ
,
.
.
.
,
;
p
p
a+c−ℓ
p
a+c−ℓ+p
p
!
ζp,1 (s)
= (1 − s)a Gℓ −a, −p, −c; (1 − s)−1 .
Proof. Straightforward, as g −1 x1 = (1 − s)−1 by (5.17).
Remark 7.1.1. For each of ℓ = 0, 1, the second identity reduces to the first
when c = 0; and as c → ∞, it reduces to a version of the first in which ℓ is
incremented by 1. Interpolation of this sort is familiar from Thm. 4.6 and will
be seen repeatedly.
38
(1)
(2)
(n)
∼ C by s ∈ P1 ,
Similarly, the parametrization of the top curve Cp,q = C1,1 =
1,1
substituted into part (ii) of each of Thms. 4.4 and 4.6, yields the following.
Theorem 7.2. As in Theorem 7.1, let
ζ1,1 (s) := −
4s
.
(1 − s)2
Then for all a ∈ C and ℓ ∈ Z, and κ = 0, 1, one has that near s = 1,
2 F1
−a + ℓ + κ2 , a +
1
2 +κ
κ
2
1
ζ1,1 (s)
κ
(−) (a + κ/2)1−ℓ−κ
κ/2
=
[−ζ1,1 (s)/4]
(a)1−ℓ
× 21 sa Fℓ −a, 12 ; s−1 + (−)κ s−a Fℓ −a, 21 ; s .
Moreover, for all a ∈ C, c ∈ C and ℓ = 0, 1, and κ = 0, 1, one has that near
s = 1,
3 F2
−a + ℓ + 1 + κ2 , a + κ2 ; −a − c + ℓ + κ2
1
−a − c + ℓ + 1 +
2 + κ;
=
κ
2
1
ζ1,1 (s)
(−)κ (a + κ/2)−ℓ−κ (a + c − ℓ − κ/2)
κ/2
[−ζ1,1 (s)/4]
(a)−ℓ (a + c − ℓ)
× 21 sa Gℓ −a, 21 , −c; s−1 + (−)κ s−a Gℓ −a, 12 , −c; s .
Proof. Straightforward, as β −1/2 x1 = s−1/2 and β −1/2 x2 = −s1/2 .
The cases ℓ = 0, 1 of the identities of Thm. 7.1 were previously obtained
by Gessel and Stanton using series manipulations [12]. (See their Eqs. (5.10)–
(5.13),(5.15).) When p = 2, the ℓ = 0, 1 cases of the first identity become
one-parameter specializations of Bailey’s first cubic transformation of 3 F2 and
its companion.
By comparison, the two identities of Thm. 7.2 are elementary. It should
be possible to derive them, or at least the first, by using contiguous function
relations or other classical hypergeometric techniques.
7.2. Parametrizations by t
(q)
(2)
The rational parametrization of the curve Cp,q = Cp,2 by t ∈ P1 given in
Eq. (5.3) and Lemma 5.5, and the fact that [x1 : x2 ] = [t + 1 : t − 1], when
substituted into part (i) of each of Thms. 4.4 and 4.6, yield the following.
Theorem 7.3. For each odd p > 1, with n := p + 2, define a map P1t → P1ζ by
p
ζ = ζp,2 (t) :=
4nn t2 (1 − t2 )2p [(1 + t)p + (1 − t)p ]
.
n
pp
[(1 + t)n + (1 − t)n ]
39
Then for all a ∈ C and ℓ ∈ Z, and κ = 0, 1, one has that near t = 0,
n Fn−1
a
+ κ2 , . . . , a+(n−1)
+ κ2
n
n
a−ℓ+p
a−ℓ+1
κ
+
,
.
.
.
,
+ κ2 ; 12
p
2
p
+κ
ζp,2 (t)
!
−κ/2
(−)κ (a + nκ/2)1−ℓ−κ 4pp
ζp,2 (t)
=
(a)1−ℓ
nn
(1 + t)p + (1 − t)p
× 21 (1 + t)−2a Fℓ −a, −p/2; (1 + t)2
(1 + t)n + (1 − t)n
(1 + t)p + (1 − t)p
+ (−)κ (1 − t)−2a Fℓ −a, −p/2; (1 − t)2
(1 + t)n + (1 − t)n
n
n a
(1 + t) + (1 − t)
.
×
(1 + t)p + (1 − t)p
Moreover, for all a ∈ C, c ∈ C and ℓ = 0, 1, and κ = 0, 1, one has that near
t = 0,
n+1 Fn
a
+ κ2 , . . . , a+(n−1)
+ κ2 ;
n
n
a−ℓ+(p−1)
a−ℓ
κ
+ 2,...,
+ κ2 ; 12
p
p
a+c−ℓ
+ κ2
p
a+c−ℓ+p
+ κ2
p
!
ζp,2 (t)
+ κ;
−κ/2
(−)κ (a + nκ/2)−ℓ−κ (a + c − ℓ + pκ/2) 4pp
ζ
(t)
=
p,2
(a)−ℓ (a + c − ℓ)
nn
p
p
2 (1 + t) + (1 − t)
−2a
1
× 2 (1 + t)
Gℓ −a, −p/2, −c; (1 + t)
(1 + t)n + (1 − t)n
(1 + t)p + (1 − t)p
+ (−)κ (1 − t)−2a Gℓ −a, −p/2, −c; (1 − t)2
(1 + t)n + (1 − t)n
n
n a
(1 + t) + (1 − t)
.
×
(1 + t)p + (1 − t)p
Remark 7.3.1. The even function ζ = ζp,2 (t) is a degree-[n(n − 1)] Belyı̆ map.
The case ℓ = 0, κ = 0 of the first identity appeared in § 1 as the sample
result (1.2). Note that if κ = 0, there is simplification: the right-hand prefactor
becomes unity.
(2)
(n)
(3)
Similarly, the rational parametrization of the top curve Cp,q = C ∼
=C
1,2
1,2
given in (6.16) by t̃ ∈ P1 (related to t by a Möbius transformation), substituted
into part (ii) of each of Thms. 4.4 and 4.6, yields the following.
Theorem 7.4. Define a degree-6 Belyı̆ map P1t̃ → P1ζ by
ζ = ζ1,2 (t̃) :=
(1 + t̃3 )2
(1 − t̃3 )2
=
1
+
.
4t̃3
4t̃3
40
Then for all a ∈ C and ℓ ∈ Z, and κ = 0, 1, 2, one has that near t̃ = 0,
−a + ℓ + κ3 ;
b1 (κ), b2 (κ)
a
2
κ a+1
, 2
3
κ
3
1
ζ1,2 (t̃)
κ/3
4
(−)κ (1)κ (a + 2κ/3)1−ℓ−κ
=
ζ1,2 (t̃)
(a)1−ℓ
27
(
−a
−a
3
(1 + t̃)
(1 + t̃)3
(1 + ω t̃)3
(1 + ω t̃)3
1
Fℓ −a, 31 ;
Fℓ −a, 13 ;
+ ω̄ κ
×
3
3
3
3
3
1 + t̃
1 + t̃
1 + t̃
1 + t̃
)
−a
3
3
(1 + ω̄ t̃)
1 (1 + ω̄ t̃)
;
+ ωκ
F
.
−a,
ℓ
3
1 + t̃3
1 + t̃3
3 F2
+
+
Moreover, for all a ∈ C, c ∈ C and ℓ = 0, 1, 2, and κ = 0, 1, one has that
near t̃ = 0,
−a + ℓ + 1 +
b1 (κ), b2 (κ);
κ a
;
3 2
κ a+1
, 2
3
1
−a − c + ℓ + κ3
κ
−a − c + ℓ + 1 + 3
ζ1,2 (t̃)
κ/3
κ
(−) (1)κ (a + 2κ/3)−ℓ−κ (a + c − ℓ − κ/3) 4
=
ζ1,2 (t̃)
(a)−ℓ (a + c − ℓ)
27
(
−a
3 −a
(1 + t̃)3
(1 + t̃)3
(1 + ω t̃)3
1
1
1
κ (1 + ω t̃)
G
G
,
−c;
,
−c;
−a,
−a,
+
ω̄
×
ℓ
ℓ
3
3
3
1 + t̃3
1 + t̃3
1 + t̃3
1 + t̃3
−a
)
3
(1 + ω̄ t̃)3
κ (1 + ω̄ t̃)
1
+ω
Gℓ −a, 3 , −c;
.
1 + t̃3
1 + t̃3
4 F3
+
+
κ
;
3
In these two identities, the lower parameters b1 , b2 are 13 , 23 ; or 32 , 43 ; or 34 , 53 ;
depending on whether κ = 0, 1, or 2. Also, ω := ε3 = exp(2πi/3) and ω̄ := ε3 2 .
Proof. Straightforward, as β = σ3 = x1 x2 x3 = 1 + t̃3 by (6.16).
One can also derive identities from the parametrization by t̃ ∈ P1 of the bira(2)
(3)
tionally equivalent top curve C2,1 ∼
= C2,1 , given along with the parametrization
(2)
(3)
of C ∼
= C in (6.16). Details are left to the reader.
1,2
1,2
7.3. Parametrizations by u
(q)
(3)
The rational parametrization of the plane curve Cp,q = C1,3 by u ∈ P1 given
in (6.20), in Example 6.7, when substituted into part (i) of each of Thms. 4.4
and 4.6, yields the following.
Theorem 7.5. Define a degree-24 Belyı̆ map P1u → P1ζ by
ζ = ζ1,3 (u) := −256
u3 (1 − u3 )3 (1 + 8u3 )3
(1 + 8u6 )2 (1 + 88u3 − 8u6 )2
=
1
−
.
(1 − 20u3 − 8u6 )4
(1 − 20u3 − 8u6 )4
41
Then for all a ∈ C and ℓ ∈ Z, and κ = 0, 1, 2, one has that near u = 0,
a
4
+ κ3 , . . . , a+3
+ κ3
4
ζ
(u)
1,3
a − ℓ + 1 + κ3 ; b1 (κ), b2 (κ)
−κ/3
(−)κ (1)κ (a + 4κ/3)1−ℓ−κ
27
=
−
ζ1,3 (u)
(a)1−ℓ
256
(
3
3
3
3 −a
(1 − u) (1 + 2u)
1
1 (1 − u) (1 + 2u)
;
−a,
−
F
×
ℓ
3
3
1 − 20u3 − 8u6
1 − 20u3 − 8u6
−a
3
3
(1 − ωu)3 (1 + 2ωu)3
1 (1 − ωu) (1 + 2ωu)
−a,
−
F
+ ω̄ κ
;
ℓ
3
1 − 20u3 − 8u6
1 − 20u3 − 8u6
)
3
3 −a
3
3
κ (1 − ω̄u) (1 + 2ω̄u)
1 (1 − ω̄u) (1 + 2ω̄u)
Fℓ −a, − 3 ;
.
+ω
1 − 20u3 − 8u6
1 − 20u3 − 8u6
4 F3
Moreover, for all a ∈ C, c ∈ C and ℓ = 0, 1, and κ = 0, 1, 2, one has that
near u = 0,
a
4
a + c − ℓ + κ3
ζ1,3 (u)
κ
a+c−ℓ+1+ 3
−κ/3
κ
(−) (1)κ (a + 4κ/3)−ℓ−κ (a + c − ℓ + κ/3)
27
=
−
ζ1,3 (u)
(a)−ℓ (a + c − ℓ)
256
(
3
3 −a
(1 − u) (1 + 2u)
(1 − u)3 (1 + 2u)3
1
1
−a,
−
G
,
−c;
×
ℓ
3
3
1 − 20u3 − 8u6
1 − 20u3 − 8u6
−a
(1 − ωu)3 (1 + 2ωu)3
(1 − ωu)3 (1 + 2ωu)3
1
+ ω̄ κ
−a,
−
,
−c;
G
ℓ
3
1 − 20u3 − 8u6
1 − 20u3 − 8u6
)
−a
(1 − ω̄u)3 (1 + 2ω̄u)3
(1 − ω̄u)3 (1 + 2ω̄u)3
1
−a,
−
G
,
−c;
.
+ ωκ
ℓ
3
1 − 20u3 − 8u6
1 − 20u3 − 8u6
5 F4
+ κ3 , . . . , a+3
+ κ3 ;
4
a − ℓ + κ3 ; b1 (κ), b2 (κ);
In both identities, the lower parameters b1 , b2 are defined as in the previous
theorem.
Proof. Straightforward, as g = σ3 = x1 x2 x3 = 1 − 20u3 − 8u6 by (6.20).
One can similarly derive identities from the rational parametrizations of the
(3)
(3)
(4)
(4)
pair of top curves C1,3 ∼
= C3,1 , also given in Example 6.7, by
= C1,3 and C3,1 ∼
substituting them into Thms. 4.4(ii) and 4.6(ii). The resulting identities involve
4 F3 , 5 F4 , like the preceding.
8. Identities without free parameters
In this final section an alternative approach to the parametrizing of certain
F
n n−1 ’s, based on computation in rings of symmetric polynomials, is sketched.
The need for a strengthened approach is indicated by the nature of the preceding
results. Each identity in § 7 was derived from either Thm. 4.4 or Thm. 4.6: it
had a free parameter a ∈ C and was based on a (rational) parametrization of a
42
(q)
(n)
Schwarz curve, either Cp,q or Cp,q . But not many such curves are of zero or low
genus, making it difficult to derive large numbers of hypergeometric identities,
even if multivalued parametrizations with radicals are allowed.
In § 8.1, it is shown that if a ∈ Z, the relevant curve is a quotiented Schwarz
(n;symm.)
(q;symm.)
; and if qa ∈ Z, resp. na ∈ Z, it is another such
or Cp,q
curve Cp,q
(n;cycl.)
(q;cycl.)
. It is actions of the symmetric group Sq resp. Sn
, resp. Cp,q
curve Cp,q
and the cyclic group Cq resp. Cn that are quotiented out, and quotienting may
lower the genus; which facilitates explicit parametrization, rational or otherwise.
The several interesting examples worked out in § 8.2 and § 8.3 illustrate this,
though they provide representations of algebraic n Fn−1 ’s arising from En ’s that
are not of the imprimitive irreducible type characterized in Thm. 2.3, which was
previously the focus. (The difference in type is due to reducible monodromy
and/or a lower hypergeometric parameter being a non-positive integer.)
8.1. Theory
Theorem 8.2 below is a restatement of the ‘Birkeland’ ℓ = 0 case of Thm. 4.4,
which is the simplest because when ℓ = 0, the right-hand function Fℓ degenerates
to unity. The restatement emphasizes the role played by symmetric functions
of the trinomial roots x1 , . . . , xn . As usual, σl = σl (x1 , . . . , xn ) denotes the
(n)
l’th elementary symmetric polynomial, and Cp,q ⊂ Pn−1 comprises all points
[x1 : . . . : xn ] ∈ Pn−1 such that σ1 , . . . , σq−1 and σq+1 , . . . , σn−1 equal zero.
Definition 8.1. For each p, q > 1 with gcd(p, q) = 1 and n := p + q, let
(q;κ)
Fp,q
(a;
ζ) := n Fn−1
a
+ κq , . . . , a+(n−1)
+ κq
n
n
a+1
+ κq , . . . , a+p
+ κq ; 1q
p
p
\
+ κq , . . . , q−κ
+ κq , . . . ,
q
q
q
+
ζ
κ
q
!
for κ = 0, . . . , q − 1, and
(n;κ)
Fp,q
(a;
ζ) := n Fn−1
−a
+ nκ , . . . , −a+(p−1)
+ nκ ; aq + nκ , . . . , a+(q−1)
p
p
q
1
κ
κ
n
κ
n−κ
\
+
,
.
.
.
,
+
,
.
.
.
,
+
n
n
n
n
n
n
+
κ
n
ζ
!
for κ = 0, . . . , n − 1. For all a ∈ C, each function is defined and analytic in a
neighborhood of ζ = 0, unless (in the first) a lower parameter is a non-positive
integer. If exactly one lower parameter equals such an integer, −m, and an
upper parameter equals an integer −m′ with −m 6 −m′ 6 0, then the function
can be defined as a limit, with the aid of Lemma 2.1.
By a congruential argument, the differential equation En corresponding to
each of these n Fn−1 ’s will have reducible monodromy, i.e., have an upper and
a lower parameter that differ by an integer, iff qa ∈ Z, resp. na ∈ Z. In fact,
the En will have ζ = 0, resp. ζ = ∞ as a logarithmic point, i.e., have a pair of
lower, resp. upper parameters that differ by an integer, iff qa ∈ Z, resp. na ∈ Z.
(n)
Theorem 8.2. (i) Near the “ ζ = 0” point [x1 : . . . : xn ] ∈ Cp,q with
(
ε−(j−1)
,
j = 1, . . . , q,
q
xj =
0,
j = q + 1, . . . , n,
43
for each κ = 0, . . . , q − 1 one has
(−)κ (1)κ (a + nκ/q)1−κ
(a)1
"
#
−κ/q
q
−qa
p q
a 1 X
q−1
−n (j−1)κ
(j−1)
qp q
ζ
(−)
σq
.
(εq )
εq
xj
× (−)
nn
q j=1
(q; κ)
Fp,q
(a; ζ) =
(n)
(ii) Near the “ ζ = ∞” point [x1 : . . . : xn ] ∈ Cp,q with
xj = ε−(j−1)
,
n
j = 1, . . . , n,
for each κ = 0, . . . , n − 1 one has
(−)κ (1)κ (a + qκ/n)1−κ
(a)1
#
"
κ/n
n
−na
p q
a 1 X
−q (j−1)κ
(j−1)
n−1
qp q
(ε
)
ε
x
ζ
(−)
σn
.
× (−)
j
n
nn
n j=1 n
(n; κ)
Fp,q
(a; ζ −1 ) =
In both (i) and (ii),
ζ := (−)n
nn σn q
,
pp q q σq n
(−)q
q
pp q q
p σn
ζ
=
(−)
,
nn
σq n
as usual; and it is assumed that a ∈ C is such that the left-hand function is
defined. When κ > 0, branches must be chosen appropriately.
In parts (i) and (ii) of this theorem, the trinomial roots appear in the form
x−qa
, resp. x−na
. In consequence, the case when a ∈ Z, in particular when a is a
j
j
negative integer, is especially nice. If κ = 0, which will be assumed henceforth,
this is seen as follows. The identities of parts (i) and (ii) specialize if κ = 0 to
(q; 0)
Fp,q
(a;
q
−qa
a 1 X
q−1
ε(j−1) xj
ζ) = (−) σq ·
,
q j=1 q
n
−na
a 1 X
(n; 0)
.
εn(j−1) xj
Fp,q
(a; ζ −1 ) = (−)n−1 σn ·
n j=1
(8.1a)
(8.1b)
When a ∈ Z, these reduce to
q
a 1 X
(q; 0)
x−qa ,
Fp,q
(a; ζ) = (−)q−1 σq ·
q j=1 j
(8.2a)
n
a 1 X
(n; 0)
Fp,q
(a; ζ −1 ) = (−)n−1 σn ·
x−na .
n j=1 j
(8.2b)
(q)
The right side of (8.2a), resp. (8.2b), is single-valued on Cp,q ⊂ Pq−1 , resp.
(n)
(q; 0)
Cp,q ⊂ Pn−1 . Therefore, when a ∈ Z the algebraic functions ζ 7→ Fp,q (a; ζ)
(n; 0)
(q)
(n)
and ζ 7→ Fp,q (a; ζ −1 ) are respectively uniformized by Cp,q , Cp,q .
44
In fact, a stronger result is true. There is an obvious permutation symmetry
in (8.2ab) under Sq , resp. Sn , which can be quotiented out, as will now be
(k)
explained. Consider the action of Sk on the curve Cp,q ⊂ Pk−1 , for k = 2, . . . , n.
(k)
It was shown in § 5 that Cp,q is defined by a system of k − 2 equations, each of
which sets to zero a homogeneous polynomial in the invariants σ̄1 , . . . , σ̄k , the
(k)
elementary symmetric polynomials in x1 , . . . , xk . The function field of Cp,q is
the field of rational functions in x1 , . . . , xk that are homogeneous of degree zero,
with these defining equations quotiented out.
(k)
The function field of Cp,q contains a subfield of Sk -stable functions, comprising all rational functions in σ̄1 , . . . , σ̄k that are homogeneous (in x1 , . . . , xk )
of degree zero; again, with the defining equations quotiented out. One element
of this subfield is the function ζ, which gives the degree-(n − k + 1)k cover(k)
(k)
ing πp,q : Cp,q → P1ζ . Up to birational equivalence, define a quotiented curve
(k;symm.)
(k)
Cp,q
:= Cp,q /Sk that has this subfield of Sk -stable functions as its function field, so that
(k;symm.)
(8.3)
−→ P1ζ
C(k)
p,q −→ Cp,q
(k)
is a decomposition of πp,q into maps of respective degrees k!, nk .
(2)
The genus-zero case k = 2 is illustrative. It is the case that Cp,q ∼
= P1t ,
where by convention t = (x1 + x2 )/(x1 − x2 ) is the rational parameter; and
(2;symm.) ∼
Cp,q
= P1v , where v := t2 . This is because the only nontrivial element
of S2 is the involution x1 ↔ x2 , which on account of the parametrization
(2;symm.)
[x1 : x2 ] = [t+1 : t−1] is the involution t 7→ −t; so the function field of Cp,q
(k;symm.)
comprises only functions even in t. When k = 1 or 0, Cp,q
will be defined
(1;symm.) ∼ (1) ∼ 1
(0;symm.) ∼ (0) ∼ 1
as follows. By convention, Cp,q
= Cp,q = Ps and Cp,q
= Cp,q = Pζ .
(k;symm.) ∼ (n−k;symm.)
Lemma 8.3. Cp,q
for k = 0, . . . , n, due to the respective
= Cp,q
function fields being isomorphic. Here n := p + q, as usual.
Proof. To prove the case k = n (or k = 0), observe that the function field
(n;symm.)
of Cp,q
comprises all rational functions of σn q /σq n ∝ ζ. (Also, Sn is
(n)
(n)
the covering group of πp,q : Cp,q → P1ζ .) If 0 < k < n, use the fact that any
(n−k;symm.)
element of Cp,q
can be viewed as a quotient of two symmetric polynomials in xk+1 , . . . , xn that are homogeneous and of the same degree. Any
homogeneous symmetric polynomial in xk+1 , . . . , xn can be expressed as a quotient of two such polynomials in x1 , . . . , xk by the elimination procedure of § 5;
cf. Lemma 5.4.
Theorem 8.4. For each a ∈ Z,
(q; 0)
(q;symm.)
(i) the algebraic function ζ 7→ Fp,q (a; ζ) is uniformized by Cp,q
(n; 0)
.
(n;symm.)
(ii) the algebraic function ζ 7→ Fp,q (a; ζ −1 ) is uniformized by Cp,q
Proof. Immediate, by the permutation symmetry of formulas (8.2a),(8.2b).
45
.
The following is a specialization of Thm. 8.4, with a constructive proof.
Theorem 8.5. For each a ∈ Z,
(q; 0)
(i) if p or q equals 1, resp. 2, the algebraic function ζ 7→ Fp,q (a; ζ) can
be rationally parametrized by s, resp. v := t2 , where s, t are the rational
(1)
(2)
(q;symm.) ∼ 1
parameters of Cp,q ∼
= P1s and Cp,q ∼
= P2t . (This is because Cp,q
= Ps ,
1
resp. Pv .)
(n; 0)
(ii) the algebraic function ζ 7→ Fp,q (a; ζ −1 ) is in fact rational, i.e., is uni(0)
formized by Cp,q := P1ζ .
In part (i), the rational parametrization of the argument of the algebraic func(1)
(2)
tion, whether ζ = πp,q (s) or ζ = πp,q (t), was already given in Theorem 5.6; the
latter is even in t, and hence, single-valued in v.
(n;0)
Proof. Part (ii) is trivial, since Fp,q (a; ζ) will be a polynomial in ζ if a ∈ Z,
because one of the upper hypergeometric parameters will then be a nonpositive
integer (see Defn. 8.1). Also, the cases q = 1, 2 of part (i) are familiar from § 5.
If q = 1 then the right side of (8.2a) equals (1−s)a , and if q = 2 then one can use
the parametrization [x1 : x2 ] = [t + 1 : t − 1] and the formula for σq = σq (x1 , x2 )
given in Lemma 5.5 to express the right side of (8.2a) as a rational function of
the parameter t; in fact, of v = t2 .
The cases p = n − q = 1, 2 of part (i) are more interesting. They can be
(q;symm.) ∼ (p;symm.)
viewed as consequences of Cp,q
, which comes from Lemma 8.3.
= Cp,q
But a constructive rather than an abstract proof will be given, in the eliminationtheoretic spirit of Lemma 5.4. Suppose the integer γ = −qa is positive. Write
(
q
X
pγ − xγn ,
p = 1,
γ
xj =
(8.4)
γ
γ
p
−
x
−
x
,
p = 2,
γ
n
n−1
j=1
Pn
γ
where pγ :=
j=1 xj is a symmetric polynomial in x1 , . . . , xn , the so-called
γ’th power-sum symmetric polynomial. The elementary symmetric polynomials
{σl }nl=1 in x1 , . . . , xn form an algebraic basis for the ring of symmetric ones, so
n
the {pγ }∞
γ=1 can be expressed as polynomials in the {σl }l=1 ; e.g., by inverting
the Newton–Girard formula [24, § I.2]
σl = l−1
l
X
(−)γ−1 pγ σl−γ ,
(8.5)
γ=1
which holds for all l > 1, it being understood that σl = 0 if l > n. On the curve
(n)
Cp,q , each σl other than σ0 , σq , σn equals zero, which introduces simplifications.
Now, introduce an alternative sequence of subsidiary Schwarz curves
(n−1) ′
′
(1) ′
1
C(n)
−→ · · · −→ C(2)
p,q −→ Cp,q
p,q −→ Cp,q −→ Pζ ,
46
(8.6)
based on the successive elimination of x1 , . . . , xn rather than of xn , . . . , x1 , so
(1)
that Cp,q ′ ∼
= P1s′ , where
σq = (−)q−1 xqn · (1 − s′ ),
σn = (−)n−1 xnn · s′ ,
(8.7)
(2)
(cf. (5.17)), and Cp,q ′ ∼
= P1t′ is parametrized by t′ := (xn + xn−1 )/(xn − xn−1 ), so
′
that [xn : xn−1 ] = [t + 1 : t′ − 1]. By using (8.4),(8.5) and (8.7), one can write
2
the right side of (8.2a) as a rational function of s′ , resp. of t′ (in fact of v ′ := t′ ,
by invariance under xn ↔ xn−1 ). Moreover, the expressions for ζ in terms of
s′ , t′ are the same as those for ζ in terms of s, t.
The proof when γ = −qa < 0 is an easy modification of the preceding.
Remark 8.5.1. Illustrations of Thm. 8.5(i), showing how the algorithm embedded in its proof is implemented, are given in § 8.2 below.
Now consider what the two identities of Thm. 8.2 amount to, when a is not
m
an integer, but nonetheless qa ∈ Z, resp. na ∈ Z. If a = m
q , resp. a = n , with
m ∈ Z, the κ = 0 specializations (8.1ab) reduce to
q
m/q 1 X
(q; 0) m
·
ε−(j−1)m x−m
,
Fp,q
( q ; ζ) = (−)q−1 σq
j
q j=1 q
n
m/n 1 X
(n; 0) m
ε−(j−1)m x−m
.
·
Fp,q
( n ; ζ −1 ) = (−)n−1 σn
j
n j=1 n
(8.8a)
(8.8b)
h
iq
in
h
(q; 0)
(n; 0) m
−1
It follows that the power Fp,q ( m
;
ζ)
;
ζ
)
,
resp.
F
(
, is a ratiop,q
q
n
(q)
nal function of x1 , . . . , xq , resp. x1 , . . . , xn , i.e., is in the function field of Cp,q ,
(n)
resp. Cp,q . Equivalently, these powers of algebraic functions are uniformized
(q)
(n)
−(j−1)m −(j−1)m
by Cp,q , Cp,q . But because of the presence of the coefficients εq
, εn
,
they are not stable under the action of Sq on x1 , . . . , xq , resp. Sn on x1 , . . . , xn .
Rather, they are stable under the associated subgroups of cyclic permutations.
The group of cyclic permutations of x1 , . . . , xk will be denoted by Ck , and the
group of dihedral ones (if k > 2) by Dk . The relevant order and subgroup
indices are |Ck | = k and (Dk : Ck ) = 2, (Sk : Dk ) = (k − 1)!/2.
(k;cycl.)
(k;dihedr.)
(k;symm.)
Definition 8.6. The quotient curves Cp,q
, Cp,q
, Cp,q
are defined
(k)
to be Cp,q / Γ with Γ = Ck , Dk , Sk , so that if k > 2 one has the sequence of
coverings
(k;cycl.)
(k;dihedr.)
C(k)
−→ Cp,q
−→ C(k;symm.)
−→ P1ζ ,
p,q
p,q −→ Cp,q
which have respective degrees k, 2, (k − 1)!/2, and nk , as a decomposition of
(k)
(k)
the degree-(n − k + 1)k covering map πp,q : Cp,q → P1ζ . If k = 1, 2 then there is
(k;cycl.)
no dihedral curve, but Cp,q
(k;symm.)
→ Cp,q
Theorem 8.7. For each m ∈ Z,
47
has degree (k − 1)! in all cases.
iq
h
(q;cycl.)
(q; 0)
;
ζ)
is uniformized by Cp,q
.
(i) the algebraic function ζ 7→ Fp,q ( m
q
h
in
(n; 0)
(n;cycl.)
−1
(ii) the algebraic function ζ 7→ Fp,q ( m
.
) is uniformized by Cp,q
n; ζ
Proof. Immediate, by the cyclic permutation symmetry of (8.8a),(8.8b).
The following is a specialization of Theorem 8.7(i), with a constructive proof.
Theorem 8.8. For each m = −1, −2, . . . ,
h
iq
(q; 0)
(i) for each q > 1, the algebraic function ζ 7→ F1,q ( m
;
ζ)
has a (q − 1)!q
(q;symm.) ∼ (1) ∼
valued parametrization by s, the rational parameter of C1,q
= C1,q =
1
Ps . That is, it satisfies a degree-(q − 1)! polynomial equation with coefficients in Z[s].
iq
h
(q; 0)
has
(ii) for each odd q > 1, the algebraic function ζ 7→ F2,q ( m
q ; ζ)
a (q − 1)!-valued parametrization by v = t2 , the rational parameter of
(2)
(q;symm.) ∼ 1
C2,q
= P1t .
= Pv , i.e., by the square of the rational parameter of C1,q ∼
That is, it satisfies a degree-(q − 1)! polynomial equation with coefficients
in Z[v].
(2)
(1)
The rational parametrizations ζ = π1,q (s) and ζ = π2,q (t) of the argument of
the algebraic function, for (i),(ii), were already given in Theorem 5.6; the latter
is even in t, and hence, single-valued in v.
Proof. The rings of symmetric polynomials in x1 , . . . , xq and in x1 , . . . , xn have
(n)
(n)
(q)
(q)
algebraic bases {σl }ql=1 or {pγ }qγ=1 , resp. {σl }nl=1 or {pγ }nγ=1 , of elementary or power-sum symmetric polynomials. Consider
q
q
Y
X
Gq,−m (y) :=
εq−(j−1)m x−m
y−
,
(8.9)
χ(j)
χ
j=1
the product being taken over one representative χ of each of the (q − 1)! cosets
of Cq in Sq , acting on x1 , . . . , xq . By Eq. (8.8a),
iq
h
m
(q;0) m
y0 ,
(8.10)
( q ; ζ) = q −q (−)q−1 σq
Fp,q
where y0 is a certain root of Gq,−m (y). Each coefficient of the degree-(q − 1)!
polynomial Gq,−m (y) is stable under the action of Sq on x1 , . . . , xq , and can
(q)
therefore be expressed as a polynomial in the {pγ }qγ=1 . But,
pγ(q)
=
q
X
j=1
xγj
=
(
γ
p(n)
γ − xn ,
p(n)
γ
−
48
xγn
−
p = 1,
xγn−1 ,
p = 2.
(8.11)
(n)
Hence, each coefficient of Gq,−m (y) is expressible in terms of the {σl }nl=1
(n)
(n) (n)
(n)
and xn , resp. xn , xn−1 . But on Cp,q , each of the {σl }nl=1 other than σq , σn
equals zero. By introducing the alternative sequence (8.6) of subsidiary Schwarz
(n) (n)
curves, one can express σq , σn and xn , resp. xn , xn−1 , in terms of s′ resp. t′ ,
(1)
(2)
the rational parameter of C1,p ′ resp. C2,p ′ . And the expressions for ζ in terms
′ ′
of s , t are the same as those for ζ in terms of s, t. Each rational function of t′
encountered is a function of v ′ := (t′ )2 , by invariance under xn ↔ xn−1 .
Remark 8.8.1. Illustrations of Thm. 8.8, showing how the algorithm embedded
in its proof is implemented, are given in § 8.3 below.
8.2. Parametrizations when a ∈ Z
Theorems 8.9 and (especially) 8.11 below are sample applications of the
algorithm embedded in the proof of Thm. 8.5(i). They show how it is possible
rationally to parametrize many algebraic n Fn−1 ’s with a ∈ Z, due to their
(q;symm.)
that are of genus zero. These
being uniformized by quotient curves Cp,q
examples also illustrate how lower parameters that are non-positive integers can
be handled by taking a limit, and applying the auxiliary Lemma 2.1.
(q)
The first theorem is relatively simple, since the governing curve is Cp,q =
(1)
Cn−1,1 , which is of genus zero without quotienting; though Lemma 2.1 is used.
Theorem 8.9. For each n > 2,
1 1
(n − 1) + s
s
nn
− n ; n , . . . , n−2
n
=
−
n−1 Fn−2
n−2
1
n−1 (1 − s)n
,
.
.
.
,
(n
−
1)
(n
− 1)(1 − s)
n−1
n−1
in a neighborhood of s = 0.
(q;0)
Proof. As mentioned in the proof of Thm. 8.5(i), if q = 1 then Fp,q (a; ζ) equals
(1 − s)a , where
nn
sq
ζ = (−)q p q
(8.12)
p q (1 − s)n
(1) ∼ 1
(1;0)
is the usual map φ(1) : Cp,q =
Ps → P1ζ . Hence, Fn−1,1 (−1; ζ) equals (1 − s)−1 .
Formally,
−1 0 1
, n , n . . . , n−2
(1;0)
n
n
ζ ,
(8.13)
Fn−1,1 (−1; ζ) = n Fn−1
0
1
n−2
n−1 , n−1 . . . , n−1
but the right side is undefined, on account of the presence of a non-positive
integer −m = 0 as a lower parameter (accompanied, fortunately, by a matching
upper parameter, −m′ = 0). One must interpret this in a limiting sense, i.e.,
!
a a+1 a+2
, n , n . . . , a+(n−1)
(1;0)
n
n
Fn−1,1 (−1; ζ) = lim n Fn−1
ζ .
(8.14)
a+(n−1)
a+1 a+2
a→−1
n−1 , n−1 . . . ,
n−1
By Lemma 2.1, this limit equals
1
n−1
+
· n−1 Fn−2
n
n
49
− n1 ; n1 , . . . , n−2
n
1
n−2
n−1 , . . . , n−1
ζ ,
(8.15)
and the theorem now follows by a bit of algebra.
Remark 8.9.1. Theorem 8.9 can also be viewed as a degenerate case of Thm. 7.1.
Explicit representations for this algebraic n−1 Fn−2 , for low n, were recently
given by Dominici [8]. The differential equation En−1 of which this n−1 Fn−2 is a
solution has monodromy group H < GLn−1 (C) isomorphic to Sn , by Thm. 2.4.
The following formula, of Girard type, facilitates computation in the function
fields of any top Schwarz curve and its subsidiaries.
Lemma 8.10. Consider the subset of Cn , coordinatized by x1 , . . . , xn , on which
the elementary symmetric polynomials σ1 , . . . , σq−1 and σq+1 , . . . , σn−1 equal
(n)
.) On this
zero. (The image of this subset under Cn → Pn−1 is the top curve Cp,qP
subset, one can express any power-sum symmetric polynomial pγ = nj=1 xj γ ,
γ > 1, in terms of σq and σn by
X
pγ =
cmq ,mn σq mq σn mn ,
where the sum is over all mq , mn > 0 with mq q + mn n = γ, and
mq + mn − 1
mq + mn − 1
χ(q)mq +χ(n)mn
q
cmq ,mn = (−)
+n
,
mn
mq
with χ(l) := 1, 0 if l ≡ 0, 1 (mod 2).
Proof. This comes, e.g., from the determinantal formula
σ1
2σ2
..
.
1
σ1
..
.
0
1
..
.
...
...
0
0
..
.
γσγ
σγ−1
σγ−2
...
σ1
pγ =
,
(8.16)
with σl = 0 if l > n. It is inverse to the Newton–Girard formula [24, § I.2].
Each hypergeometric function in the following theorem, with a ∈ Z, is al(q;symm.)
(3;symm.)
gebraic with governing curve Cp,q
= C2,3
. This quotient curve is of
(q)
(3)
genus zero, although the unquotiented curve Cp,q = C2,3 , which sextuply covers
it, is of genus 3 by the formula of Thm. 6.3. The point of the theorem is that if
a ∈ Z, a uniformization by rational functions becomes possible.
(3;0)
Theorem 8.11. For each integer a 6 −1, the function F2,3 (a; ζ) satisfies
(3;0)
F2,3 (a; ζ(t)) =
1
3
[s3 (t)]
a
Pa (s3 (t), s5 (t)) − (t + 1)−3a − (t − 1)−3a
in a neighborhood of t = 0, where
1 + 10t2 + 5t4
(1 − t2 )2 (1 + 3t2 )
,
s5 (t) = −
,
2t
2t
55 t2 (1 − t2 )6 (1 + 3t2 )3
55 s5 3 (t)
= 3
,
ζ(t) = − 2 3 5
2 3 s3 (t)
3
(1 + 10t2 + 5t4 )5
s3 (t) =
50
(8.17)
and Pa ∈ Z[s3 , s5 ] is defined by, e.g.,
( −a
3 s3 ,
Pa (s3 , s5 ) =
3 s3 5 + 5 s5 3 ,
a = −1, −2, −3, −4;
a = −5.
(3;0)
For each a, the special function F2,3 (a; ζ) can be expressed in terms of nondegenerate hypergeometric functions by Lemma 2.1; e.g.,
4 6 7 8
22 33
(3;0)
5; 5, 5, 5; 1
F2,3 (−1; ζ) = 1 − 5 ζ · 5 F4
ζ
,
3 4 5
5
2; 3, 3; 2
1 1 2 3
3 2
−5; 5, 5, 5
ζ
,
= + · 4 F3
1 1 2
5 5
2; 3, 3
11 12 13 14
22 32
26 39
(3;0)
5 , 5 , 5 , 5 ; 2
F2,3 (−5; ζ) = 1 − 4 ζ − 14 ζ 3 · 5 F4
ζ
.
3 10 11
5
5
2; 3 , 3 ; 4
Proof. This is the (p, q) = (2, 3) case of Thm. 8.5(i), made explicit, and the
(q;0)
identity (8.17) is a rational parametrization of the formula (8.2a) for Fp,q (a, ζ)
3
5
q
when a ∈ Z. The functions s3 (t), s5 (t) are σ3 /x5 , σ5 /x5 , i.e., σq /xn , σn /xnn ,
expressed in terms of t, the rational parameter of the alternative Schwarz curve
(p)
(2)
(2)
Cp,q ′ = C2,3 ′ . The given expressions and the formula for ζ = πp,q (t) come
from Lemma 5.5 and Thm. 5.6; though x1 , x2 are to be replaced by x5 , x4 , and
‘t’ is to be understood as (x5 + x4 )/(x5 − x4 ), not (x1 + x2 )/(x1 − x2 ), so that
[x5 : x4 ] = [t + 1 : t − 1].
For each integer a 6 −1, the quantity Pa = Pa (s3 , s5 ) is a normalized version
of the power-sum symmetric polynomial pγ , i.e.,
pγ =
n
X
j=1
xγj =
5
X
xγj ,
γ := −qa = −3a,
(8.18)
j=1
expressed as a polynomial in σ3 , σ5 by the Girard formula of Lemma 8.10. The
subtracted terms (t + 1)−3a , (t − 1)−3a in (8.17) come from subtracting xγ4 , xγ5
from pγ to obtain xγ1 + xγ2 + xγ3 , as in (8.4).
(3,0)
The removal from F2,3 (a; ζ), which is a 5 F4 (ζ), of lower hypergeometric
parameters that are non-positive integers, is a straightforward application of
Lemma 2.1. E.g., for a = −1, −2, −3, −4, −5, the relevant lower/upper parameters (−m, −m′ ) are respectively (0, 0), (0, 0), (−1, 0), (−1, 0), (−2, −1). Only if
a = −1 or −2 does −m = −m′ , permitting the 5 F4 to be reduced to a 4 F3 .
Remark 8.11.1. In the two cases a = −1, −2 in which an upper and a lower
(3,0)
parameter of F2,3 (a; ζ) can be cancelled, reducing it to a 4 F3 (ζ), the E4 of
which this 4 F3 is a solution has monodromy group H < GL4 (C) isomorphic
to S5 , by Thm. 2.4.
(3,0)
If a 6 −3 then F2,3 (a; ζ) is essentially a 5 F4 (ζ), which is the solution of
an E5 that has reducible monodromy, due to an upper and a lower parameter
differing by an integer. (See the typical case a = −5, below.)
51
Corollary 8.12. Define a degree-10 Belyı̆ map P1v → P1ζ by
55 v(1 − v)6 (1 + 3v)3
33 (1 + 10v + 5v 2 )5
(1 − 35v − 125v 2 − 225v 3 )2 (27 + 115v + 25v 2 + 25v 3 )
.
=1−
27 (1 + 10v + 5v 2 )5
ζ = ζ2,3 (v) :=
Then in a neighborhood of v = 0,
1 1 2 3
3 + 5v 2
−5; 5, 5, 5
F
,
ζ
(v)
=
4 3
2,3
1 1 2
3(1 + 10v + 5v 2 )
2; 3, 3
5 F4
=
11 12 13 14
5 , 5 , 5 , 5 ;
3 10 11
2; 3 , 3 ; 4
2
ζ2,3 (v)
(3 + v)(5 + 10v + v 2 )(1 + 28v + 134v 2 + 92v 3 + v 4 )(1 + 10v + 5v 2 )10
.
15 (1 − v)18 (1 + 3v)9
Proof. These are the rational parametrizations of the nondegenerate hypergeometric functions obtained in the theorem (when a = −1, resp. −5). They are
even in t and expressible in terms of v := t2 , as expected.
Remark 8.12.1. The uniformizing parameter v is a rational parameter for the
(q;symm.)
(3;symm.)
genus-0 quotient curve Cp,q
= C2,3
. The reason why the above 4 F3 (ζ)
and 5 F4 (ζ) are 10-branched functions of ζ is that the maps
(3)
(3;symm)
C2,3 → C2,3
∼
= P1v → P1ζ
(8.19)
have
degrees
6, 10. The 10-sheetedness of the latter map comes from
respective
n
p+q
5
=
=
=
10.
q
q
3
(3)
The subsidiary curve C2,3 ⊂ P2 is of genus 3 and is a smooth quartic through
the fundamental points of P2 , with defining equation
(x1 + x2 + x3 )(x1 x2 x3 ) + (x1 x2 + x2 x3 + x3 x1 )2
− (x1 + x2 + x3 )2 (x1 x2 + x2 x3 + x3 x1 ) = 0, (8.20)
(3)
as follows from Thm. 6.5. A formula for the degree-6 quotient map C2,3 →
(3;symm.) ∼ 1
C
= P can be derived by eliminating x4 , x5 from the defining equations
2,3
v
(5)
2
σ1 = 0, σ2 = 0, σ4 = 0 of the top curve C2,3 ⊂ P4 , using v = t′ with t′ =
(x5 + x4 )/(x5 − x4 ). It is
v=
(x1 + x2 + x3 )2
.
4(x1 x2 + x2 x3 + x3 x1 ) − 3(x1 + x2 + x3 )2
52
(8.21)
8.3. Parametrizations when qa ∈ Z
Theorems 8.13 and 8.14 below are sample applications of the algorithm embedded in the proof of Thm. 8.8, to the cases (p, q) = (2, 3) and (1, 4) with
a = − q1 . They show how one can construct parametrizations of many algebraic n Fn−1 ’s with qa ∈ Z, due to powers of the n Fn−1 ’s being uniformized by
(q;cycl.)
quotient curves Cp,q
that if not rational, are at least low-degree covers of
rational ones. In these two theorems the rational lower curves will respectively
(q;symm.)
(q;dihedr.)
be Cp,q
and Cp,q
, and the parametrizations of the n Fn−1 ’s will for
the first time involve radicals.
Theorem 8.13. Define a degree-10 Belyı̆ map P1v → P1ζ by the rational formula
for ζ = ζ2,3 (v) given in Corollary 8.12. Then in a neighborhood of v = 0,
1
2
8 11
− 15
, 15
, 15
, 15
1 2 5
,
,
3 3 6
n
5
5 2
1
1
= 2 + 3 v − 54 v + 2 (1 + 3v) 1 +
(1 + 10v + 5v 2 )1/3 4 F3
ζ2,3 (v)
115
27 v
+
25 2
27 v
+
25 3
27 v
1/2 o1/3
. (8.22)
Proof. This is the (p, q) = (2, 3) case of Thm. 8.8(ii), made explicit, and the
(q;0)
q
(3;0)
3
identity (8.22) is a rational parametrization of Fp,q (a, ζ) = F2,3 (a, ζ)
when m = −1, i.e., a = −1/3 and qa = −1.
(n)
As in Thm. 8.11, the relevant top and subsidiary Schwarz curves are Cp,q =
(5)
(q)
(3)
(p)
(2)
C2,3 and Cp,q = C2,3 , and the complementary subsidiary curve is Cp,q ′ = C2,3 ′ .
(3)
(2)
The homogeneous coordinates on C2,3 ⊂ P2 are x1 , x2 , x3 and those on C2,3 ′ ∼
= P1t
are x4 , x5 . The rational parameter on the latter curve is t := (x5 +x4 )/(x5 −x4 ),
and if s3 := σ3 /x35 , s5 := σ5 /x55 , then
s3 =
1 + 10t2 + 5t4
1 x55 − x54
=
,
3
2
2
x5 x5 − x4
2t
ζ=−
s5 = −
x24 x35 − x34
(1 − t2 )2 (1 + 3t2 )
=−
,
3
2
2
x5 x5 − x4
2t
55 s35
55 t2 (1 − t2 )6 (1 + 3t2 )3
=
5
22 33 s3
33 (1 + 10t2 + 5t4 )5
follow from Lemma 5.5 (with x1 , x2 relabelled as x5 , x4 ). This was the origin of
the formula ζ = ζ2,3 (v) given in Corollary 8.12, with v := t2 .
By the formula in Eq. (8.8a),
(3;0)
F2,3 (− 31 ; ζ) := 5 F4
1
− 15
,
2
,
15
1 5
, ;
3 6
1 8 11
, ,
3 15 15
1 2
,
3 3
ζ
= 4 F3
1
− 15
,
2
, 8 , 11
15 15 15
1 2 5
, ,
3 3 6
ζ
(8.23)
(an upper and a lower parameter being cancelled) has the representation
(3;0)
F2,3 (− 13 ; ζ) =
h
i3
(3;0)
F2,3 (− 13 ; ζ) =
σ3 −1/3 x1 + ε3 x2 + ε23 x3 ,
3
−1
1
x1 + ε3 x2 + ε23 x3 .
33 σ3
1
3
53
(8.24a)
(8.24b)
Define a polynomial G3,1 (y), symmetric in x1 , x2 , x3 , one of the roots of which
is the final factor in (8.24b), as a product over the two cosets of C3 in S3 , i.e.,
n
3 o
3 o n
y − x1 + ε23 x2 + ε3 x3
G3,1 (y) = y − x1 + ε3 x2 + ε23 x3
(8.25)
3
= y 2 + −2σ̂13 + 9σ̂1 σ̂2 − 27σ̂3 y + σ̂12 − 3σ̂2 ,
as one finds by a bit of computation. Here, {σ̂l }3l=1 are the elementary symmetric
polynomials in x1 , x2 , x3 alone; the ones in x4 , x5 will be denoted by {σ̄l }2l=1 .
Each coefficient in G3,1 can be expressed rationally and symmetrically in
(3;symm.) (2;symm.)
terms of x4 , x5 , as the function fields of C2,3
, C2,3
are isomorphic.
One way to do this is to exploit the formula given in Lemma 5.4. Another
uses the structure of the ring of symmetric polynomials in x1 , . . . , x5 , and was
sketched in the proof of Thm. 8.8. One can express {σ̂l }3l=1 in terms of the
power-sum symmetric polynomials {p̂γ }3γ=1 , which can be expressed in terms of
the overall power-sum symmetric polynomials {pγ }5γ=1 in x1 , x2 , x3 , x4 , x5 , along
with x4 , x5 . But the {pγ }5γ=1 can be expressed in terms of {σl }5l=1 . Of these,
the only nonzero ones are σ3 , σ5 , which were expressed above in terms of x4 , x5 .
Regardless of which technique one uses, one finds that
3
−7σ̄14 + 36σ̄12 σ̄2 − 27σ̄22
2
(8.26)
y + −2σ̄12 + 3σ̄2 .
G3,1 (y) = y +
σ̄1
Let F3 denote the left side of Eq. (8.24b), so that F3 = y/33 σ3 , where y is a
root of G3,1 . It follows from (8.26) and the formula for s3 = σ3 /x35 in terms of t,
and [x5 : x4 ] = [t + 1 : t − 1], that F3 satisfies
F3 2 −
27 + 90v − 5v 2
4v(3 + 5v)3
= 0,
F
−
3
27 (1 + 10v + 5v 2 )
729 (1 + 10v + 5v 2 )2
(8.27)
where v = t2 . The theorem now follows from the quadratic formula.
(3;0)
Remark 8.13.1. The cube of F2,3 (− 13 ; ζ), i.e., of the 4 F3 in Thm. (8.13), is
(3;cycl.)
(3;symm.) ∼
uniformized by C2,3
, which doubly covers C2,3
= P1v . (This is clear
from (8.24b).) In fact, the statement of the theorem implicitly contains a plane
(3;cycl.)
(3;symm.) ∼ 1
model of C2,3
as a double cover of C2,3
= Pv , namely
w2 = (1 + 3v) 1 +
115
27 v
+
25 2
27 v
(3;symm.)
+
25 3
27 v
.
(8.28)
This affine quartic C2,3
∋ (v, w) is elliptic (of genus 1), with Klein–Weber
invariant j = −212 52 /3. It is triply covered by the unquotiented Schwarz curve
(3)
C2,3 ⊂ P2 , which is of genus 3 and has defining equation (8.20).
In summary, the maps
(3)
(3;cycl.)
C2,3 −→ C2,3
(3;symm.) ∼
−→ C2,3
54
= P1v −→ P1ζ
(8.29)
have respective degrees 3, 2, 10. This diagram contrasts with
(1)
(3)
(2)
C2,3 −→ C2,3 ∼
= P1s −→ P1ζ ,
= P1t −→ C2,3 ∼
(8.30)
in which the maps have respective degrees 3, 4, 5.
An S3 -invariant formula v = v(x1 , x2 , x3 ) for the degree-6 quotient map
(3)
(3;symm.) ∼ 1
C2,3 → C2,3
= Pv , which is the composition of the first two maps in (8.29),
(3)
(3;cycl.)
was given in Eq. (8.21). The first of the two, the degree-3 map C2,3 → C2,3
,
is of the form (x1 , x2 , x3 ) 7→ (v, w). A rational formula for w = w(x1 , x2 , x3 ),
C3 -invariant rather than S3 -invariant, can also be worked out.
Theorem 8.14. Define a degree-15 Belyı̆ map P1x → P1ζ by
ζ = ζ1,4 (x) : =
1 x(1 − 5x)4 (5 + 6x + 5x2 )4
55
s4
−(1 − 5x)(5 + 6x + 5x2 )
= 4
◦
5
2
5
5
4 (1 − x) (1 + 10x + 5x )
4 (1 − s)
64 x
=1−
(4 − 5x − 10x2 − 5x3 )(1 − 55x − 5x2 − 5x3 )2 (1 + 5x2 + 10x3 )2
.
4(1 − x)5 (1 + 10x + 5x2 )5
Then in a neighborhood of x = 0,
3
7 11
1
, 20
, 20
, 20
− 20
ζ
(x)
1,4
1 1 3
4, 2, 4
n
1/2 o1/4
5
5 3
= 12 − 16
x − 85 x2 − 16
x + 21 1 − 54 x − 25 x2 − 45 x3
. (8.31)
(1 − x)1/4 (1 + 10x + 5x2 )1/4 4 F3
Proof. This is the (p, q) = (1, 4) case of Thm. 8.8(i), made explicit, and the iden (q;0)
q (4;0)
4
tity (8.31) is a rational parametrization of Fp,q (a, ζ) = F1,4 (a, ζ) when
m = −1, i.e., a = −1/4 and qa = −1. The relevant top and subsidiary Schwarz
(q)
(4)
(n)
(5)
curves are Cp,q = C1,4 and Cp,q = C1,4 , and the complementary subsidiary curve
(p)
(1)
is Cp,q ′ = C1,4 ′ ∼
= P1s′ .
The computations resemble those in the proof of Thm. 8.13 and will only be
sketched. One defines a degree-6 polynomial G4,1 (y), symmetric in x1 , x2 , x3 , x4 ,
as a product over the six cosets of C4 in S4 . Each coefficient in G4,1 can
be expressed rationally in x5 and the rational parameter s′ = σ5 /x55 . This
leads to a degree-6 polynomial equation for F4 , defined as the fourth power of
(4;0)
F1,4 (− 14 ; ζ), i.e., of the 4 F3 in the theorem. The coefficients of the degree-6
polynomial are polynomial in s′ . By examination, if one substitutes
s′ = s′ (x) =
−(1 − 5x)(5 + 6x + 5x2 )
,
64 x
(8.32)
this polynomial will factor. The only relevant factor is a quadratic one, namely
F4 2 −
25 x2 (1 + x)4
8 − 5x − 10x2 − 5x3
F4 +
= 0.
2
8(1 − x)(1 + 10x + 5x )
256(1 − x)2 (1 + 10x + 5x2 )2
The theorem follows from Eq. (8.33) by the quadratic formula.
55
(8.33)
(4;0)
Remark 8.14.1. The fourth power of F1,4 (− 14 ; ζ), i.e., of the 4 F3 in Thm. (8.14),
(4;cycl.)
(4;dihedr.)
is uniformized by C1,4
, which doubly covers C1,4
, which in turn, triply
(4;symm.) ∼ (1;symm.) ∼ 1
covers C1,4
= Ps′ . Hence, x can be identified as a rational
= C1,4
(4;dihedr.)
parameter for the genus-0 curve C1,4
(4;cycl.)
C1,4
. The statement of the theorem im(4;dihedr.) ∼ 1
as a double cover of C
=P ,
plicitly contains a plane model of
namely
w2 = 1 − 54 x − 25 x2 − 45 x3 .
(4;cycl.)
This affine cubic C1,4
2
invariant j = −5 /2, like
1,4
x
(8.34)
∋ (x, w) is elliptic (of genus 1), with Klein–Weber
(3)
C1,4 .
(See Example 6.8; the equality of the j-invariants
(5)
remains to be investigated.) It is quadruply covered by the top curve C1,4 ∼
=
(4)
C1,4 ⊂ P3 , which is of genus 4.
In summary, the maps
(4;symm.) ∼ 1
(4)
(4;cycl.)
(4;dihedr.) ∼ 1
(5)
−→ C1,4
C1,4 ∼
= Ps′ −→ P1ζ
= Px −→ C1,4
= C1,4 −→ C1,4
(8.35)
have respective degrees 4, 2, 3, 5. This diagram contrasts with
(1)
(4)
(3)
(2)
(5)
C1,4 ∼
= P1s −→ P1ζ ,
= P1t −→ C1,4 ∼
= C1,4 −→ C1,4 −→ C1,4 ∼
(8.36)
in which the maps have respective degrees 2, 3, 4, 5.
(4)
(4;symm.) ∼
An S4 -invariant formula for the degree-24 quotient map C1,4 → C1,4
=
1
Ps′ , which is the composition of the first three of the four maps in (8.35), can
be derived by eliminating x5 from the defining equations σ1 = 0, σ2 = 0, σ3 = 0
(5)
of the top curve C1,4 ⊂ P4 , using s′ = σ5 /x55 . It is
s′ =
x1 x2 x3 x4
.
(x1 + x2 + x3 + x4 )4
(8.37)
Acknowledgements. We thank the anonymous referee for many suggestions that led to
improvements in the exposition.
References
[1] G. Belardinelli, Risoluzione analitica delle equazioni algebriche generali,
Rend. Sem. Mat. Fis. Milano 29 (1959) 13–45.
[2] B.C. Berndt, Ramanujan’s Notebooks, Part I, Springer-Verlag, New
York/Berlin, 1985.
[3] F. Beukers, G. Heckman, Monodromy for the hypergeometric function
n Fn−1 , Invent. Math. 95 (1989) 325–354.
[4] R. Birkeland, Über die Auflösung algebraischer Gleichungen durch hypergeometrische Funktionen, Math. Z. 26 (1927) 566–578.
56
[5] C.A. Charalambides, A new kind of numbers appearing in the n-fold convolution of truncated binomial and negative binomial distributions, SIAM
J. Appl. Math. 33 (1977) 279–288.
[6] Wenchang Chu, Binomial convolutions and hypergeometric identities,
Rend. Circ. Mat. Palermo (2) 43 (1995) 333–360.
[7] L. Comtet, Advanced Combinatorics, Reidel, Boston/Dordrecht, 1974.
[8] D. Dominici, Asymptotic analysis of generalized Hermite polynomials, Analysis (Munich) 28 (2008) 239–261, available on-line as
arXiv:math/0606324.
[9] A. Eagle, Series for all the roots of a trinomial equation, Amer. Math.
Monthly 46 (1939) 422–425.
[10] A. Erdélyi, W. Magnus, F. Oberhettinger, F.G. Tricomi (Eds.), Higher
Transcendental Functions, McGraw–Hill, New York, 1953–55. Also known
as The Bateman Manuscript Project .
[11] H. Fell, The geometry of zeros of trinomial equations, Rend. Circ. Mat.
Palermo (2) 29 (1980) 303–336.
[12] I. Gessel, D. Stanton, Strange evaluations of hypergeometric series, SIAM
J. Math. Anal. 13 (1982) 295–308.
[13] M.L. Glasser, Hypergeometric functions and the trinomial equation,
J. Comput. Appl. Math. 118 (2000) 169–173.
[14] H.W. Gould, Some generalizations of Vandermonde’s convolution, Amer.
Math. Monthly 63 (1956) 84–91.
[15] H.W. Gould, A series transformation for finding convolution identities,
Duke Math. J. 28 (1961) 193–202.
[16] H.W. Gould, Combinatorial Identities: A Standardized Set of Tables Listing 500 Binomial Coefficient Summations, Morgantown, WV, 1972.
[17] G.R. Greenfield, D. Drucker, On the discriminant of a trinomial, Linear
Algebra Appl. 62 (1984) 105–112.
[18] N.A. Hall, The solution of the trinomial equation in infinite series by the
method of iteration, Natl. Math. Mag. 15 (1941) 219–229.
[19] C. Jordan, Calculus of Finite Differences, 3rd ed., Chelsea, New York, 1965.
[20] M. Kato, Minimal Schwarz maps of 3 F2 with finite irreducible monodromy
groups, Kyushu J. Math. 60 (2006) 27–46.
[21] M. Kato, M. Noumi, Monodromy groups of hypergeometric functions satisfying algebraic equations, Tohoku Math. J. (2) 55 (2003) 189–205.
57
[22] T. Kimura, K. Shima, A note on the monodromy of the hypergeometric
differential equation, Japan. J. Math. (N.S.) 17 (1991) 137–163.
[23] P. Lefton, A trinomial discriminant formula, Fibonacci Quart. 20 (1982)
363–365.
[24] I.G. Macdonald, Symmetric Functions and Hall Polynomials, 2nd ed., Oxford Univ. Press, 1995. With contributions by A. Zelevinsky.
[25] D. Merlini, R. Sprugnoli, M.C. Verri, Lagrange inversion: When and how,
Acta Appl. Math. 94 (2006) 233–249.
[26] E.N. Mikhalkin, On solutions of general algebraic equations by means of
integrals of elementary functions, Siberian Math. J. 47 (2006) 301–306.
Russian original in Sibirsk. Mat. Zh. 47 (2006) 365–371.
[27] M. Passare, A. Tsikh, Algebraic equations and hypergeometric series, in:
O.A. Laudal, R. Piene (Eds.), The Legacy of Niels Henrik Abel, SpringerVerlag, New York/Berlin, 2004, pp. 653–672.
[28] G. Pólya, Sur les séries entières, dont la somme est une fonction algébrique,
Enseignement Math. 22 (1921–22) 38–47.
[29] D. Zeitlin, A new class of generating functions for hypergeometric polynomials, Proc. Amer. Math. Soc. 25 (1970) 405–412.
58
| 0math.AC
|
Construction of Directed 2K Graphs
Bálint Tillman
Athina Markopoulou
University of California, Irvine
[email protected]
University of California, Irvine
[email protected]
Carter T. Butts
Minas Gjoka
University of California, Irvine
[email protected]
Google
[email protected]
arXiv:1703.07340v1 [cs.SI] 21 Mar 2017
ABSTRACT
We study the problem of constructing synthetic graphs that resemble real-world directed graphs in terms of their degree correlations.
We define the problem of directed 2K construction (D2K) that takes
as input the directed degree sequence (DDS) and a joint degree
and attribute matrix (JDAM) so as to capture degree correlation
specifically in directed graphs. We provide necessary and sufficient
conditions to decide whether a target D2K is realizable, and we
design an efficient algorithm that creates realizations with that
target D2K. We evaluate our algorithm in creating synthetic graphs
that target real-world directed graphs (such as Twitter) and we
show that it brings significant benefits compared to state-of-the-art
approaches.
CCS CONCEPTS
•Mathematics of computing → Graph algorithms;
KEYWORDS
Directed graphs, graph realizations, construction algorithms
1
INTRODUCTION
It is often desirable to generate synthetic graphs that resemble realworld networks w.r.t. certain properties of interest. For example,
researchers often want to simulate a process on a realistic network
topology, and they may not have access to a real-world network,
or they may want to generate several different realizations of the
graphs of interest. In this paper, we focus specifically on directed
graphs that appear in many application scenarios including, but
not limited to, online social networks such as Twitter (e.g., referring
to the follower relations or to actual communication among users).
There is a large body of work, both classic ([10],[22],[21],[30])
and recent (dK-series [26],[27], PAM [11]) that has studied the problem of constructing realizations of undirected graphs that exhibit
exactly some target structural properties such as a given degree
distribution or a given joint degree matrix. In this paper, we adopt
the dK-series framework [26],[27], which provides an elegant way
to trade accuracy (in terms of graph properties) for complexity (in
generating graph realizations). Construction of dK-graphs is well
understood (i.e., efficient algorithms and realizability conditions
are known) for 1K (graphs with a given degree distribution) and
2K ( graphs with a given joint degree matrix). For d > 2 (which
is necessary for capturing the clustering exhibited in social networks), we recently proved that the problem is NP-hard [7] but we
also developed efficient heuristics [19]. In contrast, construction
is not well-understood for directed graphs: results are known for
construction of graphs with a target directed degree sequence [17],
[16], but there is no notion of directed degree correlation or directed
dK-series for d ≥ 2.
In this paper, we address this problem. We define two notions of
degree correlation in directed 2K graphs, namely directed 2K (D2K),
and its special case D2Km. D2K includes the notion of directed
degree sequence and builds on an old trick (mapping directed graphs
to bipartite undirected graphs) to also express degree-correlation via
a joint degree-attribute matrix (JDAM) for the bipartite graph. This
problem definition lends itself naturally to techniques we previously
developed for undirected 2K [19], an observation that we exploit to
develop (i) necessary and sufficient realizability conditions and (ii)
an efficient algorithm that constructs D2K realizations. Our D2K
approach advances the state-of-the-art in modeling and simulating
realistic directed graphs, especially in the context of online social
networks.
The outline of the rest of the paper is as follows. Section 2 summarizes related work. Section 3 defines the Directed 2K problem
(D2K and its special case D2Km). Section 4 provides realizability
conditions for D2K and an efficient algorithm for constructing such
realizations. Section 5 applies the algorithm to construct Directed
2K graphs that resemble real world graphs, and demonstrates the advantages of our approach compared to state-of-the-art approaches.
Section 6 concludes the paper.
2
RELATED WORK
We adopt the systematic framework of dK-series [26], which characterizes the properties of a graph using a series of probability
distributions specifying all degree correlations within d-sized, simple, and connected subgraphs of a given graph G. In this framework,
higher values of d capture progressively more properties of G at
the cost of more complex representation of the probability distribution. The dK-series exhibit two desired properties: inclusion (a dK
distribution includes all properties defined by any d 0 K distribution,
∀d 0 < d and convergence (nK, where n = |V | specifies the entire
graph, within isomorphism).
We focus on graph construction approaches that produce simple
graphs with prescribed target distributions, unlike the stochastic
approach presented by [9] or the configuration model in [1]. Algorithms of known time complexity exist for d ≤ 2, and Monte Carlo
Markov Chain (MCMC) approaches are used for d > 2.
0K Construction. 0K describes graphs with prescribed number
of nodes and edges. This notion translates to simple Erdős-Rényi
(ER) graphs with fixed number of edges. There is a simple extension
of ER graphs to generate not just undirected but directed graphs as
well, which we will use in our evaluation.
Construction of Directed 2K Graphs
1K Construction. Degree sequences are equivalent to 1K as
defined in the dK-series. Because degree sequences have been
studied since the 1950s, we only focus on the most relevant results.
The realizability conditions for degree sequences were given by the
Erdős-Gallai theorem [10], and first algorithm to produce a single
realization by Havel-Hakimi [22],[21]. More recently, importance
sampling algorithms were proposed in [4] and [6].
2K Construction. Joint Degree Matrix (JDM) is given by the
number of edges between nodes of degree i (Vi ) and j (Vj ) as in [2]:
Õ Õ
J DM(i, j) =
1 {(u,v)∈E }
(1)
u ∈Vi v ∈Vj
Realizability conditions were given by [2] and several other
algorithms to produce single realizations for given JDM [5], [19] and
[29]. The algorithms presented in [2] and [29] are designed to only
produce more restricted realizations with a property called Balanced
Degree Invariant, while in [5] and [19], the algorithms have nonzero probability to produce any realization of a 2K distribution. An
importance sampling algorithm was introduced by Bassler et. al.
[3].
Similarly an extension of JDM, Joint Degree and Attribute Matrix
(JDAM) is given in graph with a single node attribute by the number
of edges between nodes of degree i, attribute value p (V {i,p } ) and
degree j, attribute value q (V {i,q } ) as in [19]. Known results can be
easily applied from the JDM problem including sampling.
Õ
Õ
J DAM({i, p}, {j, q}) =
1 {(u,v)∈E }
(2)
u ∈V{i,p } v ∈V{j,q}
dK, d > 2 Construction. While several attempts were made to
find polynomial time algorithms to produce 3K graphs [26] or 2K
realization with prescribed (degree-dependent) clustering coefficient in [18],[19] and [27], it was recently proved that even the
realizability check for these inputs is NP-Complete in [7].
Annotated graph construction was proposed in [8] that considered degree correlations, however the proposed construction
method will generate graphs with self-loops or multi-edges initially.
An additional step removes these extra edges to make the graph
simple and finally the largest connected component of the graph is
returned as the constructed realization.
The space of simple realizations of 1K distributions is connected
over double edge swaps preserving degrees [30] and a similar result
was shown in [5] or [2] for 2K with double edge swaps that preserve
degrees and the joint degree matrix. These swaps allow the use of
MCMC to generate approximate probability samples for 1K and
2K. However, fast mixing for the MCMC has not been proved in
general, only for special classes of realizations [15] [14].
We highlight related work for bipartite and directed degree sequences in details in the following section as part of our description
of directed dK-series.
3
DIRECTED 2K PROBLEM DEFINITION
We are able to define analogous distributions for the dK-series to
capture directed graphs with given properties.
Directed 0K. As mentioned in Section 2, it is trivial to extend
ER model for directed graphs. In addition, we consider another
model called UMAN [23], that captures the number of mutual,
asymmetric, and null dyads in a graph. We can think about UMAN
B. Tillman et al.
as 0K distribution with fixed numbers of mutual and unreciprocated
edges.
Directed 1K. In an undirected graph G, a node v has degree d(v),
and
the
degree
sequence
is
simply
DS = {d 1 , d 2 , ...d |V | }. In a directed graph, a node v has both in and
out degree and the directed degree sequence is DDS = {(dvin , dvout ), v ∈
V }; an example is shown on Fig. 1-top left.
It is well known from Gale’s work [17], that any directed graph
can be 1-1 mapped to an undirected bipartite graph, where each
node v of the directed graph is split in two nodes vin and vout ,
and the undirected edges across the two (in and out) partitions of
the bipartite graph correspond to the directed edges in the directed
graph. A self loop (v, v) in the directed graph corresponds to a “nonchord”(vin , vout ) in the bipartite graph, and is shown in dashed
line on Fig. 1-left-bottom.
Construction algorithms are known for a bipartite degree sequence with [17], or without non-chords, and therefore of the
corresponding directed graphs with or without [16] self loops, respectively. More recently, an importance sampling algorithm was
shown in [24].
Directed 2K. Our goal in this paper is to go beyond directed
degree sequence and capture degree correlation, and there are two
ways to go about it.
D2Km:
Joint
(Directed)
Degree
Matrix
out )).
J DM((diin , diout ), (d in
,
d
j
j
One way is to work directly with the directed graph, see Fig.
1-top row. We can partition nodes by both their in and out degrees
(dvin , dvout ), and we can define the joint degree matrix to capture
out
the number of edges J DM((diin , diout ), (d in
j , d j )), between nodes
out
with (diin , diout ) and (d in
j , d j ). This is shown on the top of Fig. 1.
This is a natural extension of the JDM in the undirected case, and
expresses a restrictive notion of degree correlation. However, there
is no known algorithm that can provably generate this target JDM.
For example our own 2K construction algorithm for undirected
graphs [19] does not work all the time, although it is still a good
heuristic for sparse graphs.
D2K: Joint Degree-Attribute Matrix J DAM(deдree, in or out)
An alternative approach is to work with the equivalent representation as an undirected bipartite graph without non-chords (Fig.
1-bottom left), and define degree correlation there. We partition in
and out nodes by their degree, essentially considering that nodes
in the bipartite graph can have an attribute that takes two values, “in” or “out.” We can now define degree correlation using
the Joint Degree-Attribute Matrix (JDAM, which we first defined
in [19]), as shown on Fig. 1- bottom right. This leads to a JDAM
with two attribute values, such that ∀i, j = 1, ..., dmax degrees
and p ∈ {in, out } attribute values J DAM({i, p}, {j, p}) = 0, i.e., because the bipartite graph has no edges between two “in” or ”‘out”
nodes. Furthermore, the number of non-chords will be noted as
f ({i, p}, {j, q}), where i, j ∈ {1, ..., dmax } and p , q ∈ {in, out }; f
can be computed by passing through the directed degree sequence
once and counting the number entries with in-degree i and outdegree j.
This notion of bipartite J DAM has all the properties known for
J DM and J DAM, since it is a special case of J DAM that we first
defined in [19]. This includes sufficient and necessary conditions for
Construction of Directed 2K Graphs
B. Tillman et al.
Figure 1: Defining Directed 2K, to capture degree correlations in a directed graph. Top left, Directed 1K: Directed graph with
a given degree sequence (DDS). Bottom left, Bipartite 1K: Mapping of the previous to a Bipartite undirected graph with a
given bipartite degree sequence; non-chords in the bipartite graph (shown in dashed line) correspond to self-loops in the
directed graph. Bottom right, Directed 2K (D2K): Joint-Degree-Attribute Matrix (JDAM), where nodes of the bipartite graph
are partitioned by their degree-and-(in or out) attribute. Top right: Directed 2Km: Joint Degree Matrix (JDM) for directed
graphs, where nodes are partitioned according to their (in degree, out degree).
realizability, construction algorithms, existence of Balanced Degree
Invariant realizations, importance sampling algorithm extensions
from J DM, connectivity of space of realizations over J DAM preserving double-edge swaps and MCMC properties. However, we
have to show for D2K, that the non-chords described by the directed
degree sequence can be added as well.
Relation of the two problems. An overview of the problems
of interest is provided on Fig. 1. One difference between the two 2K
problems is that D2Km provides a more restrictive notion of degree
correlation than D2K since it partitions nodes by two numbers
(d in , d out )) vs. one number dvin or dvout . D2Km can essentially be
obtained as a special case of D2K by further partitioning nodes with
the same d in by their out degree as well. Therefore, D2Km can be
solved by the same algorithm that solves D2K. For the rest of the
paper, we will consider the D2K problem.
Directed 2K (D2K) Problems: Given targets JDAM and DDS:
• Realizability: Decide whether this D2K is realizable, i.e.,
whether there exist graphs with these properties.
• Construction: Design an algorithm that constructs at
least one such realization.
• Sampling: Sample from the space of all graph realizations
with the target D2K.
4
REALIZABILITY AND ALGORITHM
In this section, we take as input the two target properties, namely
the target JDAM({i, p}, {j, q}) with two attribute values and the
DDS , and construct a directed 2K-graph with N nodes that exhibits exactly these target properties. In this section, we use the
bipartite representation of directed graphs as in Figure 1; this enables us to simplify our algorithm description, loose directionality
of the edges and only handle a partition with non-chords.
Recall that in the D2K definition, nodes are partitioned into K
parts Vk , k = 1...K, according to the distinct d in = i or d out = j
they exhibit and JDAM({i, p}, {j, q}) is indexed accordingly. For
example, on Fig. 1 bottom-right, each node belongs to one of four
parts V {0,in } = {v ∈ V : d in = 0}, V {1,out } = {v ∈ V : d out = 1},
V {1,in } = {v ∈ V : d in = 1}, V {2,in } = {v ∈ V : d in = 2}, and the
JDAM is 3x3 (by removing rows and columns corresponding to any
V {0,p } , since there are no edges using these parts of any partition).
4.1
Realizability
Not all target properties are realizable (or “graphical”): there does
not always exist at least one simple directed graph with those exact
properties. Necessary and sufficient conditions for a target D2K, i.e.,
JDAM({i, p}, {j, q}) and DDS , to be realizable are the following.
I ∀i, j, p : J DAM({i, p}, {j, p}) = 0
II ∀i, j, p, q,
if J DAM({i, p}, {j, q})
>
0,
J DAM({i, p}, {j, q}) +f ({i, p}, {j, q}) ≤ |Vk | · |Vl |
Í
J DAM ({i,p }, {j,q })
III ∀i, p : |V {i,p } | = {j,q }
= number of
i
times i appears in DDS as p and it is an integer.
These are generalizations of the conditions for an undirected
JDM, JDAM to be realizable, and they are clearly necessary. The first
condition states that the target JDAM is bipartite, i.e., there should
B. Tillman et al.
Construction of Directed 2K Graphs
be no edges between two nodes both in “in” or “out” parts. The second condition considers edges between two (“in” and “out” ) parts
and states that the number of edges defined by the J DAM({i, p}, {j, p})
plus the number of non-chords should not exceed the total number
of edges possible in a complete bipartite graph across the two parts.
The last condition ensures that the target JDAM and the target DDS
are consistent: the number of nodes with in (or out) degree i should
be the same whether computed using the JDAM or the DDS. The
conditions are shown to be sufficient via the constructive proof
of the algorithm. Necessity of these conditions for simple graph
construction are trivial.
4.2
Algorithm
Algorithm 1
Input: DDS , J DAM
Initialization:
a: Create nodes, partition, stubs using DDS
b: Add non-chords to G using DDS
Add Edges:
1: for every pair ({i, p}, {j, q}) of partition:
2:
while J DAM({i, p}, {j, q}) < J DAM ({i, p}, {j, q})
3:
Pick any nodes u (from V {i, p}), v (from V {j, q})
s.t. (u, v) is not a non-chord or existing edge
4:
if u does not have free stubs:
5:
u 0 : node in V {i, p} with free stubs
6:
neighbor switch for u using u 0
if neighbor switch fails, u := u 0
7:
if v does not have free stubs:
8:
v 0 : node in V {j, q} with free stubs
9:
neighbor switch for v using v 0
if neighbor switch fails, v := v 0
10:
add edge between (u, v)
11:
J DAM({i, p}, {j, q}) ++; J DAM({j, q}, {i, p}) ++;
12: Transform bipartite G to directed graph
Output: simple directed graph G
First, we create a set of nodes V , where |V | = 2 · |DDS|, we
assign stubs to each node and partition nodes, as specified in the
target directed degree sequence DDS . The stubs are originally free,
i.e., they are only connected to one node. We also initialize all
entries of JDAM to 0 and the non-chords between nodes according
to DDS . Then the algorithm proceeds by connecting two nodes
(one from “in” and one from “out” side), thus adding one edge (u, v)
at a time, that (i) are not previously connected to each other (ii) do
not have a non-chord between them (to avoid self loops) and (iii)
for whom the corresponding entry in the JDAM has not reached
its target. The challenge lies in showing that the algorithm will
always be able to make progress, by adding one edge at a time,
until all entries of the JDAM reach their target, when the algorithm
terminates. Indeed, there may be cases (2-5 in Fig.2), where u, v
or both do not have free stubs. Even in those cases, however, we
will be able to perform JDAM-preserving edge rewirings (called
neighbor switch [19]: remove a neighbor t of v such that t is not
a neighbor of v 0 and add edge (t, v 0 )) to free stubs and then add
the edge (u, v) (cases 2-3); or we will be able to add another edge
(u 0, v 0 ) (cases 4-5a). Next, we prove that this is indeed always the
case.
Proof. {Here, we follow the style of proof from Gjoka et. al.
[19]. However, we would like to point out that other similar results
for JDM construction could be extended for directed 2K as well,
such as the proof in [5].}
Condition II. guarantees that two nodes can be always chosen to
add an edge and Condition III. ensures that at least one node exist
in every part of the partition as long as J DAM(i, j) < J DAM (i, j)1 .
Now, we simply show that every iteration can proceed by adding a
new edge to the graph:
Case 1. Add a new edge between two nodes w/free stubs, no
local rewiring needed.
Case 2. Add a new edge between a node v w/out free stubs and a
node u w/free stubs where neighbor switch is possible for v without
using any non-chords.
Case 3. Add a new edge between two nodes w/out free stubs
where neighbor switches are possible for both nodes without using
any non-chords.
Case 4. Add a new edge between a node v w/out free stubs
and a node u w/free stubs (or w/out free stubs where neighbor
switch is possible) where neighbor switch is not possible for v
using v 0 without using any non-chords. In this case v 0 has the
same neighbors as v except the one for which it has an assigned
non-chord. In this case v 0 is not connected to u and it is possible to
add {v 0, u} edge ({v 0, u} is clearly not an edge since then u would
be also connected to v or v could have done a neighbor switch).
Case 5. Add a new edge between two nodes (u, v) w/out free stubs,
where neither can do a neighbor switch with u 0 and v 0 respectively.
We have to break this case into two subcases, based on whether
two nodes u 0, v 0 w/out free stubs form a non-chord or not.
Case 5a. u 0, v 0 is not a non-chord. This means that we can add
a new edge between u 0, v 0 . It is easy to see that u 0, v 0 edge is not
already present, because otherwise u and v could have performed
a neighbor switch.
Case 5b. u 0, v 0 is a non-chord. This case is not possible when u, v
are not able to perform neighbor switches at the same time. Without
loss of generality, let’s say that u connects to all the neighbors of
u 0 and node v 0 . This means that no neighbor switch is available
for u. Now, if we want to construct v such that it can’t perform a
neighbor switch with v 0 , we need v to connect all of the neighbors
of v 0 ; however, this would include u as well, and clearly that edge
doesn’t exist. Contradiction.
This concludes our proof and shows that the algorithm will termiÍ
nate and generate a bipartite graph after adding J DAM (i, j)/2 =
|E| edges.
Running Time. The time complexity of the above algorithm
is O(|E| · dmax ). In each iteration of the while loop, one edge
is always added, until we add all |E| edges. However, we have
to consider how much time it takes to pick such nodes. There
could be neighbor switches that remove previously added edges
or add edges between the two parts. If we naively looked up node
pairs, it would become an issue for dense graphs. A simple solution is to keep track of J DAM ({i, p}, {j, q}) − J DAM({i, p}, {j, q})
1 For
more details please read Lemma 2. and 3. in [19].
Construction of Directed 2K Graphs
B. Tillman et al.
Figure 2: Different possible cases, while attempting to add (u, v) edge in Algorithm 1.
many node pairs where edges can be added in a set P. For every
pair of {i, p}, {j, q}, it is possible to initialize P by passing through
O(J DAM ({i, p}, {j, q}) + f ({i, p}, {j, q})) node pairs. A new (u, v)
node pair can simply be picked as a random element from P. If
a neighbor switch for u ∈ V {i,p } (and similarly to v), rewires a
neighbor t ∈ V {j,q } , then P = P \ {u 0, t } ∪ {u, t } maintains available node pairs in P. Note: {u 0, t } might not be in P. This ensures
that |P | ≥ J DAM({i, p}, {j, q}) − J DAM 0 ({i, p}, {j, q}). These simple set operations can be done in constant time, and building P
takes O(E +V ) time over all partition class pairs. Finally we remove
(u, v) from P, which could be different from the starting pair if Case
4 or 5a occurs.
It is also possible to keep track of nodes with free stubs in a
queue for each part of the partition. Once a node has no free stubs,
it will remain so, except during neighbor switches. This allows
selection of candidates for neighbor switches, or new edges when
neighbor switches are not possible, in constant time. However, it
still takes O(dmax ) to check whether a node with free stubs is a
good candidate for neighbor switch, because the sets of neighbors
can be almost the same length, which takes linear time in the size of
sets. In the worst case, there is a possibility for at most two neighbor
switches per new edge, hence the running time is O(|E| · dmax ).
The directed graph can be constructed from the bipartite representation by collapsing nodes with non-chords and assigning
directions to edges appropriately, this takes O(E + V ) time.
4.3
Space of Realizations
The order in which the algorithm adds edges is unspecified. The
algorithm can produce any realization of a realizable D2K, with a
non-zero probability. Considering all possible edge permutations
as the order in which to add the edges, the ones where no neighbor
switch is required correspond to all the possible realizations. Unfortunately, the remaining orderings are difficult to quantify, thus the
current algorithm cannot sample uniformly from all realizations
with a target D2K during construction.
Another way to sample from the space of graph realizations
with a given D2K is by edge rewiring. This method is typically used
by MCMC approaches that transform one realization to another
by rewiring edges so as to present the target properties. On the
positive side, D2K is a special case of an undirected JDAM, and
thus inherits the property that JDAM realizations are connected via
2K-preserving double-edge swaps [5],[2] if non-chords are allowed
Figure 3: Two realizations with the same degree sequence
and JDAM. There is no JDAM preserving double-edge swap
that would not use any self-loops and the C 6 swaps are not
preserving JDAMs. This shows that the edges along the 4directed-cycle must change their direction simultaneously.
(equivalently, self-loops in directed graphs). On the negative side,
we cannot use the known swaps to sample from the space of simple
directed graphs.
The space of simple realizations of directed degree sequences is
connected over double edge swaps, that preserve (in and out) degrees, and triangular C 6 swaps. Triangular C 6 swaps are necessary
in some cases where the difference between two realization is the
orientation of a directed three-cycle: in this case, the orientation
of the cycle has to be reversed in a single step. The sufficiency of
only these two types of swap was shown in [12]. The necessity of
these swaps also carries over to (simple) directed 2K realizations.
However, Fig. 3 shows a counterexample (a directed 4-cycle) where
the classic swaps are not sufficient to transform one realization to
the other, thus requiring a more complex swap. We leave it as an
open question whether tight upper bounds can be derived on the
swap size for Directed 2K realizations. 2
5 EVALUATION ON REAL-WORLD GRAPHS
5.1 Datasets
We have used examples of directed graphs for our experiments
from SNAP [25]: p2p-Gnutella08, Wiki-Vote, AS-Caida, Twitter. We
2 There
are possibly other cases where swaps must be more complex and include more
edges at once, for example larger directed cycles with specific in/out degree order. In
this paper, we do not provide tight upper bounds on the number of self-loops or the
size of swaps required, but we do emphasize that no multi-edges are required and the
number of self-loops are of course bounded by |N | .
B. Tillman et al.
Construction of Directed 2K Graphs
Table 1: Graphs
Name
p2p-Gnutella08
Wiki-Vote
AS-Caida
Twitter
#Nodes
6,301
7,115
26,475
81,306
#Edges
20,777
103,689
57,582
1,768,135
have chosen these networks in order to represent several different
sizes and generating processes for directed graphs.
We have removed any present self-loops or multi-edges from
these graphs, since our goal is to produce simple graphs and this
step ensures that the measured inputs from these graphs will be
realizable using different directed graph construction methods (0K,
UMAN, 1K, 2K, 2Km).
AS-Caida has edge labels according to the relationship between
two ASes (peer, sibling, provider, customer), however provider and
customer edges describe the same relation from the opposite point
of view. We have modified the AS-Caida network by removing
customer relations between ASes. The affect of the mutual edges which would be present using both provider and customer relations
- will be visited again in our discussion. In [8], this graph was
also considered with the relations included, however during our
construction, we only use directed dK-series as described before in
Section 3.
An overview of the final graphs used in our experiments is shown
in Table 1.
5.2
Properties
In our results, we are going to evaluate the performance of graph
generators in terms of properties associated with directed graphs.
While the size of the generated graphs are maintained (number
of nodes and edges), there are many other properties one could
investigate.
To evaluate correctness of our implementation of D2K, first we
use Degree Distributions and Degree Correlations. These are the
ones expected to be exactly matched by definition.
In addition, we consider additional properties such as shortest
paths, spectrum, strongly connected components, betweenness
centrality, and k-core distribution. Finally, to measure how some
of the local structures are preserved, we use dyad and triad censuses, dyad-wise shared partners, average neighbor degree, and
expansion.
The dyad census counts the different configurations for every
pair of nodes: ”mutual” - edges in both direction, ”asymmetric”
- edge only in one direction and ”null” - no edge present. The
triad census counts the non-isomorphic configuration for every
triplet of nodes. A complete list of configurations and naming
conventions can be found in [23]. Configurations are identified by
three numbers (mutual, asymmetric, and null counts) and a letter
in case of different non-isomorphic configuration with the same
number of edges. For example ”003” is a triplet of nodes where
none of the edges are present, ”030C” is a directed 3-cycle and ”300”
is a triplet of nodes where all directed edges are present.
Shared partners for pairs of nodes can be defined in three ways
for directed graphs: using independent two-paths, using shared
outgoing neighbors or using shared incoming neighbors between
pairs of nodes [28]. Dyad-wise shared partners (DSP) count node
pairs by the number of shared partners appearing in a network.
Average neighbor degree captures the average degree of a nodes’
neighbors, and we split this property for in - and out degrees. Similarly, we refer to expansion for directed graphs as the ratio of the
first hop and second hop neighborhoods’ sizes going out, or coming
in to a node. These properties capture similar aspects of a network,
but expansion excludes any mutual edges or edges between nodes
in the first hop neighbors.
Some of above properties are also used by Orsini et. al. [27] to
study the convergence of dK-series over different types of undirected networks. However, we have included a few properties more
natural to discuss for directed graphs, such as the triad census.
5.3
Implementation
While our current code is not available online, the undirected version of our algorithm is available in NetworkX [20] and easy to
modify for D2K or D2Km using the description in Section 4. We
plan to release our code for D2K and D2Km as part of the NetworkX
library in the near future. Until then, please contact the first author
for the implementation.
The implementation of the used graph properties is available as
part of NetworkX or trivial to implement using the above description.
5.4
Results
We compare realizations generated by Directed ER (D0K), UMAN,
Directed Degree Sequence (D1K), Directed 2K, Directed 2Km with
the corresponding target properties captured on input graph (G).
We use 20 random instances for every construction method and
then average our results for each specific property.
Due to space constraints, we provide detailed results only for
the Twitter graph in Figure 4, 5 and 6 and a brief overview of our
observations over different input graphs.
First, we can observe in Figure 4 that Directed Degree Distributions and Degree Correlations are captured by D2K, D2Km as
expected by definition. This shows that our implementation is correct. On the other hand, D0K, D1K and UMAN capture Degree
Correlations poorly, thus D2K graphs have a possibility to capture
other properties more accurately than D0K or D1K.
Dyad Census is not well captured for Twitter, as we can see in
Figure 5. However, there are order of magnitude improvements
in the number of mutual edges between D2Km (123,040.4) and
D2K (3,628.7), D1K (2,155.95) or D0K (233.05). Of course, UMAN
preserves this property by definition.
Triad Census is surprisingly well captured by UMAN, the reason being the exact match for the Dyad Census in the previous
point. On the other hand, a convergence can be seen between dKseries generators with significant improvements in dense triadic
structures from D1K to D2K and from D2K to D2Km.
Betweenness Centrality CDF has no significant improvements
after matching degree distributions with D1K in Twitter; other examples reached target closer with D1K. Interestingly UMAN performs almost identically to D0K, even though the number of mutual
edges is significantly different.
Construction of Directed 2K Graphs
B. Tillman et al.
Figure 4: Results for Twitter graph: Directed Degree Distribution and Degree Correlation
Figure 5: Results for Twitter graph: Dyad-, Triad Census, Shortest Path Distribution, K-core distribution, Betweenness Centrality
Figure 6: Results for Twitter graph: Expansion, Average Neighbor Degree, DSP and top 20 Eigenvalues
Shortest Path Distribution has slow convergence to target
across different methods, but the average shortest path is shorter
than the observed in G.
Strongly Connected Components (SCC) are not well captured by any of the dK-series generators and they tend to produce
realizations with a single giant SCC and many one-node components without any intermediate sizes of SCCs.
K-Core Distribution is best captured by D2Km, and there is
a small improvement from D1K to D2K using Twitter. However,
the dense core using D1K or D2K is almost an order of magnitude
lower core index.
B. Tillman et al.
Construction of Directed 2K Graphs
Table 2: Summary of results: showing improvements by fixing more properties. Labels: ”.” - no improvement, ”-” - decreased accuracy, ”+” - increased accuracy, ”Exact” - matched
by definition.
Property
Degree Distribution
Degree Correlation
Dyad Census
Triad Census
Betweenness Centrality
Shortest Path Distribution
Eigenvalues
DSP
Expansion
Avg. Neighbor degrees
S. Connected Components
K-Core Distribution
UMAN→D1K
Exact
+
+
+
+
+
+
+
+
.
+
D1K→D2K
Exact
Exact
+
+
.
+
+
+
+
+
.
.
D2K→D2Km
Exact
Exact
+
+
.
+
+
+
+
Exact
.
+
Eigenvalues of Twitter is again best targeted by D2Km. There
is a difference between leading eigenvalues in graph realizations
of the other methods but starting at the second eigenvalue the
difference between D1K and D2K quickly decreases.
Dyad-wise Shared Partners follow similar trends to other
properties, such that D2Km is significantly more accurate than D1K
and D2K. D2K improves over D1K in terms of ”outgoing shared partners” but that improvement decreases at ”independent two-paths”
and disappears at ”incoming shared partners”.
Expansion property is again best approximated by D2Km and
D2Km even matches Average Neighbor Degree exactly if marginalized by degrees as in Figure 6. D2K also follows the general shape
of these distributions but includes larger error, while D1K has systematic difference compared to G.
Table 2 gives an overview of how network properties are affected
by the different dK graph construction methods for the other remaining networks. The Twitter network showcased most of our
general findings, but individually some of these networks have
characteristics that makes them different from Twitter, e.g. p2pGnutella08 does not contain any mutual edges. The most interesting
question is whether D2K or D2Km capture network properties more
accurately. The answer is yes in most cases, but it might not be a
significant improvement in targeting certain properties.
Local structures are generally better captured by D2K and even
more precisely for D2Km, but global properties might not be significantly affected depending on the original network. However,
this result is not surprising, since one of the main assumptions of
the dK-series is that it is not necessary to target high d values for
every graph [27].
5.5
Discussion
We have used several construction algorithms to generate random
graphs and we have observed convergence for most properties in
the directed dK-series. However, certain properties are not well
captured by these methods, such as the Strongly Connected Components. This shows that further extensions or heuristics would be
practical to extend the directed dK-series.
For some networks D1K could be a good choice, since it captures
many network properties. However, we have shown that even
in those cases properties related to first hop neighbors are not
expected to be captured.
While for most metrics the degree labeled construction performs
reasonably well -starting from D1K- it is important to notice that
these methods create a low number of mutual edges. On one side,
UMAN generates graphs with prescribed number of mutual edges
and otherwise ER random graph-like structures. On the other hand,
for larger (sparse) networks the degree labeled construction only
achieves a fraction of the target number of mutual edges. A solution
to this problem would be to generate D2K graphs with a number
of mutual edges. It is possible to design heuristics for this problem
but exact solutions might be difficult to achieve. In addition, we
could consider for D2Km two matrices: one describing asymmetric
and another for mutual edges between nodes with given (in, out)
degrees.
We have also shown that further partitioning of nodes can help
to better describe graphs by their mixing. D2Km is a very specific
partition, that preserves average neighbor degrees for directed
graphs, which is a property that is given by 2K in the undirected
case. In the limit of the possible partitions the graphs could be
exactly fixed; however, in our example this is not the case. While
there are parts of the partition with only a single node in them,
most of their edges go to other parts of the partition with multiple
nodes. This results in a chance to construct distinct realizations
with minimum number of fixed edges across different realizations.
It is an interesting question, how different partitions would affect
the number of realizations. We leave it as future work for both
undirected and directed cases.
The limitations of degree labeled construction for undirected
graphs have been shown recently, such that the realizability problem of undirected 2K with fixed number of triangles, 3K [7], and
second order degree sequences (degree and number of two hop
neighbors) [13] are NP-Complete and a relaxation of JDM partitions called PAM [11] is believed to be NP-Complete and only
solved for special cases. We can say with good confidence that
more restrictive models are likely to lead to NP-Complete problems;
however, different partitions can be solved as long as they produce
a valid JDAM for the D2K problem.
6
CONCLUSION
We have shown a new approach for directed graph construction by
considering in and out degree correlations in addition to directed
degree sequences. This extension enabled us to build a framework
for directed graphs similar to the dK-series, using bipartite graphs.
We solve the problem of generating bipartite graphs with prescribed
degree correlations using the Joint Degree and Attribute Matrix
construction algorithms. Following in the footsteps of classic work,
we use this property of the JDAM problem and defined directed 2K
as the combination of JDAM and a directed degree sequence. To
solve our proposed problem, we provide the necessary and sufficient
conditions for realizability of such inputs and a simple, efficient
algorithm to generate simple realizations (without self-loops) as
well. The uniform sampling from the space of these realizations is
not trivial, as we have shown an example for a necessary complex
Construction of Directed 2K Graphs
edge swap in section 4.3. However, we see an opportunity for
further research in edge swaps for an MCMC approach or a variant
of the importance sampling algorithm from [3].
In addition to directed 2K, we have defined D2Km, a special case
using additional node attributes. D2Km provides a more restricted
notion of directed 2K, that exactly captures average neighbor degree of nodes marginalized by degree. In our experiments, we have
shown convergence for degree labeled directed graph construction
similar to the undirected case [27]. This result is similar to [27] in
nature but it shows that degree correlations capture more information about graph structure and enables us to generate graphs
which more closely resemble real-life networks.
We have identified that directed dK-series can benefit directly by
prescribing the number of mutual edges. We consider this extension the most straightforward step to improve directed dK-series.
However, even this problem could turn out to be NP-Complete.
REFERENCES
[1] William Aiello, Fan Chung, and Linyuan Lu. 2000. A random graph
model for massive graphs. In Proceedings of the thirty-second annual
ACM symposium on Theory of computing. Acm, 171–180.
[2] Georgios Amanatidis, Bradley Green, and Milena Mihail. 2015. Graphic
realizations of joint-degree matrices. arXiv preprint arXiv:1509.07076
(2015).
[3] Kevin E Bassler, Charo I Del Genio, Péter L Erdős, István Miklós, and
Zoltán Toroczkai. 2015. Exact sampling of graphs with prescribed
degree correlations. New Journal of Physics 17, 8 (2015), 083052.
[4] Joseph Blitzstein and Persi Diaconis. 2011. A sequential importance
sampling algorithm for generating random graphs with prescribed
degrees. Internet Mathematics 6, 4 (2011), 489–522.
´ Czabarka, Aaron Dutle, Péter L Erdős, and István Miklós. 2015.
[5] Eva
On realizations of a joint degree matrix. Discrete Applied Mathematics
181 (2015), 283–288.
[6] Charo I Del Genio, Hyunju Kim, Zoltán Toroczkai, and Kevin E Bassler.
2010. Efficient and exact sampling of simple graphs with given arbitrary degree sequence. PloS one 5, 4 (2010), e10012.
[7] William Devanny, David Eppstein, and Bálint Tillman. 2016. The
computational hardness of dK-series. In NetSci 2016.
[8] Xenofontas Dimitropoulos, Dmitri Krioukov, Amin Vahdat, and
George Riley. 2009. Graph Annotations in Modeling Complex Network
Topologies. ACM Trans. Model. Comput. Simul. 19, 4, Article 17 (Nov.
2009), 29 pages. DOI:http://dx.doi.org/10.1145/1596519.1596522
[9] SN Dorogovtsev. 2003. Networks with desired correlations. arXiv
preprint cond-mat/0308336 (2003).
[10] P Erdős and T Gallai. 1960. Gráfok előı́rt fokú pontokkal. Mat. Lapok
11 (1960), 264–274.
[11] Péter L Erdős, Stephen G Hartke, Leo van Iersel, and István Miklós.
2015. Graph realizations constrained by skeleton graphs. arXiv preprint
arXiv:1508.00542 (2015).
[12] Péter L Erdős, Zoltán Király, and István Miklós. 2013. On the swapdistances of different realizations of a graphical degree sequence. Combinatorics, Probability and Computing 22, 03 (2013), 366–383.
[13] Péter L Erdős and István Miklós. 2016. Not all simple looking degree
sequence problems are easy. arXiv preprint arXiv:1606.00730 (2016).
[14] Péter L Erdos, István Miklós, and Zoltán Toroczkai. 2015. A decomposition based proof for fast mixing of a Markov chain over balanced
realizations of a joint degree matrix. SIAM Journal on Discrete Mathematics 29, 1 (2015), 481–499.
[15] Péter L Erdős, István Miklós, and Zoltán Toroczkai. 2016. New classes
of degree sequences with fast mixing swap Markov chain sampling.
B. Tillman et al.
arXiv preprint arXiv:1601.08224 (2016).
[16] Delbert Ray Fulkerson and others. 1960. Zero-one matrices with zero
trace. Pacific J. Math 10, 3 (1960), 831–836.
[17] David Gale and others. 1957. A theorem on flows in networks. Pacific
J. Math 7, 2 (1957), 1073–1082.
[18] Minas Gjoka, Maciej Kurant, and Athina Markopoulou. 2013. 2.5 kgraphs: from sampling to generation. In INFOCOM, 2013 Proceedings
IEEE. IEEE, 1968–1976.
[19] Minas Gjoka, Bálint Tillman, and Athina Markopoulou. 2015. Construction of simple graphs with a target joint degree matrix and beyond.
In 2015 IEEE Conference on Computer Communications (INFOCOM).
IEEE, 1553–1561.
[20] Aric A. Hagberg, Daniel A. Schult, and Pieter J. Swart. 2008. Exploring
network structure, dynamics, and function using NetworkX. In Proceedings of the 7th Python in Science Conference (SciPy2008). Pasadena,
CA USA, 11–15.
[21] S Louis Hakimi. 1962. On realizability of a set of integers as degrees of
the vertices of a linear graph. I. J. Soc. Indust. Appl. Math. 10, 3 (1962),
496–506.
[22] Václav Havel. 1955. Poznámka o existenci konecnych grafu. Časopis
pro pěstovánı́ matematiky 80, 4 (1955), 477–480.
[23] Paul W Holland and Samuel Leinhardt. 1976. Local structure in social
networks. Sociological methodology 7 (1976), 1–45.
[24] Hyunju Kim, Charo I Del Genio, Kevin E Bassler, and Zoltán Toroczkai.
2012. Constructing and sampling directed graphs with given degree
sequences. New Journal of Physics 14, 2 (2012), 023012.
[25] Jure Leskovec and Andrej Krevl. 2014. SNAP Datasets: Stanford
Large Network Dataset Collection. http://snap.stanford.edu/data.
(June 2014).
[26] Priya Mahadevan, Dmitri Krioukov, Kevin Fall, and Amin Vahdat. 2006.
Systematic topology analysis and generation using degree correlations.
In ACM SIGCOMM Computer Communication Review, Vol. 36. ACM,
135–146.
[27] Chiara Orsini, Marija M Dankulov, Pol Colomer-de Simón, Almerima
Jamakovic, Priya Mahadevan, Amin Vahdat, Kevin E Bassler, Zoltán
Toroczkai, Marián Boguñá, Guido Caldarelli, and others. 2015. Quantifying randomness in real networks. Nature communications 6 (2015).
[28] Tom A. B. Snijders, Philippa E. Pattison, Garry L. Robins, and Mark S.
Handcock. 2006. New Specifications for Exponential Random Graph
Models. Sociological Methodology 36 (2006), 99–154.
[29] Isabelle Stanton and Ali Pinar. 2012. Constructing and sampling graphs
with a prescribed joint degree distribution. Journal of Experimental
Algorithmics (JEA) 17 (2012), 3–5.
[30] R Taylor. 1980. Constrained switchings in graphs. University of Melbourne, Department of Mathematics.
| 8cs.DS
|
Analytical formula for the roots of the general complex cubic polynomial
Ibrahim Baydoun1
1
ESPCI ParisTech, PSL Research University, CNRS, Univ Paris Diderot, Sorbonne Paris Cité, Institut
Langevin, 1 rue Jussieu, F-75005, Paris, France
January 22, 2018
arXiv:1512.07585v2 [math.GM] 19 Jan 2018
Abstract
We present a new method to calculate analytically the roots of the general complex polynomial of degree three. This
method is based on the approach of appropriated changes of variable involving an arbitrary parameter. The advantage
of this method is to calculate the roots of the cubic polynomial as uniform formula using the standard convention of the
square and cubic roots. In contrast, the reference methods for this problem, as Cardan-Tartaglia and Lagrange, give the
roots of the cubic polynomial as expressions with case distinctions which are incorrect using the standar convention.
keyword: cubic polynomial roots, appropriated change of variable, analytical uniform formula.
1
Introduction
The problem of solution formulas of polynomial equations is fundamental in algebra [1, 2]. It is useful, by example, for
the eigenvalues perturbation [3] and the solution of a system of nonlinear equations [4]. Moreover, it has many physical
applications, as the study of the singularities of the surfaces of refractive indices in crystal optics [5] and the diagonal
eigenvalues perturbation [6]. Another example is the calculation of the sound velocity anisotropy in cubic crystals [7],
or the diagonalization of Christoffel tensor to calculate the velocities of the three quasi-modes of the elastic waves in
anisotropic medium [8]. An explicit calculation for the particular case of 3 × 3 real symmetric matrices has been studied
in [9].
For a cubic equation, Cardano-Tartaglia and Lagrange’s formulas are subject to different interpretations depending on
the choices for the values of the square and cubic roots of the formulas [10, 11]. The correct interpretation is generally
derived from conditions between the roots and the coefficients of the cubic polynomial. It is a fundamental question to
be able to define uniform formula that yields correct solutions for all cubic equations. Ting Zhao’s thesis and related
work recently brought an important advance on Lagrange’s formulas with uniform solution formulas for cubic and quartic
equations with real coefficients. The results was obtained by introducing a new convention for the argument of the cubic
root of a complex number [12]. For cubic equations, we present in this paper a uniform formula that is valid with complex
coefficients. This more general result are uniquely defined by the standard convention of the square and cubic roots.
To illustrate the problem, we introduce Cardan-Tartaglia formula and afterward we give two examples. Indeed, CardanTartaglia calculate the 3 solutions of the equation
(C) : z 3 + pz + q = 0
as
v
u
u1
√
2π
3
zn := exp n
−1 t
3
2
where ∆C is defined as
−q +
r
∆C
27
!
v
u
u1
√
2π
3
+ exp −n
−1 t
3
2
−q −
r
∆C
27
!
; n ∈ {0, 1, 2} ,
∆C := − 4p3 + 27q 2 .
We show by the following examples that the standard convention of the cubic roots
√
1
3
X = arg (X) ; X ∈ C
arg
3
is not relevant in the above Cardan-Tartaglia’s expression. Indeed:
Example 1.1. The solution of (C) for the case p = 1 and q = 2 is
√
−1 − −3
; n ∈ {0, 1, 2} ,
zn
2
instead of zn .
Example 1.1 requires to shift the standard convention by −2π/3 in order to be suitable for the Cardan-Tartaglia’s solutions.
Another example requires another convention:
1
√
Example 1.2. The solution of (C) for the case p = 1 and q = −1 is
√
−1 + −3
zn
; n ∈ {0, 1, 2} ,
2
instead of zn .
Therefore, a natural question is how we can calculate the roots of (C) using uniquely the standard convention?
To overcome this problem, a mathematical software as Matlab generates the symbolic solutions of (C) in terms of one
cubic root RC as follows
RC −
p
,
3RC
√
−3
2
p
+ RC
3RC
+
p
RC
−
,
6RC
2
−
√
−3
2
p
+ RC
3RC
+
p
RC
−
,
6RC
2
where RC is given by
RC =
sr
3
q
−∆C
− .
22 33
2
On the one hand, this formulation hid the problem of the convention of the cubic roots. But, on the other hand, it involves
another problem by dividing by RC which leads to a singularity in the formulation if RC is zero. So, the above question
can be reworded by the following manner: how we calculate the roots of (C) using uniquely the standard convention and
without any singularity?
The response to this question is the objective of this paper.
We present a uniform analytical formula for the roots of the general complex cubic polynomial. Our method is based
on the approach of changes of variable involving an arbitrary parameter, which rests on two principal ideas to diagonalize
the general 3 × 3 complex matrix:
Firstly, we introduce one change of variable involving an arbitrary parameter. So, the characteristic equation of this matrix
can be replaced by another non-polynomial equation in terms of the first new variable.
Secondly, we introduce another change of variable involving another arbitrary parameter. The second variable replaces
the first and the second arbitrary parameter can be chosen as required to make the form of the new equation as the sum
of a cube of a certain monomial and a term independent of the unknown second variable.
Therefore, we can solve easily this new polynomial equation and consequently deduce the solution of the original equation,
which is the eigenvalues of the general 3 × 3 complex matrix.
In the sequel, we apply this approach to calculate the roots of the general complex cubic polynomial. In fact, we
construct a particular matrix such that its characteristic polynomial will coincide with this cubic polynomial. Consequently,
the eigenvalues of this particular matrix become identical to the roots of this cubic polynomial.
The rest of the paper is structured as follows. In section 2, we derive the approach of appropriated changes of variable
involving an arbitrary parameter. Consequently in section 3, we apply it on the general complex cubic polynomial.
2 Approach of appropriated changes of variable involving an arbitrary parameter
Let M be the general 3 × 3 complex matrix:
m1
M := m4
m7
m2
m5
m8
m3
m6 .
m9
In this section, we aim to calculate the spectrum of M with this approach. We detail it by the following four steps. The
first step consists of performing a change of variable for the eigenvalues of M so that we get a non-polynomial equation.
In the second step, we will deal with this equation in order to simplify it by some algebraic manipulations. The third
step will introduce a second change of variable in order to replace the non-polynomial equation by another polynomial
equation. Finally, in the fourth step, we will solve the latter polynomial equation to deduce the spectrum of M.
Starting the first step by introducing the following definition which will be useful in the sequel:
Definition 2.1. We associate to M the following expressions in terms of its components:
m3 m8
,
c2 := c21 + 4 (m2 m4 + m3 m7 ) ,
m2
1
3
3
1
(m1 − c1 ) − Tr (M) ,
e2 := (m1 − c1 )2 + (c2 − c1 )2 − (m1 − c1 ) Tr (M) +
[Tr (M)]2 − Tr M2 ,
2
2
8
4
1
1
(2m1 − c1 )3 + 3 (2m1 − c1 ) c2 −
(2m1 − c1 )2 + c2 Tr (M) + (2m1 − c1 ) [Tr (M)]2 − Tr M2 − det (M) ,
4
4
1
1
3 (2m1 − c1 )2 + c2 − (2m1 − c1 ) Tr (M) +
[Tr (M)]2 − Tr M2 .
2
4
c1 := m1 − m5 −
3
2
1
e3 :=
8
1
e4 :=
8
e1 :=
Tr (M) and det (M) stand respectively for the trace and the determinant of matrix M.
Now, we introduce the first change of variable involving an arbitrary parameter so that the eigenvalues of M can be
rewritten in terms of this new variable b̃ such that b̃ verifies a non-polynomial equation like the following proposition:
2
Proposition 2.2. An eigenvalue k of matrix M can be written in terms of two unknowns ã and b̃ as follows:
√ i
ã − b̃
1h
2m1 − c1 + ∆k ,
+
2
2
k=
(1)
where ã can be chosen arbitrary and b̃ is governed by the following non-polynomial equation:
√
√
1 2√
b̃ ∆k − b̃3 = 0,
A + B b̃ + C b̃2 + D ∆k + E b̃ ∆k +
2
(2)
where ∆k is defined by
∆k := b̃2 + 2 (c1 − ã) b̃ + ã2 − 2c1 ã + c2 =
2
m3 m8
+ 4 (m2 m4 + m3 m7 ) .
ã − b̃ − m1 + m5 +
m2
(3)
The coefficients of Eq. (2) depend on parameter ã as follows:
A=
ã3
+ e1 ã2 + e2 ã + e3 ,
2
B=−
3ã2
− 2e1 ã − e2 ,
2
3ã
+ e1 ,
2
c1
ã2
ã + e4 ,
+ e1 +
D=
2
2
C=
where the coefficients of Eq. (4) are given by definition 2.1.
c1
E = −ã − e1 +
,
2
Proof:
Firstly, we decompose the matrix M as the sum of two matrix M1 and M2 as follows:
ã 0
0
m1 − ã
m2
m3
M = 0 b̃ −m0 + m4
m5 − b̃ m6 + m0 ,
0 0
0
m7
m8
m9
|
{z
} |
{z
}
M1
(4)
(5)
M2
where m0 is chosen such that M2 has an eigenvector of the form [0, y0 , z0 ]. The advantage of this choice is to construct
an eigenvector of M in terms of these of M2 which will be elaborated in the sequel. Indeed, if we take m0 such that
m3
m8
− m9
− m6 ,
m0 = m5 − b̃ + m3
m2
m2
then M2 have the eigenvalue
k0 = −m3
m8
+ m9
m2
(6)
relative to the eigenvector
m3
v0 := 0, −
,1 .
m2
Equation 6 gives one eigenvalue k0 of M2 , then the characteristic polynomial of M2
1 0 0
det X 0 1 0 − M2
0 0 1
is divided in C [X] by the polynomial X − k0 . Consequently, we deduce that the two other eigenvalues of M2 verify a
second order polynomial and they are given by:
k̃ =
√ i
−ã − b̃
1h
2m1 − c1 ∓ ∆k .
+
2
2
(7)
Secondly, we search an eigenvector for M of the form
v0 + vv; v ∈ C,
where v̂ := [v1 , v2 , v3 ] is an unit eigenvector of M2 relative to k̃. Indeed, using
M 2 v 0 = k0 v 0 ,
M2 v = k̃v,
we obtain from Eq. (5) that the vectorial equation
M (v0 + vv) = k (v0 + vv) ; k ∈ C,
is equivalent to the following system:
v1v ã + k̃ −k = 0,
3
(S)
− m0 (vv3 + 1) + v k̃v2 −
b̃ vv2 − m
m2
v k̃v3 + k0 − k (vv3 + 1) = 0.
3
m3
k
m2 0
− k vv2 −
m3
m2
= 0,
The impact of the choice of m0 appears in the fact that the first equation of (S) is independent of v0 + vv and it can be
verified by taking k = ã + k̃. Thus, by taking v as
v=
1 k − k0
,
v3 k̃ − k
the third equation of (S) is verified. It remains to verify the second equation of (S). Indeed, k = ã + k̃ and Eq. (7) implies
Eq. (1) where k depends only on ã − b̃ since ∆k depends only on ã − b̃. Therefore, by inserting Eq. (1) in the characteristic
equation of M, we get
k3 − Tr (M) k2 +
1
{[Tr (M)]2 − Tr M2 }k − det(M) = 0,
2
(8)
so that we can choose arbitrarily among ã and b̃ one parameter as required, while the other parameter will be the unique
unknown of Eq. (8), as well as k is an eigenvalue of M. We obtain that if ã and b̃ satisfy Eq. (8), then the second equation
of (S) is verified. To end the proof, we develop Eq. (8) to deduce that Eq. (8) and Eq. (2) are equivalent.
In the second step, we aim to simplify Eq. (2) of the unknown b̃, where ã can be chosen as required. So, we derive from
Eq. (2) two equations of b̃ given by the following two propositions:
Proposition 2.3. For fixed ã, b̃ satisfies the following equation:
√
√
d1 + d2 b̃ + d3 b̃2 + d4 ∆k − d3 b̃ ∆k = 0,
(9)
where the coefficients of Eq. (9) are given by:
d1 := d3 ã2 + s1 ã + s2 ,
d2 := −2d3 ã − s1 ,
d3 := m2 m4 + m3 m7 ,
Here s1 , s2 and s3 are given using definition 2.1 as follows:
c1 c2
s1 :=
+ c1 e2 + c2 e1 − 2c1 e4 − e3 ,
2
d4 := d3 ã + s3 .
s2 := c1 e3 + c2 e4 ,
s3 := e3 + c1 e4 .
Proof:
√
Eq. (9) is deduced by multiplying Eq. (2) by b̃ + ∆k − ã + c1 .
(10)
Proposition 2.4. For fixed ã, b̃ satisfies the following equation:
√
√
f1 + f2 b̃ + f3 b̃2 + f4 ∆k + f3 b̃ ∆k = 0,
(11)
where the coefficients of Eq. (11) are given by:
f1 := f3 ã2 + s4 ã + s5 ,
f3 := (m5 − m1 ) m3 m7 + (m9 − m1 ) m2 m4 − m2 m6 m7 − m3 m4 m8 ,
f2 := −2f3 ã − s4 ,
f4 := −f3 ã + s6 .
Here s4 , s5 and s6 are given using definition 2.1 as follows:
s4 := c2 d3 − 2c21 e4 − 2c1 e3 − 2d3 e2 ,
s5 := c2 e3 − 2d3 e3 + c1 c2 e4 ,
s6 := c1 e3 + c2 e4 − 2d3 e4 .
(12)
Proof:
Similary
to the proof of proposition 2.3, on one hand we multiply Eq. (2) by 2d3 and on the other hand we multiply Eq. (9)
√
by ∆k . Consequently, we subtract the obtained equations to deduce Eq. (11).
Now, we show that equations (9) and (11) imply the two equations (13) and (14) given by the following corollary:
Corollary 2.5. For fixed ã, b̃ satisfies the following equations:
√
√
r1 + r2 b̃ + r3 ∆k − 2b̃ ∆k = 0,
√
r4 + r5 b̃ + r6 ∆k + 2b̃2 = 0,
(13)
(14)
where the coefficients of Eq. (13) and Eq. (14) are given by:
s1
s4
+ ,
d3
f3
r1 := −r2 ã + s9 ,
r2 := −
r4 := 2ã2 + s7 ã + s8 ,
r5 := −4ã − s7 ,
r3 := 2ã + s10 ,
r6 :=
s6
s3
+
d3
f3
(15)
Here, the coefficients of Eq. (15) are defined using definition 2.1, Eq. (10) and Eq. (12) as follows:
s7 :=
s1
s4
+ ,
d3
d3
s8 :=
s2
s5
+ ,
d3
f3
s9 :=
s2
s5
− ,
d3
f3
s10 :=
s3
s2 − 2d3 e4
−
.
d3
f3
Proof:
We divide Eq. (9) and Eq. (11) respectively by d3 and f3 . Then, the subtraction and the addition of the obtained equations
imply Eq. (13) and Eq. (14).
In the third step, we introduce the second change of variable
√
l := − b̃ + o ∆k
(16)
in order to deduce from the intractable equations of b̃ another polynomial quation of the new variable l, where o is an
arbitrary parameter can be chosen as required:
4
Proposition 2.6. Let o ∈ C. Then l verifies the following equation:
2l3 + Bl l2 + Cl l + Dl = 0,
(17)
where the coefficients of Eq. (17) are given by:
Bl := −l1 o + 6ã − l2 ,
3
2
2
3
Cl := l3 o2 − 2 (l1 ã + l4 ) o + 6ã2 − 2l2 ã + l5 ,
2
Dl := −l6 o + (l3 ã − l7 ) o − l1 ã + 2l4 ã + l8 o + 2ã − l2 ã + l5 ã + l9 .
Here Bl , Cl and Dl are defined using definition 2.1, Eq. (10), Eq. (12) and Eq. (15) as follows:
s1
− s7 + r 6 ,
d3
s2
l4 := 2 − s8 − c1 r6 ,
d3
r6 s1 − (s7 + 4c1 ) s3
,
d3
s1 (s8 − 2c2 ) − (4c1 + s7 ) s2
l6 := 2c1 s8 + c2 s7 +
,
d3
s7 s2 − s1 s8
s8 s3 − r 6 s2
l8 := c2 r6 +
,
l9 :=
.
d3
d3
s3
,
d3
s7 s3 − r 6 s1
l5 := s8 +
,
d3
(s8 − 2c2 ) s3 − r6 s2
l7 := 2c1 s8 + c2 (s7 + r6 ) +
,
d3
l2 := r6 − s7 − 2
l1 := 2
l3 := s8 − 2c2 + 2c1 r6 +
Proof:
√
The first part of the proof consists in getting expressions for b̃2 and ∆k just in terms of l and o. Indeed, Eq. (16) implies:
√
b̃ = − o ∆k + l .
(18)
We take the square of Eq. (18) and we replace ∆k by its value given in Eq. (3). Then, we insert Eq. (18) in the obtained
equation to deduce an expression of b̃2 . This chain of operations is illustrated as follows:
b̃2 = o2
∆k
|{z}
b̃2 +2(c1 −ã)b̃+ã2 −2c1 ã+c2
√
+l2 + 2ol ∆k ⇒ 1 − o2 b̃2 = 2o2 (c1 − ã)
− o
b̃
|{z}
√
∆k +l
√
+o2 ã2 − 2c1 ã + c2 + l2 + 2ol ∆k
√
2o l − (c1 − ã) o2
∆k + ã2 − 2c1 ã + c2 o2 + l2 − 2 (c1 − ã) o2 l
⇒ b̃ =
.
1 − o2
2
(19)
2
So,
√ inserting Eq. (18) and Eq. (19) in Eq. (14) in order to substitute b̃ and b̃ and consequently to get the expression of
∆k as follows:
r4 + r5
− o
√
+r6 ∆k + 2
b̃
|{z}
√
∆k +l
b̃2
|{z}
=0
o
n
√
(1−o2 )−1 2o[l−(c1 −ã)o2 ] ∆k +(ã2 −2c1 ã+c2 )o2 +l2 −2(c1 −ã)o2 l
1 − o2 (r4 − r5 l) + 2 ã2 − 2c1 ã + c2 o2 − 2 (c1 − ã) o2 l + l2
. (20)
⇒ ∆k =
(1 − o2 ) (r5 o − r6 ) + 2 [2 (c1 − ã) o3 − 2ol]
√
The second part of the proof consists to √
use the previous expressions of b̃2 and ∆k in order to deduce Eq. (17). Indeed,
we take the square of the equation b̃ + o √∆k + l = 0 deduced from Eq. (16). Then, we insert Eq. (3) and Eq. (9) in the
obtained equation to substitute ∆k and b̃ ∆k . Afterwards we insert Eq. (18) to substitute b̃. This chain of operations is
illustrated as follows:
b̃2 + o2
√
∆k
|{z}
√
b̃ ∆k
| {z }
+l2 + 2o
b̃2 +2(c1 −ã)b̃+ã2 −2c1 ã+c2
√
d−1
d1 +d2 b̃+d3 b̃2 +d4 ∆k
3
d2
(1 + o)2 b̃2 + 2 (c1 − ã) o2 + l + o
d3
− o
b̃
|{z}
√
∆k +l
√
+2lo ∆k + 2lb̃ = 0 ⇒
√
d1
d4
+2 l +
o ∆k + ã2 − 2c1 ã + c2 o2 + l2 + 2 o = 0 ⇒
d
d
3
3
√
d1
d2
d4
d2
− (c1 − ã) o2 − o o ∆k − 2 (c1 − ã) o2 + l + o l + ã2 − 2c1 ã + c2 o2 + l2 + 2 o = 0.
(1 + o)2 b̃2 + 2
d3
d3
d3
d3
2
Finally,
√ to complete the proof, it is enough to insert Eq. (19) and Eq. (20) in the above equation in order to substitute b̃
and ∆k and consequently to deduce Eq. (17).
Since o is an arbitrary parameter, then we can choose o as required to solve equation (17) of the unknown l:
Lemma 2.7. If o verifies the following polynomial equation of degree two:
l12 − 6l3 o2 + 2 (l1 l2 + 6l4 ) o + l22 − 6l5 = 0,
then l = −ã + rm , where
rm =
l1 o + l2
+ msl ;
6
m∈
−1 ,
1−
√
2
−3
,
1+
(21)
√
2
Here sl is given by:
sl :=
s
3
l1 o + l2
6
3
−
1
(l6 o3 + l7 o2 + l8 o − l9 ).
2
5
−3
.
Proof:
Equation (17) can be rewritten as follows
3
l12 − 6l3 o2 + 2 (l1 l2 + 6l4 ) o + l22 − 6l5
Bl
Cl
Bl
l3 + 3 l2 + 3 l +
ã = 0.
+ s3l −
6
6
6
12
The condition Bl2 = 6Cl , which is equivalent to Eq. (21), allows us to factorise Eq. (22) as follows:
#
"
3
2
Bl
Bl
Bl
Bl
2
3
l+
l+
+ sl
+ sl = 0.
+ sl = l +
− sl l +
6
6
6
6
So, Eq. (23) can be solved easily, wich ends the proofs.
(22)
(23)
Lemma 2.7 gives the expressions for the set of solutions of l. Then, we deduce from the change of variable (18) the following
polynomial equation of b̃, which will be useful to give an explicit formula for the set of solutions of b̃:
Corollary 2.8. If o verifies Eq. (21), then b̃ verifies the following equation
1 − o2 b̃2 + 2 rm − ã + o2 (ã − c1 ) b̃ + (rm − ã)2 − o2 ã2 − 2c1 ã + c2 = 0,
(24)
where
rm ∈ Sl :=
l1 o + l2
− sl ;
6
√
l1 o + l2
1 − −3
+
sl ;
6
2
√
l1 o + l2
1 + −3
+
sl .
6
2
Proof:
We have from Eq. (16) that
√
∆k = −
1
b̃ + l .
o
(25)
The proof ends by taking the square of Eq. (25) and afterwards replacing ∆k by its value given by Eq. (3) in the obtained
equation, where l is replacing by rm − ã from lemma 2.7.
Finally, in the fourth step, we can determine the set of solutions of b̃ by solving the nonlinear system constructed
by Eq. (24) and one of the following equations (9), (11), (13) or (14) of b̃. Thus, we can solve this nonlinear system by
choosing ã as required to simplify the calculation. Consequently, we deduce from Eq. (1) the spectrum of M:
Corollary 2.9. The spectrum of
1
1+
k=
2
matrix M is given by:
2
rm
+ s3 (1 − o) rm − s2 o (1 − o) − d3 c2 o2
1 d3 r m
+
2m
−
c
−
,
1
1
o (1 − o) (s3 + s1 o) − 2d3 c1 o2 + d3 (1 + o) rm
o
(26)
where rm ∈ Sl and o verifies Eq. (21).
Proof:
Inserting Eq. (25) in Eq. (9), we get an equation of b̃ and b̃2 . Then, we substitute b̃2 using Eq. (24) to obtain:
2
d3 r m
+ (1 − o) s3 rm − s2 o (1 − o) − d3 c2 o2
.
(27)
(1 − o) (s3 + s1 o) − 2d3 c1 o2 + d3 (1 + o) rm
√
Finally, to deduce Eq. (26), we insert Eq. (25) in Eq. (1) in order to substitute ∆k and afterwards we insert Eq. (27) in
the obtained equation in order to substitute b̃.
b̃ = ã −
The denumerator of the fraction in Eq. (26) contains rm which involves a cubic root. So, we calculate another expression
for spectrum of M in order to rationalize this denumerator. The two expressions of spectrum of M will be useful in the
proof of theorem 3.3.
Corollary 2.10. The spectrum of matrix M is given by:
1
rm
1
s9 o2 − s8 o + (r6 − s10 o) rm
k=
+
2m
−
c
−
1+
,
1
1
2
o (s7 − s10 ) o + r2 o2 + r6 + 2orm
o
(28)
where rm ∈ Sl and o verifies Eq. (21).
Proof:
√
We insert Eq. (25) in both Eq. (9) and Eq. (11) to replace ∆k by its value in terms of b̃, l and o. Then, by substituting
2
b̃ from one obtained equation in the other, we deduce that:
s9 o2 − s8 o + (r6 − s10 o) rm
.
(29)
(s7 − s10 ) o + r2 o2 + r6 + 2orm
√
Finally, to deduce Eq. (28), we insert Eq. (25) in Eq. (1) to substitute ∆k and afterwards we insert Eq. (29) in the
obtained equation to substitute b̃.
b̃ = ã −
6
3
Application on the cubic polynomial
Let (P) be the general complex polynomial of degree three in C [x]:
(P) : x3 + bx2 + cx + d = 0; b, c, d ∈ C.
We aim to calculate analytically the roots of (P) by applying the results of section 2. So, we introduce the following
definitions which will be useful to find the expressions for the roots of (P):
Definition 3.1. We associate to (P) the following expressions in terms of its coefficients:
∆l := 2c3 8b6 + 132b3 d + 36d2 + c3 + 33b2 c2 − 66bcd + 12b4 c d2 − 7c3 − b2 c2 d 24b3 + 291d + d3 144bc − 2b3 − 27d ,
∆o := −4b3 d + b2 c2 + 18bcd − 4c3 − 27d2 ,
and
do := 4b4 c2 − 4b3 cd − 14b2 c3 + b2 d2 + 28bc2 d + c4 − 12cd2 .
Definition 3.2. We associate to (P) the following expressions in terms of these given by definition 3.1:
√
√
−3
δl := (d − bc) ∆o 4b2 c2 − 4bcd + 2c3 + d2 +
∆l ,
9
and
√
√
2 −3
4b3 c − 2db2 − 13bc2 + 15dc + 2c ∆o ,
3
√
√
A2 := 8b5 c2 − 8b4 cd − 40b3 c3 + 2b3 d2 + 116b2 c2 d + 23bc4 − 99bcd2 − 21c3 d + 27d3 − −3 8b2 c2 − 10bcd + c3 + 3d2
∆o .
A1 := −
Now, we present the principal result of the paper which is the formula of the roots of (P). This formula is in fact a
Lagrange-type [11], where the square and cubic are given by the standard convention, contrary to Cardan-Tartaglia and
Lagrange formulas which are incorrect with the standard convention. Indeed, the difference with the previous works is the
two terms α1 and α2 of Eq. (30) which were missed in Cardan-Tartaglia and Lagrange formulas:
Theorem 3.3. Let the two cubic roots R1 and R2 :
r√
2b3 − 9cb + 27d
−3 √
∆o +
,
9
27
r√
2b3 − 9cb + 27d
−3 √
3
R2 :=
∆o −
.
9
27
R1 :=
3
The set of solutions of (P) is given in terms of R1 and R2 as follows:
√
1 − −3
b
m ∈ −1 ,
,
xm = mα1 R1 + m2 α2 R2 − ;
3
2
1+
√
2
−3
,
(30)
where α1 and α2 are defined as follows:
√
3
io
n√ h
p
4
−1 arg A1 3 δl − arg (−do R1 ) ,
exp
2
√
q
3
√
4
−1 arg A2 3 δl2 − arg d2o R2
.
α2 :=
exp
2
α1 :=
Here, ∆o , do , δl , A1 and A2 are given by definitions 3.1 and 3.2.
Proof:
Firstly, we construct the following matrix
−b
A := 0
−d
1
0
0
d−1 c
1 ,
0
which has (P) as its characteristic polynomial. Consequently, the roots of (P) are identical to the spectrum of A, which
can be calculated from corollaries 2.9 and 2.10. But, these corollaries require firstly to make explicitly o and rm relative
to matrix A. Indeed, lemma 2.7 gives Eq. (21) of o and the expression of l in terms of the components of M. So, we
identify the components of M with these of A so that M = A. In other words, we replace in expressions of section 2 the
components of M by their corresponding values in terms of b, c and d. Thus, we deduce that
√
√
b2 c3 − c4 − 3cd2 + b2 d2 − 2b3 cd ± −3c(d − bc) ∆o
o=
,
(31)
do
√
√
since Eq. (21) is a quadratic polynomial equation for o. In the sequel, we take the positive sign before −3c(d − bc) ∆o
in Eq. (31). Also, the applying of lemma 2.7 on matrix A implies that
√
√
√
1 + −3
c √ p
1 − −3
3
3
r m = lo − m 4
,
.
∆o δl ; m ∈ −1 ,
do
2
2
7
Here, lo is defined as follows:
√
√
−3
2
2
4
3 3
3 2
2 2
4
2
3
∆
c
b
c
−
2bd
−
c
2b
cd
+
b
c
−
b
d
−
5b
c
d
−
4bc
+
6bcd
+
5c
d
+
lo := d−1
.
o
o
3
Secondly, since we have the expressions of o and rm relative to matrix A, then by applying corollaries 2.9 and 2.10 on
matrix A, we deduce that the roots of (P) are given by
xm =
1
2
"
#
2
2
2
1 −crm + d (1 − o) rm + bdo (1 − o) + c b − 4c o
rm
=
−
b
−
o
(1 − o) [d + (bc − d)o] − 2bco2 − c (1 + o) rm
o
"
#
bd(bc − d) + cd b2 − 2c − d 2c2 − bd o−1 + do−1 do−1 − d + 2bc rm
1
rm
1
1+
−b−
.
2
o 2b2 c2 − 4bcd − 2c3 + d2 − 2 (c3 − bcd + d2 ) o−1 + d2 o−2 − 2c(d − bc)o−1 rm
o
1+
(32)
Then, we aim to simplify the expression of xm in Eq. (32), indeed Eq. (32) itself implies that
2
−crm
+ d (1 − o) rm + bdo (1 − o) + c b2 − 4c o2
=
(1 − o) [d + (bc − d)o] − 2bco2 − c (1 + o) rm
bd(bc − d) + cd b2 − 2c − d 2c2 − bd o−1 + do−1 do−1 − d + 2bc rm
.
2b2 c2 − 4bcd − 2c3 + d2 − 2 (c3 − bcd + d2 ) o−1 + d2 o−2 − 2c(d − bc)o−1 rm
(33)
So, the following property allows us to eliminate, in the denominator of the fractions of Eq. (33), the cubic root concealed
in rm :
C
A
C
MA + N C
A
=
⇒
=
=
; ∀A, B, C, D, M, N ∈ C∗ .
B
D
B
D
MB + N D
(34)
By applying property (34) on Eq. (33) for M = 2o−1 c(d − bc) and N = −c (1 + o), we just obtain a square root in the
denominator as follows:
p
√
p
√
m2 3 42 δl2
−3
b
xm = − +
(35)
4b3 c − 2db2 − 13bc2 + 15dc + c ∆o − m 3 4δl ×
3
12δl
3
√
√
∆o
8b5 c2 − 8b4 cd − 40b3 c3 + 2b3 d2 + 116b2 c2 d + 23bc4 − 99bcd2 − 21c3 d + 27d3 + −3 8b2 c2 − 10bcd + c3 + 3d2
.
18δl
Therefore, we rationalize Eq. (35) using the following identities:
8b5 c2 − 8b4 cd − 40b3 c3 + 2b3 d2 + 116b2 c2 d + 23bc4 − 99bcd2 − 21c3 d + 27d3
√
√
9A1 δl
+ −3 8b2 c2 − 10bcd + c3 + 3d2
,
∆o =
4do
√
√
2 −3
3A2 δl
4b3 c − 2db2 − 13bc2 + 15dc + 2c ∆o =
3
d2o
and consequently we deduce that
xm = −
p
√
√
3
4 mdo A1 3 δl − m2 4A2 3 δl2
b
− .
8
d2o
3
√
3
Thirdly, in order to simplify d2o in the denominator of Eq. (36), we prove that:
n√ h
io
p 3
p
p
−1 arg A1 3 δl − arg (−4do R1 ) (−4do R1 ) ,
A1 3 δl = (−4do R1 )3 ⇒A1 3 δl = exp
q 3 √
q
q
3
√
√
√
3
3
3
A2 3 δl2
42 d2o R2 ⇒A2 3 δl2 = exp
−1 arg A2 3 δl2 − arg
42 d2o R2
42 d2o R2 .
=
Then, by inserting Eq. (37) in Eq. (36), we deduce Eq. (30).
(36)
(37)
Acknowledgments
I would like to thank, in particularly, Dalia Ibrahim for her support during the difficult moment to work this paper. Also,
I wish to thank the reviewers of the first submission for their useful comments.
References
[1] B. van der Waerden, A History of Algebra, Springer, 1985.
[2] S. M. Lane, G. Birkhoff, Algebra, American Mathematical Society, 1999.
[3] W. W. Hager, R. N. Pederson, Perturbation in eigenvalues, Linear Algebra and its Applications 42 (1982) 39 – 55.
doi:http://dx.doi.org/10.1016/0024-3795(82)90137-9.
8
[4] I. C. F. Ipsen, B. Nadler, Refined perturbation bounds for eigenvalues of hermitian and non-hermitian matrices, SIAM
Journal on Matrix Analysis and Applications 31 (2009) 40 – 53.
[5] O. N. Kirillov, A. A. Mailybaev, A. P. Seyranian, Unfolding of eigenvalue surfaces near a diabolic point due to a
complex perturbation, Journal of Physics A: Mathematical and General 38 (24) (2005) 5531.
[6] J. Anderson, A secular equation for the eigenvalues of a diagonal matrix perturbation, Linear Algebra and its Applications 246 (1996) 49 – 70. doi:http://dx.doi.org/10.1016/0024-3795(94)00314-9.
[7] T. Tsang, H.-Y. Park, Sound velocity anisotropy in cubic crystals, Physics Letters A 99 (8) (1983) 377 – 380.
doi:http://dx.doi.org/10.1016/0375-9601(83)90297-9.
[8] I. Baydoun, É. Savin, R. Cottereau, D. Clouteau, J. Guilleminot, Kinetic modeling of multiple scattering of elastic
waves in heterogeneous anisotropic media, Wave Motion 51 (8) (2014) 1325–1348.
[9] T. Mensch, P. Rasolofosaon, Elastic-wave velocities in anisotropic media of arbitrary symmetry-generalization of
thomsen’s parameters, Geophysical Journal International 128 (1997) 43 – 64.
[10] N. Bourbaki, Éléments d’histoire des mathématiques, Springer, 2006.
[11] J. L. Lagrange, Réflexions sur la résolution algébrique des équations, In: Nouveaux Mémoires de l’Académie royale
des Sciences et Belles-Lettres de Berlin, Oeurves complètes, tome 3, 1770.
[12] T. Zhao, D. Wang, H. Hong, Solution formulas for cubic equations without or with constraints, Journal of Symbolic
Computation 46.
9
| 0math.AC
|
arXiv:1704.05304v2 [math.GR] 24 Aug 2017
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES
FOR FAMILIES OF SUBGROUPS
TIMM VON PUTTKAMER AND XIAOLEI WU
Abstract. Given a group G and a family of subgroups F , we consider its classifying space
EF G with respect to F . When F = VCyc is the family of virtually cyclic subgroups, JuanPineda and Leary conjectured that a group admits a finite model for this classifying space
if and only if it is virtually cyclic. By establishing a connection to conjugacy growth we
can show that this conjecture holds for linear groups. We investigate a similar question that
was asked by Lück–Reich–Rognes–Varisco for the family of cyclic subgroups. Finally, we
construct finitely generated groups that exhibit wild inner automorphims but which admit
a model for EVCyc (G) whose 0-skeleton is finite.
Introduction
Given a group G, a family F of subgroups of G is a set of subgroups of G which is
closed under conjugation and taking subgroups. We denote by EF (G) a G-CW-model for
the classifying space for the family F . The space EF (G) is characterized by the property
that the fixed point set EF (G)H is contractible for any H ∈ F and empty otherwise. Recall
that a G-CW-complex X is said to be finite if it has finitely many orbits of cells. Similarly,
X is said to be of finite type if it has finitely many orbits of cells of dimension n for any n.
We abbreviate EF (G) by EG for F = VCyc the family of virtually cyclic subgroups, EG
for F = F in the family of finite subgroups and EG for F the family consisting only of the
trivial subgroup. In [JL06, Conjecture 1], Juan-Pineda and Leary formulated the following
conjecture:
Conjecture A. [JL06, Juan-Pineda and Leary] Let G be a group admitting a finite model
for EG. Then G is virtually cyclic.
This conjecture has been proven for hyperbolic groups in the same paper and its validity
has been extended, for example to elementary amenable groups [KMN11], acylindrically
hyperbolic groups, 3-manifold groups, one-relator groups and CAT(0) cube groups [vW].
Note that a group G admits a model for EG with a finite 0-skeleton if and only if the
following holds
Date: August, 2017.
2010 Mathematics Subject Classification. 20B07, 20J05.
Key words and phrases. Classifying space, virtually cyclic subgroups, cyclic subgroups, conjugacy growth,
linear groups, relatively hyperbolic groups.
1
2
TIMM VON PUTTKAMER AND XIAOLEI WU
(BVC) G has a finite set of virtually cyclic subgroups {V1 , V2 , . . . , Vn } such that every
virtually cyclic subgroup of G is conjugate to a subgroup of some Vi .
Following Groves and Wilson, we shall call this property BVC and the finite set {V1 , V2 ,
. . . , Vn } a witness to BVC for G. So far almost all the proofs of Conjecture A boil down to
verifying whether the group has BVC. In fact, the following conjecture was made in [vW,
Conjecture B]
Conjecture B. Let G be a finitely presented group which has BVC. Then G is virtually
cyclic.
Note that there are many finitely generated torsion-free groups with BVC that are not
virtually cyclic, some groups with additional properties are constructed in Section 4. By
establishing a connection between conjugacy growth and BVC, our first theorem verifies
Conjecture B and hence Conjecture A for linear groups.
Theorem I. [2.11] A finitely generated linear group has BVC if and only if it is virtually
cyclic.
The theorem also puts very strong restrictions on linear representations of a group with
BVC.
Corollary. [2.12] Let ϕ : G → L be a surjective homomorphism where G is a finitely
generated group with BVC and L is linear. Then L is virtually cyclic.
Note that, strictly speaking, this does not directly follow from Theorem I if the linear
group G is not virtually torsion-free, since the BVC property is not inherited by quotients
in general (see Corollary 4.22). This was one reason for us to introduce the weaker notion
bVCyc, or more generally bF for a family of subgroups F , that is inherited by quotients,
see Section 1.
When F = Cyc, the family of cyclic subgroups, a similar question was asked by Lück–
Reich–Rognes–Varisco [Lüc+, Question 4.9]:
Question A. Does a group G have a model of finite type for ECycG if and only if G is finite,
cyclic or infinite dihedral?
The case of the infinite dihedral group was missing in the original formulation of the
question in [Lüc+] but has to be included as we show in Lemma 3.9.
Similar to the case of the family of virtually cyclic subgroups, a group G admits a model
for ECycG with a finite 0-skeleton if and only if the following holds
(bCyc) G has a finite set of cyclic subgroups {C1 , C2 , . . . , Cn } such that every cyclic subgroup of G is conjugate to a subgroup of some Ci .
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS
3
It turns out that once we answer Question A positively for virtually cyclic groups, most
of the proofs of Conjecture A also apply to Question A. In fact, we prove in Section 3 the
following
Theorem II. For the following classes of groups, Question A has a positive answer.
(a) elementary amenable groups
(b) one-relator groups
(c) acylindrically hyperbolic groups
(d) 3-manifold groups
(e) CAT(0) cube groups and
(f) linear groups
In analogy to Conjecture B, we can phrase the following
Question B. Suppose G is a finitely presented group which has bCyc. Is G necessarily
finite, cyclic, or infinite dihedral?
Once more, we obtain a proof of Theorem II essentially by answering Question B with
the exception of the class of elementary amenable groups. Here, we additionally rely on
stronger finiteness conditions that are imposed on the group by having a model of finite
type for ECyc (G).
In the last section of this paper we construct groups which can serve as counterexamples
for various reasonable questions regarding the BVC resp. bCyc property. For example, in
Lemma 2.4 we show that a finitely generated group with BVC whose infinite cyclic subgroups are quasi-isometrically embedded can have at most linear conjugacy growth. Part
(c) of the following theorem shows that one cannot dispense with the assumption on quasiisometrically embedded infinite cyclic subgroups. The main tools for the constructions
are HNN extensions and small cancellation theory over relatively hyperbolic groups as
developed in [Osi10] and [HO13].
Theorem III.
(a) There exists a finitely generated torsion-free group G with a finite
index subgroup H such that H has bCyc but G does not. [4.6]
(b) There exists a finitely generated torsion-free group G = H ⋊Z such that G has bCyc
but H does not. [4.18]
(c) There exists a finitely generated torsion-free group G with bCyc that has exponential conjugacy growth. [4.20]
Note that Leary and Nucinkis [LN03] give examples of groups H ≤ G where H is a
finite index subgroup of G such that H has a finite model for EH (= EH here) but G has
infinitely many conjugacy classes of finite subgroups.
Results of this paper will also appear as part of the first author’s thesis.
Acknowledgements. The first author was supported by an IMPRS scholarship of the
Max Planck Society. The second author would like to thank the Max Planck Institute for
4
TIMM VON PUTTKAMER AND XIAOLEI WU
Mathematics at Bonn for its support. We also would like to thank Denis Osin for helpful
discussions regarding the proof of [HO13, Theorem 7.2].
1. Groups Admitting a Finite Model for EG and Property bF
In this section we first review some properties of and results on groups admitting a finite
model for EG. Most of this material is taken directly from [vW, Section 1]. Afterwards,
we introduce the notion bF for a family of subgroups F , which, for F = VCyc is slightly
weaker but more flexible than the BVC notion.
We summarize the properties of groups admitting a finite model for EG as follows
Proposition 1.1. Let G be a group admitting a finite model for EG, then
(a) G has BVC.
(b) G admits a finite model for EG.
(c) For every finite subgroup of H ⊂ G, the Weyl group WG H is finitely presented and
of type FP∞ . Here WG H = NG (H)/H, where NG (H) is the normalizer of H in G.
(d) G admits a model of finite type for EG. In particular, G is finitely presented.
Remark 1.2. If one replaces finite by finite type in the assumptions of the above proposition, then the conclusions still hold if one also replaces finite by finite type in (b).
Lemma 1.3. [vW, Lemma 1.3] Let G be a group. There is a model for EG with finite
0-skeleton if and only if G has BVC.
The following structure theorem about virtually cyclic groups is well known, see for
example [JL06, Proposition 4] for a proof.
Lemma 1.4. Let G be a virtually cyclic group. Then G contains a unique maximal normal
finite subgroup F such that one of the following holds
(a) the finite case, G = F;
(b) the orientable case, G/F is the infinite cyclic group;
(c) the nonorientable case, G/F is the infinite dihedral group.
Note that the above lemma implies that a torsion-free virtually cyclic group is either
trivial or infinite cyclic. Thus we have the following
Corollary 1.5. Let G be a torsion-free group, then G has BVC if and only if there exist
elements g1 , g2 , . . . gn in G such that every element in G is conjugate to a power of some gi .
Lemma 1.6. [vW, Lemma 1.6] Let V be a virtually cyclic group and let g, h ∈ V be two
elements of infinite order, then there exist p, q ∈ Z such that g p = hq . Furthermore, there
exists v0 ∈ V such that for any v ∈ V of infinite order there exist nonzero p0 , p such that
v0p0 = v p with pp0 ∈ Z.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS
5
Lemma 1.7. [vW, Lemma 1.7] If a group G has BVC, then G has finitely many conjugacy
classes of finite subgroups. In particular, the order of finite subgroups in G is bounded.
In a group G, we call an element g primitive if it cannot be written as a proper power.
Note that a primitive element is necessarily of infinite order. Corollary 1.5 implies the
following
Lemma 1.8. Let G be a torsion-free group. If G has infinitely many conjugacy classes of
primitive elements, then G does not have BVC.
Lemma 1.9. [KMN11, Lemma 5.6] If a group G has BVC, then any finite index subgroup
also has BVC.
Definition 1.10. Let F be a family of subgroups. For a natural number n ≥ 1, we say that
a group G has property nF if there are H1 , . . . , Hn ∈ F such that any cyclic subgroup of G
is contained in a conjugate of Hi for some i. We say that G has bF if G has nF for some
n ∈ N.
We are mostly interested in bVCyc as both bCyc and BVC imply bVCyc, which leads
to unified proofs when we deal with finiteness properties of ECyc (G) and EG. Note that for
torsion-free groups all three notions agree. However, as the following two examples show,
they generally do not coincide.
Example 1.11. Consider the group Z × Z/2. It is virtually cyclic, thus has bVCyc as
well as BVC. But it does not have bCyc. In fact, for any n ≥ 1, let Cn be the subgroup
generated by (2n , 1). Then for any m , n, Cm cannot be conjugate to a subgroup of Cn .
See Proposition 3.8 for more information.
Example 1.12. Taking G = (Z/2)∞ and applying [Osi10, Thereom 1.1], one embeds G
into a finitely generated group with only three conjugacy classes. In particular, this group
has bCyc and bVCyc. But it cannot have BVC, since the orders of finite subgroups are not
bounded (Lemma 1.7).
The following is immediate:
Lemma 1.13. Suppose the family F is closed under quotients. If π : G → Q is an epimorphism and G has bF , then Q has bF .
Lemma 1.14. Let K ≤ G be a finite index subgroup and suppose G has bF . Then K also
has bF .
Proof.
Let m = [G : K] and let Kgi for 1 ≤ i ≤ m be the right cosets. Furthermore
let {H1 , . . . , Hn } be a witness of bF for G. Then consider the following finite collection of
subgroups of K which lie in F :
{gi H j g−1
i ∩ K | 1 ≤ j ≤ n, 1 ≤ i ≤ m}
6
TIMM VON PUTTKAMER AND XIAOLEI WU
We claim that these form a witness to bF for K. Let C ≤ K be some cyclic subgroup, then
there exists some g ∈ G such that C ≤ gH j g−1 for some j. Write g = kgi for some i and
some k ∈ K. Then k−1 Ck ≤ gi H j g−1
i ∩ K.
Theorem 1.15. Let G be a finitely generated group in one of the following classes
(a) virtually solvable groups,
(b) one-relator groups,
(c) acylindrically hyperbolic groups,
(d) 3-manifold groups,
(e) CAT(0) cube groups.
If G has bVCyc, then G is virtually cyclic.
Proof.
(a) The proof works the same as [GW13] without much change. In fact [GW13,
Lemma 2.1] is available for bVCyc by Lemma 1.14. [GW13, Lemma 2.2] is still
available via almost the same proof. Moreover, [GW13, Lemma 2.3] can be easily
deduced from Lemma 1.13. [GW13, Lemma 2.4] can be directly applied to the
bVCyc case, since BVC is the same as bVCyc for a torsion-free group.
(b) If G contains torsion, the group is hyperbolic by Newman’s Spelling Theorem
[New68] and the claim follows from (c) below. Otherwise G is torsion-free and
the result follows from [vW, Theorem 2.12]. One can also check that the proof of
[vW, Theorem 2.12] indeed works for bVCyc.
(c) The proof of [vW, Proposition 3.2] showed that G does not have bVCyc. In fact, it
was shown that there are infinitely many primitive conjugacy classes of elements in
G such that any two of them cannot be conjugated into a common virtually cyclic
subgroup.
(d) The proof is almost the same as [vW, Proposition 3.6]. [vW, Corollary 3.4] now is
replaced by Lemma 1.13 which is true for all groups.
(e) The proof given in [vW, Section 4] can be carried over, observing that [vW, Lemma
4.6] still holds with BVC replaced by bVCyc.
Corollary 1.16. If G has bVCyc and surjects onto a finitely generated group Q that lies
in one of the classes described in Theorem 1.15, then Q is virtually cyclic. In particular,
the abelianization H1 (G, Z) is finitely generated of rank at most one.
2. Conjugacy Growth of Groups and BVC for Linear Groups
In this section, we establish a connection between BVC and conjugacy growth. As
an application, we show that a finitely generated linear group has BVC if and only if it
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS
7
is virtually cyclic. The proof is based on Breuillard–Cornulier–Lubotzky–Meiri’s results
[Bre+13] on the conjugacy growth of linear groups.
2.1. Growth of Groups and BVC. Let G be a group with a finite symmetric generating
set S . We define the word metric on G as follows
dS (g, h) = min{n | g−1 h = s1 s2 · · · sn , si ∈ S }.
This can also be seen as the metric on the Cayley graph Cay(G, S ) of G with respect to S ,
where we assign edges unit length. For any g ∈ G, we define the word length of g via
|g|S = dS (e, g),
where e is the identity element of G. Now given n > 0, we denote by Bn (G, S ) the ball
of radius n with respect to the word metric. The word growth function is the function
that maps n > 0 to |Bn(G, S )|, i.e. the number of elements of distance at most n from the
identity.
Similarly, for an element g ∈ G, we define its length up to conjugacy by
n
o
|g|cS = min hgh−1 S | h ∈ G .
By definition, this number only depends on the conjugacy class [g] of g.
Now given n > 0, we can consider the ball of radius n under the length up to conjugacy
in the set of conjugacy classes of G,
Bcn (G, S ) = {[g] | g ∈ G, |g|cS ≤ n}
The conjugacy growth function gc (n) assigns to n > 0 the number |Bcn (G, S )|, i.e. the
number of conjugacy classes which intersect Bn (G, S ). For f, g : N → N, we write f g
if there is some constant C ∈ N such that f (n) ≤ g(Cn) for all n ∈ N. If f g and g f ,
we say that f and g are equivalent and write f ∼ g. Under this equivalence relation, the
conjugacy growth function is independent of the choice of generating set. We say that a
group has linear (resp. at most linear) conjugacy growth if gc (n) ∼ n (resp. gc (n) n), and
we say that a group has exponential conjugacy growth if gc (n) ∼ 2n or equivalently if
log |Bcn (G, S )|
> 0.
n→∞
n
For more information about conjugacy growth, we refer to [GS10] and [HO13]. To build
the connection between conjugacy growth and BVC, we need the notion of distortion, see
lim inf
[CF06, Section 2] for more information.
Definition 2.1. [BH99, III.Γ.3.13] Let G be a finitely generated subgroup, and let S be a
symmetric finite generating set as above. The algebraic translation number of an element
h ∈ G, denoted by ||h||S , is the limit
|hn |S
n→∞ n
An element h ∈ G is undistorted if the translation number ||h||S is positive.
||h||S := lim
8
TIMM VON PUTTKAMER AND XIAOLEI WU
Note that an element h of infinite order is undistorted in G if and only if the infinite
cyclic subgroup hhi is quasi-isometrically embedded in G.
Remark 2.2. The length ||h||S depends only on the conjugacy class of h and ||hm || = |m|||h||S
for every m ∈ Z [BH99, III.Γ.3.14]. The property of being undistorted is independent of
the generating set S [CF06, Remark 2.6].
Lemma 2.3. Let G be a group with a finite generating set S . Then
|h|S ≥ ||h||S
for all h ∈ G.
Proof. For any h ∈ G we have
dS (e, hn ) ≤ dS (e, h) + dS (h, h2 ) + · · · + dS (hn−1 , hn ).
Since dS (hi , hi+1 ) = dS (e, h) for any i, it follows
dS (e, hn )
|hn |S
=
≤ dS (e, h) = |h|S .
n
n
Letting n go to infinity, we have ||h||S ≤ |h|S .
Lemma 2.4. Let G be a finitely generated group and V be a virtually cyclic subgroup of
G. If V contains an infinite order element that is undistorted, then the conjugacy growth of
V in G is linear, i.e. the limit
|Bc (G, S ) ∩ V|
lim n
n→∞
n
exists and is greater than 0. Therefore, if G has bVCyc and every infinite order element of
it is undistorted, then it has at most linear conjugacy growth.
Proof. Note first that a virtually cyclic group has only finitely many conjugacy classes
of finite subgroups. So to count the conjugacy growth of V in G, we only need to count
the elements of infinite order in G. Let v0 be the infinite order element in V determined by
Lemma 1.6 and let v be the element that is undistorted. Then there are nonzero integers m0
m0
m
0
m
and m such that vm
0 = v and m ∈ Z. In particular, we have ||v0 ||S = | m0 |||v||S > 0. In fact,
this also says that any infinite order element in V is undistorted using Lemma 1.6.
Now by Remark 2.2 and Lemma 2.3, we have
|vn0 |S ≥ ||vn0 ||S = n||v0||S .
Since the algebraic translation number is invariant under conjugation, we have |gvn0 g−1 |S ≥
n||gv0g−1 ||S = n||v0 ||S , for any g ∈ G. Thus
n||v0 ||S ≤ |vn0 |cS ≤ n|v0 |S .
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS
9
This says that the conjugacy growth of the cyclic group generated by v0 in G is linear.
As for the whole group V, note that V satisfies the following short exact sequence by
Lemma 1.4
1→F →V →C→1
where C is isomorphic to Z or the infinite dihedral group Z ⋊ Z/2. In the case C Z ⋊ Z/2,
the image of all the infinite order elements under the quotient map π : V → C must lie in
Z ⋊ {0} ≤ Z ⋊ Z/2. Hence |π(v)| makes sense in either case. And since |π(v0 )| = 1, one sees
that for any infinite order element v, we have
||v||S = |π(v)|||v0||S .
Thus given n, there are at most
|F|·n
||v0 ||S
many conjugacy classes of elements in Bcn (G, S ) ∩ V.
Therefore the conjugacy growth of V in G is linear.
There is a notion of a semihyperbolic group which generalizes hyperbolic as well as
CAT(0) groups, see [BH99, Section III.Γ.4] for more details. By [BH99, III.Γ.4.19], every
element of infinite order in a semihyperbolic group is undistorted, hence we have
Corollary 2.5. Let G be a semihyperbolic group with bVCyc, then it has at most linear
conjugacy growth. In particular, if it is hyperbolic, then it is virtually cyclic.
Proof. We only need to show that a hyperbolic group has linear conjugacy growth if and
only if it is virtually cyclic. This can be found for example in [HO13, Theorem 1.1], see
also [CK02].
Remark 2.6. We believe that a semihyperbolic group has at most linear conjugacy growth
if and only if it is virtually cyclic. Unfortunately, we do not know how to show this.
Later in Section 2.3, we will see that if a group acts properly on a CAT(0) space via
semi-simple isometries, then any infinite order element is also undistorted.
2.2. Linear Groups and BVC. Recall that a linear group is a subgroup of the general
linear group GLd (K), where K is a field. In [GS10] it was conjectured that a non-virtually
solvable finitely generated linear group has exponential conjugacy growth, which was later
proven in [Bre+13]. In order to show Theorem I we rely on the following stronger result:
Theorem 2.7. [Bre+13, Theorem 1.2] For every d, there exists a constant c(d) > 0 such
that if K is a field and S is a finite symmetric subset of GLd (K) generating a non-virtually
solvable subgroup, then
1
lim inf log χS (n) ≥ c(d),
n→∞ n
where χS (n) is the number of elements in K[X] appearing as characteristic polynomials of
elements of Bn (G, S ).
10
TIMM VON PUTTKAMER AND XIAOLEI WU
In general, an element of infinite order in a linear group may be distorted, thus we
cannot apply Lemma 2.4 directly.
Definition 2.8. Let K be a field. An absolute value on K is a function
| · | : K → R+ = {r ≥ 0 | r ∈ R}
that satisfies the following conditions:
(a) |x| = 0 if and only if x = 0.
(b) |xy| = |x||y| for all x, y ∈ K.
(c) |x + y| ≤ |x| + |y| for all x, y ∈ K.
We learned the following two lemmas from Yves de Cornulier’s answer on Mathoverflow*
Lemma 2.9. Let G be finitely generated subgroup of GLd (K) where K is some field. Let
g ∈ G be an element such that at least one of its eigenvalues is not a root of unity. Then g
is undistorted.
Proof. First of all we can assume that K is a finitely generated field. Let λ be an eigenvalue of g which is not a root of unity. By [Tit72, Lemma 4.1], up to replacing g by its
inverse, there exists an absolute value | · | on a field extension K′ of K such that |λ| > 1. Let
|| · || be a submultiplicative matrix norm on the vector space of n × n matrices over K′ . For
example, we can take
||A|| = max Σnj=1 |ai j |
i
where A = (ai j ). Then for A, B two n × n matrices over K′ we have ||AB|| ≤ ||A|| · ||B|| and
|µ| ≤ ||A|| for any eigenvalue µ of A.
Now, let s1 , . . . , sm be the elements of some finite symmetric generating set S for G and
let s = max1≤i≤m ||si ||. If g can be written as a word of length l in the generators S , then
||g|| ≤ sl . Hence 1 < |λ| ≤ sl , so l ≥ log s |λ|. Since gk has eigenvalue λk , we have
|gk |S ≥ logs |λ|k ,
thus
|gk |S
≥ log s |λ|.
k→∞ k
||g||S = lim
Therefore g is undistorted.
Lemma 2.10. Let G be finitely generated subgroup of GLd (K) where K is a field of positive
characteristic. Then any element of infinite order in G is undistorted.
*see http://mathoverflow.net/questions/143305/distortion-of-cyclic-subgroups-of-linear-groups
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 11
Proof. Let g ∈ G, by Lemma 2.9 we can assume that all eigenvalues of g are roots
of unity. But then g must have finite order. In fact, some power h of g will only have
eigenvalues equal to one, i.e. h is unipotent in GLd (K). So (h − I)m = 0 for some m.
Choose n with pn ≥ m, where p is the characteristic of K. Then
n
n
0 = (h − I) p = h p − I,
thus h has finite order, so g has finite order.
Theorem 2.11. Let G ≤ GLd (K) be a finitely generated group where K is a field. If G has
BVC or more generally just bVCyc, then G is virtually cyclic.
Proof. Since a virtually solvable group has bVCyc if and only if it is virtually cyclic by
Theorem 1.15, we can assume that G is not virtually solvable. Thus Theorem 2.7 applies
and it follows that the characteristic polynomial growth of G is exponential.
If K has characteristic zero, then by Selberg’s Lemma, G is virtually torsion-free. By
Lemma 1.9, up to replacing G by a finite index subgroup, we can assume that G is torsionfree. Let g1 , g2 , . . . , gm , g′1 , g′2 , . . . , g′k be generators of the infinite cyclic witnesses for
bVCyc, where the gi have at least one eigenvalue which is not a root of unity and all
eigenvalues of the g′j are roots of unity. Note that there are only finitely many characteristic polynomials corresponding to the elements in the cyclic subgroup generated by the
elements g′j . Thus these elements essentially do not contribute to the characteristic polynomial growth function χΣ (n). By Lemma 2.9, the elements gi are undistorted. Thus by
the proof of Lemma 2.4 the number of characteristic polynomials with word length ≤ n in
G is linear in n, contradicting Theorem 2.7.
If K has positive characteristic, by Lemma 2.10, it follows that any infinite order element
in G is undistorted. Then by Lemma 2.4, G has at most linear conjugacy growth, which
also contradicts Theorem 2.7.
Since the property bVCyc passes to quotients by Lemma 1.13, we can immediately
conclude that representations of finitely generated groups having BVC are rather trivial:
Corollary 2.12. Let ϕ : G → L be a surjective homomorphism where G is a finitely generated group with BVC and L is linear. Then L is virtually cyclic.
2.3. Groups Acting on CAT(0) Spaces and BVC. In the following we provide an alternative proof of Theorem I for finitely generated linear groups over a field of positive
characteristic using metric spaces of non-positive curvature.
Definition 2.13. Let G be a group acting by isometries on a metric space X. The action
is called proper if for each x ∈ X there exists a number r > 0 such that the set {g ∈ G |
B(x, r) ∩ gB(x, r)} is finite, where B(x, r) is the ball of radius r centered at x.
12
TIMM VON PUTTKAMER AND XIAOLEI WU
Definitions 2.14. [BH99, II.6.1] Let X be a metric space and let g be an isometry of X.
The displacement function of g is the function dg : X → R+ = {r ≥ 0 | r ∈ R} defined
by dg (x) = d(gx, x). The translation length of g is the number |g| := inf{dg (x) | x ∈ X}.
The set of points where dg attains this infimum will be denoted by Min(g). More generally,
T
if G is a group acting by isometries on X, then Min(G) := g∈G Min(g). An isometry g
is called semi-simple if Min(g) is non-empty. An action of a group by isometries of X is
called semi-simple if all of its elements are semi-simple.
It is clear that the translation length is invariant under conjugation, i.e. |hgh−1| = |g| for
any g, h ∈ G. Moreover, if G acts properly on a CAT(0) space via semi-simple isometries,
we have that |gn | = |n| · |g| for any n ∈ Z, e.g. by the Flat Torus Theorem [BH99, II.7.1].
It turns out that the translation length can also be defined by the following limit for g a
semi-simple isometry
d(x, gn x)
,
|g| = lim
n→∞
n
where x is an arbitrary point of the CAT(0) space X [BH99, II.6.6]. The following example
shows that this is not true if X is not CAT(0).
Example 2.15. Take S 1 to be the standard circle of radius 1. Let the cyclic group ht | tn =
1i act on S 1 via rotation by the angle 2π
n for n ≥ 2. The action is semi-simple and |t| =
1
n
But limn→∞ n d(x, t x) = 0 for any x ∈ S 1 .
2π
n .
Proposition 2.16. Let G ≤ GLd (K) be a finitely generated subgroup where K is a field of
positive characteristic. Suppose that there is a bound on the orders of finite subgroups in
G. Then G acts on a CAT(0) space properly via semi-simple isometries.
Proof. This is essentially due to Degrijse and Petrosyan [DP15, Section 5.4]. Note that
they obtained an action of G on a CAT(0) space X, where X is a product of Euclidean
buildings. Moreover, X is separable and G acts discretely on X with countable locally
finite stabilizers via semi-simple isometries [BH99, II.6.6 (2)]. By our assumption on the
orders of finite subgroups on G, it follows that all point stabilizers are finite. Hence (by for
example [DP15, 2.2(2)]) the action is proper.
Lemma 2.17. [BH99, I.8.18] Let (X, d) be a metric space. Let G be a group with a finite
generating set S and associated word metric dS . If G acts by isometries on X, then for any
choice of basepoint x0 ∈ X there exists a constant C > 0 such that
dX (gx0 , hx0 ) ≤ CdS (g, h)
for all g, h ∈ G.
Lemma 2.18. Let G be a finitely generated group acting on a CAT(0) space X properly
via semi-simple isometries. Then every infinite order element in G is undistorted.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 13
Proof. Let S be a finite symmetric generating set for G and let g ∈ G be an element of infinite order. Since the action is proper, the translation length |g| is positive. By Lemma 2.17,
for any fixed point x0 in X, there exists C > 0 such that
|gn |S = dS (gn , 1) ≥ CdX (gn x0 , x0 ) ≥ C|gn | = nC|g|
This implies that g is undistorted.
Now applying Lemma 2.4, we have
Corollary 2.19. Let G be a finitely generated group acting on a CAT(0) space X properly
via semi-simple isometries. Then if G has bVCyc, the conjugacy growth function of G is
at most linear.
This also leads to a different proof for Theorem 2.11 in the positive characteristic case
under the slightly stronger assumption of having BVC instead of bVCyc.
Theorem 2.20. Let G ≤ GLd (K) be a finitely generated subgroup where K is a field of
positive characteristic. Then G has BVC if and only if it is virtually cyclic.
Proof. By Theorem 1.15, we can assume that G is not virtually solvable. By Lemma 1.7,
if G has BVC, then it has only finitely many conjugacy classes of finite subgroups. Hence
by Proposition 2.16, G acts properly on a CAT(0) space via semi-simple isometries. Thus,
G has linear conjugacy growth, which contradicts Theorem 2.7.
3. The Classifying Space for the Family of Cyclic Subgroups
In this section, we study the classifying space for the family of cyclic subgroups, denoted by ECyc (G) for a group G. The results on its finiteness properties are largely analogous to those obtained for the classifying space for the family of virtually cyclic subgroups.
We will first study for which groups G the G-CW-complex ECyc (G) can be of finite type.
As it turns out most proofs given in the literature can be generalized slightly such that their
conclusions apply to both ECyc (G) as well as EG. Second, we will deal with questions
concerning the finite-dimensionality of ECyc (G) and relate it to previously known results
on the finite-dimensionality of EG.
Proposition 3.1. If a group G has a model for ECycG of finite type, then it has a model for
EG of finite type. In particular, G is finitely presented.
Proof.
Note that there are models of finite type for EC where C is a cyclic group.
Then the claim follows from the transitivity principle [LW12, Proposition 5.1].
Similar to [vW, Lemma 1.3], one obtains the following using the construction in [Lüc89,
Proposition 2.3]
Lemma 3.2. A group G admits a finite 0-skeleton for ECycG if and only if G has bCyc.
14
TIMM VON PUTTKAMER AND XIAOLEI WU
Example 3.3. Let D∞ = Z ⋊ Z/2 = ht, s | s2 = 1, sts = t−1 i be the infinite dihedral group.
Then hti, hsi and htsi are witnesses to bCyc for D∞ since tst−1 = t1−2n s. A straightforward
calculation also shows that there cannot be fewer witnesses.
In [GW13] Lemma 2.2 played an important role for the proof of Conjecture A for solvable groups. The statements of Lemma 2.2 hold with BVC replaced by bCyc. For the
convenience of the reader, we present part (a) of this lemma in a more general context.
Lemma 3.4. Let G be a group satisfying bF where F is a family of Noetherian subgroups,
i.e. any subgroup of an element H ∈ F is finitely generated. Then G satisfies the ascending
chain condition on normal subgroups.
S S
Proof.
The group G can be written as G = ni=1 g∈G Hig where {Hi | 1 ≤ i ≤ n}
is a witness to bF . But then any normal subgroup N of G can be likewise expressed as
S S
N = ni=1 g∈G (N ∩ Hi )g .
Let (N j ) be an ascending chain of normal subgroups of G. For any i, the chain (N j ∩ Hi ) j
has to stabilize since Hi was Noetherian, i.e. there exists ji such that N j ∩ Hi = N j+1 ∩ Hi
for all j ≥ ji . Then the original chain stabilizes at jmax = max1≤i≤n ji .
Remark 3.5. Observe that bCyc fails to pass to finite index overgroups, a counterexample
can be given by Z ≤ Z × Z/2.
Note that there is no assumption about absence of torsion in the following (cf. Lemma 1.8)
Observation 3.6. If G has bCyc, then G has only finitely many primitive conjugacy classes.
Lemma 3.7. Let F be a finite group and suppose that V = F ⋊ϕ Z is a group with bCyc.
Then F = 1.
Proof.
Let d be the order of ϕ. Then F × dZ is a subgroup of index d in F ⋊ϕ Z.
Hence by Lemma 1.14, we can assume V = F × Z.
Now assume that F is nontrivial. Let c be an element of maximal order in F and let p
be a prime that divides its order. Then for any n ≥ 1, gn = (c, pn ) ∈ F × Z is primitive in
F × Z. If fact, if (x, k)m = (xm , mk) = (c, pn ), for some m > 1, then p divides m. On the
other hand since xm = c, c lies in the subgroup generated by x. But since c has maximal
order, we have x and c generate the same cyclic subgroup in F. But since p divides the
order of c and m, this cannot happen. When n , m, gn is not conjugate to gm since the
second coordinate differs. Thus by Observation 3.6, the claim follows.
Proposition 3.8. A virtually cyclic group V has bCyc if and only if V is finite, infinite cyclic
or infinite dihedral.
Proof. By Lemma 3.7 and Lemma 1.4, the only case left to consider is if V is nonorientable, i.e. there is an exact sequence
1 → F → V → D∞ → 1
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 15
with F finite. But then V has a finite index subgroup isomorphic to F ⋊ Z, hence F = 1
and the claim follows from Example 3.3.
Lemma 3.9. There exists a model of finite type for ECyc D∞ and any model for ECyc D∞ has
to be infinite-dimensional.
Proof.
Let D∞ = Z ⋊ Z/2 = hs, t | s2 = 1, sts = t−1 i. We claim that the join
E = Z ∗ EZ/2, given an appropriate action, is a model for ECyc D∞ . We write [x, y, q] for
an element in E, where x ∈ Z, y ∈ EZ/2 and q ∈ [0, 1]. Note that [x, y, 0] = [x, y′ , 0] and
[x, y, 1] = [x′ , y, 1] for all x, x′ ∈ Z and y, y′ ∈ EZ/2. We then define the action as follows:
t · [x, y, q] = [x + 2, y, q]
s · [x, y, q] = [−x, s · y, q]
Then one observes that the stabilizer of [x, y, q] with 0 < q < 1 is trivial. The stabilizer of
[x, y, 0] is equal to ht x si and the stabilizer of [x, y, 1] equals hti. One furthermore checks
that for n , 0
n
E ht i = EZ/2 ≃ ∗
and for n arbitrary
E ht
n
si
= Zht
n
si
= {n}.
Since E itself is contractible as well, it follows that E is a model for ECyc (G) of finite type.
The claim about the infinite-dimensionality of any model for ECyc (G) follows from
Lemma 3.12 below by noting that hti is a normal maximal cyclic subgroup of D∞ . Alternatively observe that E/D∞ is homotopy equivalent to the suspension of RP∞ .
Corollary 3.10. Let G be a virtually cyclic group, then it has a model for ECyc (G) of finite
type if and only if it is finite, infinite cyclic or infinite dihedral.
Proof. By Proposition 3.8, we only need to prove that there is a model of finite type
for ECyc (G) if G is finite, infinite cyclic or infinite dihedral. If G is a finite group, then
the standard bar-construction provides such a model. If G is infinite cyclic, we can take
ECyc (G) to be a point. In case G is infinite dihedral Lemma 3.9 provides a model of finite
type.
Proof of Theorem II: By Theorem 1.15 and Corollary 3.10 and the fact that both bCyc
and bVCyc implies BVC, we only need to show if an elementary amenable group has a
finite type model for ECyc (G), then it is virtually cyclic. Since Kropholler’s class of groups
LH F contains the class of elementary amenable groups, [Kro93, Theorem B] implies that
an elementary amenable group of type FP∞ has a bound on the orders of its finite subgroups and finite Hirsch length. Hence by Proposition 3.1, G cannot have a model of
finite type for ECyc (G) if it has infinite Hirsch length. When G has finite Hirsch length, by
[HL92], it is locally finite by virtually solvable, i.e. it has a virtually solvable quotient with
16
TIMM VON PUTTKAMER AND XIAOLEI WU
locally finite kernel. Thus by Kropholler’s result, G is finite by virtually solvable. Now
by Lemma 1.13 and Theorem 1.15 (a), G is finite by virtually cyclic. Hence G is virtually
cyclic.
Even though EG is conjecturally never of finite type except in trivial cases, finitedimensional models for EG abound for reasonable classes of groups. For example, finitedimensional models for EG have been constructed for hyperbolic groups [JL06], CAT(0)
groups [Lüc09], elementary amenable groups of finite Hirsch length [FN13; DP13] and
also certain linear groups [DKP15] such as discrete subgroups of GLd (R). The question
arises in which cases we can expect a finite-dimensional model for ECyc (G). We will first
settle the question in the case that G is virtually cyclic and then draw a couple of immediate
conclusions in the general case. Let us first record the following simple observation.
Observation 3.11. Let H be a subgroup of a group G and let X be a model for ECyc (G),
then resGH X is a model for ECyc (H).
Recall that the classifying space EF of a non-trivial finite group F cannot be finitedimensional [Bro82, VIII.2.5].
Lemma 3.12. Suppose H ≤ G is a maximal cyclic subgroup and assume that [NG (H) : H]
is finite but not equal to one, where NG (H) is the normalizer of H in G. Then any model
for ECyc (G) has to be infinite-dimensional.
Proof.
Let X be a model for ECyc (G). Since H is cyclic, the CW-complex X H is
contractible. Observe that all isotropy groups of X H are equal to H since H was maximal
cyclic. This implies that the Weyl group NG (H)/H of H acts freely on X H . Since NG (H)/H
is non-trivial finite, X H has to be infinite-dimensional.
Proposition 3.13. Let G be a finite group with a finite-dimensional model for ECyc (G).
Then G is already cyclic.
Proof.
By Observation 3.11 we only need to consider minimal non-cyclic groups,
i.e. finite groups such that every proper subgroup is cyclic. Thus by [MM03] G is solvable. Moreover, since the proper subgroup [G, G] has to be cyclic, G has derived length at
most 2. Let H be a maximal cyclic subgroup containing [G, G], then NG (H) = G. Then by
Lemma 3.12, H = G and hence G is cyclic.
Proposition 3.14. Let V be virtually cyclic. Then ECyc V is finite-dimensional if and only
if V is cyclic.
Proof. By Proposition 3.13 we only need to prove the claim if V is infinite. Suppose
V is orientable, i.e. V F ⋊ϕ Z for some finite group F and some ϕ ∈ Aut(F) and assume
that ECyc V is finite-dimensional. If n denotes the order of ϕ then F × Z F ⋊ϕ nZ and by
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 17
Observation 3.11 also F × Z has a finite-dimensional classifying space. But Z ≤ F × Z is
a normal maximal cyclic subgroup, thus F = 1 by Lemma 3.12.
Now, suppose V was non-orientable having a finite-dimensional model for ECyc V. Let
F be the maximal normal finite subgroup of V, then F ⋊ Z is a subgroup of V. By the
above, it follows that F = 1, so V is infinite dihedral. But this is impossible by Lemma 3.9.
Corollary 3.15. A virtually cyclic group V has a finite or finite-dimensional model for
ECyc V if and only if V is cyclic.
From Proposition 3.14 we immediately obtain the following
Observation 3.16. If G is a group having a finite-dimensional model for ECyc (G), then
there is a finite-dimensional model for EG. Conversely, suppose that G is a group having
a finite-dimensional model for EG. Then G admits a finite-dimensional model for ECyc (G)
if and only if Cyc(G) = VCyc(G).
Obviously the condition Cyc(G) = VCyc(G) holds whenever the group G is torsionfree. However, this is not necessary, even for virtually free groups. For example, groups of
the form G = ∗ni=1 Z/ni Z where ni ≥ 0 admit a finite-dimensional model for ECyc (G) if and
only if ni , 2 for all i or G Z/2 by the Kurosh subgroup theorem.
Lemma 3.17. Let A be an abelian group with ECyc (A) finite-dimensional. Then A is cyclic,
torsion-free or locally finite cyclic.
Proof. By Proposition 3.13 we can assume in the following that A is infinite. If A is
finitely generated, we can write A Zn × F with F finite abelian and n ≥ 1. In particular, A
contains Z × F, so F = 1, i.e. A is torsion-free. More generally, if A contains an element of
infinite order x, then any finite set {y1 , . . . , yn } ⊂ A together with the element x will generate
an infinite abelian subgroup, which must be torsion-free by the previous observation. The
only case that remains is A being an infinite torsion group. But since any finite subgroup
has to be cyclic, it follows that A is locally finite cyclic.
For example, the Prüfer p-group P is a countable locally cyclic infinite abelian p-group.
By [LW12, Theorem 4.3] it follows that the minimal dimension of a model for ECyc (P)
equals one.
Lemma 3.18. Let G be elementary amenable and suppose that there is a finite-dimensional
model for ECyc (G). Then G is virtually solvable.
Proof.
Since ECyc (G) has a finite-dimensional model, so does EF Cyc (G) by [LW12,
Proposition 5.1], where F Cyc denotes the family of finite cyclic subgroups. It follows that
the Hirsch length h(G) of G is finite, since h(G) ≤ cdQ (G) ≤ gdF Cyc (G) < ∞. The first
inequality follows from [Hil91, Lemma 2]. For the second inequality note that Q[G/F] is
18
TIMM VON PUTTKAMER AND XIAOLEI WU
a projective QG-module for F finite [Bro82, I.8 Ex. 4] and thus the cellular chain complex
of EF Cyc (G) yields a projective resolution of Q over QG. Moreover, note that any locally
finite subgroup H of G has to be locally cyclic, in particular H is abelian. Combined with
the structure theorem of elementary amenable groups of finite Hirsch length [HL92], it
follows that G is virtually solvable.
Lemma 3.19. A hyperbolic group G has a finite-dimensional classifying space ECyc (G) if
and only if it does not contain the infinite dihedral group as a subgroup and the normalizers
of all non-trivial finite subgroups are finite cyclic. Moreover, if a subgroup K is maximal
finite, then NG (K) = K.
Proof.
Note that a hyperbolic group G has a finite-dimensional model for EG [JL06,
Remark 7 + Proposition 8]. Hence by Observation 3.16, G has a finite-dimensional classifying space ECyc (G) if and only if VCyc(G) = Cyc(G).
Suppose first that ECyc (G) is finite-dimensional, which is equivalent to saying that
VCyc(G) = Cyc(G). In particular, the infinite dihedral group does not appear as a subgroup. Now let K be a non-trivial finite subgroup of G. If NG (K) was infinite, it would
contain an element g ∈ NG (K) of infinite order. But then hg, Ki ≤ NG (K) would be an infinite virtually cyclic group that is not cyclic. Thus NG (K) is finite and thus also cyclic. The
claim about maximal finite subgroups being self-normalizing follows from Lemma 3.12.
For the converse, let V ≤ G be a non-trivial virtually cyclic subgroup. If V is finite, it is
cyclic, being the subgroup of its finite cyclic normalizer. Now suppose that V is infinite and
let K denote its unique maximal normal finite subgroup. Since K is normal in V, NG (K) is
infinite. It follows that K must be trivial. Since V cannot be the infinite dihedral group, it
follows that V is infinite cyclic.
Note that in the proof we have only used hyperbolicity of G to conclude that it has a
finite-dimensional model for EG and any infinite subgroup of it contains an element of
infinite order.
Proposition 3.20. Let G be a finitely generated linear group over a field of characteristic
zero. If G has a finite-dimensional model for ECyc (G), then the virtual cohomological
dimension of G is finite. More precisely,
vcd(G) ≤ gdCyc (G) + 1.
Proof. Let n = gdCyc (G) be the minimal dimension of a model for ECyc (G). Moreover
let H ≤ G be a torsion-free subgroup of finite index and note that ECyc (G) is also a model
for ECyc (H) via restriction. By [LW12, Proposition 5.1] there exists a model for EH of
dimension n + 1, since EZ has a one-dimensional model.
Linear groups of finite virtual cohomological dimension have been characterized by
Aplerin and Shalen in terms of the Hirsch lengths of their finitely generated unipotent
subgroups [AS82].
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 19
4. Groups with Wild Inner Automorphisms and BVC
As noted in the introduction, essentially all proofs of Conjecture A rely on understanding the BVC property for manageable classes of groups. In this section we want to construct groups outside of these realms that exhibit wild behaviour with respect to the BVC
property. For constructing finitely generated examples we will make use of small cancellation theory over relatively hyperbolic groups as developed in [Osi10] and [HO13]. In the
following we content ourselves with setting up some notation that is being used later on
without repeating the definition of relatively hyperbolic groups [Osi06]. Note that in this
section we will use the convention that gw = w−1 gw.
If G is a group that is hyperbolic relative to a family {Hλ }λ∈Λ of subgroups, we call
an element g ∈ G parabolic, if it is conjugate to an element lying in one of the parabolic
subgroups Hλ . Non-parabolic elements of infinite order are called loxodromic. For g ∈ G
a loxodromic element, there exists a unique maximal virtually cyclic subgroup EG (g) that
contains g. It is given by
EG (g) = {h ∈ G | ∃ m ∈ N \ {0} such that h−1 gm h = g±m }.
Recall that two elements a, b ∈ G are called commensurable if there exists k, l ∈ Z \ {0}
such that ak is conjugate to bl . A subgroup S of G is called suitable if there exist two
loxodromic elements s1 , s2 ∈ S that are not commensurable such that EG (s1 ) ∩ EG (s2 ) = 1.
Given a group H and an isomorphism θ : A → B between two subgroups A and B of H,
we can define a new group H∗θ = H∗At =B , called the HNN extension of H along θ, by the
presentation hH, t | t−1 xt = θ(x), x ∈ Ai. The letter t is called stable letter.
A sequence g0 , tǫ1 , g1 , . . . , tǫn , gn of elements with gi ∈ H and ǫi ∈ {−1, +1} is said to
be reduced if there is no pinch, where we define a pinch to be a consecutive sequence
t, gi , t−1 with gi ∈ B or t−1 , g j , t with g j ∈ A. Britton’s Lemma states that if the sequence
g0 , tǫ1 , g1 , . . . , tǫn , gn is reduced and n ≥ 1, then g0 tǫ1 g1 · · · tǫn gn , 1 in H∗θ . The number of
occurrences of t or t−1 in a reduced expression for an element g ∈ H∗θ is called its t-length,
it is independent of the choice of reduced word.
We will make use of the following two lemmas repeatedly. For the convenience of the
reader, we recall their statements:
Lemma 4.1 ([Osi06, Theorem 1.4]). Let G be a group hyperbolic relative to a collection
g
of subgroups {Hλ }λ∈Λ . Then for every λ ∈ Λ and g ∈ G \ Hλ the intersection Hλ ∩ Hλ is
finite.
In particular, if G is torsion-free and f, g ∈ Hλ are not conjugate in Hλ , then they are
not conjugate in G.
Lemma 4.2 ([HO13, Lemma 2.14]). Let G be a group and A, B two isomorphic subgroups.
Let f ∈ G be an element that is not conjugate to any element of A ∪ B. Then in the HNN
extension G∗At =B
20
TIMM VON PUTTKAMER AND XIAOLEI WU
(a) f is conjugate to some g ∈ G in G∗At =B if and only if f and g are conjugate in G.
(b) If f is primitive in G, then it is primitive as an element of G∗At =B .
Lemma 4.3. Let G be hyperbolic relative to a subgroup H and suppose that G is torsionfree. Let h ∈ H be primitive as an element of H. Then h is primitive in G.
Proof.
First note that non-trivial powers of loxodromic elements are loxodromic
again. In fact, let g be loxodromic and suppose that gn is parabolic for some n , 0, i.e.
gn = αyα−1 , where y ∈ H and α ∈ G. Then it would follow that 0 = τrel (y) = τrel (gn ) =
|n|τrel (g) > 0 by [Osi06, Lemma 4.24, Theorem 4.25], where τrel (g) denotes the relative
translation number of g.
So if h = gn for some g ∈ G and n ≥ 1, then g has to be parabolic, i.e. g = αxα−1 for
some x ∈ H and α ∈ G. By Lemma 4.1 it follows that α ∈ H, thus n = 1, since h was
primitive in H.
Lemma 4.4. Let G be a group and let a, b ∈ G be two primitive elements. Then any
primitive element g ∈ G, considered as an element of the HNN extension G∗at =b is still
primitive.
Proof.
Let g ∈ G be a primitive element. If g is neither conjugate to an element of hai
nor hbi, then Lemma 4.2 implies that g is primitive as an element of the HNN extension.
So we can assume without loss of generality that g = a. Let a = ωn with n ≥ 1, where
ω is of t-length 2, so ω = g0 tǫ g1 t−ǫ g2 for g0 , g1 , g2 ∈ G and ǫ ∈ {±1}. For the equality
a = ωn to hold, the expression t−ǫ g0 g2 tǫ must be a pinch, so say g2 g0 = bm for some m.
Then t−ǫ g2 g0 tǫ = am . Expanding the power of ω, we obtain
ωn = g0 tǫ g1 (am g1 )n−1 t−ǫ g2
This implies that g1 (am g1 )n−1 ∈ hai, say g1 (am g1 )n−1 = ak . Thus
k+m −1
g0
a = ωn = g0 bk g2 = g0 bk g2 g0 g−1
0 = g0 b
But since a is primitive in G, it follows that k + m ∈ {±1}. The equation g1 (am g1 )n−1 = ak is
equivalent to (am g1 )n = am ak = am+k , which implies that n = 1, again since a is primitive.
The induction step works as in the proof of [HO13, Lemma 2.14].
Lemma 4.5. Let G be a group and let a, b ∈ G be arbitrary non-trivial elements. Then
any primitive element g ∈ G, considered as an element of the HNN extension G∗(an )t =bm is
primitive as long as |n|, |m| ≥ 2.
Proof.
Let g ∈ G be a primitive element. Then g is neither conjugate to an element
of han i nor hbm i. Thus the claim follows from Lemma 4.2.
The next proposition shows that the converse of Lemma 1.9 does not hold.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 21
Proposition 4.6. There exists a finitely generated torsion-free group G with a finite index
subgroup H such that H has bCyc, but G does not.
Proof. The proofs of Lemma 7.1 and Theorem 7.2 in [HO13] apply. By Lemma 4.2
it follows that the elements a f with f ∈ F in the statement of Lemma 7.1 are primitive in
C. In the last part of the proof of Theorem 7.2, one only needs to observe that the elements
a f with f ∈ F are primitive in G(i) by Lemma 4.3 since they lie in the parabolic subgroup
C. Since a f1 and a f2 lie in different conjugacy classes for any f1 , f2 in F (this was stated
at the end of the proof of [HO13, Theorem 7.2]), G does not have bCyc by Lemma 1.8.
Lemma 4.7. Let H be a torsion-free countable group. There exists a 2-generated torsionfree group G that contains H as a subgroup such that
(a) Any g ∈ G is conjugate to an element of H.
(b) If h ∈ H is primitive, then it is primitive as an element of G.
(c) If h, h′ ∈ H are not conjugate, then they are not conjugate in G.
Proof.
Let
G(0) = H ∗ F(x, y)
and enumerate elements of H = {1 = h0 , h1 , h2 , . . .} as well as G(0) = {1 = g0 , g1 , g2 , . . .}.
Suppose the group G(i) has been constructed such that G(i) is hyperbolic relative to H,
hx, yi is a suitable subgroup and h1 , . . . , hi lie in hx, yi and g j for 1 ≤ j ≤ i is conjugate to
an element of H. Construct G(i + 1) out of G(i) as follows: If gi+1 is parabolic, then let
G′ (i) = G(i), otherwise let ι : EG(i) (gi+1 ) → hh1 i be an isomorphism and form
G′ (i) = hGi , t | et = ι(e) for e ∈ E(G(i) (gi+1 )i
By [HO13, Corollary 2.16], G′ (i) is still hyperbolic relative to H and hx, yi is again
suitable. Now apply [HO13, Theorem 6.2] to the suitable subgroup hx, yi, words {hi+1 , t}
resp. {hi+1 } to obtain G(i + 1). Observe that there is a canonical quotient map G(i) →
G(i + 1). Here we do not distinguish H and its image in G′ (i) or G(i + 1). Note that G(i + 1)
is also hyperbolic relative to H, hx, yi is again suitable.
Letting G be the direct limit of the G(i), it follows that G will be two-generated and any
element of G will be conjugate to an element of H. Moreover, if h ∈ H is primitive, then
it will remain primitive in G(i) by Lemma 4.3 for any i, thus it will be primitive in G. By
Lemma 4.1 non-conjugate elements in H remain non-conjugate in G.
Proposition 4.8. There exists a finitely generated torsion-free group G which has exactly
three conjugacy classes {(1), (x), (x2)}, where x ∈ G is a primitive element.
Proof.
We first construct a countable group as follows. We let G0 = hxi Z.
Of course, the only primitive elements in G0 are x and x−1 . Suppose we have already constructed a chain G0 ≤ G1 ≤ . . . ≤ Gn of countable torsion-free groups such that the element
22
TIMM VON PUTTKAMER AND XIAOLEI WU
x, viewed as an element of Gn , is primitive. To construct Gn+1 out of Gn , we first enumerate
all primitive elements of Gn by {p1 , p2 , . . .} and enumerate all non-trivial elements that are
non-primitive by {g1 , g2 , . . .}. Secondly, we form the multiple HNN extension
Gn+1 = hGn , {si }i∈N , {ti }i∈N | pisi = x, gtii = x2 i.
By an inductive application of Lemma 4.4 and Lemma 4.5 it follows that x remains primS
itive in Gn+1 . Finally, we let G = n≥0 Gn . By construction, G satisfies our desired properties. Note that since x is primitive, x and x2 are non-conjugate. To obtain a finitely
generated example we can apply Lemma 4.7.
Remark 4.9. If we let G be a group as constructed in Proposition 4.8 and x ∈ G a primitive
element, then G × Z does not have bCyc, since it has infinitely many primitive conjugacy
classes {(x, n) | n ∈ Z}. On the other hand, G × Z has bCyc if G is a torsion-free group with
exactly two conjugacy classes.
As we have noted, if G is a torsion-free group with exactly two conjugacy classes G × Z
has bCyc. However, the situation changes if we allow for a semidirect product:
Proposition 4.10. There exists a countable torsion-free group H with exactly two conjugacy classes such that a certain extension H ⋊ Z does not have bCyc.
Proof.
Let G0 = ha, bi be a free group of rank 2, let ε0 : G0 → Z be defined by
mapping a to 0 and b to 1. Note that for any m ∈ Z, the element abm is primitive and
ε0 (abm ) = m.
Suppose Gn and εn : Gn → Z have been constructed. To obtain Gn+1 , enumerate all
non-trivial elements of ker εn by {g1 , g2 , . . .} and form the multiple HNN extension
Gn+1 = hGn , {ti }i∈N | gtii = ai.
We can extend εn to Gn+1 to define εn+1 : Gn+1 → Z by arbitrarily assigning a value to the
stable letters ti . Now note that for m , 0, abm ∈ G0 ≤ Gn is neither conjugate to an element
of hgi i, nor to an element of hai. Thus, by Lemma 4.2, the elements abm are primitive in
Gn+1 for m , 0.
Let G be the direct limit of the Gn , and let ε : G → Z be induced by the epimorphisms
εn . Any non-trivial element of ker(ε) is then conjugate to a. However, since for m , 0, the
elements abm are primitive and obviously in different conjugacy classes as ε(abm ) = m, it
follows that G does not have bCyc.
Proposition 4.11. For any m ≥ 1, there exists a finitely generated torsion-free group G
that has exactly n + 1 conjugacy classes (1), (x1 ), . . . , (xm ) such that xki is conjugate to xi
for any i and any k , 0.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 23
Proof.
Note that any torsion-free group with exactly two conjugacy classes will have
the property that xk is conjugate to x as long as x is non-trivial. So we will demonstrate the
claim for m = 2, the general case follows from an analogous argument. Let G−1 = ha, bi
be a free group of rank two, and define
G0 = hG−1 , {si }i∈Z\{0} , {ti }i∈Z\{0} | (ai ) si = a, (bi)ti = bi.
Now observe that the elements a and b are not conjugate in G0 by repeated application of
Lemma 4.2. If we form an HNN extension with relation (ai ) si = a we apply Lemma 4.2 to
the element b. For relations of the type (bi )ti = b we apply the same lemma to the element
a.
We proceed constructing countable torsion-free groups Gn for n ≥ 1 inductively. First,
observe that we can write Gn \ {1} = S 0 ⊔ S a ⊔ S b , where S a resp. S b are those elements of
Gn which have a non-trivial power that is conjugate to a resp. b, and S 0 being defined as
the complement of S a ∪ S b . Note that S a ∩ S b = ∅: If g ∈ S a ∩ S b , then gk ∼ a and gl ∼ b
for some k, l , 0. But then gkl ∼ al ∼ a, and at the same time gkl ∼ bk ∼ b. But this is
impossible since a and b are not conjugate in Gn by induction. Our construction proceeds
in two steps:
Step 1. Enumerate all element of S a ∪ S b = {g1 , g2 , . . .}. We form the multiple HNN
extension
Q = hGn , {ti }i∈N | gtii = b if gi ∈ S b , otherwise gtii = ai.
In other words Q is the direct limit of a sequence of HNN extensions Q0 ≤ Q1 ≤ Q2 ≤ . . .
where Q0 = Gn and
Qi = hQi−1 , ti | gtii = b if gi ∈ S b , otherwise gtii = ai.
Now we prove by induction that a and b are not conjugate in Qi for any i. Suppose the
claim is true for Qi−1 . If gi ∈ S a then apply Lemma 4.2 to the element b. Note that b is not
conjugate to a power of a in Qi−1 by induction. Moreover b is not conjugate to gni for any
n
n , 0 in Qi−1 . Otherwise, it would follow that b ∼ bk ∼ gnk
i ∼ a ∼ a in Qi−1 since there
is some k , 0 such that gki ∼ a. Interchanging the roles of a and b, we see that the same
conclusion holds if gi ∈ S b . Hence it follows that a and b are non-conjugate in Q.
Step 2. Enumerate all elements of S 0 = {h1 , h2 , . . .}. Again we construct a sequence of
HNN extensions P0 ≤ P1 ≤ . . ., starting with P0 = Q. Suppose Pi−1 has been constructed
and a, b are non-conjugate in Pi−1 . To form Pi , we consider the following cases:
(a) If a non-trivial power of hi is conjugate to a in Pi−1 , we form the HNN extension
Pi = hPi−1 , si | hisi = ai.
We can again employ Lemma 4.2 to the element b to prove that a and b are nonconjugate in Pi using the same argument as in step 1.
24
TIMM VON PUTTKAMER AND XIAOLEI WU
(b) If a non-trivial power of hi is conjugate to b in Pi−1 , we let
Pi = hPi−1 , si | hisi = bi.
Again, interchanging the roles of a and b one sees that these two elements remain
non-conjugate in Pi .
(c) In the remaining case we can choose
Pi = hPi−1 , si | hisi = ai.
and observe that Lemma 4.2 can be applied.
S
We let Gn+1 = i≥0 Pi . Note that a and b are non-conjugate in Gn+1 and all elements of
Gn \ {1} are either conjugate to a or b in Gn+1 .
S
Finally, we define G = n≥0 Gn . Again, Lemma 4.7 yields a finitely generated example.
We see in particular that for a group G as constructed in the previous proposition, the
group G × Z has bCyc.
Example 4.12. Letting G be an infinite 2-generated group of exponent p with exactly p
conjugacy classes (for example, as constructed in [Ols91, Theorem 41.2]), then the group
G × Z does not have bCyc. This can be seen by considering the elements (x, pn ) for n ∈ N
where x ∈ G is non-trivial.
It is easy to see that a torsion-free group with infinitely many commensurability classes
cannot have bCyc. The following shows that a converse to Lemma 1.8 does not hold:
Proposition 4.13. There exists a finitely generated torsion-free group G without primitive
elements that does not have bCyc.
Proof.
We first prove that there exists a torsion-free countable group without prim-
itive elements and infinitely many commensurability classes. We start by an inductive
procedure, letting G0 = F2 be the non-abelian free group on two generators. In fact, any
countable torsion-free group with infinitely many commensurability classes of elements
would work as well. Suppose Gn has been constructed, enumerate all non-trivial elements
Gn \ {1} = {g1 , g2 , . . .} and define Gn+1 as the following multiple HNN extension:
Gn+1 = hGn , {ti }i∈N | gtii = g2i i
Suppose x, y ∈ Gn \ {1} are not commensurable. For any i ∈ N and for any k, l ∈ Z \ {0} it
follows that xk is not conjugate to an element of hgi i or yl is not conjugate to an element of
hgi i. By Lemma 4.2 it follows that x and y are not commensurable in Gn+1 . Letting G be
the direct limit of the Gn , we have that there are infinitely many commensurability classes
in G and any non-trivial element of G can be written as a proper power of a conjugate of
itself. To obtain a finitely generated example, apply Lemma 4.7.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 25
Lemma 4.14. There exists a finitely generated torsion-free group G without bCyc and
exactly two commensurability classes.
Proof.
Let G0 = ha, bi be a free group of rank two. Suppose Gn has been constructed,
then enumerate all elements {g1 , g2 , . . .} of Gn \ {1} that are not primitive. We form the
multiple HNN extension
Gn+1 = hGn , {ti }i∈N | gtii = b2 i
By Lemma 4.5 any primitive element of Gn stays primitive in Gn+1 . Similarly, also nonconjugate primitive elements g, h ∈ Gn will stay non-conjugate in Gn+1 by Lemma 4.2. If
we let G be the direct limit of the Gn , then G contains infinitely many primitive conjugacy
classes, thus G fails to have bCyc. However, given any non-trivial element g ∈ G, by construction g2 will be conjugate to b2 . Thus there are precisely two commensurability classes.
Finally, to obtain a finitely generated group with the same properties, apply Lemma 4.7.
Lemma 4.15. Let G be a group such that centralizers of all non-trivial elements are infinite
cyclic. Let a, b ∈ G and suppose that an = bm for some m, n , 0. If a is primitive, then
b ∈ hai.
Proof.
Note that CG (an ) = hxi for some x ∈ G and a ∈ CG (an ) and b ∈ CG (an ), thus
a = xk for some k and b = xl for some l. Since a is primitive it follows that k = ±1. Hence
b ∈ hai.
In a hyperbolic group the centralizers of infinite order elements are virtually cyclic
[BH99, III.Γ.3.10]. In particular if G is torsion-free hyperbolic and g is a primitive element,
then CG (hgi) = NG (hgi) = hgi.
Lemma 4.16. Let G be a torsion-free hyperbolic group and let a, b be two primitive elements such that a is not conjugate to b. Then the HNN extension G∗at =bk is hyperbolic for
any nonzero interger k.
Proof.
By [KM98, Corollary 1], we only need to show that the HNN extension
G∗at =bk is separated. Recall that a subgroup U of G is called conjugate separated if the set
{u ∈ U | u x ∈ U} is finite for all x ∈ G \ U. And an HNN extension G∗θ for an isomorphism
θ : U → V is called separated if either U or V is conjugate separated, and the set U ∩ V g is
finite for all g ∈ G. Now hai is conjugate separated since CG (hai) = NG (hai) = hai. And if
hai ∩ hg−1 bk gi was non-empty, then an = (g−1 bg)km for some m, n , 0. By Lemma 4.15 it
follows that b is conjugate to a which contradicts our assumptions.
Proposition 4.17. There exists a countable torsion-free group G and an epimorphism
ε : G → Z such that G has bCyc but ker(ε) does not.
Proof.
Let G0 = ha, c, di, and ε0 : G0 → Z be defined by mapping a to 1 and the
other free generators to 0. Moreover, choose a bijection φ0 : N0 → G0 \ {1}.
26
TIMM VON PUTTKAMER AND XIAOLEI WU
For each n > 0 we construct a countable torsion-free group Gn , an epimorphism εn : Gn →
Z and choose a bijection φn : N0 → Gn \ {1} such that
(a) Gn is either an HNN extension of Gn−1 through the stable letter tn or equals Gn−1 .
Moreover, Gn is a hyperbolic group.
(b) εn |Gn−1 = εn−1
(c) Any primitive element x ∈ ker(εn−1 ) is primitive as an element of Gn .
(d) If two primitive elements x, y ∈ ker(ε0 ) are conjugate in Gn , then the tn−1 -length of
ω is at most one.
Furthermore we choose the bijection φ : N0 → N0 × N0 which enumerates the elements
of N0 × N0 diagonally, i.e. {(0, 0), (1, 0), (0, 1), (0, 2), (1, 1), . . .}.
Now, suppose Gn has been constructed. Let (i, j) = φ(n) and let gn = φi ( j) ∈ Gn . In Gn ,
if gn is not primitive, or if it is conjugate to an element of hai or hdi, then set Gn+1 = Gn ,
εn+1 = εn and φn+1 = φn . Otherwise we construct Gn+1 as an HNN extension depending on
the value of εn (g) as follows:
(i) If εn (gn ) , 0, we set
Gn+1 = hGn , tn | gtnn = aεn (gn ) i.
(ii) If gn ∈ ker(εn ), we define
Gn+1 = hGn , tn | gtnn = di.
Note that in both cases Lemma 4.16 applies and thus Gn is hyperbolic.
If gn is conjugate in Gn to an element of hc, di, say gn = α−1
n xn αn for some xn ∈ hc, di, we
define εn+1 : Gn+1 → Z by tn 7→ |εn (αn )| + 4ε(tn−1 ) + 1, otherwise we let tn 7→ εn (tn−1 ) + 1.
Here, we interpret ε(t−1 ) = 1. Note that αn and xn are not unique here, but we fix our
choices for each such gn .
Furthermore we choose a bijection φn+1 : N0 → Gn+1 \ {1}.
Proof of (c). Let x ∈ ker(εn ) be primitive. The claim in case (i) follows directly from
Lemma 4.2 since x is not conjugate to any element in hgn i ∪ hai. In case (ii) we can apply
Lemma 4.4 since since gn is primitive in Gn by assumption and d ∈ ker(ε0 ) is primitive in
Gn by induction.
Proof of (d). Let x, y ∈ ker(ε0 ) ≤ Gn+1 be primitive and let ω ∈ Gn+1 be a reduced word
such that ω−1 xω = y. If ω contains no tn or tn−1 , then we are certainly done, thus we can
moreover assume that we are in case (ii). Suppose the tn -length of ω is at least two, then we
can write ω = ω1 tn±1 ω2 tn±1 ω3 as a reduced expression, where ω2 ∈ Gn and ω1 , ω3 ∈ Gn+1 .
−1
If ω = ω1 tn ω2 tn ω3 we know that tn−1 ω−1
1 xω1 tn has to be a pinch, so ω1 xω1 ∈ hgn i, i.e.
k
ω−1
1 xω1 = gn . Since x is primitive as an element of G n+1 , it follows that k = ±1. But we
k
−1 k
m
also know that tn−1 ω−1
2 d ω2 tn has to be a pinch, thus ω2 d ω2 = gn where m = ±1. Since
ω2 ∈ Gn , this is impossible by our choice of gn . Similarly the case that ω = ω1 tn−1 ω2 tn−1 ω3
is impossible.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 27
If ω contains two adjacent stable letters whose exponents are different, we first consider
−1
k
the case that ω = ω1 tn ω2 tn−1 ω3 . Then tn−1 ω−1
1 xω1 tn has to be a pinch, thus ω1 xω1 = gn
k
−1
with k = ±1, so that tn−1 gkn tn = dk . Now we also know that tn ω−1
2 d ω2 tn has to be a pinch,
k
thus ω−1
2 d ω2 ∈ hdi, so ω2 ∈ hdi since G n is hyperbolic. But then ω was not a reduced
word to begin with. Thus we have shown that the reduced element ω has tn -length at most
one. If ω = ω1 tn−1 ω2 tn ω3 then an analogous argument applies since hgn i is self-normalizing
as well since gn is primitive in case (ii).
We now define G as the direct limit of the Gn . Note that the εn induce an epimorphism
ε : G → Z.
G has bCyc. Let g ∈ G be a non-trivial element. We claim that it is conjugate to an
element in hai ∪ hdi. We can find n such that g ∈ Gn . Since Gn is hyperbolic we can find a
primitive element h in Gn such that g is some power of h. If h is conjugate to an element of
hai ∪ hdi or if h is conjugate to gn+1 in Gn+1 we are done. Otherwise, h remains primitive in
Gn+1 by Lemma 4.2. On the other hand, our enumeration function φ and the construction
guarantees that in the end any primivite element will be conjugate to an element in hai∪hdi.
ker(ε) does not have bCyc. We first prove that if x, y are two primitive elements in
hc, di ≤ ker(ε0 ) and there is some ω ∈ Gn+1 such that ω−1 xω = y, then |ε(ω)| ≤ 2ε(tn ),
where we interpret ε(t−1 ) = 1 as above. Let n = −1, and note that the conjugating element
ω lies in hc, di since G0 is free. In particular, ε(ω) = 0 ≤ 2ε(t−1 ). Now suppose n ≥ 0
and let ω ∈ Gn+1 such that ω−1 xω = y. By (d) it follows that the tn -length of ω is at most
one. If the tn -length equals zero, we are done by induction since ε(tn−1 ) < ε(tn ). If the
tn -length is non-zero we can assume without loss of generality that ω = ω1 tn ω2 where ω1
and ω2 are elements in Gn . Since ω−1 xω = y, it follows that ω−1
1 xω1 ∈ hgn i and similarly
−1 ±1
ω2 d ω2 = y. Here, gn is the element in the kernel of εn that was used to construct Gn+1
from Gn via an HNN extension.
±1
−1
Suppose ω−1
1 xω1 = gn . Then we also know that gn = αn xn αn , thus by induction
−1
ω1 αn
we know that |ε(ω1 α−1
= x±1
n )| ≤ 2ε(tn−1 ) since x
n , so |ε(ω1 )| ≤ |ε(αn )| + 2ε(tn−1 ).
Moreover, by induction we also conclude that |ε(ω2 )| ≤ 2ε(tn−1 ). Altogether we obtain
|ε(ω)| ≤ |ε(ω1 )| + ε(tn ) + |ε(ω2 )|
≤ 2ε(tn ) .
We are now ready to show that ker(ε) contains infinitely many primitive conjugacy
classes. Let x, y be two primitive elements in subgroup hc, di ≤ ker(ε0 ) such that x is
not conjugate to y in ker(ε0 ). Suppose x and y are conjugate via some ω ∈ ker(ε), i.e.
ω−1 xω = y. There is some n ∈ N such that x, y, ω ∈ Gn+1 . As in the proof above we can
±1
assume without loss of generality that ω = ω1 tn ω2 with ω1 , ω2 ∈ Gn and ω−1
1 xω1 = gn ,
ω2 d±1 ω2 = y such that |ε(ω2 )| ≤ 2ε(tn−1 ) and |ε(ω1 )| ≤ |ε(αn )| + 2ε(tn−1 ). But since we
chose ε(tn ) > |εn (αn )| + 4ε(tn−1 ), we see that ω cannot lie in the kernel of ε.
28
TIMM VON PUTTKAMER AND XIAOLEI WU
Theorem 4.18. There exists a finitely generated torsion-free group G = H ⋊ Z such that G
has bCyc but H does not.
Proof. Let C be a torsion-free countable group having bCyc and admitting an epimorphism ε : C → Z such that ker(ε) contains infinitely many conjugacy classes that are
primitive in C, see Proposition 4.17. Let
G(0) = C ∗ F(x, y).
We extend ε to α0 : G(0) → Z by mapping x and y to 1 ∈ Z. Moreover, we enumerate the
elements of C = {c0 = 1, c1 , c2 , . . .} and those of G(0) = {g0 = 1, g1 , g2 , . . .}. Essentially
the same argument as in the proof of [HO13, Theorem 7.2] applies. There is a sequence
of groups G(i) and epimorphisms αi : G(i) → Z such that G(i + 1) is a quotient of G(i)
and αi descends to an epimorphism αi+1 : G(i + 1) → Z. Under the quotient maps the
group C embeds into G(i) such that G(i) is hyperbolic relative to C. Moreover, in G(i) : (a)
the elements c1 , . . . , ci are contained in hx, yi and (b) the elements g1 , . . . , gi are conjugate
to elements of C. If we define G as the direct limit of the G(i), we obtain an induced
epimorphism α : G → Z. By (a) G is 2-generated and by (b) G has bCyc since C has
bCyc. Since G(i) is hyperbolic relative to C, if elements c, c′ ∈ C are conjugate in G(i), i.e.
wcw−1 = c′ for some w ∈ G(i), then w ∈ C by Lemma 4.1. Thus Lemma 4.3 together with
the previous observation imply that ker(α) contains infinitely many primitive conjugacy
classes.
Proposition 4.19. There exists a countable torsion-free group G and an epimorphism
ε : G → Z such that both G and ker(ε) have bCyc, and ker(ε) contains a non-abelian
free subgroup.
Proof.
We will first inductively construct a specific sequence G0 ≤ G1 ≤ . . . of
groups, together with epimorphisms εn : Gn → Z such that εn+1 |Gn = εn as follows: We let
G0 = ha, b, ci be a non-abelian free group and let ε0 : G0 → Z be defined by mapping a to
1 and b, c to 0. Now, suppose Gn has been constructed such that Gn−1 ≤ Gn , together with
an epimorphism εn : Gn → Z. We enumerate all elements of Gn = {1 = g0 , g1 , g2 , . . .} and
define Gn+1 as the following multiple HNN extension
Gn+1 = hGn , {ti }i∈N | gtii = aεn (gi ) if εn (gi ) , 0 and gtii = b otherwise i.
We can then extend εn to Gn+1 to define εn+1 , by mapping the stable letters ti to 0 ∈ Z.
Finally, we let G be the direct limit of the Gn . Observe that G has bCyc with witnesses
hai and hbi and that there is an induced epimorphism ε : G → Z such that hb, ci ≤ ker(ε).
Moreover, since the stable letters of the HNN extensions are contained in ker(ε), also ker(ε)
has bCyc with only one cyclic group hbi as the witness.
Using arguments as in the proof of [HO13, Theorem 7.2] one can construct a group as
in the previous proposition that is additionally finitely generated.
LINEAR GROUPS, CONJUGACY GROWTH, AND CLASSIFYING SPACES FOR FAMILIES OF SUBGROUPS 29
Proposition 4.20. There exists a torsion-free finitely generated group G of exponential
conjugacy growth that has bCyc .
Proof.
As a first step, we will prove that there exists a torsion-free countable group
G that contains an infinite cyclic subgroup hai such that G has bCyc and the elements a2k+1
are pairwise non-conjugate. Moreover, there is an element t ∈ G, such that at = a2 . In
the end, the latter property will ensure that the word length of an element am is of the
order of log(|m|). We will construct G as direct limit of countable groups G0 ≤ G1 ≤ . . . as
follows: We let G0 = ha, t | at = a2 i be the Baumslag-Solitar group BS (1, 2). Note that this
group has exponential conjugacy growth, as the elements a2k+1 are pairwise non-conjugate,
see [GS10, Example 2.3]. To construct Gn+1 from Gn inductively, we first enumerate all
S
elements in Gn \ ( g∈G hag i) = {g1 , g2 , . . .} and then form the multiple HNN extension
Gn+1 = hGn , {si }i∈N | gisi = ti.
It follows inductively from Lemma 4.2 that the elements a2k+1 are pairwise non-conjugate
S
viewed as elements of Gn+1 . The direct limit G = n≥0 Gn then satisfies the previously
S
required properties as any element in G \ ( g∈G hag i) will be conjugate to the element
t ∈ G0 .
In the second step we want to construct a group with the same properties but which
is additionally finitely generated. In the following we will write C instead of G for the
previously constructed countable group. We let
G(0) = C ∗ F(x, y)
and enumerate the elements of C = {c0 = 1, c1 , c2 , . . .} resp. G(0) = {g0 = 1, g1 , g2 , . . .}.
Note that G(0) is hyperbolic relative to C and x, y generate a suitable subgroup of G(0).
In the following we will not distinguish notationally between elements of G(0) and their
representatives in quotients of G(0). We will inductively construct quotient groups G(i) of
G(0) such that
(a) the subgroup C embeds under the quotient map into G(i). Again, we will not
distinguish between C and its image in G(i).
(b) G(i) is torsion-free and hyperbolic relative to C. Moreover, x and y generate a
suitable subgroup of G(i).
(c) The elements c j for 1 ≤ j ≤ i, considered as elements in G(i), lie in the subgroup
generated by x and y.
(d) In G(i), for 1 ≤ j ≤ i, the elements g j are conjugate to elements in C.
(e) The elements a2k+1 are pairwise non-conjugate in G(i).
To construct G(i + 1) from G(i), we proceed as follows: If gi+1 is parabolic, we set
G′ (i) = G(i), otherwise we choose an isomorphism ι : EG(i) (gi+1 ) → hti and form the HNN
extension
G′ (i) = hG(i), s | e s = ι(e), e ∈ EG(i) (gi+1 )i.
30
TIMM VON PUTTKAMER AND XIAOLEI WU
Note that hx, yi is still a suitable subgroup of G′ (i) [HO13, Corollary 2.16].
In the next step, we apply [HO13, Theorem 6.2] to G′ (i) and the elements {s, ci+1 } resp.
{ci+1 } in the case that gi+1 is parabolic, to obtain a quotient G(i + 1) of G′ (i). Since s will
lie in the subgroup hx, yi, we obtain a quotient map G(i) → G(i + 1). By construction, (d)
holds for G(i + 1). The properties (a)-(c) follow directly from [HO13, Theorem 6.2]. The
last statement (e) is a consequence of Lemma 4.1 and the fact that G(i + 1) is torsion-free.
Finally, let G be defined as the direct limit of the G(i). By (c) it follows that G is 2generated, property (d) implies that G has bCyc since the same was already true for C.
Moreover, by (e), the elements a2k+1 ∈ C are pairwise non-conjugate in G as well. Thus G
has exponential conjugacy growth.
Proposition 4.21. Let Q be a countable group with n conjugacy classes. Then there exists
a torsion-free countable group G with n + 1 conjugacy classes such that Q is a quotient of
G.
Proof.
Let {q0 = 1, q1 , . . . , qn−1 } ⊂ Q be representatives of conjugacy classes of
elements in Q. We let G0 be a countable free group and ε0 : G0 → Q be an epimorphism.
We can moreover assume that ker(ε0 ) is non-trivial. Choose preimages g1 , . . . , gn−1 of
q1 , . . . , qn−1 under ε0 and choose some non-trivial element g0 ∈ ker(ε0 ).
We now inductively define countable torsion-free groups Gm together with epimorphisms εm : Gm → Q such that Gm−1 ≤ Gm and εm |Gm−1 = εm−1 .
Suppose Gm and εm have already been constructed. Enumerate all non-trivial elements
{h1 , h2 , . . .} of Gm and form the multiple HNN extension
Gm+1 = hGm , {ti }i∈N | relations explained below i.
We know that εn (hi ) is conjugate to q ji for some ji . If q ji = 1, then we impose the relation
htii = g0 . Otherwise there is some αi ∈ Gn such that εm (α−1
i hi αi ) = q ji . We then impose
−1
ti
the relation (αi hi αi ) = gi . With these choices we can extend εm to εm+1 : Gm+1 → Q by
mapping all stable letters ti to the trivial element in Q.
Finally, we let G be the direct limit of the groups Gm . By construction, representatives
of the conjugacy classes of G are {1, g0 , g1 , . . . , gn−1 }.
Note that the above construction is optimal in the sense that if Q has n conjugacy classes
and contains torsion then any torsion-free group G that surjects onto Q must have at least
n + 1 conjugacy classes.
Now let Q be a countable group with finitely many conjugacy classes but infinitely many
conjugacy classes of finite subgroups. Then Q has bCyc but not BVC, see Example 1.12
for the construction of such a group Q. Thus the previous proposition also shows:
Corollary 4.22. There exists a group G with BVC such that a quotient of G fails to have
BVC.
REFERENCES
31
References
[AS82]
Roger C. Alperin and Peter B. Shalen. “Linear groups of finite cohomological
dimension.” In: Invent. Math. 66 (1982), pp. 89–98.
[BH99]
Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature. Berlin: Springer, 1999.
[Bre+13]
Emmanuel Breuillard, Yves Cornulier, Alexander Lubotzky, and Chen Meiri.
“On conjugacy growth of linear groups.” In: Math. Proc. Cambridge Philos.
[Bro82]
Soc. 154.2 (2013), pp. 261–277.
Kenneth S. Brown. Cohomology of groups. Graduate Texts in Mathematics,
[CF06]
87. New York-Heidelberg-Berlin: Springer-Verlag. X, 306 p. (1982). 1982.
Danny Calegari and Michael H. Freedman. “Distortion in transformation groups
(with an appendix by Yves de Cornulier).” In: Geom. Topol. 10 (2006), pp. 267–
293.
[CK02]
Michel Coornaert and Gerhard Knieper. “Growth of conjugacy classes in Gromov hyperbolic groups.” In: Geom. Funct. Anal. 12.3 (2002), pp. 464–478.
[DKP15]
Dieter Degrijse, Ralf Köhl, and Nansen Petrosyan. “Classifying spaces with
virtually cyclic stabilizers for linear groups.” In: Transform. Groups 20.2 (2015),
pp. 381–394.
[DP13]
[DP15]
Dieter Degrijse and Nansen Petrosyan. “Commensurators and classifying spaces
with virtually cyclic stabilizers.” In: Groups Geom. Dyn. 3 (2013), pp. 543–
555.
Dieter Degrijse and Nansen Petrosyan. “Bredon cohomological dimensions
for groups acting on CAT(0)-spaces.” In: Groups Geom. Dyn. 9.4 (2015),
pp. 1231–1265.
[FN13]
Martin G. Fluch and Brita E.A. Nucinkis. “On the classifying space for the
family of virtually cyclic subgroups for elementary amenable groups.” In:
[GS10]
Proc. Amer. Math. Soc. 141.11 (2013), pp. 3755–3769.
Victor Guba and Mark Sapir. “On the conjugacy growth functions of groups.”
[GW13]
In: Illinois J. Math. 54.1 (2010), pp. 301–313.
J.R.J. Groves and John S. Wilson. “Soluble groups with a finiteness condition
arising from Bredon cohomology.” In: Bull. Lond. Math. Soc. 45.1 (2013),
pp. 89–92.
[Hil91]
Jonathan A. Hillman. “Elementary amenable groups and 4-manifolds with Eu-
[HL92]
ler characteristic 0.” In: J. Aust. Math. Soc., Ser. A 50.1 (1991), pp. 160–170.
Jonathan A. Hillman and Peter A. Linnell. “Elementary amenable groups of
finite Hirsch length are locally-finite by virtually-solvable.” In: J. Aust. Math.
Soc., Ser. A 52.2 (1992), pp. 237–241.
32
REFERENCES
[HO13]
Michael Hull and Denis Osin. “Conjugacy growth of finitely generated groups.”
[JL06]
In: Adv. Math. 235 (2013), pp. 361–389.
Daniel Juan-Pineda and Ian J. Leary. “On classifying spaces for the family of
virtually cyclic subgroups.” In: Recent developments in algebraic topology. A
conference to celebrate Sam Gitler’s 70th birthday, San Miguel de Allende,
México, December 3–6, 2003. Providence, RI: American Mathematical Society (AMS), 2006, pp. 135–145.
[KM98]
Olga Kharlampovich and Alexei Myasnikov. “Hyperbolic groups and free
constructions.” In: Trans. Amer. Math. Soc. 350.2 (1998), pp. 571–613.
[KMN11] Dessislava H. Kochloukova, Conchita Martı́nez-Pérez, and Brita E.A. Nucinkis.
“Cohomological finiteness conditions in Bredon cohomology.” In: Bull. Lond.
[Kro93]
Math. Soc. 43.1 (2011), pp. 124–136.
Peter H. Kropholler. “On groups of type (FP)∞.” In: J. Pure Appl. Algebra
90.1 (1993), pp. 55–67.
[LN03]
Ian J. Leary and Brita E.A. Nucinkis. “Some groups of type VF.” In: Invent.
Math. 151.1 (2003), pp. 135–165.
[Lüc+]
Wolfgang Lück, Holger Reich, John Rognes, and Marco Varisco. Assembly
maps for topological cyclic homology of group algebras. arXiv: 1607.03557.
[Lüc09]
Wolfgang Lück. “On the classifying space of the family of virtually cyclic
subgroups for CAT(0)-groups.” In: Münster J. Math. 2.1 (2009), pp. 201–214.
[Lüc89]
Wolfgang Lück. Transformation groups and algebraic K-theory. Berlin etc.:
Springer-Verlag, 1989, pp. xii + 443.
[LW12]
Wolfgang Lück and Michael Weiermann. “On the classifying space of the
family of virtually cyclic subgroups.” In: Pure Appl. Math. Q. 8.2 (2012),
[MM03]
pp. 497–555.
George A. Miller and Halcott C. Moreno. “Non-Abelian groups in which every
[New68]
subgroup is Abelian.” In: Trans. Amer. Math. Soc 4 (1903), pp. 398–404.
Bill B. Newman. “Some results on one-relator groups.” In: Bull. Amer. Math.
Soc. 74 (1968), pp. 568–571.
[Ols91]
Alexander Yu. Ol’shanskij. Geometry of defining relations in groups. Dordrecht etc.: Kluwer Academic Publishers, 1991.
[Osi06]
Denis Osin. “Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems.” In: Mem. Amer. Math. Soc. 843 (2006),
[Osi10]
p. 100.
Denis Osin. “Small cancellations over relatively hyperbolic groups and em-
[Tit72]
bedding theorems.” In: Ann. of Math. (2) 172.1 (2010), pp. 1–39.
Jacques Tits. “Free subgroups in linear groups.” In: J. Algebra 20 (1972),
pp. 250–270.
REFERENCES
[vW]
33
Timm von Puttkamer and Xiaolei Wu. On the finiteness of the classifying
space for the family of virtually cyclic subgroups. arXiv: 1607.03790.
University of Bonn, Mathematical Institute, Endenicher Allee 60, 53115 Bonn, Germany
E-mail address: [email protected]
Max Planck Institut für Mathematik, Vivatsgasse 7, 53111 Bonn, Germany
E-mail address: [email protected]
| 4math.GR
|
arXiv:1706.02393v1 [cs.CV] 7 Jun 2017
ShiftCNN: Generalized Low-Precision Architecture
for Inference of Convolutional Neural Networks
Denis A. Gudovskiy, Luca Rigazio
Panasonic Silicon Valley Lab
Cupertino, CA, 95014
{denis.gudovskiy,luca.rigazio}@us.panasonic.com
Abstract
In this paper we introduce ShiftCNN, a generalized low-precision architecture for
inference of multiplierless convolutional neural networks (CNNs). ShiftCNN is
based on a power-of-two weight representation and, as a result, performs only
shift and addition operations. Furthermore, ShiftCNN substantially reduces computational cost of convolutional layers by precomputing convolution terms. Such
an optimization can be applied to any CNN architecture with a relatively small
codebook of weights and allows to decrease the number of product operations by at
least two orders of magnitude. The proposed architecture targets custom inference
accelerators and can be realized on FPGAs or ASICs. Extensive evaluation on
ImageNet shows that the state-of-the-art CNNs can be converted without retraining
into ShiftCNN with less than 1% drop in accuracy when the proposed quantization
algorithm is employed. RTL simulations, targeting modern FPGAs, show that
power consumption of convolutional layers is reduced by a factor of 4 compared to
conventional 8-bit fixed-point architectures.
1
Introduction
Current achievements of deep learning algorithms made them an attractive choice in many research
areas including computer vision applications such as image classification (He et al., 2015), object
detection (Redmon et al., 2016), semantic segmentation (Shelhamer et al., 2016) and many others.
Until recently real-time inference of deep neural networks (DNNs) and, more specifically, convolutional neural networks (CNNs) was only possible on power-hungry graphics processing units (GPUs).
Today numerous research and implementation efforts (Sze et al., 2017) are conducted in the area of
computationally inexpensive and power-efficient CNN architectures and their implementation for
embedded systems as well as data center deployments.
We propose a generalized low-precision CNN architecture called ShiftCNN. ShiftCNN substantially
decreases computational cost of convolutional layers which are the most computationally demanding
component during inference phase. The proposed architecture combines several unique features. First,
it employs a hardware-efficient power-of-two weight representation which requires performing only
shift and addition operations. State-of-the-art CNNs can be converted into ShiftCNN-type of model
without network retraining with only a small drop in accuracy. Second, a method to precompute
convolution terms decreases the number of computations by at least two orders of magnitude for
commonly used CNNs. Third, we present ShiftCNN design which achieves a factor of 4 reduction
in power consumption when implemented on a FPGA compared to conventional 8-bit fixed-point
architectures. In addition, this architecture can be extended to support sparse and compressed models
according to conventional (LeCun et al., 1990) or more recent approaches (Han et al., 2016).
The rest of the paper is organized as follows. Section 2 reviews related work. Section 3 gives theoretical background. Section 4 proposes a generalized low-precision quantization algorithm. Section 5
introduces a method to decrease computational cost of convolutional layers and its corresponding
architecture. Section 6 presents experimental results on ImageNet, quantization and complexity
analysis including power consumption projections for FPGA implementation. The last section draws
conclusions.
2
2.1
Related Work
Low-Precision Data Representation
Conventional CNN implementations on GPUs use a power-inefficient floating-point format. For
example, (Horowitz, 2014) estimated that 32-bit floating-point multiplier consumes 19× more power
than 8-bit integer multiplier. Several approaches for a more computationally-efficient data representation within CNNs have been proposed recently. One direction is to convert (quantize) a baseline
floating-point model into a low-precision representation using quantization algorithms. Such quantization algorithms can be classified into the following categories: fixed-point quantization (Courbariaux
et al., 2015), binary quantization (Hubara et al., 2016; Rastegari et al., 2016), ternary quantization (Li
and Liu, 2016) and power-of-two quantization (Miyashita et al., 2016; Zhou et al., 2017). Moreover,
these quantization algorithms can be selectively applied to various parts of the network e.g. weights,
inputs and feature maps.
According to the latest published experimental results (Zhou et al., 2017), out of the referenced
quantization algorithms only models converted to fixed-point and power-of-two representations
achieve baseline accuracy without significant accuracy drop. In addition, binary and ternary-quantized
models require a subset of convolutional layers, e.g. first and last layers, to be in the other, usually
fixed-point, format. In this paper, we select power-of-two weight representation for building truly
multiplierless CNNs and, unlike (Miyashita et al., 2016; Zhou et al., 2017), employ a more general
quantization algorithm which allows to achieve minimal accuracy drop without time-consuming
retraining. In our architecture, inputs and feature maps are in the dynamic fixed-point format.
2.2
Convolution Computation Algorithms
There is a significant body of work on various convolution computation algorithms e.g. frequencydomain convolution or Winograd minimal filtering convolution summarized in (Lavin, 2015). Unfortunately, they cannot be universally applied for all CNN architectures. For example, frequency-domain
methods are not suitable for modern CNNs with small convolution filter sizes and Winograd algorithm
achieves moderate speed-ups (6-7×) only with batch sizes more than one which increases processing
latency. Other algorithms suited for GPUs with optimized vector operations using BLAS libraries are
not compatible with custom ASIC or FPGA accelerators.
Another interesting approach was proposed by (Bagherinezhad et al., 2016) where convolutions are
optimized by lookups into a shared trained dictionary assuming that weights can be represented by a
trained linear combinations of that dictionary. While this approach achieves substantial speed-ups
during inference, it significantly degrades accuracy of networks due to inherent weight approximation
assumptions. Instead, our approach employs powerful distribution-induced representation to form a
naturally small codebook (dictionary) .
3
Note on Quantization Optimality
Consider a general feedforward non-linear neural network that maps an input x to an output y
according to
y = f (x ; θ),
(1)
where θ is the vector of all trainable parameters and f is the function defined by the network structure.
Typically, the parameters θ are found using maximum log-likelihood criteria with regularization by
solving optimization problem defined by
arg max log p(y |x ; θ) − αR(θ),
θ
(2)
where R(θ) is a regularization term that penalizes parameters and α ≥ 0 is a hyperparameter that
defines a tradeoff between estimator bias and variance. Usually, only a subset of θ i.e. the weights
w of the model are penalized. Hence, the most widely used L2 and L1 regularization types define
2
regularization term as R(θ) = kw k22 and R(θ) = kw k1 , respectively. It is well-known (Goodfellow
et al., 2016) that regularization places a prior on weights which induces normal distribution for L2
regularization and Laplace distribution for L1 regularization.
Based on the known probability density function (PDF) fW (w), we can estimate quantization
distortion by
J Z wj
X
D=
|w − w̃j |2 fW (w)dw,
(3)
j=1
wj−1
where J is the size of quantizer codebook. Then, the optimal quantizer in the mean squared error
(MSE) sense minimizes (3). By taking into account (2), the distortion measure D is minimized when
a non-uniform quantizer is used presuming selected regularization method. Interestingly, a similar
reasoning can be applied to network feature maps. Although, we derived the optimal quantizer in the
MSE sense, some network parameters may have more effect on the optimization objective function (2)
than others. For example, Choi et al. (2017) proposed to use a Hessian-weighted distortion measure
in order to minimize the loss due to quantization assuming a second-order Taylor series expansion.
Hence, an additional objective function-dependent term can be added to (3) as well.
4
ShiftCNN Quantization
The input of each convolutional layer in commonly used CNNs can be represented by an input tensor
X ∈ RC×H×W , where C, H and W are the number of input channels, the height and the width,
respectively. The input X is convolved with a weight tensor W ∈ RC̃×C×Hf ×Wf , where C̃ is the
number of output channels, Hf and Wf are the height and the width of filter kernel, respectively.
A bias vector b ∈ RC̃ is added to the result of convolution operation. Then, the c̃th channel of an
output tensor Y ∈ RC̃×H̃×W̃ before applying non-linear function can be computed by
Yc̃ = Wc̃ ∗ X + b c̃ ,
(4)
where ∗ denotes convolution operation.
Based on conclusions in Section 3, we approximate W by a low-precision weight tensor Ŵ using
hardware-efficient non-uniform quantization method. Each entry ŵi (i ∈ {c̃, c, hf , wf }) of Ŵ is
defined by
N
X
ŵi =
Cn [idxi (n)],
(5)
n=1
where for ShiftCNN codebook set Cn = {0, ±2−n+1 , ±2−n , ±2−n−1 , . . . , ±2−n−bM/2c+2 } and a
codebook entry can be indexed by B-bit value with total M = 2B − 1 combinations. Then, each
weight entry can be indexed by (N B)-bit value. For example, consider a case with N = 2 and B = 4.
Then, the weight tensor is indexed by 8-bit value with codebooks C1 = {0, ±20 , ±2−1 , . . . , ±2−6 }
and C2 = {0, ±2−1 , ±2−2 , . . . , ±2−7 }. Note that for N = 1, (5) can be simplified to binaryquantized (B = 1, 0 ∈
/ C1 ) model or ternary-quantized (B = 2) model. Moreover, fixed-point integers
can be represented by (5) as well when N 1. Similar to (Zhou et al., 2017), we normalize W by
max (abs(W)), where abs(·) is a element-wise operation, before applying (5). Then, N × 1 weight
index vector can be calculated by Algorithm 1. Unlike the quantization algorithms in (Miyashita
et al., 2016; Zhou et al., 2017) for N = 1, the proposed algorithm is more general and suitable for
cases N > 1.
Effectively, the non-uniform quantizer in Algorithm 1 maps the weight tensor W
∈
C̃×C×H ×W
f
f
RC̃×C×Hf ×Wf → Ŵ ∈ CN
, where CN is a set formed by sum of Cn sets. Figure 1 illustrates possible CN sets i.e. weight fields. Interestingly, binary-quantized filters contribute the same
unit magnitude to the output feature map with either positive or negative sign, while ternary-quantized
weights allow to mute that contribution in some dimensions of weight vector space. In other words,
output feature map heavily depend on input magnitudes. At the same time, ShiftCNN is more flexible
and allows to trade off computational complexity and cardinality of weight field CN while reusing
the same arithmetic pipeline.
In this work, we mostly concentrate on a weight approximation task for hardware-efficient CN ,
while (Hubara et al., 2016; Rastegari et al., 2016; Li and Liu, 2016; Zhou et al., 2017) empirically
3
Algorithm 1 Quantization to ShiftCNN weight representation
1: Initialize: w̃i = 0, r = wi / max (abs(W)), where i ∈ {c̃, c, hf , wf }
2: for n = 1 . . . N do
3:
qsgn = sgn(r)
4:
qlog = log2 |r|
5:
qidx = bqlog c
6:
blog = qidx + log2 1.5
7:
if qlog > blog then
8:
qidx ++
9:
q = qsgn 2qidx
10:
idxi (n) = qsgn (−n − qidx + 2)
11:
if |idxi (n)| > bM/2c then
12:
q=0
13:
idxi (n) = 0
14:
r -= q
15:
w̃i += q
1
1
-1
1
1
-1
-1
(a)
1
1
1
-1
-1
(b)
-1
1
-1
(c)
-1
(d)
Figure 1: Weight vector in 2D space for models: (a) binary-quantized (N = 1, B = 1, 0 ∈
/ C1 ), (b)
ternary-quantized (N = 1, B = 2), (c) (N = 1, B = 3) and (d) (N = 2, B = 3).
employ a small subset of CN and, in order to restore desired accuracy, rely on a constrained retraining
process i.e. introduce additional regularization. Though the topic of a high-dimensional classifier
over reduced field needs further research, we can derive simple dimension relationships between CN
V
subfields using field extensions. For example, vector w ∈ CN
dimension can be rewritten over a
ternary field T = {0, ±1} as
dimCN (w ) = dimCN (T ) dimT (w ) .
| {z } | {z }
V
bM/2c + N − 1
5
(6)
ShiftCNN Architecture
Input Tensor
Input Tensor
Right-Shift
>>1
Shift
Arithmetic Unit
Sign
Flip
Array of
Adders
Right-Shift
>>1
...
Precomputed Tensor
Precomputed Tensor
Weight Tensor
...
(b)
Bias
Output Tensor
(a)
Figure 2: (a) High-level structure. (b) Shift Arithmetic Unit (ShiftALU).
4
Sign
Flip
We propose to precompute all possible P = (M + 2(N − 1)) combinations of product terms for each
element in the input tensor X when (4) is applied and, then, accumulate results in the output tensor
Y according to the weight tensor Ŵ indices. Assuming for simplicity H̃ = H, W̃ = W for filter
kernels with stride 1, conventional convolutional layer performs (C̃Hf Wf CHW ) multiplications.
In the proposed approach, a convolutional layer computes only (P CHW ) product operations, which
for ShiftCNN are power-of-two shifts. For the state-of-the-art CNNs P (C̃Hf Wf ) and, therefore,
the computational cost reduction in terms of the number of product operations is quantified by a
(C̃Hf Wf /P ) ratio. Note that the number of addition operations is N -times higher for the proposed
approach. In addition, a new memory buffer for precomputed tensor P ∈ RP ×C×H×W is implicitly
introduced. Details on further microarchitecture-level optimizations including the size of the P are
given below.
Precomputed Tensor
. . .
0
M
U
X
Output Tensor
Precomputed Tensor
0
Control
Logic
M
U
X
MUX
Weight Tensor
Bias
Figure 3: Array of adders.
Figure 2(a) illustrates a high-level structure of the proposed ShiftCNN inference architecture. It
computes (4) according to assumptions in (5) by precomputing tensor P on a special shift arithmetic
unit (ShiftALU). The ShiftALU block showed in Figure 2(b) consists of a shifter and a sign flipping
logic. The shifter shifts input values to the right and writes each shift-by-one result as well as
pass-through input value to the memory buffer to keep precomputed P tensor. In addition, each of
bP/2c precomputed terms is passed through the sign flipping logic and, then, written into P memory
buffer. Mathematically, ShiftALU computations are equivalent to replacing the weight tensor W in
(4) convolution with the codebook CN in order to generate precomputed convolution terms.
Algorithm 2 Scheduling and control logic
1: for h = 1 . . . H and w = 1 . . . W do
2:
read C × 1 input vector x = Xh,w
3:
precompute (P − 1) × C matrix P = Ph,w and write into the memory buffer
4:
for c̃ = 1 . . . C̃ do
5:
write bias b c̃ to the output tensor Yc̃,h,w
6:
for n = 1 . . . N do
7:
for hf = 1 . . . Hf and wf = 1 . . . Wf do
8:
select value from {0, P} using idxi (n), where i ∈ {c̃, c, hf , wf }
9:
add value to the output tensor Yc̃,h,w
Then, precomputed convolution terms are fetched from the memory buffer according to the weight
tensor indices using a multiplexer as showed in Figure 3 and processed on an array of adders. The
multiplexer selects either one of (P − 1) memory outputs or zero value using B-bit index. Next, the
multiplexer outputs are accumulated and bias term added according to scheduling and control logic.
Lastly, the result is stored in the output tensor. One of possible scheduling and control logic schemes
for ShiftCNN is presented in Algorithm 2 pseudo-code. It extracts computational parallelism along
C-dimension of the input tensor. We define a parallelization level C̄ as the number of concurrently
computed convolution terms. For the sake of simplicity, we consider a case with stride 1 filter kernel
5
and, therefore, H̃ = H, W̃ = W and choose the parallelization level C̄ = C. Therefore, we are
able to decrease size of the precomputed memory to ((P − 1)C) entries. In order to provide enough
bandwidth for the selected parallelization level, the memory buffer can be organized as an array of
(P − 1) length-C shift registers which are written along C dimension and read (multiplexed) along
P -dimension. Then, the array of adders for the parallelized case is naturally transformed into an
adder tree.
We can consider several corner cases for ShiftCNN architecture. For a case N = 1 and B = 1, the
proposed shift arithmetic unit can be simplified into architecture for processing binary-quantized
models with a single sign flipping block only. For a case N = 1 and B = 2, ShiftALU is the
same as in the previous case, but multiplexer is extended by zero selection logic for processing
ternary-quantized models. For a case N 1, ShiftCNN can be converted into architecture for
processing fixed-point models.
6
Evaluation
6.1
ImageNet Results
To evaluate the accuracy of ShiftCNN and its baseline full-precision counterparts, we selected the
most popular up-to-date CNN models for image classification. Caffe (Jia et al., 2014) was used to run
networks and calculate top-1 and top-5 accuracies. GoogleNet and ResNet-50 pretrained models were
downloaded from the Caffe model zoo1 . SqueezeNet v1.12 and ResNet-183 were downloaded from
the publicly available sources. All the referenced models were trained on ImageNet (Russakovsky
et al., 2015) (ILSVRC2012) dataset by their authors. A script to convert baseline floating-point model
into ShiftCNN without retraining is publicly available4 . The script emulates power-of-two weight
representation by decreasing precision of floating-point weights according to Algorithm 1. We run
50,000 images from ImageNet validation dataset on the selected network models and report accuracy
results in Table 1. ResNet results in parentheses include quantized batch normalization and scaling
layers with another weight distribution as explained below.
Table 1: ImageNet accuracy of baseline models and corresponding ShiftCNN variants.
Decrease
in Top-1
Accuracy, %
Decrease
in Top-5
Accuracy, %
81.02
45.93
80.31
81.01
35.39
1.01
0.01
35.09
0.71
0.01
68.93
57.67
68.54
68.88
89.15
81.79
88.86
89.06
11.26
0.39
0.05
7.36
0.29
0.09
32
4
4
4
64.78
24.61
64.24(61.57)
64.75(64.69)
86.13
47.02
85.79(84.08)
86.01(85.97)
40.17
0.54(3.21)
0.03(0.09)
39.11
0.34(2.05)
0.12(0.16)
32
4
4
4
72.87
54.38
72.20(70.38)
72.58(72.56)
91.12
78.57
90.71(89.48)
90.97(90.96)
18.49
0.67(2.49)
0.29(0.31)
12.55
0.41(1.64)
0.15(0.16)
Network
Shifts
N
Bitwidth B
Top-1
Accuracy, %
Top-5
Accuracy, %
SqueezeNet
SqueezeNet
SqueezeNet
SqueezeNet
base
1
2
3
32
4
4
4
58.4
23.01
57.39
58.39
GoogleNet
GoogleNet
GoogleNet
GoogleNet
base
1
2
3
32
4
4
4
ResNet-18
ResNet-18
ResNet-18
ResNet-18
base
1
2
3
ResNet-50
ResNet-50
ResNet-50
ResNet-50
base
1
2
3
1
https://github.com/BVLC/caffe/wiki/Model-Zoo
https://github.com/DeepScale/SqueezeNet
3
https://github.com/HolmesShuan/ResNet-18-Caffemodel-on-ImageNet
4
https://github.com/gudovskiy/ShiftCNN
2
6
The converted models in Table 1 have a performance drop of less than 0.29% for convolutional layers
when B = 4 and N > 2. The models with B = 4 and N = 2 have a performance drop within 1%
which is a good practical trade-off between accuracy and computational complexity. We didn’t notice
any significant accuracy improvements for bit-widths B > 4 when N > 1. While in Table 1 the
cases with N = 1 do not perform well without retraining, Zhou et al. (2017) reported that the cases
with N = 1 and B = 5 perform as good as the baseline models with their retraining algorithm.
6.2
Weight Quantization Analysis
Surprisingly good results for such low-precision logarithmic-domain representation even without
retraining can be explained by looking at the distribution of the network weights. Figure 4(a) illustrates
weight (normalized to unit magnitude) PDF of a typical convolutional layer (res2a_branch1 layer
in ResNet-50) as well as a scaling layer scale2a_branch1 in ResNet-50. The typical convolutional
layer weights are distributed according to Section 3 which is well approximated by the non-uniform
quantization Algorithm 1. At the same time, batch normalization and scaling layers in ResNet
networks have an asymmetric and a broader distributions which cannot be approximated well enough
unless N > 2.
0.30
res2a
scale2a
0.25
0.25
0.20
Density
Density
0.20
0.15
0.10
0.00
0.00
0.0
0.0
Weight Value
0.5
(a)
1.0
0.25
0.10
0.05
0.5
0.30
0.15
0.05
1.0
0.35
base
N=8, B=3
N=3, B=4
N=2, B=4
N=1, B=4
Density
0.30
N=8, B=3
N=3, B=4
N=2, B=4
N=1, B=4
0.20
0.15
0.10
0.05
0.2
0.4
0.6
Top-1 probability
(b)
0.8
1.0
0.00
1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8
Top-1 probability error
(c)
Figure 4: PDFs: (a) ResNet-50 weights, (b) GoogleNet top-1 class and (c) GoogleNet top-1 class
error.
Due to difficulty of analytical analysis of chained non-linear layers, we empirically evaluate quantization effects at the GoogleNet softmax layer output with ImageNet validation dataset input. Figure 4(b)
shows top-1 class PDF for baseline floating-point, 8-bit fixed-point (N = 8, B = 3) and other
ShiftCNN configurations. According to that PDF, all distributions are very close except the lowest
precision N = 1, B = 4 case. In order to facilitate quantization effects, we plot PDF of the probability error for top-1 class as a difference between floating-point probability and its lower precision
counterparts in Figure 4(c). Then, we can conclude that N = 3, B = 4 classifier bears slightly
lower variance than 8-bit fixed-point model and N = 2, B = 4 classifier obtains slightly higher
variance with nearly zero bias in both cases. At the same time, the lowest precision model experiences
significant divergence from the baseline without retraining process.
6.3
Complexity Analysis
Section 5 concludes that the number of product operations in each convolutional layer can be
decreased by a factor of (C̃Hf Wf /P ) with all multiplications replaced by power-of-two shifts.
Furthermore, the pipelined ShiftALU is able to compute (P − 1) outputs in a single cycle by using
(bP/2c − 1) shift-by-one and bP/2c sign flipping operations. Hence, ShiftALU further reduces the
number of cycles needed to process precomputed tensor by a factor of P . Finally, the speed-up in
terms of the number of cycles for multiplication operations for the conventional approach and the
number of ShiftALU cycles for the proposed approach is equal to (C̃Hf Wf ).
Table 2 presents a high-level comparison of how many multiplication cycles (operations) needed for
the conventional convolutional layers and how many ShiftALU cycles needed for the three analyzed
network models. It can be seen that the SqueezeNet has a 260× speed-up due to the fact that it
employs filter kernel with the dimensions Hf , Wf ≤ 3. At the same time, GoogleNet and ResNet-18
7
widely use filter kernels with the dimensions equal to 5 and 7. Therefore, the achieved speed-ups are
687× and 1090×, respectively.
Table 2: High-level complexity comparison for popular CNNs
Network
Conventional Mult.
Million Cycles
ShiftALU
Million Cycles
Speed-Up
410
1750
1970
1.58
2.55
1.81
260×
687×
1090×
SqueezeNet
GoogleNet
ResNet-18
6.4
FPGA Implementation and Power Estimates
We implemented a computational pipeline of ShiftCNN convolutional layer in RTL starting from the
ShiftALU input showed in Figure 2(b) to the adder tree output showed in Figure 3. Memories for the
input, output, weight and bias tensors are intentionally excluded from the design. For comparison, a
conventional 8-bit fixed-point (multiplier-based) arithmetic unit was implemented in RTL with the
same architectural assumptions. Then, RTL code was compiled for an automotive-grade Xilinx Zynq
XA7Z030 device running at 200 MHz clock rate. The input tensor bit-width is 8-bit which is enough
for almost lossless feature map representation in dynamic fixed-point format (Gysel et al., 2016). The
resulted outputs (before addition operations) of either ShiftALU or the multiplier-based approach are
16-bits. Bit-width B of the ShiftALU indices are 4-bits, N = 2 and parallelization level C̄ = 128.
Table 3 presents FPGA utilization and power consumption estimates of the implemented computational pipeline in terms of number of lookup tables (LUTs), flip-flops (FFs), digital signal processing
blocks (DSPs) and dynamic power measured by Xilinx Vivado tool. Note that the multiplier-based
approach is implemented both using DSP blocks and using LUTs only. According to these estimates,
ShiftCNN utilizes 2.5× less FPGA resources and consumes 4× less dynamic power. The latter can
be explained by low update rate of the precomputed tensor memory. Moreover, our power evaluation
shows that 75% of ShiftCNN dynamic power is dissipated by the adder tree and, hence, the power
consumption of product computation is decreased by a factor of 12. In other words, the power
consumption of convolutional layer is dominated by addition operations.
Table 3: FPGA utilization and power consumption estimates
7
ALU
LUTs
FFs
DSPs
Power, mW
Relative Power
Adders only
ShiftCNN
Mult. DSP
Mult. LUT
1644
4016
0
10064
1820
2219
2048
4924
0
0
191
0
76
102
423
391
75%
100%
415%
383%
Conclusions
We introduced ShiftCNN, a novel low-precision CNN architecture. Flexible hardware-efficient
weight representation allowed to compute convolutions using shift and addition operations only with
the option to support binary-quantized and ternary-quantized networks. The algorithm to quantize
full-precision floating-point model into ShiftCNN without retraining was presented. In addition, a
naturally small codebook of weights formed by the proposed low-precision representation allowed to
precompute convolution terms and, hence, to reduce the number of computation operations by two
orders of magnitude or more for commonly used CNN models.
ImageNet experiments showed that state-of-the-art CNNs when being converted into ShiftCNN
experience less than 1% drop in accuracy for a reasonable model configuration. Power analysis of the
proposed design resulted in a factor of 4 dynamic power consumption reduction of convolutional
layers when implemented on a FPGA compared to conventional 8-bit fixed-point architectures. We
believe, the proposed architecture is able to surpass existing fixed-point inference accelerators and
make a new types of deep learning applications practically feasible for either embedded systems or
data center deployments.
8
References
Hessam Bagherinezhad, Mohammad Rastegari, and Ali Farhadi. LCNN: lookup-based convolutional
neural network. arXiv preprint arXiv:1611.06473, 2016.
Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Towards the limit of network quantization.
International Conference on Learning Representations (ICLR), Apr 2017.
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Training deep neural networks with
low precision multiplications. In International Conference on Learning Representations (ICLR),
May 2015.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
Philipp Gysel, Mohammad Motamedi, and Soheil Ghiasi. Hardware-oriented approximation of
convolutional neural networks. arXiv preprint arXiv:1604.03168, 2016.
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks
with pruning, trained quantization and Huffman coding. In International Conference on Learning
Representations (ICLR), May 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. arXiv preprint arXiv:1512.03385, 2015.
Mark Horowitz. Computing’s energy problem (and what we can do about it). In IEEE Interational
Solid State Circuits Conference (ISSCC), pages 10–14, Feb 2014.
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks. In Advances in Neural Information Processing Systems 29 (NIPS), pages 4107–
4115. 2016.
Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio
Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding.
arXiv preprint arXiv:1408.5093, 2014.
Andrew Lavin. Fast algorithms for convolutional neural networks. arXiv preprint arXiv:1509.09308,
2015.
Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural
Information Processing Systems 2 (NIPS), pages 598–605. 1990.
Fengfu Li and Bin Liu. Ternary weight networks. arXiv preprint arXiv:1605.04711, 2016.
Daisuke Miyashita, Edward H. Lee, and Boris Murmann. Convolutional neural networks using
logarithmic data representation. arXiv preprint arXiv:1603.01025, 2016.
Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. XNOR-net: Imagenet
classification using binary convolutional neural networks. arXiv preprint arXiv:1603.05279, 2016.
Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified,
real-time object detection. In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), June 2016.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,
Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet
Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115
(3):211–252, 2015.
Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic
segmentation. arXiv preprint arXiv:1605.06211, 2016.
Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel Emer. Efficient processing of deep neural
networks: A tutorial and survey. arXiv preprint arXiv:1703.09039, 2017.
Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization:
Towards lossless CNNs with low-precision weights. In International Conference on Learning
Representations (ICLR), Apr 2017.
9
| 9cs.NE
|
arXiv:1604.01485v1 [cs.CV] 6 Apr 2016
A Focused Dynamic Attention Model for Visual
Question Answering
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
[email protected], {eleyans,elefjia}@nus.edu.sg
National University of Singapore
Abstract. Visual Question and Answering (VQA) problems are attracting increasing interest from multiple research disciplines. Solving VQA
problems requires techniques from both computer vision for understanding the visual contents of a presented image or video, as well as the
ones from natural language processing for understanding semantics of
the question and generating the answers. Regarding visual content modeling, most of existing VQA methods adopt the strategy of extracting
global features from the image or video, which inevitably fails in capturing fine-grained information such as spatial configuration of multiple objects. Extracting features from auto-generated regions – as some
region-based image recognition methods do – cannot essentially address
this problem and may introduce some overwhelming irrelevant features
with the question. In this work, we propose a novel Focused Dynamic
Attention (FDA) model to provide better aligned image content representation with proposed questions. Being aware of the key words in the
question, FDA employs off-the-shelf object detector to identify important
regions and fuse the information from the regions and global features via
an LSTM unit. Such question-driven representations are then combined
with question representation and fed into a reasoning unit for generating
the answers. Extensive evaluation on a large-scale benchmark dataset,
VQA, clearly demonstrate the superior performance of FDA over wellestablished baselines.
Keywords: Visual Question Answering, Attention
1
Introduction
Visual question answering (VQA) is an active research direction that lies in
the intersection of computer vision, natural language processing, and machine
learning. Even though with a very short history, it already has received great
research attention from multiple communities. Generally, the VQA investigates
a generalization of traditional QA problems where visual input (e.g., an image)
is necessary to be considered. More concretely, VQA is about how to provide a
correct answer to a human posed question concerning contents of one presented
image or video.
VQA is a quite challenging task and undoubtedly important for developing
modern AI systems. The VQA problem can be regarded as a Visual Turing Test
2
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
[1,2], and besides contributing to the advancement of the involved research areas, it has other important applications, such as blind person assistance and
image retrieval. Coming up with solutions to this task requires natural language
processing techniques for understanding the questions and generating the answers, as well as computer vision techniques for understanding contents of the
concerned image. With help of these two core techniques, the computer can
perform reasoning about the perceived contents and posed questions.
Recently, VQA is advanced significantly by the development of machine learning methods (in particular the deep learning ones) that can learn proper representations of questions and images, align and fuse them in a joint question-image
space and provide a direct mapping from this joint representation to a correct
answer.
For example, consider the following image-question pair: an image of an apple
tree with a basket of apples next to it, and a question “How many apples are in
the basket?”. Answering this question requires VQA methods to first understand
the semantics of the question, then locate the objects (apples) in the image,
understand the relation between the image objects (which apples are in the
basket), and finally count them and generate an answer with the correct number
of apples.
The first feasible solution to VQA problems was provided by Malinowski
and Fritz in [2], where they used a semantic language parser and a Bayesian
reasoning model, to understand the meaning of questions and to generate the
proper answers. Malinowski and Fritz also constructed the first VQA benchmark dataset, named as DAQUAR, which contains 1,449 images and 12,468
questions generated by humans or automatically by following a template and
extracting facts from a database [2]. Shortly after, Ren et al. [3] released the
TORONTO–QA dataset, which contains a large number of images (123,287)
and questions (117,684), but the questions are automatically generated and thus
can be answered without complex reasoning. Nevertheless, the release of the
TORONTO–QA dataset was important since it provided enough data for deep
learning models to be trained and evaluated on the VQA problem [4,5,3]. More
recently, Antol et al. [6] published the currently largest VQA dataset. It consists
of three human posed questions and ten answers given by different human subjects, for each one of the 204,721 images found in the Microsoft COCO dataset
[7]. Answering the 614,163 questions requires complex reasoning, common sense,
and real-world knowledge, making the VQA dataset suitable for a true Visual
Turing Test. The VQA authors split the evaluation on their dataset on two tasks:
an open-ended task, where the method should generate a natural language answer, and a multiple-choice task, where for each question the method should
chose one of the 18 different answers.
The current top performing methods [8,9,10] employ deep neural network
model that predominantly uses the convolutional neural network (CNN) architecture [11,12,13,14] to extract image features and a Long Short-Term Memory
(LSTM) [15] network to extract the representations for questions. The CNN and
LSTM representation vectors are then usually fused by concatenation [16,3,5] or
A Focused Dynamic Attention Model for Visual Question Answering
3
element-wise multiplication [17,18]. Other approaches additionally incorporate
some kind of attention mechanism over the image features [18,19,20].
Properly modeling the image contents is one of the critical factors for solving
VQA problems well. A common practice with existing VQA methods on modeling image contents is to extract global features for the overall image. However,
only using global feature is arguably insufficient to capture all the necessary visual information and provide full understanding of image contents such as multiple objects, spatial configuration of the objects and informative background.
This issue can be relieved to some extent by extracting features from object
proposals – the image regions that possibly contain objects of interest. However,
using features from all image regions [19,18] may provide too much noise or overwhelming information irrelevant to the question and thus hurt the overall VQA
performance.
In this work, we propose a question driven attention model that is able to
automatically identify and focus on image regions relevant for the current question. We name our proposed model Focused Dynamic Attention (FDA) for Visual
Question Answering. With the FDA model, computers can select and recognize
the image regions in a well-aligned sequence with the key words containing in
a given question. Recall the above VQA example. To answer the question of
“How many apples are in the basket?”, FDA would first localize the regions corresponding to the key words “apples” and “basket” (with the help of a generic
object detector) and extract description features from these regions of interest.
Then VQA compliments the features from selected image regions with a global
image feature providing contextual information for the overall image, and reconstruct a visual representation by encoding them with a Long Short-Term
Memory (LSTM) unit.
We evaluate and compare the performance of our proposed FDA model on
two types of VQA tasks, i.e., the open-ended task and the multiple-choice task,
on the VQA dataset – the largest VQA benchmark dataset. Extensive experiments demonstrate that FDA brings substantial performance improvement upon
well-established baselines.
The main contributions of this work can be summarized as follows:
– We introduce a focused dynamic attention mechanism that learns to use the
question word order to shift the focus from one image object, to another.
– We describe a model that fuses local and global context visual features with
textual features.
– We perform an extensive evaluation, comparing to all existing methods, and
achieve state-of-the-art accuracy on the open-ended, and on the multiplechoice VQA tasks.
The rest of the paper is organized as follows. In Section 2 we review the
current VQA models, and compare them to our model. We formulate the problem
and explain our motivation in Section 3. We describe our model in Section 4 and
in Section 5 we evaluate and compare it with the current state-of-the-art models.
We conclude our work in Section 6.
4
2
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
Related Work
VQA has received great research attention recently and a couple of methods
have been developed to solve this problem. The most similar model to ours
is the Stacked Attention Networks (SAN) proposed by Yang et al. [19]. Both
models use attention mechanism that combines the words and image regions.
However, [19] use convolutional neural network to put attention over the image
regions, based on the question word unigrams, bigrams, and trigrams. Further,
their attention mechanism is not using object bounding boxes, which makes the
attention less focused.
Another model that uses attention mechanism in solving VQA problems is
the ABC-CNN model described in [18]. ABC-CNN uses the question embedding
to configure convolutional kernels that will define an attention weighted map over
the image features. The advantage of our FDA model over ABC-CNN is two fold.
First, FDA employs an LSTM network to encode the image region features in a
order that corresponds to the question word order. Second, FDA does not put
handcrafted weights on the image features (a practice showed to hurt the learning
process in our experiments). Instead, FDA extracts CNN features directly from
the cropped image regions of interest. In this sense, FDA is more efficient than
ABC-CNN in visual contents modeling.
Yet another attention model for visual question answering is proposed in [17].
The work, is closely related to the work by [18], in that it also applies a weighted
map over the image and the question word features. However, similar to our
work, they use object proposals from [21] to select image regions instead of the
whole image. Different from that work, our proposed FDA model also employs
the information embedded in the order of the question words and focuses on the
corresponding object bounding boxes. In contrast, the model proposed in [21]
straightforwardly concatenate all the image region features with the question
word features and feed them all at once to a two layer network.
Jiang et al. propose another model that combines the CNN image features
and an LSTM network for encoding the multimodal representation, with the
addition of a Compositional Memory units which fuse the image and word feature
vectors [22].
Ma et al. in [10] take an interesting approach and use three convolutional
neural networks to represent not only the image, but also the question, and their
common representation in a multimodal space. The multimodal representation
is then fed to a SoftMax layer to produce the answer.
Another interesting approach worth mentioning is the work by Andreas et
al. [23]. They use a semantic grammar parser to parse the question and propose
neural network layouts accordingly. They train a model to learn to compose a
network from one of the proposed network layouts using several types of neural
modules, each specifically designed to address the different sub-tasks of the VQA
problem (e.g. counting, locating an object, etc.).
A Focused Dynamic Attention Model for Visual Question Answering
3
5
Method Overview
In this section, we briefly describe the motivation and give formal problem formulation.
3.1
Problem Formulation
The visual question answering problem can be represented as predicting the
best answer â given an image I and a question q. Common practice [6,16,3,19]
is to use the 1,000 most common answers in the training set and thus simplify
the VQA task to a classification problem. The following equation represents the
problem mathematically:
â = arg max p(a|I, q; θ)
(1)
a∈Ω
where Ω is the set of all possible answers and θ are the model weights.
3.2
Motivation
The baseline methods from [6] show only modest increase in accuracy when
including the image features (4.98% for open-ended questions, and 2.42% for
multiple-choice question). We believe that the image contains a lot more information and should increase the accuracy much more. Thus, we focus on improving the image features and design a visual attention mechanism, which learns to
focus on the question related image regions.
The proposed attention mechanism is loosely inspired on the human visual
attention mechanism. Humans shift the focus from one image region to another,
before understanding how the regions relate to each other and grasping the
meaning of the whole image. Similarly, we feed our model image regions relevant
for the question at hand, before showing the whole image.
4
Focused Dynamic Attention for VQA
The FDA model is composed of question and image understanding components, attention mechanism, and a multimodal representation fusion network
(Figure 1). In this section we describe them individually.
4.1
Question Understanding
Following a common practice, our FDA model uses an LSTM network to encode
the question in a vector representation [15,5,18,8]. The LSTM network learns to
keep in its state the feature vectors of the important question words, and thus
provides the question understanding component with a word attention mechanism.
6
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
How
many
apples
are
in
the
basket
?
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
LSTM
CNN
CNN
CNN
CNN
Question
Representation
Element-wise
Multiplication
Image
Representation
Fully Connected
Neural Network
SoftMax
Answer
Fig. 1. Focused dynamic attention model diagram.
4.2
Image Understanding
Following prior work [3,4,5], we use a pre-trained convolutional neural network
(CNN) to extract image feature vectors. Specifically, we use the Deep Residual
Networks model used in ILSVRC and COCO 2015 competitions, which won the
1st places in: ImageNet classification, ImageNet detection, ImageNet localization,
COCO detection, and COCO segmentation [24]. We extract the weights of the
layer immediately before the final SoftMax layer and regard them as visual
features. We extract such features for the whole image (global visual features)
and for the specific image regions (local visual features). However, contrary to
the existing approaches, we employ an LSTM network to combine the local and
global visual features into a joint representation.
4.3
Focused Dynamic Attention Mechanism
We introduce a focused dynamic attention mechanism that learns to focus on
image regions related to the question words.
The attention mechanism works as follows. For each image object1 it uses
word2vec word embeddings [26] to measure the similarity between the question
words and the object label. Next, it selects objects with similarity score greater
than 0.5 and extracts the feature vectors of the objects bounding boxes with a
pre-trained ResNet model [24]. Following the question word order, it feeds the
LSTM network with the corresponding object feature vectors. Finally, it feeds
the LSTM network with the feature vector of the whole image and it uses the
resulting LSTM state as a visual representation. Thus, the attention mechanism
1
During training we use the ground truth object bounding boxes and labels. At test
time we use the precomputed bounding boxes from [25] and classify them with [24]
to obtain the object labels.
A Focused Dynamic Attention Model for Visual Question Answering
7
enables the model to combine the local and global visual features into a single
representation, necessary for answering complex visual questions.
Figure 1 illustrates the focused dynamic attention mechanism with an example.
4.4
Multimodal Representation Fusion
We regard the final state of the two LSTM networks as a question and image
representation. We start fusing them into single representation by applying Tanh
on the question representation and ReLU 2 on the image representation 3 . We
proceed by doing an element-wise multiplication of the two vector representations and the resulting vector is fed to a fully-connected neural network. Finally
a SoftMax layer classify the multimodal representation into one of the possible4
answers.
5
Evaluation
In this section we detail the model implementation and compare our model
against the current state-of-the-art methods.
Fig. 2. Representative examples of questions (black), (a subset of the) answers given
when looking at the image (green), and answers given when not looking at the image
(blue) for two images from the VQA dataset. Examples provided by [6].
2
3
4
Defined as f (x) = max(0, x).
Applying different activation functions gave slightly worse overall results
We follow [6] and use the 1000 most common answers
8
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
5.1
Dataset
For all experiments we use the Visual Question Answering (VQA) dataset [6],
which is the largest and most complex image dataset for the visual question answering task. The dataset contains three human posed questions and ten answers
given by different human subjects, for each one of the 204,721 images found in
the Microsoft COCO dataset [7]. Figure 2 shows two representative examples
found in the dataset. The evaluation is done on following two test splits test-dev
and test-std and on following two tasks:
– An open-ended task, where the method should generate a natural language
answer;
– A multiple-choice task, where for each question the method should chose one
of the 18 different answers.
We evaluate the performance of all the methods in the experiments using the
public evaluation server for fair evaluation.
5.2
Baseline Model
We compare our model against the baseline models provided by the VQA dataset
authors [27], which currently achieve the best performance on the test-standard
split for the multiple-choice task. The model, first described in [6], is a standard
implementation of an LSTM+CNN VQA model. It uses an LSTM to encode
the question and CNN features to encode the image. To answer a question it
multiplies the last LSTM state with the image CNN features and feeds the
result into a SoftMax layer for classification into one of the 1,000 most common
answers. The implementation in [27] uses a deeper two layer LSTM network
for encoding the question, and normalized image CNN features, which showed
crucial for achieving the state-of-the-art.
5.3
Model Implementation and Training Details
We transform the question words into a vector form by multiplying one-hot
vector representation with a word embedding matrix. The vocabulary size is
12,602 and the word embeddings are 300 dimensional. We feed a pre-trained
ResNet network [24] and use the 2,048 dimensional weight vector of the layer
before the last fully-connected layer.
The word and image vectors are feed into two separate LSTM networks. The
LSTM networks are standard implementation of one layer LSTM network [15],
with a 512 dimensional state vector. The final state of the question LSTM is
passed through Tanh, while the final state of the image LSTM is passed through
ReLU5 . We do element-wise multiplication on the resulting vectors, to obtain a
multimodal representation vector, which is then fed to a fully-connected neural
network.
5
Defined as f (x) = max(0, x).
A Focused Dynamic Attention Model for Visual Question Answering
9
Table 1. Comparison between the baselines from [6], the state-of-the-art models and
our FDA model on VQA test-dev and test-standard data for the open-ended task. Results from most recent methods including CM [22], ACK [9], iBOWIMG [16], DPPnet
[8], D-NMN [23], D-LSTM [27], and SAN [19] are provided and compared with.
Method
VQA
Question
Image
Q+I
LSTM Q+I
CM
ACK
iBOWIMG
DPPnet
D-NMN
D-LSTM
SAN
FDA
All
test-dev
Y/N Other
Num
48.09
28.13
52.64
53.74
52.62
55.72
55.72
57.22
57.90
58.70
59.24
75.66
64.01
75.55
78.94
78.33
79.23
76.55
80.71
80.50
79.30
81.14
36.70
0.42
33.67
35.24
34.46
36.13
35.03
37.24
37.40
36.6
36.16
27.14
3.77
37.37
36.42
35.93
40.08
42.62
41.71
43.10
46.10
45.77
test-std
All
54.06
55.98
55.89
57.36
58.00
58.16
58.90
59.54
Table 2. Comparison between the baselines from [6], the state-of-the-art models and
our FDA model on VQA test-dev and test-standard data for the multiple-choice
task. Results from most recent methods including WR [17], iBOWIMG [16], DPPnet
[8], and D-LSTM [27] are also shown for comparison.
Method
VQA
Question
Image
Q+I
LSTM Q+I
WR
iBOWIMG
DPPnet
D-LSTM
FDA
5.4
All
test-dev
Y/N Other
Num
test-std
All
53.68
30.53
58.97
57.17
60.96
61.68
62.48
64.01
75.71
69.87
75.59
78.95
76.68
80.79
81.50
37.05
0.45
34.35
35.80
38.94
38.94
39.00
57.57
61.97
62.69
63.09
64.18
38.64
3.76
50.33
43.41
54.44
52.16
54.72
Model Evaluation and Comparison
We compare our model with the baselines provided by the VQA authors [6].
The results for the open-ended task are listed in Table 1 and the results for the
multiple-choice task are given in Table 2. In the tables, the “Question” and “Image” baselines are only using the question words and the image, respectively. The
“Q+I” is a baseline that combines the two, but do not use an LSTM network.
“LSTM Q+I” and “D-LSTM” are LSTM models, with one and two layers ac-
10
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
cordingly. Comparing the performance of baselines we can observe the accuracy
increase with the addition of information from each modality.
From Table 1, one can observe that our proposed FDA model achieves the
best performance on this benchmark dataset. It outperforms the state-of-the-art
(SAN) with a margin of around 0.6%. The SAN model also employs attention
to focus on specific regions. However, their attention model (without access to
the automatically generated bounding boxes) is focusing on more spread regions
which may include cluttered and noisy background. In contrast, FDA only focuses on the selected regions and extracts more clean information for answering
the questions. This is the main reason that FDA can outperform SAN although
these two methods are both based on attention models.
The advantage of employing focused dynamic attention in FDA is more significant when solving the multiple-choice VQA problems. From Table 2, one
can observe that our proposed FDA model achieves the best ever performance
on the VQA dataset. In particular, it improves the performance of the stateof-the-art (D-LSTM) by a margin of 1.1% which is quite significant for this
challenging task. The D-LSTM method employs a deeper network to enhance
the discriminative capacity of the visual features. However, they do not identify
the informative regions for answering the questions. In contrast, FDA incorporates the automatic region localization by employing a question-driven attention
model. This is helpful for filtering out irrelevant noise, and establishing the correspondence between regions and candidate answers. Thus FDA gives substantial
performance improvement.
5.5
Qualitative Results
We qualitatively evaluate our model on a set of examples where complex reasoning and focusing on the relevant local visual features are needed for answering
the question correctly.
Figure 3 shows particularly difficult examples (the predominant image color
is not the correct answer) of “What color” type of questions. But, by focusing
on the question related image regions, the FDA model is still able to produce
the correct answer.
In Figure 4 we show examples where the model focuses on different regions
from the same image, depending on the words in the question. Focusing on the
right image region is crucial when answering unusual questions for an image
(Row 1), questions about small image objects (Row 2), or when the most dominant image object partly occludes the question related region and can lead to a
wrong answer (Row 3).
Representative examples of questions that require image object identification
are shown in Figure 5. We can observe that the focused attention enables the
model to answer complex questions (Row 1, left) and counting questions (Row 1,
right). The question guided image object identification greatly simplifies the
answering of questions like the ones shown in Row 2 and Row 3.
A Focused Dynamic Attention Model for Visual Question Answering
What color is the frisbee?
- Red.
What color are the glass items?
- Green.
What color is the mouse?
- White.
What color is the lady’s umbrella?
- Blue
11
Fig. 3. Representative examples where focusing on the question related objects helps
FDA answer “What color” type of questions. The question words in bold have been
matched with an image region. The yellow region caption box contains the question
word, followed by the region label, and in parenthesis their cosine similarity (see Section 4.3 for more details).
12
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
Is this a birthday cake?
- Yes.
Is someone in all likelihood, a zoo
fancier? - Yes.
What fruit is by the sink?
- Apples.
Is there a cookbook in the picture?
- Yes.
What type of vehicle is pictured?
- Motorcycle.
Does the elephant have tusks?
- No.
Fig. 4. Representative examples where the model focuses on different regions from the
same image, depending on the question. The question words in bold have been matched
with an image region. The yellow region caption box contains the question word, followed by the region label, and in parenthesis their cosine similarity (see Section 4.3 for
more details).
A Focused Dynamic Attention Model for Visual Question Answering
13
Is the bus in the middle of the intersec- How many cows are present?
tion? - Yes.
- 4.
Does this dessert have any fruit with
it? - Yes.
Is this a carrot cake?
- Yes.
Is this an airport?
- Yes.
Is the individual skiing or snowboarding? - Snowboarding.
Fig. 5. Representative examples of questions that require image object identification.
The question words in bold have been matched with an image region. The yellow region
caption box contains the question word, followed by the region label, and in parenthesis
their cosine similarity (see Section 4.3 for more details).
14
6
Ilija Ilievski, Shuicheng Yan, Jiashi Feng
Conclusion
In this work, we proposed a novel Focused Dynamic Attention (FDA) model to
solve the challenging VQA problems. FDA is built upon a generic object-centric
attention model for extracting question related visual features from an image
as well as a stack of multiple LSTM layers for feature fusion. By only focusing on the identified regions specific for proposed questions, FDA was shown to
be able to filter out overwhelming irrelevant informations from cluttered background or other regions, and thus substantially improved the quality of visual
representations in the sense of answering proposed questions. By fusing cleaned
regional representation, global context and question representation via LSTM
layers, FDA provided significant performance improvement over baselines on the
VQA benchmark datasets, for both the open-ended and multiple-choices VQA
tasks. Excellent performance of FDA clearly demonstrates its stronger ability of
modeling visual contents and also verifies paying more attention to visual part
in VQA tasks could essentially improve the overall performance. In the future,
we are going to further explore along this research line and investigate different
attention methods for visual information selection as well as better reasoning
model for interpreting the relation between visual contents and questions.
References
1. Geman, D., Geman, S., Hallonquist, N., Younes, L.: Visual turing test for computer
vision systems. Proceedings of the National Academy of Sciences 112(12) (2015)
3618–3623
2. Malinowski, M., Fritz, M.: A multi-world approach to question answering about
real-world scenes based on uncertain input. In: Advances in Neural Information
Processing Systems. (2014) 1682–1690
3. Ren, M., Kiros, R., Zemel, R.: Exploring models and data for image question
answering. In: Advances in Neural Information Processing Systems. (2015) 2935–
2943
4. Gao, H., Mao, J., Zhou, J., Huang, Z., Wang, L., Xu, W.: Are you talking to a
machine? dataset and methods for multilingual image question. In: Advances in
Neural Information Processing Systems. (2015) 2287–2295
5. Malinowski, M., Rohrbach, M., Fritz, M.: Ask your neurons: A neural-based approach to answering questions about images. In: Proceedings of the IEEE International Conference on Computer Vision. (2015) 1–9
6. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Lawrence Zitnick, C., Parikh,
D.: Vqa: Visual question answering. In: The IEEE International Conference on
Computer Vision (ICCV). (December 2015)
7. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision–
ECCV 2014. Springer (2014) 740–755
8. Noh, H., Seo, P.H., Han, B.: Image question answering using convolutional neural network with dynamic parameter prediction. arXiv preprint arXiv:1511.05756
(2015)
A Focused Dynamic Attention Model for Visual Question Answering
15
9. Wu, Q., Wang, P., Shen, C., Hengel, A.v.d., Dick, A.: Ask me anything: Freeform visual question answering based on knowledge from external sources. arXiv
preprint arXiv:1511.06973 (2015)
10. Ma, L., Lu, Z., Li, H.: Learning to answer questions from image using convolutional
neural network. arXiv preprint arXiv:1506.00333 (2015)
11. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W.,
Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural
computation 1(4) (1989) 541–551
12. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems.
(2012) 1097–1105
13. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556 (2014)
14. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition. (2015) 1–9
15. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation
9(8) (1997) 1735–1780
16. Zhou, B., Tian, Y., Sukhbaatar, S., Szlam, A., Fergus, R.: Simple baseline for
visual question answering. arXiv preprint arXiv:1512.02167 (2015)
17. Shih, K.J., Singh, S., Hoiem, D.: Where to look: Focus regions for visual question
answering. arXiv preprint arXiv:1511.07394 (2015)
18. Chen, K., Wang, J., Chen, L.C., Gao, H., Xu, W., Nevatia, R.: Abc-cnn: An
attention based convolutional neural network for visual question answering. arXiv
preprint arXiv:1511.05960 (2015)
19. Yang, Z., He, X., Gao, J., Deng, L., Smola, A.: Stacked attention networks for
image question answering. arXiv preprint arXiv:1511.02274 (2015)
20. Xu, H., Saenko, K.: Ask, attend and answer: Exploring question-guided spatial
attention for visual question answering. arXiv preprint arXiv:1511.05234 (2015)
21. Zitnick, C.L., Dollár, P.: Edge boxes: Locating object proposals from edges. In:
Computer Vision–ECCV 2014. Springer (2014) 391–405
22. Jiang, A., Wang, F., Porikli, F., Li, Y.: Compositional memory for visual question
answering. arXiv preprint arXiv:1511.05676 (2015)
23. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Learning to compose neural
networks for question answering. arXiv preprint arXiv:1601.01705 (2016)
24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
arXiv preprint arXiv:1512.03385 (2015)
25. Pont-Tuset, J., Arbeláez, P., Barron, J., Marques, F., Malik, J.: Multiscale combinatorial grouping for image segmentation and object proposal generation. In:
arXiv:1503.00848. (March 2015)
26. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Advances in neural
information processing systems. (2013) 3111–3119
27. Jiasen Lu, Xiao Lin, D.B., Parikh, D.: Deeper lstm and normalized cnn visual
question answering model. https://github.com/VT-vision-lab/VQA_LSTM_CNN
(2015)
| 9cs.NE
|
CONSTRUCTING THE SET OF COMPLETE INTERSECTION NUMERICAL
SEMIGROUPS WITH A GIVEN FROBENIUS NUMBER
arXiv:1204.4258v3 [math.CO] 21 Jan 2013
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
Abstract. Delorme suggested that the set of all complete intersection numerical semigroups can
be computed recursively. We have implemented this algorithm, and particularized it to several
subfamilies of this class of numerical semigroups: free and telescopic numerical semigroups, and
numerical semigroups associated to an irreducible plane curve singularity. The recursive nature of
this procedure allows us to give bounds for the embedding dimension and for the minimal generators
of a semigroup in any of these families.
1. Introduction
Let N denote the set of nonnegative integers. A numerical semigroup Γ is a submonoid of N with
finite complement in N (this condition is equivalent to gcd(Γ) = 1). If Γ is a numerical semigroup,
the elements in N \ Γ are the gaps of Γ. The cardinality of N \ Γ is the genus of Γ, g(Γ). The
largest integer not in Γ is called the Frobenius number of Γ, and will be denoted by F(Γ). Clearly,
F(Γ) + 1 + N ⊆ Γ, and this is why c(Γ) = F(Γ) + 1 is known as the conductor of Γ.
Since for every x ∈ Γ, F(Γ) − x cannot be in Γ, we deduce that g(Γ) ≥ c(Γ)
2 . We say that Γ is
symmetric when the equality holds, or equivalently, for every integer x, x 6∈ Γ implies F(Γ) − x ∈ Γ.
In this setting, c(Γ) is an even integer, and thus F(Γ) is odd.
It can be easily proved that any numerical semigroup admits a unique minimal generating system
(every element is a linear combination of elements in this set with nonnegative integer coefficients and
none of its proper subsets fulfills this condition; see for instance [16, Chapter 1]). If A = {r0 , . . . , rh }
is the minimal generating set of Γ, then its elements are called minimal generators, and its cardinality
is the embedding dimension of Γ, e(Γ). The smallest minimal generator is the smallest positive
integer belonging to the semigroup, and it is known as the multiplicity of Γ, denoted by m(Γ).
The map
Ne(Γ) → Γ, ϕ(a0 , . . . , ah ) = a0 r0 + · · · + ah rh
is a monoid epimorphism. Hence Γ is isomorphic to Ne(Γ) / ker ϕ, where ker ϕ = {(a, b) ∈ Ne(Γ) ×
Ne(Γ) | ϕ(a) = ϕ(b)} (ker ϕ is a congruence on Ne(Γ) ).
A presentation for Γ is a set of generators of the congruence ϕ, and a minimal presentation is
a set of generators minimal with respect to set inclusion (actually, in our setting also with respect
to cardinality; see [16, Corollary 8.13]). It can be shown that the cardinality of any minimal
presentation is greater than or equal to e(Γ) − 1, [16, Theorem 9.6]. A numerical semigroup is a
complete intersection if this equality holds.
Given A a set positive integers, and A = A1 ∪ A2 a non trivial partition of A, we say that A
is the gluing of A1 and A2 if lcm(d1 , d2 ) ∈ hA1 i ∩ hA2 i, where di = gcd(Ai ) and hAi i denotes
the monoid generated by Ai , i = 1, 2. If A is the minimal system of generators of Γ, and Γi is
the numerical semigroup generated by Ai /di , i = 1, 2, we also say that Γ is the gluing of Γ1 and
The first author is partially supported by the project GDR CNRS 2945.
The second author is supported by the projects MTM2010-15595, FQM-343, FQM-5849, and FEDER funds. This
research was performed while the second author visited the Université d’Angers as invited lecturer, and he wants to
thank the Département de Mathématiques of this university for its kind hospitality. The authors would like to thank
the referees for their suggestions, comments and examples provided.
1
2
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
Γ2 . It turns out that d1 ∈ Γ2 , d2 ∈ Γ1 , gcd(d1 , d2 ) = 1, and neither d1 is a minimal generator of
Γ2 nor d2 is a minimal generator of Γ1 ([16, Section 8.3]). Delorme proved in [5, Proposition 9]
that a numerical semigroup is a complete intersection if and only if it is a gluing of two complete
intersection numerical semigroups (though with a different notation; the concept of gluing was
introduced in [13]). The gluing of symmetric numerical semigroups is symmetric ([5, Proposition
10 (iii)]), and as a consequence of this, complete intersections are symmetric.
In [17] there is a procedure to construct the set of all numerical semigroups with a given Frobenius
number. We show in this manuscript how can we use the concept of gluing to compute the set of all
complete intersection numerical semigroups with a given Frobenius number (or equivalently with
fixed genus). Recently there have been some experimental results that point out to the possibility
that the number of numerical semigroups with a fixed genus has a Fibonacci like behaviour ([4]).
Indeed, it is known that asymptotically the number of numerical semigroups with given genus grows
as the Fibonacci sequence ([18]). However there is not a proof for this for all genus, and we still
do not even have a demonstration that there are more numerical semigroups with genus g + 1
than numerical semigroups with genus g. This is not the case for complete intersection numerical
semigroups, as we see in the last section.
We also show how to calculate the set of all free (in the sense of [1]) numerical semigroups, which
is a special subclass of complete intersections, the set of all telescopic numerical semigroups (contained in the set of free numerical semigroups), and that of numerical semigroups associated to an
irreducible plane curve singularity (these are a particular case of telescopic numerical semigroups).
The recursive nature of gluing also allows us to give some bounds for the generators and embedding dimension for these families of semigroups when we fix the Frobenius number. The deeper we
go in the chain of inclusions given in the preceding paragraph, the smaller are the bounds.
2. The Frobenius number and multiplicity of a complete intersection
Let Γ be a numerical semigroup. We know that Γ is a complete intersection if and only if it is
the gluing of two complete intersections. Delorme (though with a different notation) highlighted
in [5, Section 11] that this fact can used to determine if a numerical semigroup is a complete
intersection (this idea has already been exploited in [2]; and in [15] one can find a procedure to
determine if an affine semigroup is the gluing of two affine semigroups), and also to compute the
set of all complete intersections. In order to construct the set of all complete intersection numerical
semigroups with given Frobenius number, we can proceed recursively by using the following formula
for the Frobenius number of the gluing of two numerical semigroups, which is just a reformulation
of Delorme’s description of the conductor of a gluing.
Proposition 1. Assume that Γ is a numerical semigroup minimally generated by A = A1 ∪ A2 ,
and that A is the gluing of A1 and A2 . Let d1 = gcd(A1 ) and d2 = gcd(A2 ). Define Γ1 = hA1 /d1 i
and Γ2 = hA2 /d2 i. Then
F(Γ) = d1 F(Γ1 ) + d2 F(Γ2 ) + d1 d2 .
Proof. Observe that Γ = d1 Γ1 + d2 Γ2 . By [5, Proposition 10 (i)],
(1)
c(Γ) = d1 c(Γ1 ) + d2 c(Γ2 ) + (d1 − 1)(d2 − 1).
Having in mind the relationship between Frobenius number and conductor, the formula follows
easily.
In Proposition 1, Γ = d1 Γ1 +d2 Γ2 and d = d1 d2 = d1 Γ1 ∩d2 Γ2 . The integer d is the element where
the gluing takes place. If we repeat the process with d1 Γ1 and d2 Γ2 in this result, we construct a
decomposition tree of Γ, whose leaves are copies isomorphic to of N (this was the idea followed in
[3]). Assume that d(1) , . . . , d(h) are the elements where the gluings take place in this splitting. The
P
P
Frobenius number of Γ is precisely hi=1 d(i) − a∈A a (see [5, Section 11] where it is highlighted
that this formula is a particular case of a result given in [10]).
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
3
Example 2. Let Γ = h10, 14, 15, 21i. Then Γ = h10, 15i + h14, 21i and 35 = 5 × 7 ∈ h10, 15i. We
repeat the process for h10, 15i = h10i + h15i and h14, 21i = h14i + h21i. We get 30 ∈ h10i ∩ h15i
and 42 ∈ h14i ∩ h21i. Hence the gluings take place at 35, 30 and 42. Thus F(Γ) = (35 + 30 + 42) −
(10 + 14 + 15 + 21) = 47.
Example 3. We construct a complete intersection numerical semigroup with four generators, by
gluing two embedding dimension two numerical semigroups.
gap> s:=NumericalSemigroup(10,11);;
gap> t:=NumericalSemigroup(7,9);;
gap> g:=NumericalSemigroup(16*10,16*11,21*7,21*9);;
gap> FrobeniusNumber(g);
2747
gap> 16*FrobeniusNumber(s)+21*FrobeniusNumber(t)+16*21;
2747
Remark 4. For Γ = N, we have
c(N) = 0,
F(N) = −1,
g(N) = 0,
m(N) = 1,
e(N) = 1.
Proposition 5. If Γ is a complete intersection, then
m(Γ) ≥ 2e(Γ)−1 .
Proof. Let h = e(Γ) − 1. We use induction on h. For h = 1, the statement follows trivially.
As Γ is a complete intersection, if A is its minimal set of generators, we can find a partition of
A = A1 ∪ A2 such that A is the gluing of A1 and A2 . Set as above di = gcd(Ai ), and Γi = hAi /di i.
Let hi = e(Γi )−1. Hence h = h1 +h2 +1. By induction hypothesis m(Γi ) ≥ 2hi . Recall that d1 ∈ Γ2
and d2 ∈ Γ1 , and they are not minimal generators. Thus d1 ≥ 2m(Γ2 ) ≥ 2h2 +1 , and analogously
d2 ≥ 2h1 +1 . For every a ∈ A1 , a/d1 is a minimal generator of Γ1 , whence a/d1 ≥ 2h1 . Therefore
a ≥ 2h1 +h2 +1 = 2h . The same argument shows that any element in A2 is greater than or equal to
2h .
Example 6. We construct recursively a family {Γ(n) }n∈N of complete intersection numerical semigroups reaching the bound of Proposition 5.
We start with Γ(1) = h2, 3i, and the general element in the sequence is defined as Γ(n+1) =
2Γ(n) + (2n+1 + 1)N.
For instance, Γ(2) = 2h2, 3i + 5N = h4, 5, 6i, Γ(3) = 2h4, 5, 6i + 9N = h8, 9, 10, 12i, and so on.
It is not hard to prove that
Γ(n+1) = h2n+1 , 2n+1 + 1, 2n+1 + 2, 2n+1 + 22 , . . . , 2n+1 + 2n i
= 2h2n , 2n + 1, . . . , 2n + 2n−1 i + (2n+1 + 1)N.
Notice that Γ(n+1) is a gluing of Γ(n) and N, since
• 2 ∈ N and 2 is not a minimal generator of N,
• 2n+1 + 1 is the sum of the two smallest minimal generators of Γ(n) ; thus 2n+1 + 1 belongs
to Γ(n) and it is not a minimal generator of Γ(n) ,
• gcd(2, 2n+1 + 1) = 1.
It follows that m(Γ(n) ) = 2n and e(Γ(n) ) = n + 1. Thus the bound in Proposition 5 is attained.
Corollary 7. If Γ is a complete intersection numerical semigroup other than N, then
e(Γ) ≤ log2 (c(Γ)) + 1.
Proof. By Proposition 5, 2e(Γ)−1 ≤ m(Γ). Since Γ 6= N, we have that m(Γ) ≤ c(Γ), and the bound
follows.
4
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
Remark 8. Notice that in the proof of Corollary 7 we use m(Γ) ≤ c(Γ). For Γ = h2, 3i, we get an
equality and also the bound given in this corollary is reached. If m(Γ) = c(Γ), then Γ = hm, m +
1, . . . , 2m − 1i, with m = m(Γ). Hence e(Γ) = m, that is, Γ has maximal embedding dimension (it is
easy to see that the embedding dimension of a numerical semigroup is always less than or equal to
its multiplicity; see for instance [16, Chapter 1]). It is well known that the cardinality of a minimal
presentation of a maximal embedding dimension numerical semigroup with multiplicity m is m(m−1)
2
(see for instance [16, Corollary 8.27]). Hence a maximal embedding dimension numerical semigroup
= m − 1, or equivalently, either
with multiplicity m is a complete intersection if and only if m(m−1)
2
the numerical semigroup is N or m = 2. If in addition we impose that the conductor and the
multiplicity agree, then the only two possibilities are N and h2, 3i.
From the definitions of multiplicity and conductor, it is easy to see that there is no numerical
semigroup Γ such that c(Γ) = 1 + m(Γ).
If c(Γ) = 2+ m(Γ), then Γ = hm, m + 2, m + 3, . . . , 2m − 1, 2m + 1i, which is a maximal embedding
dimension numerical semigroup. So the only complete intersection with c(Γ) = 2 + m(Γ) is h2, 5i.
The case c(Γ) = 3 + m(Γ) requires more effort. In this setting m = m(Γ) > 2. We have two
possibilities.
• Γ = hm, m + 3, m + 4, . . . , 2m − 1, 2m + 1, 2m + 2i, which has maximal embedding dimension,
and so it cannot be a complete intersection numerical semigroup, because m > 2.
• Γ = hm, m + 1, m + 3, m + 4, . . . , 2m − 1i. Here e(Γ) = m − 1 and the minimum element in
Γ congruent with 2 modulo m is 2m + 1 = (m + 1) + (m + 1). Thus in view of [14, Theorem
. We conclude that Γ is
1(2)], the cardinality of a minimal presentation for Γ is (m−1)(m−2)
2
(m−1)(m−2)
a complete intersection if and only if
= m − 2, and as m > 2, this is equivalent
2
to m = 3. Hence Γ = h3, 4i.
Therefore, if we assume that Γ 6∈ {N, h2, 3i, h2, 5i, h3, 4i}, and Γ is a complete intersection numerical semigroup, then we can assert that c(Γ) ≥ m(Γ) + 4, and the bound in Corollary 7 can be
slightly improved to
e(Γ) ≤ log2 (c(Γ) − 4) + 1.
This bound is attained for instance by h2, 7i, h4, 5, 6i and h4, 6, 7i.
By using [14, Section 1.2], we can determine those complete intersections with c(Γ) = m(Γ) + 4,
and thus obtain another small improvement of the above bound.
We can improve this bound by using a different strategy.
Proposition 9. Let Γ be a complete intersection numerical semigroup. Then
(e(Γ) − 1)2e(Γ)−1 ≤ c(Γ).
Proof. We use induction on the embedding dimension of Γ. If the embedding dimension of Γ is
either one or two, then the result holds trivially. So assume that e(Γ) ≥ 3. As Γ is a complete
intersection, we know that there exist two complete intersection numerical semigroups Γ1 and Γ2
such that Γ is the gluing of Γ1 and Γ2 . Thus there exist d1 ∈ Γ2 and d2 ∈ Γ1 , that are not minimal
generators, such that Γ = d1 Γ1 + d2 Γ2 . For sake of simplicity write c = c(Γ), e = e(Γ), ci = c(Γi )
and ei = e(Γi ), i = 1, 2. Then from the definition of gluing we already know that e = e1 + e2 .
Since e ≥ 3, we may assume without loss of generality that e1 ≥ 2. As d1 is not a minimal
generator of Γ2 , d1 ≥ 2m(Γ2 ), and as e1 ≥ 2 and d2 is not a minimal generator of Γ1 , d2 ≥ 2m(Γ1 )+1.
In view of Proposition 5, we deduce d1 ≥ 2e2 and d2 ≥ 2e1 + 1.
Now from (1), we have c = d1 c1 +d2 c2 +(d1 −1)(d2 −1). By induction hypothesis and the preceding
paragraph, we get c ≥ 2e2 (e1 − 1)2e1 −1 + (2e1 + 1)(e2 − 1)2e2 −1 + (2e2 − 1)2e1 = (e − 2)2e−1 + (e2 −
1)2e2 −1 + 2e − 2e1 ≥ (e − 1)2e−1 − 2e−1 + 2e − 2e1 = (e − 1)2e−1 + 2e−1 − 2e1 ≥ (e − 1)2e−1 .
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
5
Example 10. Let {Γ(n) }n∈N be the family of numerical semigroups presented in Example 6. By
using (1), it is not hard to check inductively that c(Γ(n) ) = n2n , and thus the bound of Proposition
9 is attained.
If we have a closer look at the proof of Proposition 9, then we easily deduce that for the bound
to be attained, the following must hold in all induction steps with e ≥ 3:
• (e2 − 1)2e2 −1 = 0 and thus e2 = 1, that is, Γ2 is N (we will study these semigroups in the
next section);
• from e2 = 1 it follows that e1 = e − 1 and 2e−1 − 2e1 = 0;
• m(Γ1 ) = 2e1 −1 and d2 = 2m(Γ1 ) + 1 = 2e1 + 1, whence m(Γ1 ) + 1 ∈ Γ1 ;
• c1 = (e1 − 1)2e1 −1 = (e − 2)2e−2 ;
• d1 = 2.
Also the only embedding dimension two numerical semigroup for which the equality holds is h2, 3i.
If follows that the family given in Example 6 contains all possible complete intersection numerical
semigroups with the property that the bound in Proposition 9 becomes an equality.
Proposition 11. Let Γ be a complete intersection numerical semigroup other than N, minimally
generated by {r0 , . . . , rh }. If m(Γ) 6= 2, for all k, rk < F(Γ).
Proof. Assume without loss of generality that r0 = m(Γ). The numerical semigroup Γ is symmetric
and thus for every i > 0, F(Γ) + r0 − ri ∈ Γ. If rk > F(Γ), for some k > 0, then F(Γ) + r0 − rk < r0 ,
which forces F(Γ) + r0 = rk .
If h > 1, choose 0 < i 6= k. Then rk − ri = F(Γ) + r0 − ri ∈ Γ, contradicting that rk is a minimal
generator. This proves rk < F(Γ), whenever h > 1.
For h = 1, F(Γ) = (r0 − 1)(r1 − 1) − 1. In this setting, Γ = h2, f + 2i has F(Γ) = f . For m(Γ) > 2,
we get F(Γ) = (r0 − 1)(r1 − 1) − 1 ≥ 2(r1 − 1) − 1 = (r1 − 1) + (r1 − 2) ≥ r1 .
Remark 12. If we want to compute the set of all complete intersection numerical semigroups with
Frobenius number f , then we can use the formula given in Proposition 1. Hence f = d1 f1 + d2 f2 +
d1 d2 , and we recursively construct all possible numerical semigroups with Frobenius number f1 , and
then the set with Frobenius number f2 . We next give some useful bounds and facts to perform this
task. Denote f + 1 by c.
i) d1 6= 1 6= d2 . This is because d1 ∈ Γ2 and it is not a minimal generator of Γ2 . The only
possibility to have d1 = 1 ∈ Γ2 would be Γ2 = N = h1i. But then d1 would be a minimal
generator. The same argument is valid for d2 .
ii) Since gcd(d1 , d2 ) = 1, we can assume without loss of generality that 2 ≤ d2 < d1 .
iii) Since f1 , f2 ≥ −1, f ≥ −d1 − d2 + d1 d2 = (d1 − 1)(d2 − 1) − 1. Hence d2 ≤ d1c−1 + 1; and
consequently, d2 ≤ min{d1 − 1, d1c−1 + 1}.
iv) f − dj fj ≡ 0 mod di , {i, j} = {1, 2}. In particular, if fj = −1, then f + dj ≡ 0 mod di .
v) d1 < f , except in the case Γ = h2 = d2 , f + 2 = d1 i.
a) If f1 = f2 = −1, then Γ1 = Γ2 = N, and Γ is hd2 , d1 i. If d2 6= 2, then Proposition 11, asserts
that d1 < f .
b) If f2 > 0, then f ≥ −d1 + d2 + d1 d2 = (d1 + 1)(d2 − 1) + 1 ≥ d1 + 2. Hence d1 ≤ f − 2.
c) If f1 > 0, then f ≥ d1 − d2 + d1 d2 = (d1 − 1)(d2 + 1) + 1 > 3(d1 − 1) ≥ d1 + 2(d1 − 1) − 1 > d1 .
vi) If f1 6= −1 6= f2 , then f − d1 d2 ∈ hd1 , d2 i. We are only interested in factorizations f − d1 d2 =
a1 d1 + a2 d2 , a1 , a2 ∈ N, with a1 ≡ a2 ≡ 1 mod 2, since the Frobenius number of a complete
intersection is an odd integer.
Example 13. We compute the set of all complete intersection numerical semigroups with Frobenius
number 11. First note that h2, 13i is in this set. The possible d1 belong to {3, . . . , 10}.
• d1 = 10. Then 2 ≤ d2 ≤ min{9, ⌊ 12
9 ⌋+ 1} = 2. Hence d2 must be 2, but then gcd(d1 , d2 ) 6= 1,
and we have no complete intersections under these conditions.
6
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
• d1 = 9. Then 2 ≤ d2 ≤ min{8, ⌊ 12
8 ⌋ + 1} = 2. This forces d2 = 2, which in addition is
coprime with 9.
⋆ 11 + 9 ≡ 0 mod 2, and thus f1 = −1 (Γ1 = N) is a possible choice. In this setting
f2 = (11 − 18 + 0)/2 = 1, whence Γ2 = h2, 3i. We obtain a new complete intersection
Γ = 9N + 2h2, 3i = h4, 6, 9i, because 9 ∈ h2, 3i is not a minimal generator.
⋆ 11 + 2 6≡ 0 mod 9, so f2 cannot be −1.
⋆ 11 − 18 6∈ h2, 9i, so we have no more complete intersections with this data.
• For d1 = 8, we have 2 ≤ d2 ≤ min{7, ⌊ 12
6 1.
7 ⌋ + 1} = 2. However gcd{d1 , d2 } =
12
• If d1 = 7, then 2 ≤ d2 ≤ min{6, ⌊ 6 ⌋ + 1} = 3.
⋆ d2 = 2.
∗ 11 + 7 ≡ 0 mod 2, and thus Γ1 can be N. But then f2 = (11 − 14 + 7)/2 = 2,
which is even. So this case cannot occur.
∗ 11 + 2 6≡ 0 mod 7, and so Γ2 will not be N.
∗ Finally, 11−14 6∈ h2, 7i, so no complete intersections can be found with properties.
⋆ d2 = 3.
∗ 11 + 7 ≡ 0 mod 3, and thus Γ1 could be N. In this setting f2 = (11 − 21 + 7)/3 =
−1, and so Γ2 is also N. We get a new complete intersection Γ = 7N + 3N = h3, 7i
with Frobenius number 11.
∗ 11 − 21 6∈ h3, 7i, so no more complete intersections are obtained for this choice of
d1 and d2 .
• For d1 = 6, 2 ≤ d2 ≤ min{5, ⌊ 12
5 ⌋ + 1} = 3, but both 2 and 3 are not coprime with 6.
• d1 = 5. Then d2 ∈ {2, 3, 4}.
⋆ d2 = 2.
∗ 11 + 5 ≡ 0 mod 2, and so Γ1 can possibly be N. Hence f2 = (11 − 10 + 5)/2 =
3. The only possible complete intersection numerical semigroup with Frobenius
number 3 is h2, 5i. But 5 is a minimal generator of this semigroup.
∗ 11 + 2 6≡ 0 mod 5.
∗ 11 − 10 6∈ h2, 5i.
⋆ d2 = 3. In this case 11 + 5 6≡ 0 mod 3, 11 + 3 6≡ 0 mod 5, and 11 − 15 6∈ h3, 5i.
⋆ d2 = 4.
∗ 11 + 5 ≡ 0 mod 4, and f2 = (11 − 20 + 5)/4 = −1. So Γ = 5N + 4N = h4, 5i is
another complete intersection with Frobenius number 11.
∗ 11 − 20 6∈ h4, 5i.
• d1 = 4, 2 ≤ d2 ≤ min{3, ⌊ 12
3 ⌋ + 1} = 3, and as gcd(2, 4) 6= 1, we get d2 = 3.
⋆ 11 + 4 ≡ 0 mod 3. So Γ1 could be N. If this is the case, f2 = (11 − 12 + 4)/3 = 1,
which forces Γ2 to be h2, 3i, and 4 ∈ Γ2 is not a minimal generator. So we obtain
Γ = 4N + 3h2, 3i = h4, 6, 9i, which was already computed before.
⋆ 11 + 3 6≡ 0 mod 4.
⋆ 11 − 12 6∈ h3, 4i
• d3 = 3 and d2 = 2.
⋆ 11 + 3 ≡ 0 mod 2, and Γ1 = N can be a possibility. Then f2 = (11 − 6 + 3)/2 = 7. If we
apply this procedure recursively for f = 7, we obtain that {h2, 9i, h3, 5i, h4, 5, 6i} is the
set of all possible complete intersection numerical semigroups with Frobenius number
7. However, 3 6∈ h2, 9i, 3 is a minimal generator of h3, 5i, and 3 6∈ h4, 5, 6i.
⋆ 11 + 2 6≡ 0 mod 3.
⋆ 11 − 6 = 5 ∈ h2, 3i, and 5 = 1 · 2 + 1 · 3 is the only factorization. So the only possible
choice for f1 and f2 is 1. This means that Γ1 and Γ2 must be h2, 3i. Again we obtain
no new semigroups, since 2 and 3 are minimal generators of h2, 3i.
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
7
Thus the set of complete intersection numerical semigroups with Frobenius number 11 is
{h2, 13i, h4, 6, 9i, h3, 7i, h4, 5i}.
3. Free numerical semigroups
Throughout this section, let Γ be the numerical semigroup Γ minimally generated by {r0 , . . . , rh }.
For k ∈ {1, . . . D
, h + 1}, set dk E= gcd({r0 , . . . , rk−1 }) (d1 = r0 ).
rk
r0
Write Γk = dk+1
, and ck = c(Γk ) for all k ∈ {1, . . . , h}. Set c = ch = c(Γ).
, . . . , dk+1
We say that Γ is free if either h = 0 (and thus r0 = 1) or Γ is the gluing of the free numerical semigroup Γh−1 and N. Free numerical semigroups were introduced in [1]. For other characterizations
and properties of free numerical semigroups see [16, Section 8.4].
Example 14. Notice that the order in which the generators are given is crucial. For instance,
S = h8, 10, 9i is free for the arrangement (8, 10, 9) but it is not free for (8, 9, 10). And a numerical
semigroup can be free for different arrangements, for example, S = h4, 6, 9i has this property.
Q
If we take c0 , . . . , ch pairwise coprime integers greater than one, and ri = hj=0,i6=j cj , j = 0, . . . , h,
then the numerical semigroup generated by {r0 , . . . , rh } is free for any arrangement of its minimal
generating set (see [9]).
According to Proposition 1, with A2 = {rh }, we obtain the following consequence.
Corollary 15. If Γ is free, then
F(Γ) = dh F(Γh−1 ) + rh (dh − 1).
In this way we retrieve Johnson’s formula ([11]). Notice also that Γh−1 is again free, so if
we expand recursively this formula we obtain the formula given by Bertin and Carbonne for free
numerical semigroups (see [1]; these authors named these semigroups in this way).
This equation can be reformulated in terms of the conductor as
(2)
c(Γ) = ch = dh ch−1 + (dh − 1)(rh − 1).
Lemma 16. If Γ is free, then
(1) gcd(dh , rh ) = 1;
(2) dh | F(Γ) + rh (consequently dh 6 | F(Γ));
dk
, k = 1, . . . , h, then ek rk ∈ hr0 , . . . , rk−1 i, for all k = 1, . . . , h; in
(3) if we define ek = dk+1
particular, ek ≥ 2;
(4) d1 > d2 > · · · > dh+1 = 1;
(5) dh ≤ rc(Γ)
+ 1;
h −1
(6) for h ≥ 1, (dh − 1)(rh − 1) ≥ 2h .
Proof.
(2)
(3)
(4)
(5)
(6)
(1) This follows from the fact that Γ is a numerical semigroup, and thus gcd(dh , rh ) =
dh+1 = 1.
F(Γh−1 ) = (F(Γ) + rh (1 − dh ))/dh = (F(Γ) + rh )/dh − 1.
rk
dk
As Γk is the gluing of Γk−1 and N, we have that dk+1
∈ Γk−1 . Hence dk−1
rk ∈ hr0 , . . . , rk−1 i.
If ek = 1, then rk ∈ hr0 , . . . , rk−1 i, contradicting that rk is a minimal generator.
dk
By definition, dk ≥ dk+1 . As ek = dk+1
≥ 2, we get dk > dk−1 .
Notice that F(Γ) ≥ (rh − 1)(dh − 1) − 1, since F(Γh−1 ) ≥ −1.
If dh = 2, then we show that rh > m(Γ). Assume to the contrary that rh = m(Γ). Then we
already proved above that eh rh ∈ hr0 , . . . , rh−1 i. Since eh = dh and ri is a minimal generator
P
P
of Γ for all i, we deduce that 2rh = h−1
a r , with h−1
a ≥ 2. As ri > rh for every
Ph−1 i=0 i i
Ph−1i=0 i
i = 0, . . . , h − 1, we get 2rh > rh i=0 ai , and thus i=0 ai < 2, a contradiction. Thus in
view of Proposition 5, we have that rh ≥ 2h , and if dh = 2, then rh ≥ 2h +1. Hence for dh = 2
8
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
the proof follows easily, and for dh > 2 we get (dh − 1)(rh − 1) ≥ 2(rh − 1) ≥ 2(2h − 1) ≥ 2h
(we are assuming h ≥ 1).
In view of Example 10, the bound proposed in Proposition 9 cannot be improved for free numerical semigroups, since the family introduced in Example 6 consists on free numerical semigroups.
However, we can use Proposition 9 to find an upper bound for rh , as we show next.
h −1)
For all h ≥ 2, ch−1 = c−(dh −1)(r
is an even integer, and c = ch ≥ h2h . In particular,
dh
−ch−1 dh ≤ −(h − 1)2h−1 dh .
Hence
(rh − 1)(dh − 1) ≤ c − (h − 1)2h−1 dh
This gives us the following upper bound for rh .
rh ≤
Corollary 17. For all h ≥ 2,
2h + 1 ≤ rh ≤
dh
c
− (h − 1)2h−1
+ 1.
dh − 1
dh − 1
c
dh
− (h − 1)2h−1
+ 1 ≤ c − (h − 1)2h−1 + 1.
dh − 1
dh − 1
Remark 18. In order to compute the set of all free numerical semigroups with a given Frobenius
number, we make use of the formula given in Corollary 15, by taking into account the restrictions
given in this section for dh and rh .
4. Telescopic numerical semigroups
We keep using the same notation as in the preceding section. We say that the numerical semigroup
Γ minimally generated by {r0 , . . . , rh } is telescopic if it is free for the arrangement of the generators
r0 < · · · < rh (see for instance [12]). This motivates the notation {r0 < · · · < rh }, that means
that the elements in the set {r0 , . . . , rh } fulfill the extra condition r0 < · · · < rh . We will also write
Γ = hr0 < · · · < rh i when {r0 , . . . , rh } is a generating system for Γ and r0 < · · · < rh .
Notice that in addition to the properties we had for free numerical semigroups, if Γ is telescopic,
then
(1) dh < rh , because dh | rh−1 < rh ;
o
n
p
c(Γ)
+
1
.
+
1,
(2) F(Γ) ≥ (rh − 1)(dh − 1) − 1 > (dh − 1)2 − 1, whence dh ≤ min rh − 1, rc(Γ)
h −1
Proposition 19. Let Γ be a telescopic numerical semigroup minimally generated by {r0 < · · · < rh }.
If h ≥ 2, then rh ≥ 2h+1 − 1.
D
E
Proof. Let h = 2, and let Γ1 = dr02 , dr12 . Since dr12 ≥ 3 and d2 ≥ 2, we have r1 ≥ 6. Besides, r2 > r1 ,
whence r2 ≥ 7. Note that this bound is attained for Γ2 = h4, 6, 7i.
E h ≥ 3, and that the formula is true for h − 1. We have rh ≥ rh−1 + 1 and rh ∈
D Assume that
r
rh−1
r0
h
h
h+1 −1.
. By induction hypothesis, we have h−1
dh , . . . , dh
dh ≥ 2 −1. Hence rh ≥ 2(2 −1)+1 = 2
Note that this bound is reached by Γh = h2h , 3 · 2h−1 , 7 · 2h−2 , . . . , (2k+1 − 1) · 2h−k , . . . , 2h+1 − 1i.
As in the free case, we can describe a bound for the embedding dimension of a telescopic numerical
semigroup.
Proposition 20. Let Γ be a telescopic numerical semigroup other than N. Then
(e(Γ) − 2)2e(Γ) + 2 ≤ c(Γ).
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
9
Proof. Assume that Γ is minimally generated by {r0 < · · · < rh }. Denote as usual c(Γ) by c. We
use once more induction on h.
The case h = 1 is evident.
Suppose that h ≥ 2, and that our inequality is true for h − 1. By (2), we have c = dh ch−1 + (dh −
1)(rh − 1). By induction hypothesis, ch−1 ≥ (h − 2)2h + 2, and as dh ≥ 2, and rh ≥ 2h+1 − 1, we
get c ≥ (h − 2)2h+1 + 4 + 2h+1 − 2 = (h − 1)2h+1 + 2.
Note that for all h ≥ 2, ch−1 =
(h − 1)2h+1 + 2. In particular
Hence
c − (dh − 1)(rh − 1)
is an even integer, and that c = ch ≥
dh
−ch−1 dh ≤ − (h − 2)2h + 2 dh .
(rh − 1)(dh − 1) ≤ c − (h − 2)2h + 2 dh .
This gives us the following upper bound for rh :
dh
c
rh ≤
− (h − 2)2h + 2
+ 1 ≤ c − (h − 2)2h − 1.
dh − 1
dh − 1
Corollary 21. For all h ≥ 2, we have
dh
c
− (h − 2)2h + 2
+ 1 ≤ c − (h − 2)2h − 1.
2h+1 − 1 ≤ rh ≤
dh − 1
dh − 1
Remark 22. For computing the set of all telescopic numerical semigroups with fixed Frobenius
number, we proceed as in the free case, ensuring that rh is larger than the largest generator of Γ1
multiplied by dh . Notice that dh must now be smaller than rh .
5. Plane curve singularities
Let Γ be the numerical semigroup minimally generated by {r0 < r1 < . . . < rh }. Let dk , Γk ,
ck , and ek be as in the preceding section. The numerical semigroup Γ is the numerical semigroup
associated to an irreducible plane curve singularity if Γ is telescopic and ek rk < rk+1 for all k =
1, . . . , h − 1 (see [19]).
Proposition 23. Let Γ be the semigroup associated to an irreducible plane curve singularity minimally generated by {r0 < · · · < rh }, with h ≥ 2. Then rh ≥ 13 (5 · 22h−1 − 1).
D
E
Proof. For h = 2, as Γ1 = dr02 < dr12 , we obtain r21 ≥ 3. Since d2 ≥ 2, we deduce that r1 ≥ 6.
The plane singularity condition implies r2 > e1 r1 ≥ 12, because we know that e1 ≥ 2 (Lemma 16).
Hence r2 ≥ 13.
Assume that h ≥ 3, and that the formula is true for h − 1. The plane singularity condition
rh−1
r
for k = h − 1 implies that rh ≥ h−1
dh dh−1 + 1. The quotient dh is the largest generator of
r
1
2(h−1)−1 − 1). By using that
Γh−1 . The induction hypothesis then asserts that h−1
dh ≥ 3 (5 · 2
ek ≥ 2 for all k (Lemma 16), we deduce that dh−1 ≥ 4. By putting all this together, we get
rh ≥ 4( 13 (5 · 22(h−1)−1 − 1)) + 1 = 13 (5 · 22h−1 − 1).
Proposition 24. Let Γ 6= N be the semigroup associated to an irreducible plane curve singularity
minimally generated by {r0 < · · · < rh } and with conductor c. Then
4
5
c ≥ 22h − 3 · 2h + .
3
3
Proof. The case h = 1 is evident.
Assume that h ≥ 2 and that our inequality holds for h−1. We have: c = dh ch−1 +(dh −1)(rh −1).
By induction hypothesis ch−1 ≥ 35 22h−2 − 3 · 2h−1 + 43 . Notice that dh ≥ 2. Thus our assertion
follows from Proposition 23.
10
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
We proceed now as we did in the telescopic case to obtain also an upper bound for rh . Note that
h −1)
for all h ≥ 2, ch−1 = c−(dh −1)(r
is an even integer, and that ch−1 ≥ 35 22h−2 − 3 · 2h−1 + 43 . Thus
dh
4
5 2h−2
h−1
2
−3·2
+
dh .
−ch−1 dh ≤ −
3
3
Hence
(rh − 1)(dh − 1) ≤ c −
5 2h−2
4
2
− 3 · 2h−1 +
3
3
dh .
This gives us the following upper bound for rh .
c
5 2h−2
4
dh
rh ≤
−
2
− 3 · 2h−1 +
+ 1.
dh − 1
3
3 dh − 1
Corollary 25. For all h ≥ 2, we have
5 2h−2
c
4
5
7
dh
5 2h−1 1
2
− ≤ rh ≤
−
2
− 3 · 2h−1 +
+ 1 ≤ c − 22h−2 − 3 · 2h−1 + .
3
3
dh − 1
3
3 dh − 1
3
3
A bound for the embedding dimension also follows from the above proposition.
Corollary 26. If h ≥ 2, then
h ≤ log2
4
3
60 c + 1 + 9
10
.
4
2
≤ c.
x = 2h , we get
n Write
o 5/3x − 3x + 3 ≤ c.
√
√
c+1+9
− c = 0, we get x ∈ − 60 c+1−9
, 60 10
. As the minimum of
10
Proof. From Proposition 24,
By solving 5/3x2 − 3x +
5 2h
32
√
− 3 · 2h +
4
3
x = 2h > 1, we have that the maximum
5/3x2 − 3x + 34 − c is reached in x = 9/10, and in our setting
√
possible x > 0 such that 35 22h − 3 · 2h +
4
3
≤ c is x =
60 c+1+9
.
10
Remark 27. The set of all numerical semiogrups with fixed Frobenius number associated to an
irreducible planar curve singularity is calculated as in the free case, by imposing the condition
ek rk < rk+1 .
6. Experimental results
With the ideas given in the preceding sections, we implemented in GAP ([8]), with the help of
the numericalsgps package ([7]), functions to compute the set of all complete intersection, free
and telescopic numerical semigroups, as well as the set of all numerical semigroups associated to
irreducible planar curve singularities with fixed Frobenius number (these functions will be included
in the next release of this package).
The following table was computed in 6932 milliseconds on a 2.5GHz desktop computer, and it
shows, for fixed genus g, the number of complete intersections (ci(g)), free (fr(g)), telescopic (tl(g)),
associated to an irreducible planar curve singularity (pc(g)) numerical semigroups, respectively.
Recall that for a symmetric numerical semigroup its conductor is twice its genus.
Observe that almost all complete intersections in this table are free. This is due to the fact that
the embedding dimension of all numerical semigroups appearing there is small, and for embedding
dimension three or less, the concepts of free and complete intersections coincide (among the complete
intersection numerical semigroups represented in the table 158 of them have embedding dimension
2, 1525 have embedding dimension 3, 1862 have embedding dimension 4, and 205 have embedding
dimension 5).
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
g
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
ci(g)
1
1
1
2
3
2
4
5
3
7
8
5
11
11
9
14
17
12
18
fr(g)
1
1
1
2
3
2
4
5
3
7
8
5
11
11
9
14
17
12
18
tl(g)
1
1
1
2
2
2
4
3
2
5
6
4
8
8
7
10
9
8
12
pc(g)
1
1
1
2
2
1
3
2
2
4
4
2
5
3
4
6
5
3
6
g
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
ci(g)
24
16
27
31
21
36
38
27
46
45
34
57
62
43
65
77
53
84
90
fr(g)
24
16
27
31
21
35
38
27
46
45
33
57
62
43
65
76
52
83
90
tl(g)
12
11
18
19
13
20
22
16
24
25
20
32
31
25
37
39
29
43
47
pc(g)
5
6
9
8
6
11
9
8
11
10
7
13
9
10
14
13
11
17
13
g
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
ci(g)
61
100
110
80
122
120
94
143
151
108
158
179
128
197
209
142
229
238
172
11
fr(g)
61
100
109
79
120
120
94
142
149
106
157
179
128
194
207
142
227
235
169
tl(g)
37
52
54
47
61
60
48
73
72
57
75
84
68
86
89
76
101
104
83
pc(g)
12
16
19
12
20
17
15
22
21
15
24
23
20
26
27
20
30
29
24
The largest genus, for which the set of numerical semigroups with this genus is known, is 55,
and the number of numerical semigroups with genus 55 is 1142140736859 ([6]), while there are just
2496 symmetric numerical semigroup with genus 55 (this last amount can be computed by using the
IrreducibleNumericalSemigroupsWithFrobeniusNumber command of the numericalsgps package). The proportion of complete intersections among symmetric numerical semigroups is small,
and tiny compared with the whole set of numerical semigroups.
complete intersections
free
telescopic
planar
250
# numerical semigroups
200
150
100
50
0
0
10
20
30
genus
40
50
60
12
A. ASSI AND P. A. GARCÍA-SÁNCHEZ
Also as one of the referees observed, local minimums in the graph are attained when the genus
is congruent with 2 modulo 3. We have not a proof for this fact. May be this behavior is inherited
from the symmetric case. The sequence
1, 1, 1, 2, 3, 3, 6, 8, 7, 15, 20, 18, 36, 44, 45, 83, 109, 101, 174, 246, 227, 420, 546, 498, 926, 1182, 1121,
2015, 2496, 2436, 4350, 5602, 5317, 8925, 11971, 11276,
represents the number of symmetric numerical semigroups with genus ranging from 0 to 35.
The following table shows that the proportion between complete intersections and free numerical
semigroups remains similar even for larger genus. Observe that for genus 310 it takes 70 minutes
to compute the set of all complete intersections, while it takes approximately 8 minutes and 30
seconds to determine all free numerical semigroups with this genus. For genus 55, computing the
set of all numerical semigroups with this genus might take several months and a few terabytes (this
was communicated to us by Manuel Delgado, see [6]).
g
220
230
240
250
260
310
ci(g)
18018
16333
24862
28934
25721
66335
milliseconds
538213
660838
924409
1167901
1389167
4206374
fr(g)
17675
16026
24359
28355
25186
64959
milliseconds
94134
108187
153069
158706
177691
509691
fr(g)/ci(g)
0.98
0.98
0.98
0.98
0.98
0.98
References
[1] J. Bertin, P. Carbonne, Semi-groupes d’entiers et application aux branches, J. Algebra 49 (1977), 81-95.
[2] I. Bermejo, I. Garcı́a-Marco, J. J. Salazar-González, An algorithm for checking whether the toric ideal of an affine
monomial curve is a complete intersection, J. Symbolic Comput. 42 (2007), 971-991.
[3] I. Bermejo, P. Gimenez, E. Reyes, R. H. Villarreal, Complete intersections in affine monomial curves, Bol. Soc.
Mat. Mexicana (3) 11 (2005) , no. 2, 191-203.
[4] M. Bras-Amors, Fibonacci-like behavior of the number of numerical semigroups of a given genus, Semigroup
Forum, 76 (2008), 379–384.
[5] C. Delorme, Sous-monoı̈des d’intersection complète de N, Ann. Scient. École Norm. Sup. (4), 9 (1976), 145-154.
[6] M. Delgado, Experiments with numerical semigroups, in preparation.
[7] M. Delgado, P.A. Garcı́a-Sánchez, and J. Morais, “numericalsgps”: a gap package on numerical semigroups,
(\protect\vrule width0pt\protect\href{http://www.gap-system.org/Packages/numericalsgps.html}{http://www.gap-syst
[8] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.4 ; 2004,
(\protect\vrule width0pt\protect\href{http://www.gap-system.org}{http://www.gap-system.org}).
[9] P. A. Garcı́a-Sánchez, I. Ojeda, J. C. Rosales, Affine semigroups having a unique Betti element, J. Algebra Appl.
12 (2013), 1250177 (11 pages).
[10] J. Herzog, E. Kunz, Die Wertehalbgruppe eines lokalen Rings der Dimension 1, S. B. Heidelberger Akad. Wiss.
Math. Natur. Kl (1971), 2767.
[11] S. M. Johnson, A linear Diophantine problem, Can. J. Math., 12 (1960), 390-398.
[12] C. Kirfel, R. Pellikaan, The minimum distance of codes in an array coming from telescopic semigroups, IEEE
Trans. Inform. Theory, 41(1995) 1720-1732. Special issue on algebraic geometry codes.
[13] J. C. Rosales, On presentations of subsemigroups of Nn , Semigroup Forum 55 (1997), 152-159.
[14] J. C. Rosales, P. A. Garcı́a-Sánchez, On numerical semigroups with high embedding dimension, J. Algebra 203
(1998), 567-578.
[15] J. C. Rosales, P. A. Garcı́a-Sánchez, On free affine semigroups, Semigroup Forum 58 (1999), no. 3, 367-385.
[16] J. C. Rosales and P. A. Garcı́a-Sánchez, Numerical Semigroups, Developments in Mathematics, 20. Springer,
New York, 2009.
[17] J.C. Rosales, P. A. Garcı́a-Sánchez, J.I. Garcı́a-Garcı́a, J.A. Jiménez-Madrid, Fundamental gaps in numerical
semigroups, J. Pure Appl. Alg. 189 (2004), 301-313.
[18] A. Zhai, Fibonacci-like growth of numerical semigroups of a given genus, arXiv:1111.3142
[19] O. Zariski, Le problème des modules pour les courbes planes, Hermann, 1986.
COMPLETE INTERSECTIONS WITH FIXED FROBENIUS NUMBER
13
Université d’Angers, Département de Mathématiques, LAREMA, UMR 6093, 2 bd Lavoisier, 49045
Angers Cedex 01, France
E-mail address: [email protected]
Departamento de Álgebra, Universidad de Granada, E-18071 Granada, España
E-mail address: [email protected]
| 0math.AC
|
Comparison of channels: criteria for
domination by a symmetric channel
Anuran Makur and Yury Polyanskiy∗
arXiv:1609.06877v2 [cs.IT] 22 Nov 2017
Abstract
This paper studies the basic question of whether a given channel V can be dominated (in the precise sense of being
more noisy) by a q-ary symmetric channel. The concept of “less noisy” relation between channels originated in network
information theory (broadcast channels) and is defined in terms of mutual information or Kullback-Leibler divergence.
We provide an equivalent characterization in terms of χ2 -divergence. Furthermore, we develop a simple criterion for
domination by a q-ary symmetric channel in terms of the minimum entry of the stochastic matrix defining the channel
V . The criterion is strengthened for the special case of additive noise channels over finite Abelian groups. Finally, it
is shown that domination by a symmetric channel implies (via comparison of Dirichlet forms) a logarithmic Sobolev
inequality for the original channel.
Index Terms
Less noisy, degradation, q-ary symmetric channel, additive noise channel, Dirichlet form, logarithmic Sobolev inequalities.
C ONTENTS
I
Introduction
I-A
Preliminaries . . . . . . . . . . . . . . .
I-B
Channel preorders in information theory
I-C
Symmetric channels and their properties
I-D
Main question and motivation . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
2
2
4
6
preorder
. . . . .
. . . . .
. . . . .
. . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
7
7
8
8
9
III
Less noisy domination and degradation regions
III-A
Less noisy domination and degradation regions for additive noise channels . . . . . . . . . . . .
III-B
Less noisy domination and degradation regions for symmetric channels . . . . . . . . . . . . . .
10
11
12
IV
Equivalent characterizations of less noisy preorder
IV-A
Characterization using χ2 -divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV-B
Characterizations via the Löwner partial order and spectral radius . . . . . . . . . . . . . . . . .
13
13
14
V
Conditions for less noisy domination over additive noise channels
V-A
Necessary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V-B
Sufficient conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
16
17
VI
Sufficient conditions for degradation over general channels
19
VII
Less noisy domination and logarithmic Sobolev inequalities
22
II
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Main results
II-A
χ2 -divergence characterization of the less noisy
II-B
Less noisy domination by symmetric channels
II-C
Structure of additive noise channels . . . . . .
II-D
Comparison of Dirichlet forms . . . . . . . . .
II-E
Outline . . . . . . . . . . . . . . . . . . . . . .
VIII Conclusion
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
26
∗ A. Makur and Y. Polyanskiy are with the Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA (e-mail: [email protected]; [email protected]).
This research was supported in part by the National Science Foundation CAREER award under grant agreement CCF-12-53205, and in part by
the Center for Science of Information (CSoI), an NSF Science and Technology Center, under grant agreement CCF-09-39370.
This work was presented at the 2017 IEEE International Symposium on Information Theory (ISIT) [1].
1
Appendix A: Basics of majorization theory
26
Appendix B: Proofs of propositions 4 and 12
27
Appendix C: Auxiliary results
28
References
30
I. I NTRODUCTION
For any Markov chain U → X → Y , it is well-known that the data processing inequality, I(U ; Y ) ≤ I(U ; X),
holds. This result can be strengthened to [2]:
I(U ; Y ) ≤ ηI(U ; X)
(1)
where the contraction coefficient η ∈ [0, 1] only depends on the channel PY |X . Frequently, one gets η < 1 and
the resulting inequality is called a strong data processing inequality (SDPI). Such inequalities have been recently
simultaneously rediscovered and applied in several disciplines; see [3, Section 2] for a short survey. In [3, Section 6],
it was noticed that the validity of (1) for all PU,X is equivalent to the statement that an erasure channel with erasure
probability 1 − η is less noisy than the given channel PY |X . In this way, the entire field of SDPIs is equivalent to
determining whether a given channel is dominated by an erasure channel.
This paper initiates the study of a natural extension of the concept of SDPI by replacing the distinguished role
played by erasure channels with q-ary symmetric channels. We give simple criteria for testing this type of domination
and explain how the latter can be used to prove logarithmic Sobolev inequalities. In the next three subsections, we
introduce some basic definitions and notation. We state and motivate our main question in subsection I-D, and present
our main results in section II.
A. Preliminaries
The following notation will be used in our ensuing discussion. Consider any q, r ∈ N , {1, 2, 3, . . . }. We let
Rq×r (respectively Cq×r ) denote the set of all real (respectively complex) q × r matrices. Furthermore, for any matrix
A ∈ Rq×r , we let AT ∈ Rr×q denote the transpose of A, A† ∈ Rr×q denote the Moore-Penrose pseudoinverse of A,
R(A) denote the range (or column space) of A, and ρ (A) denote the spectral radius of A (which is the maximum
q×q
of the absolute values of all complex eigenvalues of A) when q = r. We let R0
( Rq×q
sym denote the sets of positive
q×q
semidefinite and symmetric matrices, respectively. In fact, R0 is a closed convex cone (with respect to the Frobenius
q×q
norm). We also let PSD denote the Löwner partial order over Rq×q
sym : for any two matrices A, B ∈ Rsym , we write
q×q
A PSD B (or equivalently, A − B PSD 0, where 0 is the zero matrix) if and only if A − B ∈ R0 . To work with
probabilities, we let Pq , {p = (p1 , . . . , pq ) ∈ Rq : p1 , . . . , pq ≥ 0 and p1 + · · · + pq = 1} be the probability simplex
of row vectors in Rq , Pq◦ , {p = (p1 , . . . , pq ) ∈ Rq : p1 , . . . , pq > 0 and p1 + · · · + pq = 1} be the relative interior of
Pq , and Rq×r
sto be the convex set of row stochastic matrices (which have rows in Pr ). Finally, for any (row or column)
vector x = (x1 , . . . , xq ) ∈ Rq , we let diag(x) ∈ Rq×q denote the diagonal matrix with entries [diag(x)]i,i = xi for
each i ∈ {1, . . . , q}, and for any set of vectors S ⊆ Rq , we let conv (S) be the convex hull of the vectors in S.
B. Channel preorders in information theory
Since we will study preorders over discrete channels that capture various notions of relative “noisiness” between
channels, we provide an overview of some well-known channel preorders in the literature. Consider an input random
variable X ∈ X and an output random variable Y ∈ Y, where the alphabets are X = [q] , {0, 1, . . . , q − 1} and
Y = [r] for q, r ∈ N without loss of generality. We let Pq be the set of all probability mass functions (pmfs) of X,
where every pmf PX = (PX (0), . . . , PX (q − 1)) ∈ Pq and is perceived as a row vector. Likewise, we let Pr be the set
of all pmfs of Y . A channel is the set of conditional distributions WY |X that associates each x ∈ X with a conditional
q×r
pmf WY |X (·|x) ∈ Pr . So, we represent each channel with a stochastic matrix W ∈ Rsto
that is defined entry-wise as:
∀x ∈ X , ∀y ∈ Y, [W ]x+1,y+1 , WY |X (y|x)
(2)
where the (x + 1)th row of W corresponds to the conditional pmf WY |X (·|x) ∈ Pr , and each column of W has at least
one non-zero entry so that no output alphabet letters are redundant. Moreover, we think of such a channel as a (linear)
map W : Pq → Pr that takes any row probability vector PX ∈ Pq to the row probability vector PY = PX W ∈ Pr .
One of the earliest preorders over channels was the notion of channel inclusion proposed by Shannon in [4]. Given
s×t
two channels W ∈ Rq×r
sto and V ∈ Rsto for some q, r, s, t ∈ N, he stated that W includes V , denoted W inc V ,
2
r×t
if there exist a pmf g ∈ Pm for some m ∈ N, and two sets of channels {Ak ∈ Rsto
: k = 1, . . . , m} and
s×q
{Bk ∈ Rsto : k = 1, . . . , m}, such that:
m
X
V =
gk Bk W Ak .
(3)
k=1
Channel inclusion is preserved under channel addition and multiplication (which are defined in [5]), and the existence
of a code for V implies the existence of as good a code for W in a probability of error sense [4]. The channel inclusion
preorder includes the input-output degradation preorder, which can be found in [6], as a special case. Indeed, V is an
s×q
r×t
input-output degraded version of W , denoted W iod V , if there exist channels A ∈ Rsto
and B ∈ Rsto
such that
V = BW A. We will study an even more specialized case of Shannon’s channel inclusion known as degradation [7],
[8].
q×s
Definition 1 (Degradation Preorder). A channel V ∈ Rsto
is said to be a degraded version of a channel W ∈ Rq×r
sto
with the same input alphabet, denoted W deg V , if V = W A for some channel A ∈ Rr×s
sto .
We note that when Definition 1 of degradation is applied to general matrices (rather than stochastic matrices), it is
equivalent to Definition C.8 of matrix majorization in [9, Chapter 15]. Many other generalizations of the majorization
preorder over vectors (briefly introduced in Appendix A) that apply to matrices are also presented in [9, Chapter 15].
Körner and Marton defined two other preorders over channels in [10] known as the more capable and less noisy
preorders. While the original definitions of these preorders explicitly reflect their significance in channel coding,
we will define them using equivalent mutual information characterizations proved in [10]. (See [11, Problems 6.166.18] for more on the relationship between channel coding and some of the aforementioned preorders.) We say a
q×s
channel W ∈ Rq×r
sto is more capable than a channel V ∈ Rsto with the same input alphabet, denoted W mc V , if
I(PX , WY |X ) ≥ I(PX , VY |X ) for every input pmf PX ∈ Pq , where I(PX , WY |X ) denotes the mutual information
of the joint pmf defined by PX and WY |X . The next definition presents the less noisy preorder, which will be a key
player in our study.
q×r
Definition 2 (Less Noisy Preorder). Given two channels W ∈ Rsto
and V ∈ Rq×s
sto with the same input alphabet, let
YW and YV denote the output random variables of W and V , respectively. Then, W is less noisy than V , denoted
W ln V , if I(U ; YW ) ≥ I(U ; YV ) for every joint distribution PU,X , where the random variable U ∈ U has some
arbitrary range U, and U → X → (YW , YV ) forms a Markov chain.
An analogous characterization of the less noisy preorder using Kullback-Leibler (KL) divergence or relative entropy
is given in the next proposition.
q×r
Proposition 1 (KL Divergence Characterization of Less Noisy [10]). Given two channels W ∈ Rsto
and V ∈ Rq×s
sto
with the same input alphabet, W ln V if and only if D(PX W ||QX W ) ≥ D(PX V ||QX V ) for every pair of input
pmfs PX , QX ∈ Pq , where D(·||·) denotes the KL divergence.1
We will primarily use this KL divergence characterization of ln in our discourse because of its simplicity. Another
well-known equivalent characterization of ln due to van Dijk is presented below, cf. [12, Theorem 2]. We will derive
some useful corollaries from it later in subsection IV-B.
q×r
Proposition 2 (van Dijk Characterization of Less Noisy [12]). Given two channels W ∈ Rsto
and V ∈ Rq×s
sto with
the same input alphabet, consider the functional F : Pq → R:
∀PX ∈ Pq , F (PX ) , I(PX , WY |X ) − I(PX , VY |X ).
Then, W ln V if and only if F is concave.
The more capable and less noisy preorders have both been used to study the capacity regions of broadcast channels.
We refer readers to [13]–[15], and the references therein for further details. We also remark that the more capable and
less noisy preorders tensorize, as shown in [11, Problem 6.18] and [3, Proposition 16], [16, Proposition 5], respectively.
On the other hand, these preorders exhibit rather counter-intuitive behavior in the context of Bayesian networks (or
directed graphical models). Consider a Bayesian network with “source” nodes (with no inbound edges) X and “sink”
nodes (with no outbound edges) Y . If we select a node Z in this network and replace the channel from the parents of
Z to Z with a less noisy channel, then we may reasonably conjecture that the channel from X to Y also becomes less
noisy (motivated by the results in [3]). However, this conjecture is false. To see this, consider the Bayesian network in
1 Throughout this paper, we will adhere to the convention that ∞ ≥ ∞ is true. So, D(P W ||Q W ) ≥ D(P V ||Q V ) is not violated when
X
X
X
X
both KL divergences are infinity.
3
Fig. 1. Illustration of a Bayesian network where X1 , X2 , Z, Y ∈ {0, 1} are binary random variables, PZ|X2 is a BSC(δ) with δ ∈ (0, 1), and
PY |X1 ,Z is defined by a deterministic NOR gate.
Figure 1 (inspired by the results in [17]), where the source nodes are X1 ∼ Ber 21 and X2 = 1 (almost surely), the
node Z is the output of a binary symmetric channel (BSC) with crossover probability δ ∈ (0, 1), denoted BSC(δ), and
the sink node Y is the output of a NOR gate. Let I(δ) = I(X1 , X2 ; Y ) be the end-to-end mutual information. Then,
although BSC(0) ln BSC(δ) for δ ∈ (0, 1), it is easy to verify that I(δ) > I(0) = 0. So, when we replace the BSC(δ)
with a less noisy BSC(0), the end-to-end channel does not become less noisy (or more capable).
The next proposition illustrates certain well-known relationships between the various preorders discussed in this
subsection.
q×r
Proposition 3 (Relations between Channel Preorders). Given two channels W ∈ Rsto
and V ∈ Rq×s
sto with the same
input alphabet, we have:
1) W deg V ⇒ W iod V ⇒ W inc V ,
2) W deg V ⇒ W ln V ⇒ W mc V .
These observations follow in a straightforward manner from the definitions of the various preorders. Perhaps the
only nontrivial implication is W deg V ⇒ W ln V , which can be proven using Proposition 1 and the data processing
inequality.
C. Symmetric channels and their properties
We next formally define q-ary symmetric channels and convey some of their properties. To this end, we first introduce
some properties of Abelian groups and define additive noise channels. Let us fix some q ∈ N with q ≥ 2 and consider an
Abelian group (X , ⊕) of order q equipped with a binary “addition” operation denoted by ⊕. Without loss of generality,
we let X = [q], and let 0 denote the identity element. This endows an ordering to the elements of X . Each element
x ∈ X permutes the entries of the row vector (0, . . . , q − 1) to (σx (0), . . . , σx (q − 1)) by (left) addition in the Cayley
table of the group, where σx : [q] → [q] denotes a permutation of [q], and σx (y) = x ⊕y for every y ∈ X . So,
corresponding to each x ∈ X , we can define a permutation matrix Px , eσx (0) · · · eσx (q−1) ∈ Rq×q such that:
(4)
[v0 · · · vq−1 ] Px = vσx (0) · · · vσx (q−1)
for any v0 , . . . , vq−1 ∈ R, where for each i ∈ [q], ei ∈ Rq is the ith standard basis column vector with unity in the
(i + 1)th position and zero elsewhere. The permutation matrices {Px ∈ Rq×q : x ∈ X } (with the matrix multiplication
operation) form a group that is isomorphic to (X , ⊕) (see Cayley’s theorem, and permutation and regular representations
of groups in [18, Sections 6.11, 7.1, 10.6]). In particular, these matrices commute as (X , ⊕) is Abelian, and are jointly
unitarily diagonalizable by a Fourier matrix of characters (using [19, Theorem 2.5.5]). We now recall that given a row
vector x = (x0 , . . . , xq−1 ) ∈ Rq , we may define a corresponding X -circulant matrix, circX (x) ∈ Rq×q , that is defined
entry-wise as [20, Chapter 3E, Section 4]:
∀a, b ∈ [q], [circX (x)]a+1,b+1 , x−a⊕b .
(5)
where −a ∈ X denotes the inverse of a ∈ X . Moreover, we can decompose this X -circulant matrix as:
circX (x) =
q−1
X
xi PiT
(6)
i=0
since
write:
Pq−1
i=0
Pq−1
xi PiT a+1,b+1 =
i=0 xi eσi (a) b+1 = x−a⊕b for every a, b ∈ [q]. Using similar reasoning, we can
circX (x) = [P0 y · · · Pq−1 y] = P0 xT · · · Pq−1 xT
T
(7)
T
where y = x0 x−1 · · · x−(q−1) ∈ Rq , and P0 = Iq ∈ Rq×q is the q × q identity matrix. Using (6), we see that X circulant matrices are normal, form a commutative algebra, and are jointly unitarily diagonalizable by a Fourier matrix.
4
Furthermore, given two row vectors x, y ∈ Rq , we can define x circX (y) = y circX (x) as the X -circular convolution of
x and y, where the commutativity of X -circular convolution follows from the commutativity of X -circulant matrices.
A salient specialization of this discussion is the case where ⊕ is addition modulo q, and (X = [q], ⊕) is the
cyclic Abelian group Z/qZ. In this scenario, X -circulant matrices correspond to the standard circulant matrices which
are jointly unitarily diagonalized by the discrete Fourier transform (DFT) matrix. Furthermore, for each x ∈ [q], the
permutation matrix PxT = Pqx , where Pq ∈ Rq×q is the generator cyclic permutation matrix as presented in [19, Section
0.9.6]:
∀a, b ∈ [q], [Pq ]a+1,b+1 , ∆1,(b−a (mod q))
(8)
where ∆i,j is the Kronecker delta function, which is unity if i = j and zero otherwise. The matrix Pq cyclically shifts
any input row vector to the right once, i.e. (x1 , x2 , . . . , xq ) Pq = (xq , x1 , . . . , xq−1 ).
Let us now consider a channel with common input and output alphabet X = Y = [q], where (X , ⊕) is an Abelian
group. Such a channel operating on an Abelian group is called an additive noise channel when it is defined as:
Y =X ⊕Z
(9)
where X ∈ X is the input random variable, Y ∈ X is the output random variable, and Z ∈ X is the additive noise
random variable that is independent of X with pmf PZ = (PZ (0), . . . , PZ (q − 1)) ∈ Pq . The channel transition
probability matrix corresponding to (9) is the X -circulant stochastic matrix circX (PZ ) ∈ Rq×q
sto , which is also doubly
T
stochastic (i.e. both circX (PZ ) , circX (PZ ) ∈ Rq×q
sto ). Indeed, for an additive noise channel, it is well-known that the
pmf of Y , PY ∈ Pq , can be obtained from the pmf of X, PX ∈ Pq , by X -circular convolution: PY = PX circX (PZ ).
We remark that in the context of various channel symmetries in the literature (see [21, Section VI.B] for a discussion),
additive noise channels correspond to “group-noise” channels, and are input symmetric, output symmetric, Dobrushin
symmetric, and Gallager symmetric.
The q-ary symmetric channel is an additive noise channel on the Abelian group (X , ⊕) with noise pmf PZ =
wδ , (1 − δ, δ/(q − 1), . . . , δ/(q − 1)) ∈ Pq , where δ ∈ [0, 1]. Its channel transition probability matrix is denoted
Wδ ∈ Rq×q
sto :
h
q−1 T iT
Wδ , circX (wδ ) = wδ T PqT wδ T · · · PqT
wδ
(10)
which has 1 − δ in the principal diagonal entries and δ/(q − 1) in all other entries regardless of the choice of group
(X , ⊕). We may interpret δ as the total crossover probability of the symmetric channel. Indeed, when q = 2, Wδ
represents a BSC with crossover probability
δ ∈ [0, 1]. Although Wδ is only stochastic when δ ∈ [0, 1], we will refer to
the parametrized convex set of matrices Wδ ∈ Rq×q
sym : δ ∈ R with parameter δ as the “symmetric channel matrices,”
where each Wδ has the form (10) such that every row and column sums to unity. We conclude this subsection with a
list of properties of symmetric channel matrices.
Proposition 4 (Properties of Symmetric Channel Matrices). The symmetric channel matrices, Wδ ∈ Rq×q
sym : δ ∈ R ,
satisfy the following properties:
1) For every δ ∈ R, Wδ is a symmetric circulant matrix.
2) The DFT matrix Fq ∈ Cq×q , which is defined entry-wise as [Fq ]j,k = q −1/2 ω (j−1)(k−1) for 1 ≤ j, k ≤ q
√
where ω = exp (2πi/q) and i = −1, jointly diagonalizes
Wδ for every δ ∈ R. Moreover, the corresponding
eigenvalues or Fourier coefficients, {λj (Wδ ) = FqH Wδ Fq j,j : j = 1, . . . , q} are real:
1,
j=1
λj (Wδ ) =
δ
1 − δ − q−1 , j = 2, . . . , q
where FqH denotes the Hermitian transpose of Fq .
3) For all δ ∈ [0, 1], Wδ is a doubly stochastic matrix that has the uniform pmf u , (1/q, . . . , 1/q) as its stationary
distribution: uWδ= u.
T
δ
1
4) For every δ ∈ R\ q−1
, Wδ−1 = Wτ with τ = −δ/ 1 − δ − q−1
, and for δ = q−1
q
q , Wδ = q 11 is unit rank
T
and singular,
where 1 = [1 · · · 1] .
q−1
5) The set Wδ ∈ Rq×q
with the operation of matrix multiplication is an Abelian group.
sym : δ ∈ R\
q
Proof. See Appendix B.
5
D. Main question and motivation
As we mentioned at the outset, our work is partly motivated by [3, Section 6], where the authors demonstrate an
intriguing relation between less noisy domination by an erasure channel and the contraction coefficient of the SDPI (1).
q×(q+1)
q×s
For a common input alphabet X = [q], consider a channel V ∈ Rsto
and a q-ary erasure channel E ∈ Rsto
with
erasure probability ∈ [0, 1]. Recall that given an input x ∈ X , a q-ary erasure channel erases x and outputs e (erasure
symbol) with probability , and outputs x itself with probability 1 − ; the output alphabet of the erasure channel is
{e} ∪ X . It is proved in [3, Proposition 15] that E ln V if and only if ηKL (V ) ≤ 1 − , where ηKL (V ) ∈ [0, 1] is the
contraction coefficient for KL divergence:
ηKL (V ) ,
sup
PX ,QX ∈Pq
0<D(PX ||QX )<+∞
D (PX V ||QX V )
D (PX ||QX )
(11)
which equals the best possible constant η in the SDPI (1) when V = PY |X (see [3, Theorem 4] and the references
therein). This result illustrates that the q-ary erasure channel E with the largest erasure probability ∈ [0, 1] (or the
smallest channel capacity) that is less noisy than V has = 1 − ηKL (V ).2 Furthermore, there are several simple upper
bounds on ηKL that provide sufficient conditions for such less noisy domination. For example, if the `1 -distances between
the rows of V are bounded by 2(1 − α) for some α ∈ [0, 1], then ηKL ≤ 1 − α, cf. [22]. Another criterion follows from
Doeblin minorization [23, Remark III.2]: if for some pmf p ∈ Ps and some α ∈ (0, 1), V ≥ α 1p entry-wise, then
Eα deg V and ηKL (V ) ≤ 1 − α.
To extend these
ideas,
channel Wδ with the largest
we consider the following question: What is the q-ary symmetric
3
(or
the
smallest
channel
capacity)
such
that
W
V
?
Much
like
the bounds on ηKL in the
value of δ ∈ 0, q−1
δ
ln
q
erasure channel context, the goal of this paper is to address this question by establishing simple criteria for testing ln
domination by a q-ary symmetric channel. We next provide several other reasons why determining whether a q-ary
symmetric channel dominates a given channel V is interesting.
Firstly, if W ln V , then W ⊗n ln V ⊗n (where W ⊗n is the n-fold tensor product of W ) since ln tensorizes, and
n
n
I(U ; YW
) ≥ I(U ; YVn ) for every Markov chain U → X n → (YW
, YVn ) (see Definition 2). Thus, many impossibility
n
results (in statistical decision theory for example) that are proven by exhibiting bounds on quantities such as I(U ; YW
)
n
transparently carry over to statistical experiments with observations on the basis of YV . Since it is common to study
the q-ary symmetric observation model (especially with q = 2), we can leverage its sample complexity lower bounds
for other V .
Secondly, we present a self-contained information theoretic motivation. W ln V if and only if CS = 0, where
CS is the secrecy capacity of the Wyner wiretap channel with V as the main (legal receiver) channel and W as the
eavesdropper channel [24, Corollary 3], [11, Corollary 17.11]. Therefore, finding the maximally noisy q-ary symmetric
channel that dominates V establishes the minimal noise required on the eavesdropper link so that secret communication
is feasible.
Thirdly, ln domination turns out to entail a comparison of Dirichlet forms (see subsection II-D), and consequently,
allows us to prove Poincaré and logarithmic Sobolev inequalities for V from well-known results on q-ary symmetric
channels. These inequalities are cornerstones of the modern approach to Markov chains and concentration of measure
[25], [26].
II. M AIN RESULTS
In this section, we first delineate some guiding sub-questions of our study, indicate the main results that address
them, and then present these results in the ensuing subsections. We will delve into the following four leading questions:
1) Can we test the less noisy preorder ln without using KL divergence?
Yes, we can use χ2 -divergence as shown in Theorem 1.
2) Given a channel V ∈ Rq×q
sto , is there a simple sufficient condition for less noisy domination by a q-ary symmetric
channel Wδ ln V ?
Yes, a condition using degradation (which implies less noisy domination) is presented in Theorem 2.
3) Can we say anything stronger about less noisy domination by a q-ary symmetric channel when V ∈ Rq×q
sto is an
additive noise channel?
Yes, Theorem 3 outlines the structure of additive noise channels in this context (and Figure 2 depicts it).
2A
q-ary erasure channel E with erasure probability ∈ [0, 1] has channel
capacity
C() = log(q)(1 − ), which is linear and decreasing.
has channel capacity C(δ) = log(q) − H(wδ ), which is convex
q-ary symmetric channel Wδ with total crossover probability δ ∈ 0, q−1
q
and decreasing. Here, H(wδ ) denotes the Shannon entropy of the pmf wδ .
3A
6
4) Why are we interested in less noisy domination by q-ary symmetric channels?
Because this permits us to compare Dirichlet forms as portrayed in Theorem 4.
We next elaborate on these aforementioned theorems.
A. χ2 -divergence characterization of the less noisy preorder
Our most general result illustrates that although less noisy domination is a preorder defined using KL divergence,
one can equivalently define it using χ2 -divergence. Since we will prove this result for general measurable spaces, we
introduce some notation pertinent only to this result. Let (X , F), (Y1 , H1 ), and (Y2 , H2 ) be three measurable spaces,
and let W : H1 × X → [0, 1] and V : H2 × X → [0, 1] be two Markov kernels (or channels) acting on the same
source space (X , F). Given any probability measure PX on (X , F), we denote by PX W the probability measure on
(Y1 , H1 ) induced by the push-forward of PX .4 Recall that for any two probability measures PX and QX on (X , F),
their KL divergence is given by:
Z
dPX
dPX , if PX QX
log
D(PX ||QX ) ,
(12)
dQX
X
+∞,
otherwise
and their χ2 -divergence is given by:
Z
2
dPX
dQX − 1, if PX QX
χ2 (PX ||QX ) ,
X dQX
+∞,
otherwise
(13)
dPX
where PX QX denotes that PX is absolutely continuous with respect to QX , dQ
denotes the Radon-Nikodym
X
derivative of PX with respect to QX , and log(·) is the natural logarithm with base e (throughout this paper). Furthermore,
the characterization of ln in Proposition 1 extends naturally to general Markov kernels; indeed, W ln V if and only if
D(PX W ||QX W ) ≥ D(PX V ||QX V ) for every pair of probability measures PX and QX on (X , F). The next theorem
presents the χ2 -divergence characterization of ln .
Theorem 1 (χ2 -Divergence Characterization of ln ). For any Markov kernels W : H1 ×X → [0, 1] and V : H2 ×X →
[0, 1] acting on the same source space, W ln V if and only if:
χ2 (PX W ||QX W ) ≥ χ2 (PX V ||QX V )
for every pair of probability measures PX and QX on (X , F).
Theorem 1 is proved in subsection IV-A.
B. Less noisy domination by symmetric channels
Our remaining results are all concerned with less noisy (and degraded) domination by q-ary symmetric channels.
q×q
Suppose we are given a q-ary symmetric channel Wδ ∈ Rsto
with δ ∈ [0, 1], and another channel V ∈ Rq×q
sto with
common input and output alphabets. Then, the next result provides a sufficient condition for when Wδ deg V .
Theorem 2 (Sufficient Condition for Degradation by Symmetric Channels). Given a channel V ∈ Rq×q
sto with q ≥ 2
and minimum probability ν = min {[V ]i,j : 1 ≤ i, j ≤ q}, we have:
ν
0≤δ≤
⇒ Wδ deg V.
ν
1 − (q − 1)ν + q−1
Theorem 2 is proved in section VI. We note that the sufficient condition in Theorem 2 is tight as there exist channels
ν
V that violate Wδ deg V when δ > ν/(1 − (q − 1)ν + q−1
). Furthermore, Theorem 2 also provides a sufficient
condition for Wδ ln V due to Proposition 3.
4 Here, we can think of X and Y as random variables with codomains X and Y, respectively. The Markov kernel W behaves like the conditional
distribution of Y given X (under regularity conditions). Moreover, when the distribution of X is PX , the corresponding distribution of Y is
PY = PX W .
7
C. Structure of additive noise channels
Our next major result is concerned with understanding when q-ary symmetric channels operating on an Abelian group
(X , ⊕) dominate other additive noise channels on (X , ⊕), which are defined in (9), in the less noisy and degraded
senses. Given a symmetric channel Wδ ∈ Rq×q
sto with δ ∈ [0, 1], we define the additive less noisy domination region of
Wδ as:
Ladd
(14)
Wδ , {v ∈ Pq : Wδ = circX (wδ ) ln circX (v)}
which is the set of all noise pmfs whose corresponding channel transition probability matrices are dominated by Wδ
in the less noisy sense. Likewise, we define the additive degradation region of Wδ as:
add
DW
, {v ∈ Pq : Wδ = circX (wδ ) deg circX (v)}
δ
(15)
which is the set of all noise pmfs whose corresponding channel transition probability matrices are degraded versions
add
of Wδ . The next theorem exactly characterizes DW
, and “bounds” Ladd
Wδ in a set theoretic sense.
δ
Theorem 3 (Additive Less Noisy Domination
Degradation Regions for Symmetric Channels). Given a symmetric
and
q−1
and q ≥ 2, we have:
channel Wδ = circX (wδ ) ∈ Rq×q
sto with δ ∈ 0, q
add
DW
= conv wδ Pqk : k ∈ [q]
δ
⊆ conv wδ Pqk : k ∈ [q] ∪ wγ Pqk : k ∈ [q]
⊆ Ladd
Wδ ⊆ {v ∈ Pq : kv − uk`2 ≤ kwδ − uk`2 }
where the first set inclusion is strict for δ ∈ 0, q−1
and q ≥ 3, Pq denotes the generator cyclic permutation matrix
q
as defined in (8), u denotes the uniform pmf, k·k`2 is the Euclidean `2 -norm, and:
γ=
1−δ
.
δ
1 − δ + (q−1)
2
q×q
Furthermore, Ladd
: x ∈ X } defined
Wδ is a closed and convex set that is invariant under the permutations {Px ∈ R
add
add
in (4) corresponding to the underlying Abelian group (X , ⊕) (i.e. v ∈ LWδ ⇒ vPx ∈ LWδ for every x ∈ X ).
Theorem 3 is a compilation of several results. As explained at the very end of subsection V-B, Proposition 6 (in
subsection III-A), Corollary 1 (in subsection III-B), part 1 of Proposition 9 (in subsection V-A), and Proposition 11
(in subsection V-B) make up Theorem 3. We remark that according to numerical evidence, the second and third set
inclusions in Theorem 3 appear to be strict, and Ladd
Wδ seems to be a strictly convex set. The content of Theorem 3 and
these observations are illustrated in Figure 2, which portrays the probability simplex of noise pmfs for q = 3 and the
pertinent regions which capture less noisy domination and degradation by a q-ary symmetric channel.
D. Comparison of Dirichlet forms
As mentioned in subsection I-D, one of the reasons we study q-ary symmetric channels and prove Theorems 2 and
3 is because less noisy domination implies useful bounds between Dirichlet forms. Recall that the q-ary symmetric
channel Wδ ∈ Rq×q
sto with δ ∈ [0, 1] has uniform stationary distribution u ∈ Pq (see part 3 of Proposition 4). For any
channel V ∈ Rq×q
sto that is doubly stochastic and has uniform stationary distribution, we may define a corresponding
Dirichlet form:
1
(16)
∀f ∈ Rq , EV (f, f ) = f T (Iq − V ) f
q
T
where f = [f1 · · · fq ] ∈ Rq are column vectors, and Iq ∈ Rq×q denotes the q × q identity matrix (as shown in [25]
or [26]). Our final theorem portrays that Wδ ln V implies that the Dirichlet form corresponding to V dominates the
Dirichlet form corresponding to Wδ pointwise. The Dirichlet form corresponding to Wδ is in fact a scaled version of
the so called standard Dirichlet form:
!2
q
q
X
X
1
1
fk2 −
fk
(17)
∀f ∈ Rq , Estd (f, f ) , VARu (f ) =
q
q
k=1
k=1
which is the Dirichlet form corresponding to the q-ary symmetric channel W(q−1)/q = 1u with all uniform conditional
qδ
pmfs. Indeed, using Iq − Wδ = q−1
(Iq − 1u), we have:
∀f ∈ Rq , EWδ (f, f ) =
8
qδ
Estd (f, f ) .
q−1
(18)
Fig. 2. Illustration of the additive less noisy domination region and additive degradation region for a q-ary symmetric channel when q = 3 and
δ ∈ (0, 2/3): The gray triangle denotes the probability simplex of noise pmfs P3 . The dotted line denotes the parametrized family of noise pmfs of
3-ary symmetric channels {wδ ∈ P3 : δ ∈ [0, 1]}; its noteworthy points are w0 (corner of simplex, W0 is less noisy than every channel), wδ for
some fixed δ ∈ (0, 2/3) (noise pmf of 3-ary symmetric channel Wδ under consideration), w2/3 = u (uniform pmf, W2/3 is more noisy than every
channel), wτ with τ = 1−(δ/2) (Wτ is the extremal symmetric channel that is degraded by Wδ ), wγ with γ = (1−δ)/(1−δ +(δ/4)) (Wγ is a 3ary symmetric
by Wδ but Wδ ln Wγ ), and w1 (edge of simplex). The magenta triangle denotes the additive degradation
channel that is not degraded
regionconv wδ , wδ P3 , wδ P32 of Wδ . The green convex region denotes the additive less noisydomination region of Wδ , and the yellow region
conv wδ , wδ P3 , wδ P32 , wγ , wγ P3 , wγ P32
is its lower bound while the circular cyan region v ∈ P3 : kv − uk`2 ≤ kwδ − uk`2 (which is
a hypersphere for general q ≥ 3) is its upper bound. Note that we do not need to specify the underlying group because there is only one group of
order 3.
The standard Dirichlet form is the usual choice for Dirichlet form comparison because its logarithmic Sobolev constant
has been precisely computed in [25, Appendix, Theorem A.1]. So, we present Theorem 4 using Estd rather than EWδ .
q×q
Theorem 4 (Domination of Dirichlet Forms). Given the doubly stochastic channels Wδ ∈ Rsto
with δ ∈ 0, q−1
and
q
q×q
V ∈ Rsto , if Wδ ln V , then:
qδ
∀f ∈ Rq , EV (f, f ) ≥
Estd (f, f ) .
q−1
An extension of Theorem 4 is proved in section VII. The domination of Dirichlet forms shown in Theorem 4
has several useful consequences. A major consequence is that we can immediately establish Poincaré (spectral gap)
inequalities and logarithmic Sobolev inequalities (LSIs) for the channel V using the corresponding inequalities for
q×q
q-ary symmetric channels. For example, the LSI for Wδ ∈ Rsto
with q > 2 is:
(q − 1) log(q − 1)
D f 2 u||u ≤
EWδ (f, f )
(q − 2)δ
(19)
Pq
2
for all f ∈ Rq such that
k=1 fk = q, where we use (54) and the logarithmic Sobolev constant computed in
part 1 of Proposition 12. As shown in Appendix B, (19) is easily established using the known logarithmic Sobolev
constant corresponding to the standard Dirichlet form. Using the LSI for V that follows from (19) and Theorem 4,
we immediately obtain guarantees on the convergence rate and hypercontractivity properties of the associated Markov
semigroup {exp(−t(Iq − V )) : t ≥ 0}. We refer readers to [25] and [26] for comprehensive accounts of such topics.
E. Outline
We briefly outline the content of the ensuing sections. In section III, we study the structure of less noisy domination
and degradation regions of channels. In section IV, we prove Theorem 1 and present some other equivalent characterizations of ln . We then derive several necessary and sufficient conditions for less noisy domination among additive
9
noise channels in section V, which together with the results of section III, culminates in a proof of Theorem 3. Section
VI provides a proof of Theorem 2, and section VII introduces LSIs and proves an extension of Theorem 4. Finally,
we conclude our discussion in section VIII.
III. L ESS NOISY DOMINATION AND DEGRADATION REGIONS
In this section, we focus on understanding the “geometric” aspects of less noisy domination and degradation by
channels. We begin by deriving some simple characteristics of the sets of channels that are dominated by some fixed
channel in the less noisy and degraded senses. We then specialize our results for additive noise channels, and this
add
culminates in a complete characterization of DW
and derivations of certain properties of Ladd
Wδ presented in Theorem
δ
3.
Let W ∈ Rq×r
sto be a fixed channel with q, r ∈ N, and define its less noisy domination region:
LW , V ∈ Rq×r
(20)
sto : W ln V
as the set of all channels on the same input and output alphabets that are dominated by W in the less noisy sense.
Moreover, we define the degradation region of W :
q×r
DW , V ∈ Rsto
: W deg V
(21)
as the set of all channels on the same input and output alphabets that are degraded versions of W . Then, LW and DW
satisfy the properties delineated below.
Proposition 5 (Less Noisy Domination and Degradation Regions). Given the channel W ∈ Rq×r
sto , its less noisy
domination region LW and its degradation region DW are non-empty, closed, convex, and output alphabet permutation
symmetric (i.e. V ∈ LW ⇒ V P ∈ LW and V ∈ DW ⇒ V P ∈ DW for every permutation matrix P ∈ Rr×r ).
Proof.
Non-Emptiness of LW and DW : W ln W ⇒ W ∈ LW , and W deg W ⇒ W ∈ DW . So, LW and DW are
non-empty.
Closure of LW : Fix any two pmfs PX , QX ∈ Pq , and consider a sequence of channels Vk ∈ LW such that Vk → V ∈
Rq×r
sto (with respect to the Frobenius norm). Then, we also have PX Vk → PX V and QX Vk → QX V (with respect to
the `2 -norm). Hence, we get:
D (PX V ||QX V ) ≤ lim inf D (PX Vk ||QX Vk )
k→∞
≤ D (PX W ||QX W )
where the first line follows from the lower semicontinuity of KL divergence [27, Theorem 1], [28, Theorem 3.6,
Section 3.5], and
the second line holds because Vk ∈ LW . This implies that for any two pmfs PX , QX ∈ Pq , the set
S (PX , QX ) = V ∈ Rq×r
sto : D (PX W ||QX W ) ≥ D (PX V ||QX V ) is actually closed. Using Proposition 1, we have
that:
\
LW =
S (PX , QX ).
PX ,QX ∈Pq
So, LW is closed since it is an intersection of closed sets [29].
Closure of DW : Consider a sequence of channels Vk ∈ DW such that Vk → V ∈ Rq×r
sto . Since each Vk = W Ak
r×r
for some channel Ak ∈ Rr×r
sto belonging to the compact set Rsto , there exists a subsequence Akm that converges by
r×r
(sequential) compactness [29]: Akm → A ∈ Rsto
. Hence, V ∈ DW since Vkm = W Akm → W A = V , and DW is a
closed set.
Convexity of LW : Suppose V1 , V2 ∈ LW , and let λ ∈ [0, 1] and λ̄ = 1 − λ. Then, for every PX , QX ∈ Pq , we have:
D(PX W ||QX W ) ≥ D(PX (λV1 + λ̄V2 )||QX (λV1 + λ̄V2 ))
by the convexity of KL divergence. Hence, LW is convex.
r×r
Convexity of DW : If V1 , V2 ∈ DW so that V1 = W A1 and V2 = W A2 for some A1 , A2 ∈ Rsto
, then λV1 + λ̄V2 =
W (λA1 + λ̄A2 ) ∈ DW for all λ ∈ [0, 1], and DW is convex.
Symmetry of LW : This is obvious from Proposition 1 because KL divergence is invariant to permutations of its input
pmfs.
Symmetry of DW : Given V ∈ DW so that V = W A for some A ∈ Rr×r
sto , we have that V P = W AP ∈ DW for
every permutation matrix P ∈ Rr×r . This completes the proof.
10
While the channels in LW and DW all have the same output alphabet as W , as defined in (20) and (21), we may
extend the output alphabet of W by adding zero probability letters. So, separate less noisy domination and degradation
regions can be defined for each output alphabet size that is at least as large as the original output alphabet size of W .
A. Less noisy domination and degradation regions for additive noise channels
Often in information theory, we are concerned with additive noise channels on an Abelian group (X , ⊕) with X = [q]
and q ∈ N, as defined in (9). Such channels are completely defined by a noise pmf PZ ∈ Pq with corresponding channel
q×q
transition probability matrix circX (PZ ) ∈ Rq×q
sto . Suppose W = circX (w) ∈ Rsto is an additive noise channel with
noise pmf w ∈ Pq . Then, we are often only interested in the set of additive noise channels that are dominated by W .
We define the additive less noisy domination region of W :
Ladd
W , {v ∈ Pq : W ln circX (v)}
(22)
as the set of all noise pmfs whose corresponding channel transition matrices are dominated by W in the less noisy
sense. Likewise, we define the additive degradation region of W :
add
DW
, {v ∈ Pq : W deg circX (v)}
(23)
as the set of all noise pmfs whose corresponding channel transition matrices are degraded versions of W . (These
definitions generalize (14) and (15), and can also hold for any non-additive noise channel W .) The next proposition
add
illustrates certain properties of Ladd
W and explicitly characterizes DW .
Proposition 6 (Additive Less Noisy Domination and Degradation Regions). Given the additive noise channel W =
q×q
circX (w) ∈ Rsto with noise pmf w ∈ Pq , we have:
add
q×q
1) Ladd
: x ∈ X } defined
W and DW are non-empty, closed, convex, and invariant under the permutations {Px ∈ R
add
add
add
add
in (4) (i.e. v ∈ LW ⇒ vPx ∈ LW and v ∈ DW ⇒ vPx ∈ DW for every x ∈ X ).
add
2) DW
= conv ({wPx : x ∈ X }) = {v ∈ Pq : w X v}, where X denotes the group majorization preorder as
defined in Appendix A.
To prove Proposition 6, we will need the following lemma.
Lemma 1 (Additive Noise Channel Degradation). Given two additive noise channels W = circX (w) ∈ Rq×q
sto and
V = circX (v) ∈ Rq×q
sto with noise pmfs w, v ∈ Pq , W deg V if and only if V = W circX (z) = circX (z)W for some
z ∈ Pq (i.e. for additive noise channels W deg V , the channel that degrades W to produce V is also an additive
noise channel without loss of generality).
Proof. Since X -circulant matrices commute, we must have W circX (z) = circX (z)W for every z ∈ Pq . Furthermore,
V = W circX (z) for some z ∈ Pq implies that W deg V by Definition 1. So, it suffices to prove that W deg V implies
V = W circX (z) for some z ∈ Pq . By Definition 1, W deg V implies that V = W R for some doubly stochastic
q×q
channel R ∈ Rsto
(as V and W are doubly stochastic). Let r with rT ∈ Pq be the first column of R, and s = W r
T
with s ∈ Pq be the first column of V . Then, it is straightforward to verify using (7) that:
V = s P1 s P2 s · · · Pq−1 s
= W r P1 W r P2 W r · · · Pq−1 W r
= W r P1 r P2 r · · · Pq−1 r
where the third equality holds because {Px : x ∈ X } are X -circulant matrices which commute with W . Hence, V is
the product of W and an X -circulant stochastic matrix, i.e. V = W circX (z) for some z ∈ Pq . This concludes the
proof.
We emphasize that in Lemma 1, the channel that degrades W to produce V is only an additive noise channel
without loss of generality. We can certainly have V = W R with a non-additive noise channel R. Consider for instance,
V = W = 11T /q, where every doubly stochastic matrix R satisfies V = W R. However, when we consider V = W R
with an additive noise channel R, V corresponds to the channel W with an additional independent additive noise term
associated with R. We now prove Proposition 6.
Proof of Proposition 6.
add
Part 1: Non-emptiness, closure, and convexity of Ladd
W and DW can be proved in exactly the same way as in Proposition
5, with the additional observation that the set of X -circulant matrices is closed and convex. Moreover, for every x ∈ X :
W ln W Px = circX (wPx ) ln W
W deg W Px = circX (wPx ) deg W
11
where the equalities follow from (7). These inequalities and the transitive properties of ln and deg yield the invariance
add
q×q
of Ladd
: x ∈ X }.
W and DW with respect to {Px ∈ R
add
Part 2: Lemma 1 is equivalent to the fact that v ∈ DW
if and only if circX (v) = circX (w) circX (z) for some z ∈ Pq .
add
This implies that v ∈ DW if and only if v = w circX (z) for some z ∈ Pq (due to (7) and the fact that X -circulant
matrices commute). Applying Proposition 14 from Appendix A completes the proof.
We remark that part 1 of Proposition 6 does not require W to be an additive noise channel. The proofs of closure,
add
add
convexity, and invariance with respect to {Px ∈ Rq×q : x ∈ X } hold for general W ∈ Rq×q
sto . Moreover, LW and DW
add
add
are non-empty because u ∈ LW and u ∈ DW .
B. Less noisy domination and degradation regions for symmetric channels
Since q-ary symmetric channels for q ∈ N are additive noise channels, Proposition 6 holds for symmetric channels.
In this subsection, we deduce some simple results that are unique to symmetric channels. The first of these is a
specialization of part 2 of Proposition 6 which states that the additive degradation region of a symmetric channel can
be characterized by traditional majorization instead of group majorization.
Corollary 1 (Degradation Region of Symmetric Channel). The q-ary symmetric channel Wδ = circX (wδ ) ∈ Rq×q
sto for
δ ∈ [0, 1] has additive degradation region:
add
DW
= {v ∈ Pq : wδ maj v} = conv wδ Pqk : k ∈ [q]
δ
where maj denotes the majorization preorder defined in Appendix A, and Pq ∈ Rq×q is defined in (8).
Proof. From part 2 of Proposition 6, we have that:
add
DW
= conv ({wδ Px : x ∈ X }) = conv wδ Pqk : k ∈ [q]
δ
= conv wδ P : P ∈ Rq×q is a permutation matrix
= {v ∈ Pq : w maj v}
where the second and third equalities hold regardless of the choice of group (X , ⊕), because the sets of all cyclic
or regular permutations of wδ = (1 − δ, δ/(q − 1), . . . , δ/(q − 1)) equal {wδ Px : x ∈ X }. The final equality follows
from the definition of majorization in Appendix A.
With this geometric characterization of the additive degradation region, it is straightforward
to find the extremal
.
Indeed,
we compute τ by
symmetric channel Wτ that is a degraded version of Wδ for some fixed δ ∈ [0, 1]\ q−1
q
using the fact that the noise pmf wτ ∈ conv wδ Pqk : k = 1, . . . , q − 1 :
wτ =
q−1
X
λi wδ Pqi
(24)
i=1
for some λ1 , . . . , λq−1 ∈ [0, 1] such that λ1 + · · · + λq−1 = 1. Solving (24) for τ and λ1 , . . . , λq−1 yields:
τ =1−
and λ1 = · · · = λq−1 =
1
q−1 ,
δ
q−1
(25)
which means that:
q−1
1 X
wτ =
wδ Pqi .
(26)
q − 1 i=1
q−1
This is illustrated in Figure 2 for the case where δ ∈ 0, q−1
and τ > q−1
, the symmetric
q
q > δ. For δ ∈ 0, q
channels that are degraded versions of Wδ are {Wγ : γ ∈ [δ, τ ]}. In particular, for such γ ∈ [δ, τ ], Wγ = Wδ Wβ with
δ
β = (γ − δ)/(1 − δ − q−1
) using the proof of part 5 of Proposition 4 in Appendix B.
In the spirit of comparing symmetric and erasure channels as done in [15] for the binary input case, our next result
shows that a q-ary symmetric channel can never be less noisy than a q-ary erasure channel.
q×(q+1)
Proposition 7 (Symmetric Channel 6ln Erasure Channel). For q ∈ N\{1}, given a q-ary erasure channel E ∈ Rsto
with erasure probability ∈ (0, 1), there does not exist δ ∈ (0, 1) such that the corresponding q-ary symmetric channel
Wδ ∈ Rq×q
sto on the same input alphabet satisfies Wδ ln E .
12
Proof. For a q-ary erasure channel E with ∈ (0, 1), we always have D(uE ||∆0 E ) = +∞ for u, ∆0 = (1, 0, . . . , 0) ∈
Pq . On the other hand, for any q-ary symmetric channel Wδ with δ ∈ (0, 1), we have D(PX Wδ ||QX Wδ ) < +∞ for
every PX , QX ∈ Pq . Thus, Wδ 6ln E for any δ ∈ (0, 1).
In fact, the argument for Proposition 7 conveys that a symmetric channel Wδ ∈ Rq×q
sto with δ ∈ (0, 1) satisfies
Wδ ln V for some channel V ∈ Rq×r
only
if
D(P
V
||Q
V
)
<
+∞
for
every
P
,
Q
X
X
X
X ∈ Pq . Typically, we are only
sto
interested in studying q-ary symmetric channels with q ≥ 2 and δ ∈ 0, q−1
.
For
example,
the BSC with crossover
q
probability δ is usually studied for δ ∈ 0, 12 . Indeed, the less noisy domination characteristics of the extremal q-ary
q×q
symmetric channels with δ = 0 or δ = q−1
q are quite elementary. Given q ≥ 2, W0 = Iq ∈ Rsto satisfies W0 ln V ,
q×r
and W(q−1)/q = 1u ∈ Rq×q
sto satisfies V ln W(q−1)/q , for every channel V ∈ Rsto on a common input alphabet.
q×(q+1)
For the sake of completeness, we also note that for q ≥ 2, the extremal q-ary erasure channels E0 ∈ Rsto
and
q×(q+1)
q×r
E1 ∈ Rsto
, with = 0 and = 1 respectively, satisfy E0 ln V and V ln E1 for every channel V ∈ Rsto
on a
common input alphabet.
The result that the q-ary symmetric channel with uniform noise pmf W(q−1)/q is more noisy than every channel on
the same input alphabet has an analogue concerning additive white Gaussian noise (AWGN) channels. Consider all
additive noise channels of the form:
Y =X +Z
(27)
where X, Y ∈ R,
the 2input X is uncorrelated with the additive noise Z:2 E [XZ] = 0, and the noise Z has power
for some fixed σZ > 0. Let X = Xg ∼ N (0, σX ) (Gaussian distribution with mean 0 and
constraint E Z 2 ≤ σZ
2
) for some σX > 0. Then, we have:
variance σX
I Xg ; Xg + Z ≥ I Xg ; Xg + Zg
(28)
2
), Zg is independent of Xg , and equality occurs if and only if Z = Zg in distribution [28, Section
where Zg ∼ N (0, σZ
4.7]. This states that Gaussian noise is the “worst case additive noise” for a Gaussian source. Hence, the AWGN
channel is not more capable than any other additive noise channel with the same constraints. As a result, the AWGN
channel is not less noisy than any other additive noise channel with the same constraints (using Proposition 3).
IV. E QUIVALENT CHARACTERIZATIONS OF LESS NOISY PREORDER
Having studied the structure of less noisy domination and degradation regions of channels, we now consider the
problem of verifying whether a channel W is less noisy than another channel V . Since using Definition 2 or Proposition
1 directly is difficult, we often start by checking whether V is a degraded version of W . When this fails, we typically
resort to verifying van Dijk’s condition in Proposition 2, cf. [12, Theorem 2]. In this section, we prove the equivalent
characterization of the less noisy preorder in Theorem 1, and then present some useful corollaries of van Dijk’s
condition.
A. Characterization using χ2 -divergence
Recall the general measure theoretic setup and the definition of χ2 -divergence from subsection II-A. It is wellknown that KL divergence is locally approximated by χ2 -divergence, e.g. [28, Section 4.2]. While this approximation
sometimes fails globally, cf. [30], the following notable result was first shown by Ahlswede and Gács in the discrete
case in [2], and then extended to general alphabets in [3, Theorem 3]:
ηKL (W ) = ηχ2 (W ) ,
sup
PX ,QX
0<χ (PX ||QX )<+∞
χ2 (PX W ||QX W )
χ2 (PX ||QX )
(29)
2
for any Markov kernel W : H1 × X → [0, 1], where ηKL (W ) is defined as in (11), ηχ2 (W ) is the contraction coefficient
for χ2 -divergence, and the suprema in ηKL (W ) and ηχ2 (W ) are taken over all probability measures PX and QX on
(X , F). Since ηKL characterizes less noisy domination with respect to an erasure channel as mentioned in subsection I-D,
(29) portrays that ηχ2 also characterizes this. We will now prove Theorem 1 from subsection II-A, which generalizes
(29) and illustrates that χ2 -divergence actually characterizes less noisy domination by an arbitrary channel.
Proof of Theorem 1. In order to prove the forward direction, we recall the local approximation of KL divergence using
χ2 -divergence from [28, Proposition 4.2], which states that for any two probability measures PX and QX on (X , F):
2
lim+ 2 D λPX + λ̄QX ||QX = χ2 (PX ||QX )
(30)
λ→0 λ
13
where λ̄ = 1 − λ for λ ∈ (0, 1), and both sides of (30) are finite or infinite together. Then, we observe that for any
two probability measures PX and QX , and any λ ∈ [0, 1], we have:
D λPX W + λ̄QX W ||QX W ≥ D λPX V + λ̄QX V ||QX V
since W ln V . Scaling this inequality by
2
λ2
and letting λ → 0 produces:
χ2 (PX W ||QX W ) ≥ χ2 (PX V ||QX V )
as shown in (30). This proves the forward direction.
To establish the converse direction, we recall an integral representation of KL divergence using χ2 -divergence
presented in [3, Appendix A.2] (which can be distilled from the argument in [31, Theorem 1]):5
Z ∞ 2
χ (PX ||QtX )
D(PX ||QX ) =
dt
(31)
t+1
0
1
t
PX + t+1
QX for t ∈ [0, ∞), and both
for any two probability measures PX and QX on (X , F), where QtX = 1+t
sides of (31) are finite or infinite together (as a close inspection of the proof in [3, Appendix A.2] reveals). Hence, for
every PX and QX , we have by assumption:
χ2 PX W ||QtX W ≥ χ2 PX V ||QtX V
which implies that:
Z
0
∞
Z ∞ 2
χ2 (PX W ||QtX W )
χ (PX V ||QtX V )
dt ≥
dt
t+1
t+1
0
⇒ D(PX W ||QX W ) ≥ D(PX V ||QX V ) .
Hence, W ln V , which completes the proof.
B. Characterizations via the Löwner partial order and spectral radius
We will use the finite alphabet setup of subsection I-B for the remaining discussion in this paper. In the finite alphabet
q×s
setting, Theorem 1 states that W ∈ Rq×r
sto is less noisy than V ∈ Rsto if and only if for every PX , QX ∈ Pq :
χ2 (PX W ||QX W ) ≥ χ2 (PX V ||QX V ) .
(32)
This characterization has the flavor of a Löwner partial order condition. Indeed, it is straightforward to verify that for
any PX ∈ Pq and QX ∈ Pq◦ , we can write their χ2 -divergence as:
−1
χ2 (PX ||QX ) = JX diag(QX )
T
JX
.
(33)
where JX = PX − QX . Hence, we can express (32) as:
−1
JX W diag(QX W )
T
W T JX
≥ JX V diag(QX V )
−1
T
V T JX
(34)
for every JX = PX − QX such that PX ∈ Pq and QX ∈ Pq◦ . This suggests that (32) is equivalent to:
−1
W diag(QX W )
−1
W T PSD V diag(QX V )
VT
(35)
for every QX ∈ Pq◦ . It turns out that (35) indeed characterizes ln , and this is straightforward to prove directly. The
next proposition illustrates that (35) also follows as a corollary of van Dijk’s characterization in Proposition 2, and
presents an equivalent spectral characterization of ln .
q×s
q×r
and V ∈ Rsto
Proposition 8 (Löwner and Spectral Characterizations of ln ). For any pair of channels W ∈ Rsto
on the same input alphabet [q], the following are equivalent:
1) W ln V
2) For every PX ∈ Pq◦ , we have:
W diag(PX W )
5 Note
−1
−1
W T PSD V diag(PX V )
that [3, Equation (78)], and hence [1, Equation (7)], are missing factors of
14
1
t+1
VT
inside the integrals.
−1
−1
3) For every PX ∈ Pq◦ , we have R V diag(PX V ) V T ⊆ R W diag(PX W ) W T and:6
†
−1 T
−1
T
V diag(PX V ) V
= 1.
ρ W diag(PX W ) W
Proof. (1 ⇔ 2) Recall the functional F : Pq → R, F (PX ) = I(PX , WY |X ) − I(PX , VY |X ) defined in Proposition 2,
cf. [12, Theorem 2]. Since F : Pq → R is continuous on its domain Pq , and twice differentiable on Pq◦ , F is concave
if and only if its Hessian is negative semidefinite for every PX ∈ Pq◦ (i.e. −∇2 F (PX ) PSD 0 for every PX ∈ Pq◦ )
q×q
[32, Section 3.1.4]. The Hessian matrix of F , ∇2 F : Pq◦ → Rsym
, is defined entry-wise for every x, x0 ∈ [q] as:
∂2F
(PX )
∂PX (x)∂PX (x0 )
2
∇ F (PX ) x,x0 =
where we index the matrix ∇2 F (PX ) starting at 0 rather than 1. Furthermore, a straightforward calculation shows that:
−1
∇2 F (PX ) = V diag(PX V )
V T − W diag(PX W )
−1
WT
for every PX ∈ Pq◦ . (Note that the matrix inverses here are well-defined because PX ∈ Pq◦ ). Therefore, F is concave
if and only if for every PX ∈ Pq◦ :
−1
W diag(PX W )
W T PSD V diag(PX V )
−1
V T.
This establishes the equivalence between parts 1 and 2 due to van Dijk’s characterization of ln in Proposition 2.
(2 ⇔ 3) We now derive the spectral characterization of ln using part 2. Recall the well-known fact (see [33, Theorem
1 parts (a),(f)] and [19, Theorem 7.7.3 (a)]):
†
Given positive semidefinite matrices A, B ∈ Rq×q
0 , A PSD B if and only if R(B) ⊆ R(A) and ρ A B ≤ 1.
−1
−1
Since W diag(PX W ) W T and V diag(PX V ) V T are positive semidefinite for every PX ∈ Pq◦ , applying this fact
−1
−1
shows that part 2 holds if and only if for every PX ∈ Pq◦ , we have R V diag(PX V ) V T ⊆ R W diag(PX W ) W T
and:
†
−1
−1 T
T
ρ
W diag(PX W ) W
V diag(PX V ) V
≤ 1.
−1
To prove that this inequality is actually an equality, for any PX ∈ Pq◦ , let A = W diag(PX W ) W T and B =
−1
V diag(PX V ) V T . It suffices to prove that: R(B) ⊆ R(A) and ρ A† B ≤ 1 if and only if R(B) ⊆ R(A) and
ρ A† B = 1. The converse direction is trivial, so we only establish the forward direction. Observe that PX A = 1T
†
and PX B = 1T . This implies that 1T A† B = PX (AA† )B = PX B = 1T , where
(AA )B =† B because R(B) ⊆
†
†
R(A) and AA is the orthogonal projection matrix onto R(A). Since ρ A B ≤ 1 and A B has an eigenvalue
of 1, we have ρ A† B = 1. Thus, we have proved that part 2 holds if and only if for every PX ∈ Pq◦ , we have
−1
−1
R V diag(PX V ) V T ⊆ R W diag(PX W ) W T and:
†
−1
−1 T
T
W diag(PX W ) W
V diag(PX V ) V
= 1.
ρ
This completes the proof.
The Löwner characterization of ln in part 2 of Proposition 8 will be useful for proving some of our ensuing results.
We remark that the equivalence between parts 1 and 2 can be derived by considering several other functionals. For
instance, for any fixed pmf QX ∈ Pq◦ , we may consider the functional F2 : Pq → R defined by:
F2 (PX ) = D(PX W ||QX W ) − D(PX V ||QX V )
(36)
−1
−1
2
which has Hessian matrix, ∇2 F2 : Pq◦ → Rq×q
W T − V diag(PX V ) V T , that does
sym , ∇ F2 (PX ) = W diag(PX W )
not depend on QX . Much like van Dijk’s functional F , F2 is convex (for all QX ∈ Pq◦ ) if and only if W ln V . This is
reminiscent of Ahlswede and Gács’ technique to prove (29), where the convexity of a similar functional is established
[2].
As another example, for any fixed pmf QX ∈ Pq◦ , consider the functional F3 : Pq → R defined by:
F3 (PX ) = χ2 (PX W ||QX W ) − χ2 (PX V ||QX V )
−1
(37)
2
which has Hessian matrix, ∇2 F3 : Pq◦ → Rq×q
W T − 2 V diag(QX V )
sym , ∇ F3 (PX ) = 2 W diag(QX W )
◦
does not depend on PX . Much like F and F2 , F3 is convex for all QX ∈ Pq if and only if W ln V .
6 Note
that [1, Theorem 1 part 4] neglected to mention the inclusion relation R V diag(PX V )−1 V T
15
−1
V T , that
⊆ R W diag(PX W )−1 W T .
Finally, we also mention some specializations of the spectral radius condition in part 3 of Proposition 8. If q ≥ r
and W has full column rank, the expression for spectral radius in the proposition statement can be simplified to:
−1
(38)
ρ (W † )T diag(PX W ) W † V diag(PX V ) V T = 1
using basic properties of the Moore-Penrose pseudoinverse. Moreover, if q = r and W is non-singular, then the MoorePenrose pseudoinverses in (38) can be written as inverses, and the inclusion relation between the ranges in part 3 of
Proposition 8 is trivially satisfied (and can be omitted from the proposition statement). We have used the spectral radius
condition in this latter setting to (numerically) compute the additive less noisy domination region in Figure 2.
V. C ONDITIONS FOR LESS NOISY DOMINATION OVER ADDITIVE NOISE CHANNELS
We now turn our attention to deriving several conditions for determining when q-ary symmetric channels are less
noisy than other channels. Our interest in q-ary symmetric channels arises from their analytical tractability; Proposition
4 from subsection I-C, Proposition 12 from section VII, and [34, Theorem 4.5.2] (which conveys that q-ary symmetric
channels have uniform capacity achieving input distributions) serve as illustrations of this tractability. We focus on
additive noise channels in this section, and on general channels in the next section.
A. Necessary conditions
q×q
We first present some straightforward necessary conditions for when an additive noise channel W ∈ Rsto
with
q×q
q ∈ N is less noisy than another additive noise channel V ∈ Rsto on an Abelian group (X , ⊕). These conditions can
obviously be specialized for less noisy domination by symmetric channels.
Proposition 9 (Necessary Conditions for ln Domination over Additive Noise Channels). Suppose W = circX (w) and
V = circX (v) are additive noise channels with noise pmfs w, v ∈ Pq such that W ln V . Then, the following are true:
1) (Circle Condition) kw − uk`2 ≥ kv − uk`2 .
2) (Contraction Condition) ηKL (W ) ≥ ηKL (V ).
3) (Entropy Condition) H (v) ≥ H (w), where H : Pq → R+ is the Shannon entropy function.
Proof.
Part 1: Letting PX = (1, 0, . . . , 0) and QX = u in the χ2 -divergence characterization of ln in Theorem 1 produces:
2
2
q kw − uk`2 = χ2 (w||u) ≥ χ2 (v||u) = q kv − uk`2
since uW = uV = u, and PX W = w and PX V = v using (7). (This result can alternatively be proved using part 2
of Proposition 8 and Fourier analysis.)
Part 2: This easily follows from Proposition 1 and (11).
Part 3: Letting PX = (1, 0, . . . , 0) and QX = u in the KL divergence characterization of ln in Proposition 1 produces:
log (q) − H (w) = D (w||u) ≥ D (v||u) = log (q) − H (v)
via the same reasoning as part 1. This completes the proof.
We remark that the aforementioned necessary conditions have many generalizations. Firstly, if W, V ∈
doubly stochastic matrices, then the generalized circle condition holds:
≥ V − W q−1
W − W q−1
q
Fro
q
Fro
Rq×q
sto
are
(39)
where W(q−1)/q = 1u is the q-ary symmetric channel whose conditional pmfs are all uniform, and k·kFro denotes the
Frobenius norm. Indeed, letting PX = ∆x = (0, . . . , 1, . . . , 0) for x ∈ [q], which has unity in the (x + 1)th position,
in the proof of part 1 and then adding the inequalities corresponding to every x ∈ [q] produces (39). Secondly, the
q×s
contraction condition in Proposition 9 actually holds for any pair of general channels W ∈ Rq×r
sto and V ∈ Rsto on a
common input alphabet (not necessarily additive noise channels). Moreover, we can start with Theorem 1 and take the
suprema of the ratios in χ2 (PX W ||QX W ) /χ2 (PX ||QX ) ≥ χ2 (PX V ||QX V ) /χ2 (PX ||QX ) over all PX (6= QX ) to
get:
ρmax (QX , W ) ≥ ρmax (QX , V )
(40)
for any QX ∈ Pq , where ρmax (·) denotes maximal correlation which is defined later in part 3 of Proposition 12, cf.
[35], and we use [36, Theorem 3] (or the results of [37]). A similar result also holds for the contraction coefficient for
KL divergence with fixed input pmf (see e.g. [36, Definition 1] for a definition).
16
B. Sufficient conditions
q×q
We next portray a sufficient condition for when an additive noise channel V ∈ Rsto
is a degraded version of a
q×q
symmetric channel Wδ ∈ Rsto . By Proposition 3, this is also a sufficient condition for Wδ ln V .
Proposition 10 (Degradation by Symmetric Channels). Given an additive noise channel V = circX (v) with noise pmf
v ∈ Pq and minimum probability τ = min{[V ]i,j : 1 ≤ i, j ≤ q}, we have:
0 ≤ δ ≤ (q − 1) τ ⇒ Wδ deg V
where Wδ ∈
Rq×q
sto
is a q-ary symmetric channel.
Proof. Using Corollary 1, it suffices to prove that the noise pmf w(q−1)τ maj v. Since 0 ≤ τ ≤ 1q , we must have
0 ≤ (q − 1)τ ≤ q−1
q . So, all entries of w(q−1)τ , except (possibly) the first, are equal to its minimum entry of τ . As
v ≥ τ (entry-wise), w(q−1)τ maj v because the conditions of part 3 in Proposition 13 in Appendix A are satisfied.
It is compelling to find a sufficient condition for Wδ ln V that does not simply ensure Wδ deg V (such as Proposition
10 and Theorem 2). The ensuing proposition elucidates such a sufficient condition for additive noise channels. The
general strategy for finding such a condition for additive noise channels is to identify a noise pmf that belongs to
add
add
Ladd
Wδ \DWδ . One can then use Proposition 6 to explicitly construct a set of noise pmfs that is a subset of LWδ but
add
strictly includes DWδ . The proof of Proposition 11 finds such a noise pmf (that corresponds to a q-ary symmetric
channel).
Proposition 11 (Less Noisy Dominationby Symmetric
Channels). Given an additive noise channel V = circX (v) with
we
have:
noise pmf v ∈ Pq and q ≥ 2, if for δ ∈ 0, q−1
q
v ∈ conv wδ Pqk : k ∈ [q] ∪ wγ Pqk : k ∈ [q]
then Wδ ln V , where Pq ∈ Rq×q is defined in (8), and:
γ=
1−δ
δ
,
1
.
∈
1
−
δ
q−1
1 − δ + (q−1)
2
Proof. Due to Proposition 6 and {wγ Px : x ∈ X } = {wγ Pqk : k ∈ [q]}, it suffices toprove that Wδ ln Wγ . Since
q−1
δ = 0 ⇒ γ = 1 and δ = q−1
. So, we assume that
⇒ γ = q−1
q
q , Wδ ln Wγ is certainly true for δ ∈ 0, q
δ ∈ 0, q−1
,
which
implies
that:
q
q−1
1−δ
∈
,1 .
γ=
δ
q
1 − δ + (q−1)
2
Since our goal is to show Wδ ln Wγ , we prove the equivalent condition in part 2 of Proposition 8 that for every
PX ∈ Pq◦ :
−1
Wδ diag(PX Wδ )
−1
WδT PSD Wγ diag(PX Wγ )
WγT
⇔ Wγ−1 diag(PX Wγ ) Wγ−1 PSD Wδ−1 diag(PX Wδ ) Wδ−1
⇔ diag(PX Wγ ) PSD Wγ Wδ−1 diag(PX Wδ ) Wδ−1 Wγ
−1
−21
⇔ Iq PSD diag(PX Wγ ) 2 Wτ diag(PX Wδ )Wτ diag(PX Wγ )
−1
−21
⇔ 1 ≥ diag(PX Wγ ) 2 Wτ diag(PX Wδ )Wτ diag(PX Wγ )
⇔ 1 ≥ diag(PX Wγ )
− 12
Wτ diag(PX Wδ )
op
1
2
op
where the second equivalence holds because Wδ and Wγ are symmetric and invertible (see part 4 of Proposition 4
and [19, Corollary 7.7.4]), the third and fourth equivalences are non-singular ∗-congruences with Wτ = Wδ−1 Wγ =
Wγ Wδ−1 and:
γ−δ
τ=
>0
δ
1 − δ − q−1
which can be computed as in the proof of Proposition 15 in Appendix C, and k·kop denotes the spectral or operator
norm.7
q×q
7 Note that we cannot use the strict Löwner partial order
PSD (for A, B ∈ Rsym , A PSD B if and only if A − B is positive definite) for these
equivalences as 1T Wγ−1 diag(PX Wγ ) Wγ−1 1 = 1T Wδ−1 diag(PX Wδ ) Wδ−1 1.
17
−1
1
It is instructive to note that if Wτ ∈ Rq×q
transition matrix diag(PX Wγ ) 2 Wτ diag(PX Wδ ) 2
sto , then the divergence
p
√
T
T
has right singular vector PX Wδ and left singular vector PX Wγ corresponding to its maximum singular value
of unity (where the square roots are applied entry-wise)–see [36] and the references therein. So, Wτ ∈ Rq×q
sto is a
δ
sufficient condition for Wδ ln Wγ . Since Wτ ∈ Rq×q
if
and
only
if
0
≤
τ
≤
1
if
and
only
if
δ
≤
γ
≤
1
−
sto
q−1 , the
latter condition also implies that Wδ ln Wγ . However, we recall from (25) in subsection III-B that Wδ deg Wγ for
δ
δ
, while we seek some 1 − q−1
< γ ≤ 1 for which Wδ ln Wγ . When q = 2, we only have:
δ ≤ γ ≤ 1 − q−1
γ=
1−δ
δ
=1−
=1−δ
δ
q
−
1
1 − δ + (q−1)
2
which implies that Wδ deg Wγ is true for q = 2. On the other hand, when q ≥ 3, it is straightforward to verify that:
1−δ
δ
γ=
∈ 1−
,1
δ
q−1
1 − δ + (q−1)
2
.
since δ ∈ 0, q−1
q
From the preceding discussion, it suffices to prove for q ≥ 3 that for every PX ∈ Pq◦ :
−21
diag(PX Wγ )
−21
Wτ diag(PX Wδ )Wτ diag(PX Wγ )
Since τ > 0, and 0 ≤ τ ≤ 1 does not produce γ > 1 −
strictly negative entries along the diagonal. Notice that:
δ
q−1 ,
≤ 1.
op
we require that τ > 1 (⇔ γ > 1 −
δ
q−1 )
so that Wτ has
∀x ∈ [q], diag(∆x Wγ ) PSD Wγ Wδ−1 diag(∆x Wδ ) Wδ−1 Wγ
where ∆x = (0, . . . , 1, . . . , 0) ∈ Pq denotes the Kronecker delta pmf with unity at the (x + 1)th position, implies that:
diag(PX Wγ ) PSD Wγ Wδ−1 diag(PX Wδ ) Wδ−1 Wγ
for every PX ∈ Pq◦ , because convex combinations preserve the Löwner relation. So, it suffices to prove that for every
x ∈ [q]:
− 1
− 1
≤1
diag wγ Pqx 2 Wτ diag wδ Pqx Wτ diag wγ Pqx 2
op
where Pq ∈ Rq×q is defined in (8), because ∆x M extracts the (x + 1)th row of a matrix M ∈ Rq×q . Let us define
− 1
− 1
Ax , diag wγ Pqx 2 Wτ diag wδ Pqx Wτ diag wγ Pqx 2 for each x ∈ [q]. Observe that for every x ∈ [q], Ax ∈ Rq×q
0
is
diagonalizable by the real spectral theorem [38, Theorem 7.13], and has a strictly positive eigenvector
porthogonally
wγ Pqx corresponding to the eigenvalue of unity:
q
q
∀x ∈ [q],
wγ Pqx Ax = wγ Pqx
p
so that all other eigenvectors of Ax have some strictly negative entries since they are orthogonal to wγ Pqx . Suppose Ax
is entry-wise non-negative for every x ∈ [q]. Then, the largest eigenvalue (known as the Perron-Frobenius eigenvalue)
and the spectral radius of each Ax is unity by the Perron-Frobenius theorem [19, Theorem 8.3.4], which proves that
kAx kop ≤ 1 for every x ∈ [q]. Therefore, it is sufficient to prove that Ax is entry-wise non-negative for every x ∈ [q].
− 1
Equivalently, we can prove that Wτ diag wδ Pqx Wτ is entry-wise non-negative for every x ∈ [q], since diag wγ Pqx 2
scales the rows or columns of the matrix it is pre- or post-multiplied with using strictly positive scalars.
We now show the equivalent condition below that the minimum possible entry of Wτ diag wδ Pqx Wτ is non-negative:
0 ≤ min
q
X
x∈[q]
r=1
1≤i,j≤q |
[Wτ ]i,r [Wδ ]x+1,r [Wτ ]r,j
=
=
{z
[Wτ diag(wδ Pqx )Wτ ]i,j
}
τ (1 − δ)(1 − τ ) δτ (1 − τ )
δτ 2
+
+ (q − 2)
.
2
q−1
(q − 1)
(q − 1)3
The above equality holds because for i 6= j:
q
q
δ X
δ X
[Wτ ]i,r [Wτ ]r,i ≥
[Wτ ]i,r [Wτ ]r,j
q − 1 r=1 |
{z
} q − 1 r=1
= [Wτ ]2i,r ≥ 0
18
(41)
2
δ
is clearly true (using, for example, the rearrangement inequality in [39, Section 10.2]), and adding 1−δ− q−1
[Wτ ]i,k ≥
δ
[Wτ ]i,p
0 (regardless of the value of 1 ≤ k ≤ q) to the left summation increases its value, while adding 1 − δ − q−1
[Wτ ]p,j < 0 (which exists for an appropriate value 1 ≤ p≤ q as τ > 1) to the right summation decreases its value.
As a result, the minimum possible entry of Wτ diag wδ Pqx Wτ can be achieved with x + 1 = i 6= j or i 6= j = x + 1.
δ
We next substitute τ = (γ − δ)/ 1 − δ − q−1
into (41) and simplify the resulting expression to get:
!
δ
(q − 2) δ (γ − δ)
δ
.
0 ≤ (γ − δ)
−γ
1−δ+
+
1−
2
q−1
q−1
(q − 1)
1−δ
The right hand side of this inequality is quadratic in γ with roots γ = δ and γ = 1−δ+(δ/(q−1)
2 ) . Since the coefficient
2
of γ in this quadratic is strictly negative:
δ
δ
(q − 2) δ
<0⇔1−δ+
2 − 1−δ+ q−1
2 >0
(q − 1)
(q − 1)
|
{z
}
coefficient of γ 2
the minimum possible entry of Wτ diag wδ Pqx Wτ is non-negative if and only if:
δ≤γ≤
where we use the fact that
completes the proof.
1−δ
1−δ+(δ/(q−1)2 )
1−δ
δ
1 − δ + (q−1)
2
δ
≥ 1 − q−1
≥ δ. Therefore, γ =
1−δ
1−δ+(δ/(q−1)2 )
produces Wδ ln Wγ , which
Heretofore we have derived results concerning less noisy domination and degradation regions in section III, and
proven several necessary and sufficient conditions for less noisy domination of additive noise channels by symmetric
channels in this section. We finally have all the pieces in place to establish Theorem 3 from section II. In closing this
section, we indicate the pertinent results that coalesce to justify it.
Proof of Theorem 3. The first equality follows from Corollary 1. The first set inclusion is obvious, and its strictness
follows from the proof of Proposition 11. The second set inclusion follows from Proposition 11. The third set inclusion
follows from the circle condition (part 1) in Proposition 9. Lastly, the properties of Ladd
Wδ are derived in Proposition
6.
VI. S UFFICIENT CONDITIONS FOR DEGRADATION OVER GENERAL CHANNELS
q×q
While Propositions 10 and 11 present sufficient conditions for a symmetric channel Wδ ∈
sto to be less noisy than
Rq−1
an additive noise channel, our more comprehensive objective is to find the maximum δ ∈ 0, q such that Wδ ln V
for any given general channel V ∈ Rq×r
sto on a common input alphabet. We may formally define this maximum δ (that
characterizes the extremal symmetric channel that is less noisy than V ) as:
q−1
δ ? (V ) , sup δ ∈ 0,
: Wδ ln V
(42)
q
and for every 0 ≤ δ < δ ? (V ), Wδ ln V . Alternatively, we can define a non-negative (less noisy) domination factor
function for any channel V ∈ Rq×r
sto :
µV (δ) ,
sup
PX ,QX ∈Pq :
0<D(PX Wδ ||QX Wδ )<+∞
D (PX V ||QX V )
D (PX Wδ ||QX Wδ )
(43)
with δ ∈ 0, q−1
, which is analogous to the contraction coefficient for KL divergence since µV (0) , ηKL (V ). Indeed,
q
we may perceive PX Wδ and QX Wδ in the denominator of (43) as pmfs inside the “shrunk” simplex conv({wδ Pqk :
k ∈ [q]}), and (43) represents a contraction coefficient of V where the supremum is taken over this “shrunk” simplex.8
For simplicity, consider a channel V ∈ Rq×r
sto that is strictly positive entry-wise, and has domination factor
function
q−1
+
µV : 0, q−1
→
R
,
where
the
domain
excludes
zero
because
µ
is
only
interesting
for
δ
∈
0,
, and this
V
q
q
exclusion also affords us some analytical simplicity. It is shown in Proposition 15 of Appendix C that µV is always
8 Pictorially,
the “shrunk” simplex is the magenta triangle in Figure 2 while the simplex itself is the larger gray triangle.
19
finite on 0, q−1
, continuous, convex, strictly increasing, and has a vertical asymptote at δ =
q
PX , QX ∈ Pq :
µV (δ) D (PX Wδ ||QX Wδ ) ≥ D (PX V ||QX V )
q−1
q .
Since for every
we have µV (δ) ≤ 1 if and only if Wδ ln V . Hence, using the strictly increasing property of µV : 0,
we can also characterize δ ? (V ) as:
δ ? (V ) = µ−1
V (1)
(44)
q−1
q
→ R+ ,
(45)
where µ−1
V denotes the inverse function of µV , and unity is in the range of µV by Theorem 2 since V is strictly
positive entry-wise.
q×r
We next briefly delineate how one might computationally approximate δ ? (V ) for a given general channel V ∈ Rsto
.
?
From part 2 of Proposition 8, it is straightforward to obtain the following minimax characterization of δ (V ):
δ ? (V ) =
inf
sup
PX ∈Pq◦ δ∈S(PX )
δ
(46)
−1
−1
: Wδ diag(PX Wδ ) WδT PSD V diag(PX V ) V T . The infimum in (46) can be naïvely
where S(PX ) = δ ∈ 0, q−1
q
approximated by sampling several PX ∈ Pq◦ . The supremum in (46) can be estimated by verifying collections of rational
(ratio of polynomials) inequalities in δ. This is because the positive semidefiniteness of a matrix is equivalent to the
non-negativity of all its principal minors by Sylvester’s criterion [19, Theorem 7.2.5]. Unfortunately, this procedure
appears to be rather cumbersome.
Since analytically computing δ ? (V ) also seems intractable, we now prove Theorem 2 from section II. Theorem 2
provides a sufficient condition for Wδ deg V (which implies Wδ ln V using Proposition 3) by restricting its attention
?
to the case where V ∈ Rq×q
sto with q ≥ 2. Moreover, it can be construed as a lower bound on δ (V ):
ν
(47)
δ ? (V ) ≥
ν
1 − (q − 1)ν + q−1
where ν = min {[V ]i,j : 1 ≤ i, j ≤ q} is the minimum conditional probability in V .
q×q
Proof of Theorem 2. Let the channel V ∈ Rsto
have the conditional pmfs v1 , . . . , vq ∈ Pq as its rows:
T
V = v1T v2T · · · vqT .
From the proof of Proposition 10, we know that w(q−1)ν maj vi for every i ∈ {1, . . . , q}. Using part 1 of Proposition
13 in Appendix A (and the fact that the set of all permutations of w(q−1)ν is exactly the set of all cyclic permutations
of w(q−1)ν ), we can write this as:
∀i ∈ {1, . . . , q} , vi =
q
X
pi,j w(q−1)ν Pqj−1
j=1
Pq
where Pq ∈ R
is given in (8), and {pi,j ≥ 0 : 1 ≤ i, j ≤ q} are the convex weights such that j=1 pi,j = 1 for
every i ∈ {1, . . . , q}. Defining P ∈ Rq×q
sto entry-wise as [P ]i,j = pi,j for every 1 ≤ i, j ≤ q, we can also write this
9
equation as V = P W(q−1)ν . Observe that:
!
q
X
Y
P =
pi,ji Ej1 ,...,jq
q×q
1≤j1 ,...,jq ≤q
i=1
Qq
where { i=1 pi,ji : 1 ≤ j1 , . . . , jq ≤ q} form a product pmf of convex weights, and for every 1 ≤ j1 , . . . , jq ≤ q:
T
Ej1 ,...,jq , ej1 ej2 · · · ejq
where ei ∈ Rq is the ith standard basis (column) vector that has unity at the ith entry and zero elsewhere. Hence, we
get:
!
q
X
Y
V =
pi,ji Ej1 ,...,jq W(q−1)ν .
1≤j1 ,...,jq ≤q
i=1
q×q
9 Matrices of the form V = P W
are not necessarily degraded versions of W(q−1)ν : W(q−1)ν 6deg V (although
(q−1)ν with P ∈ Rsto
we certainly have input-output degradation: W(q−1)ν iod V ). As a counterexample, consider W1/2 for q = 3, and P = [1 0 0; 1 0 0; 0 1 0],
where the colons separate the rows of the matrix. If W1/2 deg P W1/2 , then there exists A ∈ R3×3
sto such that P W1/2 = W1/2 A. However,
−1
A = W1/2
P W1/2 = (1/4) [3 0 1; 3 0 1; −1 4 1] has a strictly negative entry, which leads to a contradiction.
20
Suppose there exists δ ∈ 0, q−1
such that for all j1 , . . . , jq ∈ {1, . . . , q}:
q
∃Mj1 ,...,jq ∈ Rq×q
sto , Ej1 ,...,jq W(q−1)ν = Wδ Mj1 ,...,jq
i.e. Wδ deg Ej1 ,...,jq W(q−1)ν . Then, we would have:
X
q
Y
1≤j1 ,...,jq ≤q
i=1
V = Wδ
!
pi,ji
Mj1 ,...,jq
{z
|
}
stochastic matrix
which implies that Wδ deg V .
q×q
We will demonstrate that for every j1 , . . . , jq ∈{1, . . . , q}, there exists Mj1 ,...,jq ∈ Rsto
such that Ej1 ,...,jq W(q−1)ν =
ν
1
Wδ Mj1 ,...,jq when 0 ≤ δ ≤ ν/ 1−(q−1)ν+ q−1 . Since 0 ≤ ν ≤ q , the preceding inequality implies that 0 ≤ δ ≤ q−1
q ,
1
1
where δ = q−1
is
possible
if
and
only
if
ν
=
.
When
ν
=
,
V
=
W
is
the
channel
with
all
uniform
conditional
(q−1)/q
q
q
q
pmfs, and W(q−1)/q deg V clearly holds. Hence, we assume that 0 ≤ ν < 1q so that 0 ≤ δ < q−1
q , and establish the
equivalent condition that for every j1 , . . . , jq ∈ {1, . . . , q}:
Mj1 ,...,jq = Wδ−1 Ej1 ,...,jq W(q−1)ν
−δ
using part 4 of Proposition 4. Clearly, all
is a valid stochastic matrix. Recall that Wδ−1 = Wτ with τ = 1−δ−(δ/(q−1))
the rows of each Mj1 ,...,jq sum to unity. So, it remains to verify that each Mj1 ,...,jq has non-negative entries. For any
j1 , . . . , jq ∈ {1, . . . , q} and any i, j ∈ {1, . . . , q}:
Mj1 ,...,jq i,j ≥ ν (1 − τ ) + τ (1 − (q − 1) ν)
where the right hand side is the minimum possible entry of any Mj1 ,...,jq (with equality when j1 > 1 and j2 = j3 =
· · · = jq = 1 for example) as τ < 0 and 1 − (q − 1) ν > ν. To ensure each Mj1 ,...,jq is entry-wise non-negative, the
minimum possible entry must satisfy:
ν (1 − τ ) + τ (1 − (q − 1) ν) ≥ 0
δν
δ (1 − (q − 1) ν)
⇔ ν+
−
≥0
δ
δ
1 − δ − q−1
1 − δ − q−1
and the latter inequality is equivalent to:
δ≤
ν
1 − (q − 1) ν +
ν
q−1
.
This completes the proof.
Rq×q
sto ,
We remark that if V = E2,1,...,1 W(q−1)ν ∈
then this proof illustrates that Wδ deg V if and only if 0 ≤ δ ≤
ν
. Hence, the condition in Theorem 2 is tight when no further information about V is known.
ν/ 1 − (q − 1)ν + q−1
It is worth juxtaposing Theorem 2 and Proposition 10. The upper bounds on δ from these results satisfy:
ν
(48)
ν ≤ (q − 1) ν
| {z }
1 − (q − 1)ν + q−1
|
{z
} upper bound in
upper bound in Theorem 2
Proposition 10
where we have equality if and only if ν = 1q , and it is straightforward to verify that (48) is equivalent to ν ≤ 1q .
Moreover, assuming that q is large and ν = o (1/q), the upper bound in Theorem 2 is ν/ 1 + o (1) + o 1/q 2 = Θ (ν),
while the upper bound in Proposition 10 is Θ (qν).10 (Note that both bounds are Θ (1) if ν = 1q .) Therefore, when
q×q
V ∈ Rq×q
sto is an additive noise channel, δ = O (qν) is enough for Wδ deg V , but a general channel V ∈ Rsto requires
δ = O (ν) for such degradation. So, in order to account for q different conditional pmfs in the general case (as opposed
to a single conditional pmf which characterizes the channel in the additive noise case), we loose a factor of q in the
upper bound on δ. Furthermore, we can check using simulations that Wδ ∈ Rq×q
sto is not in general less noisy than
V ∈ Rq×q
sto for δ = (q − 1)ν. Indeed, counterexamples can be easily obtained by letting V = Ej1 ,...,jq Wδ for specific
values of 1 ≤ j1 , . . . , jq ≤ q, and computationally verifying that Wδ 6ln V + J ∈ Rq×q
sto for appropriate choices of
perturbation matrices J ∈ Rq×q with sufficiently small Frobenius norm.
10 We use the Bachmann-Landau asymptotic notation here. Consider the (strictly) positive functions f : N → R and g : N → R. The little-o notation
is defined as: f (q) = o (g(q)) ⇔ limq→∞ f (q)/g(q) = 0. The big-O notation is defined as: f (q) = O (g(q)) ⇔ lim supq→∞ |f (q)/g(q)| <
+∞. Finally, the big-Θ notation is defined as: f (q) = Θ (g(q)) ⇔ 0 < lim inf q→∞ |f (q)/g(q)| ≤ lim supq→∞ |f (q)/g(q)| < +∞.
21
We have now proved Theorems 1, 2, and 3 from section II. The next section relates our results regarding less noisy
and degradation preorders to LSIs, and proves Theorem 4.
VII. L ESS NOISY DOMINATION AND LOGARITHMIC S OBOLEV INEQUALITIES
Logarithmic Sobolev inequalities (LSIs) are a class of functional inequalities that shed light on several important
phenomena such as concentration of measure, and ergodicity and hypercontractivity of Markov semigroups. We refer
readers to [40] and [41] for a general treatment of such inequalities, and more pertinently to [25] and [26], which
present LSIs in the context of finite state-space Markov chains. In this section, we illustrate that proving a channel
q×q
W ∈ Rq×q
sto is less noisy than a channel V ∈ Rsto allows us to translate an LSI for W to an LSI for V . Thus, important
information about V can be deduced (from its LSI) by proving W ln V for an appropriate channel W (such as a
q-ary symmetric channel) that has a known LSI.
We commence by introducing some appropriate notation and terminology associated with LSIs. For fixed input and
output alphabet X = Y = [q] with q ∈ N, we think of a channel W ∈ Rq×q
sto as a Markov kernel on X . We assume
that the “time homogeneous” discrete-time Markov chain defined by W is irreducible, and has unique stationary
distribution (or invariant measure) π ∈ Pq such that πW = π. Furthermore, we define the Hilbert space L2 (X , π) of
all real functions with domain X endowed with the inner product:
X
∀f, g ∈ L2 (X , π) , hf, giπ ,
π(x)f (x)g(x)
(49)
x∈X
2
2
and induced norm k·kπ . We construe W : L (X , π) → L (X , π) as a conditional expectation operator that takes a
T
function f ∈ L2 (X , π), which we can write as a column vector f = [f (0) · · · f (q − 1)] ∈ Rq , to another function
2
q
W f ∈ L (X , π), which we can also write as a column vector W f ∈ R . Corresponding to the discrete-time Markov
chain W , we may also define a continuous-time Markov semigroup:
q×q
∀t ≥ 0, Ht , exp (−t (Iq − W )) ∈ Rsto
(50)
where the “discrete-time derivative” W −Iq is the Laplacian operator that forms the generator of the Markov semigroup.
The unique stationary distribution of this Markov semigroup is also π, and we may interpret Ht : L2 (X , π) → L2 (X , π)
as a conditional expectation operator for each t ≥ 0 as well.
In order to present LSIs, we define the Dirichlet form EW : L2 (X , π) × L2 (X , π) → R:
∀f, g ∈ L2 (X , π) , EW (f, g) , h(Iq − W ) f, giπ
(51)
which is used to study properties of the Markov chain W and its associated Markov semigroup Ht ∈ Rq×q
sto : t ≥ 0 .
(EW is technically only a Dirichlet form when W is a reversible Markov chain, i.e. W is a self-adjoint operator, or
equivalently, W and π satisfy the detailed balance condition [25, Section 2.3, page 705].) Moreover, the quadratic
form defined by EW represents the energy of its input function, and satisfies:
W + W∗
2
(52)
f, f
∀f ∈ L (X , π) , EW (f, f ) = Iq −
2
π
where W ∗ : L2 (X , π) → L2 (X , π) is the adjoint operator of W . Finally, we introduce a particularly important
Dirichlet form corresponding to the channel W(q−1)/q = 1u, which has all uniform conditional pmfs and uniform
stationary distribution π = u, known as the standard Dirichlet form:
Estd (f, g) , E1u (f, g) = COVu (f, g)
!
!
X f (x)g(x)
X f (x)
X g(x)
=
−
q
q
q
x∈X
x∈X
2
(53)
x∈X
for any f, g ∈ L (X , u). The quadratic form defined by the standard Dirichlet form is presented in (17) in subsection
II-D.
q×q
:t≥0
We now present the LSIs associated with the Markov chain W and the Markov semigroup Ht ∈ Rsto
2
it defines. The LSI for the Markov semigroup with constant α ∈ R states that for every f ∈ L (X , π) such that
kf kπ = 1, we have:
X
1
D f 2 π || π =
π(x)f 2 (x) log f 2 (x) ≤ EW (f, f )
(54)
α
x∈X
22
where we construe µ = f 2 π ∈ Pq as a pmf such that µ(x) = f (x)2 π(x) for every x ∈ X , and f 2 behaves like the
Radon-Nikodym derivative (or density) of µ with respect to π. The largest constant α such that (54) holds:
α(W ) ,
inf
f ∈L2 (X ,π):
kf kπ =1
D(f 2 π||π)6=0
EW (f, f )
D (f 2 π || π)
(55)
is called the logarithmic Sobolev constant (LSI constant) of the Markov chain W (or the Markov chain (W + W ∗ )/2).
Likewise, the LSI for the discrete-time Markov chain with constant α ∈ R states that for every f ∈ L2 (X , π) such
that kf kπ = 1, we have:
1
D f 2 π || π ≤ EW W ∗ (f, f )
(56)
α
where EW W ∗ : L2 (X , π) × L2 (X , π) → R is the “discrete” Dirichlet form. The largest constant α such that (56)
holds is the LSI constant of the Markov chain W W ∗ , α(W W ∗ ), and we refer to it as the discrete logarithmic Sobolev
constant of the Markov chain W . As we mentioned earlier, there are many useful consequences of such LSIs. For
example, if (54) holds with constant (55), then for every pmf µ ∈ Pq :
∀t ≥ 0, D (µHt ||π) ≤ e−2α(W )t D (µ||π)
(57)
where the exponent 2α(W ) can be improved to 4α(W ) if W is reversible [25, Theorem 3.6]. This is a measure of
∗
ergodicity of the semigroup Ht ∈ Rq×q
sto : t ≥ 0 . Likewise, if (56) holds with constant α(W W ), then for every pmf
µ ∈ Pq :
n
∀n ∈ N, D (µW n ||π) ≤ (1 − α(W W ∗ )) D (µ||π)
(58)
as mentioned in [25, Remark, page 725] and proved in [42]. This is also a measure of ergodicity of the Markov chain
W.
Although LSIs have many useful consequences, LSI constants are difficult to compute analytically. Fortunately, the
LSI constant corresponding to Estd has been computed in [25, Appendix, Theorem A.1]. Therefore, using the relation
in (18), we can compute LSI constants for q-ary symmetric channels as well. The next proposition collects the LSI
constants for q-ary symmetric channels (which are irreducible for δ ∈ (0, 1]) as well as some other related quantities.
q×q
Proposition 12 (Constants of Symmetric Channels). The q-ary symmetric channel Wδ ∈ Rsto
with q ≥ 2 has:
1) LSI constant:
(
δ,
q=2
α(Wδ ) =
(q−2)δ
(q−1) log(q−1) , q > 2
for δ ∈ (0, 1].
2) discrete LSI constant:
(
α(Wδ Wδ∗ ) = α(Wδ0 ) =
2δ(1 − δ),
(q−2)(2q−2−qδ)δ
(q−1)2 log(q−1) ,
q=2
q>2
qδ
for δ ∈ (0, 1], where δ 0 = δ 2 − q−1
.
3) Hirschfeld-Gebelein-Rényi maximal correlation corresponding to the uniform stationary distribution u ∈ Pq :
ρmax (u, Wδ ) = 1 − δ −
δ
q−1
q×r
for δ ∈ [0, 1], where for any channel W ∈ Rsto
and any source pmf PX ∈ Pq , we define the maximal
correlation between the input random variable X ∈ [q] and the output random variable Y ∈ [r] (with joint pmf
PX,Y (x, y) = PX (x)WY |X (y|x)) as:
ρmax (PX , W ) ,
sup
E [f (X)g(Y )].
f :[q]→R, g:[r]→R
E[f (X)]=E[g(Y )]=0
E[f 2 (X)]=E[g 2 (Y )]=1
4) contraction coefficient for KL divergence bounded by:
2
δ
δ
1−δ−
≤ ηKL (Wδ ) ≤ 1 − δ −
q−1
q−1
23
for δ ∈ [0, 1].
Proof. See Appendix B.
In view of Proposition 12 and the intractability of computing LSI constants for general Markov chains, we often
q×q
“compare” a given irreducible channel V ∈ Rq×q
sto with a q-ary symmetric channel Wδ ∈ Rsto to try and establish
an LSI for it. We assume for the sake of simplicity that V is doubly stochastic and has uniform stationary pmf (just
like q-ary symmetric channels). Usually, such a comparison between Wδ and V requires us to prove domination of
Dirichlet forms, such as:
qδ
∀f ∈ L2 (X , u) , EV (f, f ) ≥ EWδ (f, f ) =
Estd (f, f )
(59)
q−1
where we use (18). Such pointwise domination results immediately produce LSIs, (54) and (56), for V . Furthermore,
they also lower bound the LSI constants of V ; for example:
α(V ) ≥ α(Wδ ) .
(60)
This is turn begets other results such as (57) and (58) for the channel V (albeit with worse constants in the exponents
since the LSI constants of Wδ are used instead of those for V ). More general versions of Dirichlet form domination
between Markov chains on different state spaces with different stationary distributions, and the resulting bounds on
their LSI constants are presented in [25, Lemmata 3.3 and 3.4]. We next illustrate that the information theoretic notion
of less noisy domination is a sufficient condition for various kinds of pointwise Dirichlet form domination.
Theorem 40 (Domination of Dirichlet Forms). Let W, V ∈ Rq×q
sto be doubly stochastic channels, and π = u be the
uniform stationary distribution. Then, the following are true:
1) If W ln V , then:
∀f ∈ L2 (X , u) , EV V ∗ (f, f ) ≥ EW W ∗ (f, f ) .
T
T
2) If W ∈ Rq×q
0 is positive semidefinite, V is normal (i.e. V V = V V ), and W ln V , then:
3) If W = Wδ ∈ Rq×q
sto
∀f ∈ L2 (X , u) , EV (f, f ) ≥ EW (f, f ) .
and Wδ ln V , then:
is any q-ary symmetric channel with δ ∈ 0, q−1
q
∀f ∈ L2 (X , u) , EV (f, f ) ≥
qδ
Estd (f, f ) .
q−1
Proof.
Part 1: First observe that:
1 T
f Iq − W W T f
q
1 T
2
∀f ∈ L (X , u) , EV V ∗ (f, f ) = f Iq − V V T f
q
∀f ∈ L2 (X , u) , EW W ∗ (f, f ) =
where we use the facts that W T = W ∗ and V T = V ∗ because the stationary distribution is uniform. This implies
that EV V ∗ (f, f ) ≥ EW W ∗ (f, f ) for every f ∈ L2 (X , u) if and only if Iq − V V T PSD Iq − W W T , which is true if
and only if W W T PSD V V T . Since W ln V , we get W W T PSD V V T from part 2 of Proposition 8 after letting
P X = u = PX W = P X V .
Part 2: Once again, we first observe using (52) that:
1 T
W + WT
2
∀f ∈ L (X , u) , EW (f, f ) = f
Iq −
f,
q
2
V +VT
1
f.
∀f ∈ L2 (X , u) , EV (f, f ) = f T Iq −
q
2
So, EV (f, f ) ≥ EW (f, f ) for every f ∈ L2 (X , u) if and only if W + W T /2 PSD V + V T /2. Since W W T PSD
V V T from the proof of part 1, it is sufficient to prove that:
W W T PSD V V T ⇒
W + WT
V +VT
PSD
.
2
2
(61)
Lemma 2 in Appendix C establishes the claim in (61) because W ∈ Rq×q
0 and V is a normal matrix.
q−1
Part 3: We note that when V is a normal matrix, this result follows from part 2 because Wδ ∈ Rq×q
,
0 for δ ∈ 0, q
24
as can be seen from part 2 of Proposition 4. For a general doubly stochastic channel V , we need to prove that
qδ
EV (f, f ) ≥ EWδ (f, f ) = q−1
Estd (f, f ) for every f ∈ L2 (X , u) (where we use (18)). Following the proof of part 2,
it is sufficient to prove (61) with W = Wδ :11
V +VT
Wδ2 PSD V V T ⇒ Wδ PSD
2
where Wδ2 = Wδ WδT and Wδ = Wδ + WδT /2. Recall the Löwner-Heinz theorem [43], [44], (cf. [45, Section 6.6,
Problem 17]), which states that for A, B ∈ Rq×q
0 and 0 ≤ p ≤ 1:
A PSD B ⇒ Ap PSD B p
(62)
or equivalently, f : [0, ∞) → R, f (x) = xp is an operator monotone function for p ∈ [0, 1]. Using (62) with p = 21
(cf. [19, Corollary 7.7.4 (b)]), we have:
1
Wδ2 PSD V V T ⇒ Wδ PSD V V T 2
1
T 2
because the Gramian matrix V V T ∈ Rq×q
is the unique positive semidefinite square root matrix of
0 . (Here, V V
V V T .)
Let V V T = QΛQT and (V + V T )/2 = U ΣU T be the spectral decompositions of V V T and (V + V T )/2, where Q
and U are orthogonal matrices with eigenvectors as columns, and Λ and Σ are diagonal matrices of eigenvalues. Since
√
V V T and (V + V T )/2 are both doubly stochastic, they both have the unit norm eigenvector 1/ q corresponding to
the maximum eigenvalue of unity. In fact, we have:
1
1
V +VT
1
1
T 2 1
VV
and
√ =√
√ =√
q
q
2
q
q
1
1
1
where we use the fact that (V V T ) 2 = QΛ 2 QT is the spectral decomposition of (V V T ) 2 . For any matrix A ∈ Rq×q
sym , let
λ1 (A) ≥ λ2 (A) ≥ · · · ≥ λq (A) denote the eigenvalues of A in descending order. Without loss of generality, we assume
1
that [Λ]j,j = λj (V V T ) and [Σ]j,j = λj ((V + V T )/2) for every 1 ≤ j ≤ q. So, λ1 ((V V T ) 2 ) = λ1 ((V + V T )/2) = 1,
√
and the first columns of both Q and U are equal to 1/ q.
From part 2 of Proposition 4, we have Wδ = QDQT = U DU T , where D is the diagonal matrix of eigenvalues
δ
such that [D]1,1 = λ1 (Wδ ) = 1 and [D]j,j = λj (Wδ ) = 1 − δ − q−1
for 2 ≤ j ≤ q. Note that we may use either of
√
the eigenbases, Q or U , because they both have first column 1/ q, which is the eigenvector of Wδ corresponding to
λ1 (Wδ ) = 1 since Wδ is doubly stochastic, and the remaining eigenvector columns are permitted to be any orthonormal
√
δ
basis of span(1/ q)⊥ as λj (Wδ ) = 1 − δ − q−1
for 2 ≤ j ≤ q. Hence, we have:
Wδ PSD V V T
Wδ PSD
21
1
1
⇔ QDQT PSD QΛ 2 QT ⇔ D PSD Λ 2 ,
V +VT
⇔ U DU T PSD U ΣU T ⇔ D PSD Σ.
2
1
1
In order to show that D PSD Λ 2 ⇒ D PSD Σ, it suffices to prove that Λ 2 PSD Σ. Recall from [45, Corollary
3.1.5] that for any matrix A ∈ Rq×q , we have:12
1
A + AT
T 2
∀i ∈ {1, . . . , q} , λi AA
.
(63)
≥ λi
2
1
Hence, Λ 2 PSD Σ is true, cf. [46, Lemma 2.5]. This completes the proof.
Theorem 40 includes Theorem 4 from section II as part 3, and also provides two other useful pointwise Dirichlet form
domination results. Part 1 of Theorem 40 states that less noisy domination implies discrete Dirichlet form domination.
In particular, if we have Wδ ln V for some irreducible q-ary symmetric channel Wδ ∈ Rq×q
sto and irreducible doubly
stochastic channel V ∈ Rq×q
sto , then part 1 implies that:
n
∀n ∈ N, D (µV n ||u) ≤ (1 − α(Wδ Wδ∗ )) D (µ||u)
(64)
2
that (61) trivially holds for W = Wδ with δ = (q−1)/q, because W(q−1)/q = W(q−1)/q
= 1u PSD V V T implies that V = W(q−1)/q .
states that for any matrix A ∈ Rq×q , the ith largest eigenvalue of the symmetric part of A is less than or equal to the ith largest singular
value of A (which is the ith largest eigenvalue of the unique positive semidefinite part (AAT )1/2 in the polar decomposition of A) for every
1 ≤ i ≤ q.
11 Note
12 This
25
for all pmfs µ ∈ Pq , where α(Wδ Wδ∗ ) is computed in part 2 of Proposition 12. However, it is worth mentioning that
(58) for Wδ and Proposition 1 directly produce (64). So, such ergodicity results for the discrete-time Markov chain V
do not require the full power of the Dirichlet form domination in part 1. Regardless, Dirichlet form domination results,
such as in parts 2 and 3, yield several functional inequalities (like Poincaré inequalities and LSIs) which have many
other potent consequences as well.
Parts 2 and 3 of Theorem 40 convey that less noisy domination also implies the usual (continuous) Dirichlet form
domination under regularity conditions. We note that in part 2, the channel W is more general than that in part 3, but
the channel V is restricted to be normal (which includes the case where V is an additive noise channel). The proofs of
these parts essentially consist of two segments. The first segment uses part 1, and the second segment illustrates that
pointwise domination of discrete Dirichlet forms implies pointwise domination of Dirichlet forms (as shown in (59)).
This latter segment is encapsulated in Lemma 2 of Appendix C for part 2, and requires a slightly more sophisticated
proof pertaining to q-ary symmetric channels in part 3.
VIII. C ONCLUSION
In closing, we briefly reiterate our main results by delineating a possible program for proving LSIs for certain Markov
q×q
chains. Given an arbitrary irreducible doubly stochastic channel V ∈ Rsto
with minimum entry ν = min{[V ]i,j : 1 ≤
i, j ≤ q} > 0 and q ≥ 2, we can first use Theorem 2 to generate a q-ary symmetric channel Wδ ∈ Rq×q
sto with
ν
such that Wδ deg V . This also means that Wδ ln V , using Proposition 3. Moreover, the
δ = ν/ 1 − (q − 1)ν + q−1
δ parameter can be improved using Theorem 3 (or Propositions 10 and 11) if V is an additive noise channel. We can
then use Theorem 40 to deduce a pointwise domination of Dirichlet forms. Since Wδ satisfies the LSIs (54) and (56)
with corresponding LSI constants given in Proposition 12, Theorem 40 establishes the following LSIs for V :
1
D f 2 u || u ≤
EV (f, f )
(65)
α(Wδ )
1
EV V ∗ (f, f )
(66)
D f 2 u || u ≤
α(Wδ Wδ∗ )
for every f ∈ L2 (X , u) such that kf ku = 1. These inequalities can be used to derive a myriad of important facts
about V . We note that the equivalent characterizations of the less noisy preorder in Theorem 1 and Proposition 8
are particularly useful for proving some of these results. Finally, we accentuate that Theorems 2 and 3 address our
motivation in subsection I-D by providing analogs of the relationship between less noisy domination by q-ary erasure
channels and contraction coefficients in the context of q-ary symmetric channels.
A PPENDIX A
BASICS OF MAJORIZATION THEORY
Since we use some majorization arguments in our analysis, we briefly introduce the notion of group majorization
over row vectors in Rq (with q ∈ N) in this appendix. Given a group G ⊆ Rq×q of matrices (with the operation of
matrix multiplication), we may define a preorder called G-majorization over row vectors in Rq . For two row vectors
x, y ∈ Rq , we say that x G-majorizes y if y ∈ conv ({xG : G ∈ G}), where {xG : G ∈ G} is the orbit of x under the
group G. Group majorization intuitively captures a notion of “spread” of vectors. So, x G-majorizes y when x is more
spread out than y with respect to G. We refer readers to [9, Chapter 14, Section C] and the references therein for a
thorough treatment of group majorization. If we let G be the symmetric group of all permutation matrices in Rq×q ,
then G-majorization corresponds to traditional majorization of vectors in Rq as introduced in [39]. The next proposition
collects some results about traditional majorization.
Proposition 13 (Majorization [9], [39]). Given row vectors x = (x1 , . . . , xq ) , y = (y1 , . . . , yq ) ∈ Rq , let x(1) ≤ · · · ≤
x(q) and y(1) ≤ · · · ≤ y(q) denote the re-orderings of x and y in ascending order. Then, the following are equivalent:
1) x majorizes y, or equivalently, y resides in the convex hull of all permutations of x.
2) y = xD for some doubly stochastic matrix D ∈ Rq×q
sto .
3) The entries of x and y satisfy:
k
X
and
i=1
q
X
i=1
x(i) ≤
x(i) =
k
X
i=1
q
X
y(i) , for k = 1, . . . , q − 1 ,
y(i) .
i=1
When these conditions are true, we will write x maj y.
26
In the context of subsection I-C, given an Abelian group (X , ⊕) of order q, another useful notion of G-majorization
can be obtained by letting G = {Px ∈ Rq×q : x ∈ X } be the group of permutation matrices defined in (4) that is
isomorphic to (X , ⊕). For such choice of G, we write x X y when x G-majorizes (or X -majorizes) y for any two
row vectors x, y ∈ Rq . We will only require one fact about such group majorization, which we present in the next
proposition.
Proposition 14 (Group Majorization). Given two row vectors x, y ∈ Rq , x X y if and only if there exists λ ∈ Pq
such that y = x circX (λ).
Proof. Observe that:
x X y ⇔ y ∈ conv ({xPz : z ∈ X })
⇔ y = λ circX (x) for some λ ∈ Pq
⇔ y = x circX (λ) for some λ ∈ Pq
where the second step follows from (7), and the final step follows from the commutativity of X -circular convolution.
Proposition 14 parallels the equivalence between parts 1 and 2 of Proposition 13, because circX (λ) is a doubly stochastic matrix for every pmf λ ∈ Pq . In closing this appendix, we mention a well-known special case
of such
group majorization. When (X , ⊕) is the cyclic Abelian group Z/qZ of integers with addition modulo q,
G = Iq , Pq , Pq2 , . . . , Pqq−1 is the group of all cyclic permutation matrices in Rq×q , where Pq ∈ Rq×q is defined in
(8). The corresponding notion of G-majorization is known as cyclic majorization, cf. [47].
A PPENDIX B
P ROOFS OF PROPOSITIONS 4 AND 12
Proof of Proposition 4.
Part 1: This is obvious from (10).
Part 2: Since the DFT matrix jointly diagonalizes all circulant matrices, it diagonalizes every Wδ for δ ∈ R (using part
1). The corresponding eigenvalues are all real because Wδ is symmetric. To explicitly compute these eigenvalues, we
refer to [19, Problem 2.2.P10]. Observe that for any row vector x = (x0 , . . . , xq−1 ) ∈ Rq , the corresponding circulant
matrix satisfies:
!
q−1
q−1
X
X
circZ/qZ (x) =
xk Pqk = Fq
xk Dqk FqH
k=0
k=0
√
= Fq diag( q xFq ) FqH
where the first equality follows from (6) for the group Z/qZ [19, Section 0.9.6], Dq = diag((1, ω, ω 2 , . . . , ω q−1 )), and
Pq = Fq Dq FqH ∈ Rq×q is defined in (8). Hence, we have:
λj (Wδ ) =
q
X
(wδ )k ω (j−1)(k−1)
k=1
=
1,
1−δ−
δ
q−1 ,
j=1
j = 2, . . . , q
where wδ = (1 − δ, δ/(q − 1), . . . , δ/(q − 1)).
Part 3: This is also obvious from (10)–recall that a square stochastic matrix is doubly stochastic if and only if its
stationary distribution is uniform [19, Section 8.7].
−δ
by direct computation:
Part 4: For δ 6= q−1
δ
q , we can verify that Wτ Wδ = Iq when τ = 1−δ− q−1
τ
δ
[Wτ Wδ ]j,j = (1 − τ ) (1 − δ) + (q − 1)
q−1
q−1
= 1 , for j = 1, . . . , q ,
δ (1 − τ ) τ (1 − δ)
τδ
[Wτ Wδ ]j,k =
+
+ (q − 2)
2
q−1
q−1
(q − 1)
= 0 , for j 6= k and 1 ≤ j, k ≤ q .
27
The δ = q−1
q case follows from (10).
is closed over matrix multiplication. Indeed, for , δ ∈ R\ q−1
, we can
Part 5: The set Wδ : δ ∈ R\ q−1
q
q
q−1
δ
straightforwardly verify that W Wδ = Wτ with τ = + δ − δ − q−1 . Moreover, τ 6= q because Wτ is invertible
(since W and Wδ are invertible using part 4). The set also includes the identity matrix as W0 = Iq , and multiplicative
inverses (using
the associativity of matrix multiplication and the commutativity of circulant matrices
part 4). Finally,
is an Abelian group.
proves that Wδ : δ ∈ R\ q−1
q
Proof of Proposition 12.
q×q
Part 1: We first recall from [25, Appendix, Theorem A.1] that the Markov chain 1u ∈ Rsto
with uniform stationary
distribution π = u ∈ Pq has LSI constant:
( 1
q=2
Estd (f, f )
2,
α(1u) =
inf
=
.
1− q2
2 u || u)
2
D
(f
f ∈L (X ,u):
log(q−1) , q > 2
kf ku =1
D(f 2 u||u)6=0
qδ
α (1u), which proves part 1.
Now using (18), α(Wδ ) = q−1
T
2
Part 2: Observe that Wδ Wδ∗ = W
δ Wδ = Wδ = Wδ0 , where the first equality holds because Wδ has uniform
qδ
0
stationary pmf, and δ = δ 2 − q−1 using the proof of part 5 of Proposition 4. As a result, the discrete LSI constant
α(Wδ Wδ∗ ) = α(Wδ0 ), which we can calculate using part 1 of this proposition.
Part 3: It is well-known in
the literature that
ρmax (u, Wδ ) equals the second largest singular value of the divergence
√ −1
√
transition matrix diag u
Wδ diag u = Wδ (see [36, Subsection I-B] and the references therein). Hence, from
δ
part 2 of Proposition 4, we have ρmax (u, Wδ ) = 1 − δ − q−1
.
q×r
Part 4: First recall the Dobrushin contraction coefficient (for total variation distance) for any channel W ∈ Rsto
:
ηTV (W ) ,
kPX W − QX W k`1
kPX − QX k`1
PX ,QX ∈Pq :
sup
(67)
PX 6=QX
1
=
max WY |X (·|x) − WY |X (·|x0 )
2 x,x0 ∈[q]
(68)
`1
where k·k`1 denotes the `1 -norm, and the second equality is Dobrushin’s two-point characterization of ηTV [48]. Using
this characterization, we have:
ηTV (Wδ ) =
0
1
max wδ Pqx − wδ Pqx
2 x,x0 ∈[q]
= 1−δ−
`1
δ
q−1
where wδ is the noise pmf of Wδ for δ ∈ [0, 1], and Pq ∈ Rq×q is defined in (8). It is well-known in the literature
(see e.g. the introduction of [49] and the references therein) that:
2
ρmax (u, Wδ ) ≤ ηKL (Wδ ) ≤ ηTV (Wδ ) .
(69)
Hence, the value of ηTV (Wδ ) and part 3 of this proposition establish part 4. This completes the proof.
A PPENDIX C
AUXILIARY RESULTS
Proposition 15 (Properties of Domination FactorFunction). Given a channel V ∈ Rq×r
sto that is strictly positive entry+
wise, its domination factor function µV : 0, q−1
→
R
is
continuous,
convex,
and
strictly increasing. Moreover, we
q
have limδ→ q−1 µV (δ) = +∞.
q
Proof. We first prove that µV is finite on 0, q−1
. For any PX , QX ∈ Pq and any δ ∈ 0, q−1
, we have:
q
q
2
D(PX V ||QX V ) ≤ χ2 (PX V ||QX V ) ≤
2
≤
k(PX − QX )V k`2
ν
2
kPX − QX k`2 kV kop
ν
28
where the first inequality is well-known (see e.g. [36, Lemma 8]) and ν = min {[V ]i,j : 1 ≤ i ≤ q, 1 ≤ j ≤ r}, and:
1
2
k(PX − QX )Wδ k`2
2
2
1
δ
2
≥ kPX − QX k`2 1 − δ −
2
q−1
D(PX Wδ ||QX Wδ ) ≥
where the first inequality follows from Pinsker’s inequality (see e.g. [36, Proof of Lemma 6]), and the second inequality
follows from part 2 of Proposition 4. Hence, we get:
2
2 kV kop
q−1
(70)
∀δ ∈ 0,
, µV (δ) ≤
2 .
q
δ
ν 1 − δ − q−1
To prove that µV is strictly increasing, observe that Wδ0 deg Wδ for 0 < δ 0 < δ <
with:
δδ 0
+
δ0
1 − δ0 −
q−1
δ − δ0
q−1
=
∈ 0,
δ0
q
1 − δ 0 − q−1
δ0
p=δ−
0
1−δ −
δ0
q−1
+
1−
δδ 0
q−1
δ0 −
q−1
q ,
because Wδ = Wδ0 Wp
δ0
q−1
where we use part 4 of Proposition 4, the proof of part 5 of Proposition 4 in Appendix B, and the fact that Wp =
Wδ−1
0 Wδ . As a result, we have for every PX , QX ∈ Pq :
D (PX Wδ ||QX Wδ ) ≤ ηKL (Wp ) D (PX Wδ0 ||QX Wδ0 )
using the SDPI for KL divergence, where part 4 of Proposition 12 reveals that ηKL (Wp ) ∈ (0, 1) since p ∈ 0, q−1
.
q
Hence, we have for 0 < δ 0 < δ < q−1
:
q
µV (δ 0 ) ≤ ηKL (Wp ) µV (δ)
(71)
using (43), and the fact that 0 < D (PX Wδ0 ||QX Wδ0 ) < +∞ if and only if 0 < D (PX Wδ ||QX Wδ ) < +∞. This
implies that µV is strictly increasing.
We next establish that µV is convex and continuous. For any fixed PX , QX ∈ Pq such that PX 6= QX , consider
the function δ 7→ D (PX V ||QX V ) /D (PX Wδ ||QX Wδ ) with domain 0, q−1
. This function is convex, because δ 7→
q
D (PX Wδ ||QX Wδ ) is convex by the convexity of KL divergence, and the reciprocal of a non-negative convex function
is convex. Therefore, µV is convex since (43) defines it as a pointwise supremum of a collection of convex functions.
Furthermore, we note that µV is also continuous since a convex function is continuous on the interior of its domain.
Finally, observe that:
limq−1
inf µV (δ) ≥
δ→
q
=
sup
limq−1
inf
PX ,QX ∈Pq δ→
PX 6=QX
sup
q
D (PX V ||QX V )
D (PX Wδ ||QX Wδ )
D (PX V ||QX V )
lim sup D (PX Wδ ||QX Wδ )
PX ,QX ∈Pq
PX 6=QX
δ→ q−1
q
= +∞
where the first inequality follows from the minimax inequality and (43) (note that 0 < D (PX Wδ ||QX Wδ ) < +∞ for
PX 6= QX and δ close to q−1
q ), and the final equality holds because PX W(q−1)/q = u for every PX ∈ Pq .
q×q
Lemma 2 (Gramian Löwner Domination implies Symmetric Part Löwner Domination). Given A ∈ Rq×q
0 and B ∈ R
that is normal, we have:
A + AT
B + BT
A2 = AAT PSD BB T ⇒ A =
PSD
.
2
2
Proof. Since AAT PSD BB T PSD 0, using the Löwner-Heinz theorem (presented in (62)) with p = 21 , we get:
1
1
A = AAT 2 PSD BB T 2 PSD 0
29
q×q
where the first equality holds because A ∈ R0
. It suffices to now prove that (BB T )1/2 PSD B + B T /2, as
the transitive property of PSD will produce A PSD B + B T /2. Since B is normal, B = U DU H by the complex
spectral theorem [38, Theorem 7.9], where U is a unitary matrix and D is a complex diagonal matrix. Using this
unitary diagonalization, we have:
B + BT
= U Re{D} U H
2
since |D| PSD Re{D}, where |D| and Re{D} denote the element-wise magnitude and real part of D, respectively.
This completes the proof.
U |D|U H = BB T
12
PSD
ACKNOWLEDGMENT
We would like to thank an anonymous reviewer and the Associate Editor, Chandra Nair, for bringing the reference
[12] to our attention. Y.P. would like to thank Dr. Ziv Goldfeld for discussions on secrecy capacity.
R EFERENCES
[1] A. Makur and Y. Polyanskiy, “Less noisy domination by symmetric channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Aachen, Germany,
June 25-30 2017, pp. 2463–2467.
[2] R. Ahlswede and P. Gács, “Spreading of sets in product spaces and hypercontraction of the Markov operator,” Ann. Probab., vol. 4, no. 6, pp.
925–939, December 1976.
[3] Y. Polyanskiy and Y. Wu, “Strong data-processing inequalities for channels and Bayesian networks,” in Convexity and Concentration, ser. The
IMA Volumes in Mathematics and its Applications, E. Carlen, M. Madiman, and E. M. Werner, Eds., vol. 161. New York: Springer, 2017,
pp. 211–249.
[4] C. E. Shannon, “A note on a partial ordering for communication channels,” Information and Control, vol. 1, pp. 390–397, 1958.
[5] ——, “The zero error capacity of a noisy channel,” IRE Trans. Inform. Theory, vol. 2, no. 3, pp. 706–715, September 1956.
[6] J. H. Cohen, J. H. B. Kemperman, and G. Zbăganu, Comparisons of Stochastic Matrices with Applications in Information Theory, Statistics,
Economics and Population Sciences. Ann Arbor: Birkhäuser, 1998.
[7] T. M. Cover, “Broadcast channels,” IEEE Trans. Inform. Theory, vol. IT-18, no. 1, pp. 2–14, January 1972.
[8] P. P. Bergmans, “Random coding theorem for broadcast channels with degraded components,” IEEE Trans. Inform. Theory, vol. IT-19, no. 2,
pp. 197–207, March 1973.
[9] A. W. Marshall, I. Olkin, and B. C. Arnold, Inequalities: Theory of Majorization and Its Applications, 2nd ed., ser. Springer Series in Statistics.
New York: Springer, 2011.
[10] J. Körner and K. Marton, “Comparison of two noisy channels,” in Topics in Information Theory (Second Colloq., Keszthely, 1975), Amsterdam:
North-Holland, 1977, pp. 411–423.
[11] I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems, 2nd ed. New York: Cambridge University
Press, 2011.
[12] M. van Dijk, “On a special class of broadcast channels with confidential messages,” IEEE Trans. Inform. Theory, vol. 43, no. 2, pp. 712–714,
March 1997.
[13] A. El Gamal, “The capacity of a class of broadcast channels,” IEEE Trans. Inform. Theory, vol. IT-25, no. 2, pp. 166–169, March 1979.
[14] C. Nair, “Capacity regions of two new classes of 2-receiver broadcast channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Seoul, South
Korea, June 28-July 3 2009, pp. 1839–1843.
[15] Y. Geng, C. Nair, S. Shamai, and Z. V. Wang, “On broadcast channels with binary inputs and symmetric outputs,” IEEE Trans. Inform. Theory,
vol. 59, no. 11, pp. 6980–6989, November 2013.
[16] D. Sutter and J. M. Renes, “Universal polar codes for more capable and less noisy channels and sources,” April 2014, arXiv:1312.5990v3
[cs.IT].
[17] F. Unger, “Better gates can make fault-tolerant computation impossible,” Electronic Colloquium on Computational Complexity (ECCC), no.
164, November 2010.
[18] M. Artin, Algebra, 2nd ed. New Jersey: Pearson Prentice Hall, 2011.
[19] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. New York: Cambridge University Press, 2013.
[20] P. Diaconis, Group Representations in Probability and Statistics, S. S. Gupta, Ed. USA: Inst. Math. Stat. Monogr., 1988, vol. 11.
[21] Y. Polyanskiy, “Saddle point in the minimax converse for channel coding,” IEEE Trans. Inform. Theory, vol. 59, no. 5, pp. 2576–2595, May
2013.
[22] J. E. Cohen, Y. Iwasa, G. Rautu, M. B. Ruskai, E. Seneta, and G. Zbăganu, “Relative entropy under mappings by stochastic matrices,” Linear
Algebra Appl., vol. 179, pp. 211–235, January 1993.
[23] M. Raginsky, “Strong data processing inequalities and Φ-Sobolev inequalities for discrete channels,” IEEE Trans. Inform. Theory, vol. 62,
no. 6, pp. 3355–3389, June 2016.
[24] I. Csiszár and J. Körner, “Broadcast channels with confidential messages,” IEEE Trans. Inform. Theory, vol. IT-24, no. 3, pp. 339–348, May
1978.
[25] P. Diaconis and L. Saloff-Coste, “Logarithmic Sobolev inequalities for finite Markov chains,” Ann. Appl. Probab., vol. 6, no. 3, pp. 695–750,
1996.
[26] R. Montenegro and P. Tetali, Mathematical Aspects of Mixing Times in Markov Chains, ser. Found. Trends Theor. Comput. Sci., M. Sudan,
Ed. Boston-Delft: now Publishers Inc., 2006, vol. 1, no. 3.
[27] E. C. Posner, “Random coding strategies for minimum entropy,” IEEE Trans. Inform. Theory, vol. IT-21, no. 4, pp. 388–391, July 1975.
[28] Y. Polyanskiy and Y. Wu, “Lecture notes on information theory,” August 2017, Lecture Notes 6.441, Department of Electrical Engineering and
Computer Science, MIT, Cambridge, MA, USA.
[29] W. Rudin, Principles of Mathematical Analysis, 3rd ed., ser. International Series in Pure and Applied Mathematics. New York: McGraw-Hill,
Inc., 1976.
[30] V. Anantharam, A. A. Gohari, S. Kamath, and C. Nair, “On hypercontractivity and a data processing inequality,” in Proc. IEEE Int. Symp. Inf.
Theory (ISIT), Honolulu, HI, USA, June 29-July 4 2014, pp. 3022–3026.
30
[31] M.-D. Choi, M. B. Ruskai, and E. Seneta, “Equivalence of certain entropy contraction coefficients,” Linear Algebra Appl., vol. 208-209, pp.
29–36, September 1994.
[32] S. Boyd and L. Vandenberghe, Convex Optimization. New York: Cambridge University Press, 2004.
[33] C. St˛epniak, “Ordering of nonnegative definite matrices with application to comparison of linear models,” Linear Algebra Appl., vol. 70, pp.
67–71, October 1985.
[34] R. G. Gallager, Information Theory and Reliable Communication. New York: John Wiley & Sons, Inc., 1968.
[35] A. Rényi, “On measures of dependence,” Acta Math. Hungar., vol. 10, no. 3-4, pp. 441–451, 1959.
[36] A. Makur and L. Zheng, “Bounds between contraction coefficients,” in Proc. 53rd Allerton Conference, Allerton House, UIUC, Illinois, USA,
September 29-October 2 2015, pp. 1422–1429.
[37] O. V. Sarmanov, “Maximal correlation coefficient (non-symmetric case),” Dokl. Akad. Nauk, vol. 121, no. 1, pp. 52–55, 1958.
[38] S. Axler, Linear Algebra Done Right, 2nd ed., ser. Undergraduate Texts in Mathematics. New York: Springer, 2004.
[39] G. H. Hardy, J. E. Littlewood, and G. Pólya, Inequalities, 1st ed. London: Cambridge University Press, 1934.
[40] M. Ledoux, “Concentration of measure and logarithmic Sobolev inequalities,” in Séminaire de Probabilités XXXIII, ser. Lecture Notes in
Mathematics, J. Azéma, M. Émery, M. Ledoux, and M. Yor, Eds., vol. 1709. Berlin, Heidelberg: Springer, 1999, pp. 120–216.
[41] D. Bakry, “Functional inequalities for Markov semigroups,” in Probability Measures on Groups: Recent Directions and Trends, ser. Proceedings
of the CIMPA-TIFR School, Tata Institute of Fundamental Research, Mumbai, India, 2002, S. G. Dani and P. Graczyk, Eds. New Delhi,
India: Narosa Publishing House, 2006, pp. 91–147.
[42] L. Miclo, “Remarques sur l’hypercontractivité at l’évolution de éntropie pour des chaînes de Markov finies,” Séminaire de probabilités de
Strasbourg, vol. 31, pp. 136–167, 1997.
[43] K. Löwner, “Über monotone matrixfunktionen,” Math. Z., vol. 38, no. 1, pp. 177–216, 1934.
[44] E. Heinz, “Beiträge zur störungstheorie der spektralzerlegung,” Math. Ann., vol. 123, pp. 415–438, 1951.
[45] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. New York: Cambridge University Press, 1991.
[46] P. Diaconis and L. Saloff-Coste, “Nash inequalities for finite Markov chains,” J. Theoret. Probab., vol. 9, no. 2, pp. 459–510, 1996.
[47] A. Giovagnoli and H. P. Wynn, “Cyclic majorization and smoothing operators,” Linear Algebra Appl., vol. 239, pp. 215–225, May 1996.
[48] R. L. Dobrushin, “Central limit theorem for nonstationary Markov chains. I,” Theory Probab. Appl., vol. 1, no. 1, pp. 65–80, 1956.
[49] Y. Polyanskiy and Y. Wu, “Dissipation of information in channels with input constraints,” IEEE Trans. Inform. Theory, vol. 62, no. 1, pp.
35–55, January 2016.
31
| 7cs.IT
|
GLOBAL GENE EXPRESSION ANALYSIS
USING MACHINE LEARNING METHODS
XU MIN
(B.Eng. Beihang Univ., China)
A THESIS SUBMITTED
FOR THE DEGREE OF MASTER OF SCIENCE
DEPARTMENT OF INFORMATION SYSTEMS
NATIONAL UNIVERSITY OF SINGAPORE
2003
I
Acknowledgements
I am very grateful to my supervisor Dr. Rudy Setiono for his insightful suggestions in both
the content and presentation of this thesis. It was his encouragement, support and patience
that saw me through and I am ever grateful to him. I am full of gratitude to my boss, Dr. Peng
Jinrong, for his understanding and support on allowing me to take the part-time Master of
Science research work.
I would like to thank Ms. Jane Lo, Mr. Guan Bin, and other members of Lab of Functional
Genomics of Institute of Molecular and Cell Biology, Singapore, for their numerous help
throughout my research work.
For the study of the hybrid of Likelihood method and Recursive Feature Elimination method,
I would like to thank Dr. Isabelle Guyon for providing the supplementary data. I also thank
Dr. Cui Lirong, Mr. Wang Yang, Dr. Oilian Kon, Dr. Wolfgang Hartmann, and Mr. Li
Guozheng for their numerous helpful consultations.
I would also like to thank my parents for their love, encouragement, guidance and patience
throughout my studies.
II
Table of Contents
Acknowledgements..................................................................................................................... I
Table of Contents....................................................................................................................... II
List of Figures............................................................................................................................ V
List of Tables ............................................................................................................................VI
Summary ................................................................................................................................ VII
1
Introduction ..................................................................................................................... 1-1
1.1
2
3
Background ................................................................................................... 1-1
1.1.1
Functional genomics ............................................................................... 1-1
1.1.2
Microarray Technology........................................................................... 1-2
1.1.3
Machine learning methods for global analysis ......................................... 1-3
1.2
Research Objectives....................................................................................... 1-4
1.3
Organization of chapters ................................................................................ 1-4
Literature review.............................................................................................................. 2-6
2.1
Finding gene groups ...................................................................................... 2-6
2.2
Finding relationships between genes and function.......................................... 2-8
2.3
Dynamic modeling and time series analysis ................................................. 2-12
2.4
Reverse engineering and gene network inference ......................................... 2-13
2.5
Overview of the field ................................................................................... 2-17
Machine learning methods used ......................................................................................3-18
3.1
3.2
Description of problem ................................................................................ 3-18
3.1.1
Structure of microarray data.................................................................. 3-18
3.1.2
Two types of classification problems .................................................... 3-19
Feature integration and univariate feature selection methods........................ 3-20
3.2.1
Information gain ................................................................................... 3-21
3.2.2
Likelihood method (LIK) ...................................................................... 3-23
III
3.3
Classification and multivariate feature selection methods............................. 3-24
3.3.1
Neural Networks ................................................................................... 3-25
3.3.2
Ensemble of neural networks (boosting)................................................ 3-30
3.3.3
Neural network feature selector............................................................. 3-33
3.3.3.1
Disadvantage of information gain measure .................................... 3-33
3.3.3.2
Neural network feature selector..................................................... 3-35
3.3.4
Support vector machines ....................................................................... 3-44
3.3.4.1
Linearly separable learning model and its training ......................... 3-45
3.3.4.2
Linearly non-separable learning model and its training.................. 3-48
3.3.4.3
Nonlinear Support Vector Machines.............................................. 3-49
3.3.4.4
SVM recursive feature elimination (RFE)...................................... 3-51
3.3.5
Bayesian classifier ................................................................................ 3-52
3.3.6
Two discriminant methods for multivariate feature selection................. 3-53
3.3.6.1
Fisher’s linear discriminant ........................................................... 3-54
3.3.6.2
Multivariate Likelihood feature ranking ........................................ 3-55
3.3.7
Combining univariate feature ranking method and multivariate feature
selection method................................................................................................. 3-56
4
Experimental results and discussion ................................................................................4-58
4.1
Second type - high dimension problems....................................................... 4-58
4.1.1
Principle Component Analysis for feature integration ........................... 4-60
4.1.2
C4.5 for feature selection ...................................................................... 4-61
4.1.3
Neural networks with features selected using information gain.............. 4-65
4.1.4
AdaBoost.............................................................................................. 4-68
4.1.5
Neural network feature selector............................................................. 4-70
4.1.6
Hybrid Likelihood and Recursive Feature Elimination method.............. 4-72
4.1.6.1
Leukemia dataset........................................................................... 4-74
IV
4.1.6.2
Small, round blue cell tumors dataset ............................................ 4-86
4.1.6.3
Artificial datasets .......................................................................... 4-95
4.1.7
4.2
5
The combination of Likelihood method and Fisher’s method ................ 4-98
First type - low dimension problems .......................................................... 4-101
Conclusion and future work ..........................................................................................5-104
5.1
Biological knowledge discovery process.................................................... 5-104
5.2
Contribution, limitation and future work.................................................... 5-105
Reference................................................................................................................................109
V
List of Figures
Figure 1.1: A scanned microarray image ..................................................................... 1-3
Figure 3.1: AdaBoost algorithm ................................................................................ 3-32
Figure 3.2: Plot of feature values with class labels, ‘o’for ‘-1’and ‘+’for ‘+1’......... 3-35
Figure 3.3: Wrapper model framework...................................................................... 3-36
Figure 3.4: Neural network feature selector .............................................................. 3-42
Figure 3.5: Recursive feature elimination algorithm .................................................. 3-52
Figure 4.1: The combination of feature selection / integration methods and classification
methods.............................................................................................................. 4-59
Figure 4.2: Wrong prediction times. .......................................................................... 4-68
Figure 4.3: Sorted LIK score of a subset of genes in the leukemia dataset................. 4-75
Figure 4.4: Classification performance of genes selected using the hybrid LIK+RFE. .. 477
Figure 4.5: Plot of the leukemia data samples according to the expression values of the
three genes selected by the hybrid LIK+RFE. ..................................................... 4-78
Figure 4.6:
Performance of SVM and Naïve Bayesian classifiers built using genes
selected according to LIK scores. ....................................................................... 4-80
Figure 4.7: Classification performance of purely using RFE. ..................................... 4-81
Figure 4.8: SVM and Bayesian prediction accuracy on test samples using features
selected by RFE combined with baseline criterion, Fisher’s criterion and Extremal
margin method. .................................................................................................. 4-85
Figure 4.9: Sorted LIK scores of genes in the SRBCT dataset. Dots indicate
LIK EWS → Non− EWS
scores and circles indicate LIK Non− EWS → EWS scores. ......................... 4-87
Figure 4.10: Plot of RMS and non-RMS samples. Plot of all 88 RMS and non-RMS
samples according to the expression values of three of the four selected genes.... 4-89
Figure 4.11: Hierarchical clustering of SRBCT samples with selected 15 genes. ....... 4-94
Figure 4.12: SVM and Bayesian prediction accuracy when running combination of
univariate and multivariate feature selection methods. ...................................... 4-100
VI
List of Tables
Table 4.1: List of combined univariate and multivariate feature selection methods tested.
........................................................................................................................... 4-60
Table 4.2: Performance of neural network using principle components as input......... 4-61
Table 4.3: Decision trees constructed by iterative mode from all 72 samples. ............ 4-62
Table 4.4: Decision trees constructed by leave-one-out mode from all 72 samples..... 4-63
Table 4.5: Decision trees constructed by leave-one-out mode from 38 training samples
with prediction accuracy on 34 test samples........................................................ 4-63
Table 4.6: Leukemia features and their information gain of all samples, training samples
and test samples, sorted by gain of all samples.................................................... 4-64
Table 4.7: Prediction performance of neural network using features selected by
information gain. ................................................................................................ 4-66
Table 4.8: Test of neural network using features selected by information gain to identify
incorrectly predicted test samples. ...................................................................... 4-67
Table 4.9: AdaBoost test results. ............................................................................... 4-69
Table 4.10: Experiment result of neural network feature selector with summed square
error function...................................................................................................... 4-71
Table 4.11: Experiment result of neural network feature selector with cross entropy error
function. ............................................................................................................. 4-71
Table 4.12: The smallest gene set found that achieves prefect classification performance.
........................................................................................................................... 4-78
Table 4.13: The genes selected by the hybrid LIK+RFE method. The genes that have
LIK scores of at least 1500 were selected initially. RFE was then applied to select
these four genes that achieved perfect performance............................................. 4-83
Table 4.14: Experimental results for the SRBCT dataset using the hybrid LIK+RFE. 4-88
Table 4.15: The genes selected by the hybrid LIK+RFE for the four binary classification
problems............................................................................................................. 4-90
Table 4.16: The performance of SVM and Naï
ve Bayesian classifiers built using the top
genes selected according to their LIK scores....................................................... 4-91
Table 4.17: Test result of LIK, RFE and LIK+RFE on artificial datasets ................... 4-97
Table 4.18: Test of SVM with RBF kernel using different parameters. .................... 4-102
VII
Summary
Microarray is a technology to quantitatively monitor the expression of large number of genes
in parallel. It has become one of the main tools for global gene expression analysis in
molecular biology research in recent years. The large amount of expression data generated by
this technology makes the study of certain complex biological problems possible and
machine learning methods are playing a crucial role in the analysis process. At present, many
machine learning methods have been or have the potential to be applied to major areas of
gene expression analysis. These areas include clustering, classification, dynamic modeling
and reverse engineering.
In this thesis, we focus our work on using machine learning methods to solve the
classification problems arising from microarray data. We first identify the major types of the
classification problems; then apply several machine learning methods to solve the problems
and perform systematic tests on real and artificial datasets. We propose improvement to
existing methods. Specifically, we develop a multivariate and a hybrid feature selection
method to obtain high classification performance for high dimension classification problems.
Using the hybrid feature selection method, we are able to identify small sets of features that
give predictive accuracy that is as good as that from other methods which require many more
features.
1-1
1 Introduction
1.1 Background
1.1.1 Functional genomics
With the completion of Human Genome Project, biology research is entering the post genome
era. Although biologists have collected a vast amount of DNA sequence data, the details of
how these sequences function still remain largely unknown. Genomes of even the simplest
organisms are very complex. Nowadays, biologists are still trying to find answers to the
following questions (Brazma and Vilo, 2000):
•
What are the functional roles of different genes and in what cellular process do they
participate?
•
How are the genes regulated? How do the genes and gene products interact? What are
the interaction networks?
•
How does the gene expression level differ in various cell types and states? How is the
gene expression changed by the various diseases of compound treatments?
Biology used to be data-poor science. With more advanced techniques developed in recent
years, biologists are now able to transform vast amount of biological information into useful
data. This makes it possible to study gene function globally, and a new field, functional
genomics emerges. Specifically, functional genomics refers to the development and
application of global (genome-wide or system-wide) experimental approaches to assess gene
function by making use of the information and reagents provided by structural genomics. It is
characterized by high throughput or large scale experimental methodologies combined with
statistical and computational analysis of the results (Hieter and Boguski, 1997).
1-2
1.1.2 Microarray Technology
Several methods have been developed to understand the behavior of genes. Microarray
technology is an important one among them. It is used to monitor large amount of genes’
expression level in parallel. Here gene expression refers to the process to transcribe a gene's
DNA sequence into the RNA that serves as a template for protein production, and gene
expression level indicates how active a gene is in certain tissue, at certain time, or under
certain experimental condition. The monitored gene expression level provides an overall
picture of the genes being studied. It also reflects the activities of the corresponding protein
under certain conditions.
Several steps are involved in this technology. First, complementary DNA (cDNA) molecules
or oligos are printed onto slides as spots. Then, two kinds of dye labeled samples, i.e. sample
and control , are hybridized. Finally, the hybridization is scanned and stored as images (see
example in Figure 1.1, a sample from Zebra fish). Using a suitable image-processing
algorithm, these images are quantified into a set of expression values representing the
intensity of spots. Usually, the dye intensity may be biased by factors like its physical
property, experimental variability in probe coupling and processing procedures, and scanner
settings. To minimize the undesirable effects caused by this biased dye intensity,
normalization is done to balance dye intensities and make expression value comparable
across experiments (Yang et al., 2001). Here the term comparable means that the difference
of any measured expression value of a gene between two experiments should reflect the
difference of its true expression levels.
1-3
Figure 1.1: A scanned microarray image
1.1.3 Machine learning methods for global analysis
Molecular biology also used to be a data poor field, and most of gene expression analysis
work was done manually with very limited information derived from experiment. The focus
of a molecular biologist was on a few genes or proteins. With the application of large-scale
biological information quantification methods like microarray and DNA sequencing, the
behavior of genes can be studied globally. Currently, there is an increasing demand for
automatic analysis of the overall relationship hidden behind large amount of genes from their
expression.
1-4
Machine learning is the study of algorithms that could learn from experience and then
predict. The theoretical aspects of machine learning are rooted in statistics and informatics,
but computational considerations are also indispensable. Due to the complex nature of
biological information, machine learning could play an important role in the analysis process.
1.2 Research Objectives
Microarray technology based gene expression profiling is one of hottest research topics in
biology at present. The experimental part of this technology is already mature. Compared
with this, the exploration of automatic analysis methods is still at its early stage. In this thesis,
we study several machine learning approaches to solving several typical gene expression
analysis problems.
The main objectives of this research are:
•
To identify typical gene expression analysis problems from machine learning point of
view.
•
To apply suitable machine learning methods to the problems from public datasets, and
to improve these methods when necessary.
•
To find new approaches to the problems.
•
To study the experimental results.
•
To apply the methods to new datasets and validate the result.
1.3 Organization of chapters
This thesis is organized as follows. Chapter 2 provides a brief review of the current methods
that can be applied to microarray data analysis. Chapter 3 gives detailed illustrations of
1-5
several important machine learning methods that can be applied to classification using gene
expression data. In particular, we improve a neural network feature selector method,
developed multivariate likelihood feature selection method, and propose a hybrid framework
of univariate and multivariate feature selection method. Chapter 4 describes the experimental
results of these methods on two different kinds of gene expression analysis problems, and
discusses the experiment results. Specifically, we perform systematic tests of the hybrid of
Likehood method and Recursive Feature Elimination method because we have obtained very
good feature selection performance on several microarray datasets; we also apply Support
Vector Machine on a recently obtained Zebra fish dataset to perform gene function
prediction. Finally, Chapter 5 concludes the thesis and future works are illustrated.
2-6
2 Literature review
Various automatic methods have been applied or developed for the gene expression analysis.
They are basically from fields such as machine learning, statistics, signal processing, and
informatics. The following are the relevant works categorized according to the analysis tasks.
2.1 Finding gene groups
Some methods could be used to find useful information or pattern from biological data,
which indicates relationship among the genes. These methods are unsupervised (Haykin,
1999), i.e., the learning models are optimized using pre-specified task-independent measures,
which reflect the difference or similarity of the training samples. Once the model has become
tuned to the statistical regularities of the input gene expression data, it develops the ability to
form internal representations for encoding features of the input and thereby to create new
classes automatically (Becker, 1991).
Principle Component Analysis (PCA) is an exploratory multivariate statistical technique for
simplifying complex data sets (Raychaudhuri et al., 2000). Given an expression matrix with a
number of features, a set of new features is generated by PCA. These new features account
for most of the information in the original features, but the number of dimensions is smaller
than that of the original data. There are several neural network algorithms that support PCA,
these algorithms are mainly Hebbian-based algorithms (Haykin, 1999) which are selforganizing and adaptive. Singular Value Decomposition (SVD) can also be used to perform
PCA. SVD is a linear transformation that decomposes the gene expression matrix into a
product of three matrices that represent the underlying characteristics of the original matrix.
Alter et al. (2000) applied SVD for gene expression analysis. They first obtained the principal
components from the decomposed matrices by applying SVD to expression data of yeast
2-7
genes; then rejected the genes that contribute little information to the principal components.
In their work, the information contribution was measured by Shannon entropy of the
expression values of the genes, where Shannon entropy characterizes the complexity of the
expression values. Finally, the remaining genes were sorted, and the results reflect the strong
relationship between the groups of these genes and their functional categories.
Clustering is a typical way to group genes together according to their features. Certain
distance measure, which reflects the similarity of genes’expression, is needed for clustering
process. Most clustering methods that have been studied in the gene expression analysis
literature use either Euclidean distance or Pearson correlation between expression profiles as
a distance measure (D’haeseleer, 2000). Other measures include Euclidean distance between
expression profiles and slopes (Wen et al., 1998), squared Pearson correlation (D’haeseleer et
al., 1997), Euclidean distance between pairwise correlations to all other genes (Ewing et al.,
1999), Spearman rank correlation (D’haeseleer et al., 1997), and mutual information
represented by pairwise entropy (D’haeseleer et al., 1997; Michaels et al., 1998; and Butte
and Kohane, 2000).
There are two kinds of clustering methods: hierarchical and non-hierarchical ones. A
hierarchical clustering method starts from individual genes, merging them into bigger clusters
until there is only one cluster left, in an agglomerative way. The method can also divisively
start from all genes, splitting them until no two of them are together. The output of the
method is a hierarchy of clusters, where the higher-level clusters are the sum of the lowerlevel ones. On the other hand, a non-hierarchical clustering method first divides genes into a
certain number of clusters, and then iteratively refines them until certain optimization
criterion is met.
2-8
Clustering methods that have been applied for gene expression analysis were reviewed in
(D’haeseleer, 2000) and (Tibshirani et al., 1999). Because different algorithms may be
applicable to different datasets, in (Yeung et al., 2001) a data-driven method to evaluate these
algorithms was proposed.
2.2 Finding relationships between genes and function
Some methods can be used to find the relationship between gene expression and other
information, i.e. properties of genes and samples. These properties can be type of
hybridization sample, experimental condition, or the biological process that they are involved
in. These are basically supervised methods for classification and regression. The methods try
to construct learning models that could represent the relationship when given gene expression
data as input and other information as output.
Machine learning classification methods have been applied to gene expression analysis in
recent years. These methods usually employ class labels to represent different groups of
expression data. An important application is cancer tissue classification, i.e. to construct
learning model to predict whether a tissue is cancerous or to predict the type of cancer using
gene expression. Cancer tissue classification is crucial for the diagnosis of patients. It used to
be based on morphological appearances, which is often hard to measure and differentiate, and
the classification result is very subjective. With the emergence of microarray technology, the
classification is improved greatly by going to the molecular level. The various machine
learning methods applied for classification are briefly described in the following paragraphs.
2-9
Neural networks are learning models that are based on the structure and behavior of neurons
in the human brain and can be trained to recognize and categorize complex patterns (Bishop,
1995). Khan et al. (2001) used neural networks to classify cancer tissues using gene
expression data as input. In their experiments, PCA was used to select a set of candidate
genes. A number of neural networks were then trained on the training dataset. The prediction
on test samples was achieved by summarizing all the outputs of the trained neural networks.
Support Vector Machine (SVM), which is rooted in statistical learning theory, is another
method that can also be used to perform classification. It can achieve good generalization
performance by minimizing both the training error and a generalization criterion that depends
on Vapnik-Chervonenkis (VC) dimension (Vapnik, 1998). In (Brown et al., 2000), SVM was
applied to classify yeast genes according to the biological process they involve in as
represented by their expression data. In (Furey et al., 2000), it was also used to classify
cancer tissues.
Decision tree generates a tree structure consisting of leaves and decision nodes. Each leaf
indicates a class, and each decision node specifies some test to be carried out on a single
attribute value, with one branch and subtree for each possible outcome of the test (Quinlan,
1993). C4.5 is a well-known decision tree induction algorithm, which uses information gain
measure. Cai et al. (2000) used it to classify cancer tissue samples. A comparison was done
with SVM, which was found to outperform C4.5.
Naï
ve Bayes Classification is a statistical discrimination method based on Bayes rule. In
(Keller et al., 2000), this method was used to classify cancer tissues. The algorithm is simple
2-10
and can be easily extended from two-class to multi-class classification. Gaussian distribution
of the data and class independence are assumed by the method.
A related technique is Bayesian Networks, which is also based on Bayesian rule. It is a
probabilistic graphical model that represents the unique joint probability distribution of
random variables efficiently. Nodes of a Bayesian network could correspond to genes and
class labels, and represent the probability of the class label given some gene expression
levels. Hwang et al. (2001) used Bayesian Networks to classify acute leukemia samples. A
simple Bayesian Network with four gene nodes and one class label node was constructed
from gene expression data. The high prediction performance indicated that the constructed
network model can correctly represent the causal relationships of certain genes that are
relevant to the classification.
Radial Basis Function (RBF) networks is a type of neural networks whose hidden neurons
contain RBFs, a statistical transformation based on a Gaussian distribution, and whose output
neuron computes a linear combination of its inputs. Hwang et al. (2001) also used an RBF
network to classify acute leukemia samples. The network was larger than the constructed
Bayesian network, but test results showed the prediction accuracy of RBF networks was
higher.
Besides classification, feature selection i.e. the process of selecting genes that are most
relevant to the class labels is also an important task for gene expression analysis. In (Slonim
et al., 2000), a statistical method involving mean and variance was used to reflect the
relevance between individual genes and class labels. In their work, the acute leukemia
samples were divided into two groups according to their class labels. Those genes whose
2-11
expression values had small variance in both groups and big mean difference between the two
groups were selected. In (Keller et al., 2000), a likelihood gene selection method was
proposed based on likelihood. It outperformed the Baseline method in (Slonim et al., 2000)
on the same cancer dataset by choosing less number of genes while achieving similar
classification performance. In (Li, 2002), the linear relationship between the logarithm of
measurement of classification ability of genes and the logarithm of rank of classification
ability of genes was found to obey Zipf’s law (Zipf, 1965). Plots of this relationship provided
a useful tool in estimating the number of genes that is necessary for classification. Gouyon et
al. (2002) proposed a Recursive Feature Elimination method based on Support Vector
Machine. It made use of the magnitude of weights of trained SVMs as indicators of the
discrimination ability of the genes. The algorithm keeps eliminating the genes that have
relatively small contribution to the classification. In the test of the method on a leukemia
dataset, small sets of genes with high discrimination contribution is obtained.
Neural trees represent multilayer feed-forward neural networks as tree structures. They have
heterogeneous neuron types in a single network, and their connectivity is irregular and sparse
(Zhang et al., 1997). Compared with the conventional neural networks, neural trees are more
flexible. They can represent more complex relationship and permit structural learning and
feature selection. Evolutionary algorithms can be used to construct neural trees. Hwang et al.
(2001) constructed neural trees using gene expression, and selected relevant genes according
to the connections in the trees. Neural trees were found to have better classification
performance than the other two methods in the paper, i.e. Redial Basis networks and
Bayesian networks. Genes with significant contribution to the classification could also be
found from the constructed neural trees.
2-12
2.3 Dynamic modeling and time series analysis
It is also important to the study significant patterns and infer the dynamic model of gene
expression from hybridization samples collected at different experimental time points. The
dynamics can provide clues to the role of genes in the biological processes. Some of these
methods have been successfully applied to discrete signal analysis.
Filkov et al. (2001) proposed a set of analysis methods that are suitable for short-term
discrete time series data. These methods include: period detection, phase detection,
correlation significance of short sequences of different length, and edge detection of group of
regularity genes. The prediction analysis on yeast microarray data (Spellman et al., 1998)
showed that the amount of data is not sufficient for large regularity pathway inference.
Singular Value Decomposition (SVD) constructs characteristic models of gene expression.
These characteristic models can also be used to construct dynamic models of gene expression
by deducing time translational matrices (Alter et al., 2000; Dewey and Bhan 2001; Holter et
al., 2001). In (Dewey and Bhan, 2001), the change of expression level in a gene was modeled
as first order Markov process. The time translational coefficient matrix was computed using
least squares method based on a combination of SVD and linear response theory. The
network model inferred from the matrix provided a way to cluster genes using their function.
The clusters derived by applying the method to yeast time series expression data were in
agreement with previously reported experimental work.
The dynamics of gene expression could also be modeled as differential equations. In (Chen et
al., 1999) a linear transcription model is proposed, and two methods, Minimum Weight
2-13
Solutions for Linear Equations and Fourier Transform for Stable Systems, are proposed for
constructing the model.
Shannon entropy can be used as a measure of the information content or complexity of a
measurement series. It indicates the amount of information contained in expression of a gene
pattern over time, or across anatomical regions, and therefore reveals the amount of
information carried by the gene during a disease process or during normal phenotypic change.
Shannon entropy is used by Fuhrman et al. (2000) to identify the most likely drug target
candidate genes from temporal gene expression patterns.
2.4 Reverse engineering and gene network inference
Gene network inference attempts to construct and study coarse-scale network models of
regulatory interactions between genes. It shows relationship between individual genes, and
then provides a richer structure than clustering, which only reveals relationship between
groups of genes. Gene network inference requires inference of the causal relationships among
genes, i.e. reverse engineering the network architecture from its activity profiles. Reverse
engineering is generally an unsupervised system identification process, which involves the
following issues: choosing hybridization sample or expression data, choosing network model,
and choosing method to construct the model, studying the structure and dynamics of the
model. The study of network dynamics often involves time series analysis techniques
mentioned in Section 2.3.
The simplest gene network model is Boolean network proposed by Kauffman (1969). In a
Boolean network model, each node is in one of two possible states: express or not-express.
The actual state depends on the states of other nodes that are linked to it. A variety of
2-14
Boolean network construction algorithms have been developed. Somogyi et al. (1996)
employed a phylogenetic tree construction algorithm (Fitch and Margoliash, 1967) to create
and visualize the network. In (Liang et al., 1998), a more systematic and general algorithm
was developed using mutual information to identify a minimal set of inputs that uniquely
defines the output for each gene at next time step. Akutsu et al. (1999) improved Liang’s
algorithm to accept noisy expression data. Ideker et al. (2000) developed an alternative
algorithm, which introduced perturbations in the expression data to iteratively and
interactively refine the sensitivity and specificity of the constructed networks. At each of the
iterations, a set of networks was inferred according to the expression data from different
experimental perturbations. They were then discriminated using entropy based approach. The
discriminations provide guide in further experimental perturbation design. Samsonova and
Serov (1999) proposed an interactive Java applet tools for visualization and analysis of the
Boolean network constructed. Maki et al. (2001) proposed a system that uses a top-down
approach for the inference of Boolean network. The inferred networks on the simulated
expression data matched the original ones well even when one of the genes was disrupted.
The advantage of Boolean network is its low construction cost. But it has the disadvantage
being too coarse to represent the true regulation relationship between genes. Linear modeling
tries to overcome this disadvantage by using weighted sum to represent the influence of other
genes on one particular gene. In the model, the overall relationship is then represented as a
matrix. Someren et al. (2000) used linear algebra methods to construct the model. Partial
Least Squares method is a statistical method that is particularly useful for modeling large
number of variables each with few observations (Stone and Brooks, 1990). Datta (2001)
applied it to Sacccharomyces cerevisiae yeast microarray data to get linear regression model
2-15
and then predicted expression level of a gene according to that of other genes. The result
appeared to be consistent with the known biological knowledge.
The dependency of the expression of one gene on the expression of other genes can also
modeled using nonlinear functions. The nonlinear approach provides the model more ability
to reveal the biological reality. However, it often introduces more difficulty in solving the
model at the same time. Maki et al. (2001) also modeled the gene interaction as an S-system
(Savageau, 1976). S-system is one of the best formalisms to estimate the complex gene
interaction mechanisms. The disadvantage of the S-system network is the number of
parameters to be estimated is vary large compared with that of Boolean network. To analysis
large scale network, S-system approach was combined with their Boolean approach.
Bayesian networks can also be used for gene network inference. The inference process
estimates the statistical confidence of dependencies between the activities of genes. Friedman
et al. (2000) used it to analyze Sacccharomyces cerevisiae yeast microarray data from
(Spellman et al., 1998). Their order relation and Markov relation analysis showed that the
constructed Bayesian network had strong link to cell cycle regulated genes. Pe’er et al.
(2001) extended this framework by the following steps: adding new kinds of factors such as
mediator, activator, and inhibitor; enabling construction of subnets of strong confidence;
enabling handling of mutation; and employing better discretization on the data for
preprocessing. Their experiment on yeast microarray data showed that the constructed
significant subnets could reveal biological pathways.
Several works have been conducted on the study of the dynamics of constructed network
model. Huang (1999) used Boolean network to interpret gene activity profiles as entities
2-16
related to the dynamics of both the regularity network and functional cellular states. In this
approach, the dynamics were mapped into state space and the system property of the network
like stability, trajectories and attractors were studied.
This dynamics could be modeled more precisely as a set of differential equations. Neural
network is one of the methods that could effectively solve these equations. Both Vohradský
(2001) and D’haeseleer (2000) modeled gene network as recurrent neural networks.
Vohradský (2001) used recurrent back-propagation (Pineda 1987), and simulated annealing
to construct the network. However, D’haeseleer (2000) tried back-propagation through time
(Werbos, 1990) in the training process, with techniques such as weight decay and weight
elimination (Weigend et al., 1991) applied to simplify the model. Compared with simulated
annealing, back-propagation is a more effective training method, but its scalability is worse
because it attempts to unfold the temporal operation of the network into a layered feedforward network. When doing their experiments, both Vohradský and D’haeseleer did not
have microarray dataset that was large enough for the network construction. Instead, they
used artificial data for experiment. The trained networks appeared to match original ones.
Szallasi, (1999) illustrated some basic natures of gene networks that could affect modeling.
They include the stochastic nature, effective size, compartmentalization, and information
content of expression matrix. In (Wessels et al., 2001; Someren et al., 2001), different
network models were categorized and compared under criteria like inferential power,
predictive power, robustness, consistency, stability, and computational cost.
2-17
2.5 Overview of the field
Clustering based on gene expression reflects the correlation of genes. Classification links the
expression of genes to functions. To study the change of expression of genes through time,
dynamic modeling and time series analysis method have been used. In order to obtain the
causal relationship or regulation of genes globally, the gene networks are needed to be
inferred from the expression data. The inference work is a reverse engineering process
(D’haeseleer, 2000). Reverse engineering is one of the major focuses of systems biology at
present. The field of Systems biology studies biology at system level by examining the
structure and dynamics of cellular and organismal function (Kitano, 2002). When the field of
systems biology advances to the stage of trying to unify the biological knowledge across
different levels of living organisms, we expect the understanding of the inherent complexity
of living organisms will become a central issue (Michigan, 1999).
3-18
3 Machine learning methods used
Our work focuses on classification and feature selection methods for global gene expression
analysis. In Section 3.1, the classification problems are described. Section 3.2 is about feature
integration and univariate feature selection methods. Section 3.3 describes multivariate
feature selection methods and classification methods.
3.1 Description of problem
In this section, we present the gene expression data and class information in a mathematical
form, and illustrate two types of classification problems that are commonly encountered in
microarray data analysis.
3.1.1 Structure of microarray data
An expression matrix can be generated when quantified expression values of different
hybridizations are available. Suppose there are m genes and n hybridizations. The
expression matrix A is an m × n matrix
a11
a
A = 21
M
a m1
a12
a 22
M
am 2
L a1n a1
L a 2 n a 2
,
=
O M M
L a mn a m
Eq. 3.1.1.1
where aij represents the expression value of the i th gene in the j th hybridization.
Certain property of gene or hybridization sample needs to be defined, i.e. labeling is required,
in order to find the relationship between the genes and their expression matrix. The gene’s
property is represented by an m × 1 vector
3-19
x1
x
x= 2,
M
xm
Eq. 3.1.1.2
where each element represents one possible value of this property. For example, xi = +1
means the i th gene belongs to some biological process, while xi = −1 means the i th gene
does not belong to this process. Similarly, the property of hybridization sample is defined as
an n × 1 vector
y1
y
y = 2 ,
M
yn
Eq. 3.1.1.3
where each element represents one possible value of this property. For example, yi = +1
means the i th hybridization sample is cancerous, while yi = −1 means the i th hybridization
sample is non-cancerous.
3.1.2 Two types of classification problems
From a micoarray experiment, we can only obtain a very limited number of hybridizations
which involve a large number of genes. That is to say, n is usually no more than a hundred,
but m can be a few thousands. So there are basically two types of classification problems:
•
First type: large number of samples with low dimension. When the relationship
between genes’expression and their property (function) is studied, the classification
problem consists of A as input of learning model and x as output. There are n
features, each of them corresponds to one hybridization, and m samples, each of them
corresponds to one gene.
3-20
•
Second type: small number of samples with large number of features and high
dimension. When the relationship between expression of all genes under
consideration and a certain property of hybridization sample is studied, the
b1
b
T
classification problem can be expressed as B = A = 2 , which is transpose of A ,
M
b n
as input of learning model and y as output. There are m features, each corresponds
to one gene, and n samples, each corresponds to one hybridization.
In this thesis, our work is focused mainly on solving the classification problems of the second
type, because this type of problems has distinct nature from the ordinary classification
problems, and many cancer tissue classification problems based on gene expression is of this
kind. We also have obtained a newly released Zebra fish developmental microarray dataset,
which can be used to form classification problems of the first type. Because we are able to
validate our prediction results using more precise biological experiments with the help of
researchers who generated this dataset, we have applied Support Vector Machine
classification method to this dataset as well.
3.2 Feature integration and univariate feature selection methods
Preprocessing is needed for the second type of classification problem (large number of
features). The main goal of preprocessing is to reduce the number of inputs for a learning
model without much loss, or even with some improvement of classification accuracy. Two
kinds of methods can be used for the reduction: feature selection and feature integration (Liu
and Motoda, 1998). Feature selection selects a subset of features as classifier input, while
3-21
feature integration generates a new feature set from original features as input. The feature
integration method used in this thesis is Principle Component Analysis (PCA) (Muirhead,
1982); the two univariate feature selection methods used in our research are Information Gain
(Quinlan, 1993) and Likelihood method (Keller et al., 2000). Due to space limitation, only
Information gain and Likelihood method will be described in this section. Here univariate
means the selection method only takes the contribution of individual features to the
classification into consideration. The multivariate feature selection method will be described
in Section 3.3, because most of them are based on classification methods. The term
multivariate means the selection method accounts for the combinatorial effect of the features
on the classification.
3.2.1 Information gain
Information gain method can be used to rank the individual discrimination ability of the
features. It comes from information theory. In this method, information amount is measured
by entropy. Let S denote the set of all samples, | S |= n . Let k denote the number of classes,
and let Ci , i = 1,K, k denote the set of samples that belong to a class,
k
UC
i =1
i
=S ,
∀1 ≤ i < j ≤ k : C i ∩ C j = Φ . Suppose we select one sample from S and label it as a member
of class Ci . This message has probability
| Ci |
of being correct. So the information it
|S|
|C ∩ S |
. The expected information of such a message is
conveys is − log 2 i
|S |
| Ci ∩ S |
|C ∩ S |
× log 2 i
.
|S |
i =1
|S |
k
info( S ) = −∑
Eq. 3.2.1.1
3-22
Similarly, the expected information amount of any subset of S can also be determined. If S
is partitioned into l subsets Ti , i = 1,K, l ,
l
UT
i
i =1
= S , ∀1 ≤ i < j ≤ l : Ti ∩ T j = Φ . Then the
expected information requirement is
l
| Ti |
× info(Ti ) .
i =1 | S |
info X (S ) = ∑
Eq. 3.2.1.2
The difference
gain = info( S ) − info X (S )
Eq. 3.2.1.3
represents the amount of information gained from this partitioning.
Information gain tends to be greater when l becomes larger, which may not truly reflect the
quality of the partition. So it needs to be normalized by taking split information into
consideration. Split information is defined as
| Ti |
|T |
× log 2 i .
i =1 | S |
|S |
l
split info = −∑
Eq. 3.2.1.4
The normalized gain is thereby defined as
gain ratio =
gain
.
split info
Eq. 3.2.1.5
Given sample expression values and their class labels, each feature’s information gain ratio
could be calculated this way: sort samples according to expression values of this feature,
partition them and calculate gain ratio of every possible split, and choose the maximum one
as this feature’s information gain ratio. This ratio provides a measure to evaluate the
classification ability of one feature. Feature selection is done by choosing features that have
the highest gain ratios.
3-23
3.2.2 Likelihood method (LIK)
Keller et al. (2000) proposed the Maximum Likelihood gene selection (LIK) method. Denote
the event that a sample belongs to class a or class b by M a and M b , respectively. The
difference in the log likelihood is used to rank the usefulness of gene g for distinguishing the
samples of one class from the other. The LIK score is computed as follows:
LIK ag→b = log( P( M a | x ag,1 ,K, x ag,na )) − log( P( M b | xag,1 , K, x ag,na ))
Eq. 3.2.2.1
and
LIK bg→ a = log( P( M b | xbg,1 ,K , xbg,nb )) − log( P( M a | xbg,1 ,K , xbg,nb ))
Eq. 3.2.2.2
where P(M i | x gj,1 , K , x gj ,n j ) is a posteriori probability that M i is true given the expression
values of the g th gene of all the training samples that belong to class j , where n j is the
number of training samples that belong to class j . Bayes rule
P(M | X ) P( X ) = P( X | M ) P( M )
Eq. 3.2.2.3
is used, with three assumptions required by the method. First is the assumption of equal prior
probabilities of the classes
P(M a ) = P(M b ) ,
Eq. 3.2.2.4
and second is the assumption that the conditional probability of X falling within a small nonzero interval centered at x given M can be modeled by a normal distribution
1
P( x | M ) =
e
δ 2π
− ( x− µ ) 2
2δ 2
Eq. 3.2.2.5
where µ and δ are the mean and standard deviation of X respectively. The values µ and
δ can be estimated from the training data. With the third assumption that the distributions of
the expression values of the genes are independent, we obtain the LIK ranking of class a
over class b for the g th gene as follows:
3-24
LIKag→b = log( P( xag,1,K, xag, n | M a )) − log( P( xag,1,K, xag,n | M b ))
a
a
na
na
i =1
i =1
= log(∏ P( x ag,i | M a )) − log(∏ P( x ag,i | M b ))
Eq. 3.2.2.6
na
(x g − µ g )2
(x g − µ g )2
= ∑ − log(δ ag ) − a ,i g a2 + log(δ bg ) + a ,i g b2 ,
2(δ a )
2(δ b )
i =1
and similarly, the LIK ranking of class b over class a for this gene is
LIK bg→a = log( P( xbg,1 ,K , xbg,n | M b )) − log( P( xbg,1 ,K , xbg,n | M a ))
b
b
( xbg,i − µ bg ) 2
( xbg,i − µ ag ) 2
g
g
.
= ∑ − log(δ b ) −
+ log(δ a ) +
g 2
g 2
2
(
)
δ
2
(
δ
)
i =1
a
b
nb
Eq. 3.2.2.7
Genes that have higher likelihood scores are expected to have better ability to distinguish one
class from the other.
3.3 Classification and multivariate feature selection methods
Classification is also called pattern recognition. It is a process to assign one of the prescribed
number of classes (categories) given an input pattern (Haykin, 1999). Training of a classifier
is usually needed to establish the learning model that could reflect the relationship between
input patterns and class labels. This thesis employs four classification methods. They are
decision tree (Quinlan, 1993), neural networks (Haykin, 1999), support vector machines
(Vapnik, 1998) and Bayesian classification (Keller et al., 2000). Due to the space limitation,
only Neural Network, Support Vector Machines and Bayesian classification will be
described. Boosting technique, which combines output of multiple classifiers to form more
accurate hypothesis, will also be illustrated in this section. Three multivariate feature
selection methods: neural network feature selector, recursive feature elimination method and
multivariate likelihood method are also described in this section.
3-25
3.3.1 Neural Networks
A neural network is a massively parallel distributed processor made up of simple processing
units, which has a natural property for storing experimental knowledge and making it
available for future use. Similar to human brain, it acquires knowledge from environment
through a learning process, and its interneuron connection strengths store knowledge
(Haykin, 1999; and Aleksander and Morton 1990). Neural networks as an analysis method
has advantages such as nonlinearity, input-output mapping, adaptivity, evindential response,
contextual information, and fault tolerance.
Three-layer feed-forward neural network is a typical neural network architecture. It has a
simple hierarchical structure with high synaptic connections. A three-layer neural network
model consists of an input layer, a hidden layer and an output layer. Input neurons in the
input layer receive input signals, and send them to neurons in the hidden layer. Hidden
neurons are computational units. They combine their inputs as their local fields, and send
their output to neurons at the output layer. Output neurons then perform a similar
combination and their output is the output of the whole model. Each of the synaptic links in
the model has a weight associated with it, which is used to amplify/reduce the signal when
the signal is passing by.
Computationally, this model can also be described as follows: Suppose there are k 0 input
neurons, k1 hidden neurons and k 2 output neurons. The synaptic links between the input and
hidden layer can be represented as a k1 × (k 0 + 1) matrix w (1) , including biases, and the links
between hidden layer and output layer can similarly be represented as a k 2 × (k1 + 1) matrix
3-26
w ( 2 ) , also including biases. Let a vector x = [+1, x1 L xk0 ]T be the input of the network. The
neurons in the hidden layer first sum up their input together with the bias associated with the
first element of x
v (1) = w (1) x ,
Eq. 3.3.1.1
and perform a certain transformation using activation ϕ (1) function and get their output
+1
y (1) = (1) (1) = [1, y1(1) L y k(11 ) ]T .
ϕ ( v )
Eq. 3.3.1.2
Here we also add an additional element + 1 as the first element of y in order to handle bias
of the output neurons. Similarly, the neurons in the output layer perform the summation
v ( 2) = w ( 2) y (1)
and transformation using activation function ϕ
Eq. 3.3.1.3
( 2)
to get the output
y ( 2) = ϕ ( 2) ( v ( 2) ) .
Eq. 3.3.1.4
Here the activation functions ϕ usually take following forms:
•
•
•
•
Threshold function
1 if v ≥ 0
y = ϕ (v ) =
0 if v < 0
Picewise-Linear function
if v ≥ a
1
y = ϕ (v) = v if a ≥ v > −a , with a > 0
0
if v < −a
Sigmoid function (logistic function)
1
y = ϕ (v ) =
with a > 0
1 + e − av
Hyperbolic tangent function
Eq. 3.3.1.5
Eq. 3.3.1.6
Eq. 3.3.1.7
3-27
y = ϕ (v) = a tanh(bv) = a
•
e bv − e − bv
e bv + e −bv
with a > 0, b > 0
Hyperbolic tangent sigmoid function
2
y = ϕ (v ) =
−1
1 + e − 2v
Eq. 3.3.1.8
Eq. 3.3.1.9
Here v and y are scalars corresponding to the elements of the vectors.
A neural network provides a mapping from its input to its output, and this mapping is
determined by the network’s structure and synaptic weights. With a proper mapping
established, given certain inputs, the network can produce the desired outputs. This
mechanism could be used for classification. Training is needed to obtain a neural network
that can give correct predictions. The training is done using an algorithm which usually
consists of two steps: generate initial weights, and then iteratively refine these weights until
certain stopping criterion is met. In essence, these algorithms are optimization algorithms,
which attempt to meet certain criteria when refining the weights. These criteria are measured
in terms of error functions because they reflect the difference between neural network outputs
and the desired outputs.
Among many training algorithms, back-propagation is most popular (Hertz et al., 1991).
There are two types of training in back-propagation: sequential mode and batch mode. In
sequential mode, the algorithm calculates training error and updates weights each time it
receives a training sample. On the other hand, in batch mode, the algorithm calculates overall
error of all the training samples, and then updates the network’s weights. The sequential
mode is computationally slower than the batch model. But the order of training samples that
are presented to the training algorithm can be randomly assigned, and the stochastic nature of
3-28
samples is able to be modeled. Below is the description of back-propagation algorithm in
sequential mode. Its batch mode version could be easily adapted from the sequential mode.
Suppose the errors at output neurons are defined as
e = d − y ( 2 ) = [e1 Lek2 ]T .
Eq. 3.3.1.10
Back-propagation uses
E = e Te
Eq. 3.3.1.11
as the instantaneous error function of the network. The objective of back-propagation training
is to minimize the average instantaneous error function to a certain extent, given all training
samples.
Given a sample vector x with its class labels d , E can be computed using Eq. 3.3.1.1 to Eq.
3.3.1.4 and Eq. 3.3.1.10 to Eq. 3.3.1.11. To meet the objective, the correction of weights
∆w (1) and ∆w ( 2) should be proportional to the partial derivatives
∂E
∂E
and
(1)
∂w
∂w ( 2 )
respectively. According to the chain rule, we have:
∆w ( 2 ) = −η
= −η
∂E
∂w ( 2 )
∂E ∂e ∂y ( 2) ∂v ( 2 )
∂e ∂y ( 2) ∂v ( 2 ) ∂w ( 2)
T
′
= −η × e × (−1) ⊗ ϕ ( 2) ( v ( 2) ) × y (1)
′
T
= η × (e ⊗ ϕ ( 2 ) ( v ( 2 ) )) × y (1)
T
= η × δ ( 2) × y (1) ,
′
here we let δ ( 2 ) = e ⊗ ϕ ( 2 ) ( v ( 2 ) ) ; and similarly,
Eq. 3.3.1.12
3-29
∆w (1) = −η
= −η
∂E
∂w (1)
∂E ∂y (1)
∂y (1) ∂w (1)
= η × ((w
(2)T
× (e ⊗ ϕ
( 2)
′
( v ))) ⊗ ϕ
( 2)
(1)
′
Eq. 3.3.1.13
( v )) × x
(1 )
T
T
′
= η × ((w ( 2 ) × δ ( 2 ) ) ⊗ ϕ (1) ( v (1) )) × x T
= η × δ (1) × x T ,
′
T
here we let δ (1) = (w ( 2 ) × δ ( 2 ) ) ⊗ ϕ (1) ( v (1 ) ) .
In Eq. 3.3.1.12 and Eq. 3.3.1.13, η is a positive learning rate parameter which controls the
amount of weight adjustment. The operator ⊗ is an element-by-element multiplication.
Activation function ϕ (1) and ϕ ( 2) must be differentiable for back-propagation training. The
bias elements need to be included or excluded in certain steps of matrix multiplication in
order to maintain consistency.
The back-propagation algorithm could be easily extended to training multi-layer feed-forward
neural networks. Adjustment of weights involves only the neuron signals of the successive
layers they connected, so the algorithm is a local method. This also makes it computationally
efficient.
Feed-forward neural networks can be used for classification. For two class problems, a neural
network with a single output neuron is needed. Samples are labeled + 1 or − 1 . The label
value will be the desired values at the output side of the network when training the network.
For multi-class problems, a neural network that has the number of output neurons identical to
3-30
the number of classes is usually required. The desired output value in this case is a vector.
The elements of the vector are − 1 , except the one that corresponds to the class of the sample,
which is + 1 . If the training samples are biased, for example, the number of positive samples
is much less than that of the negative ones, label values other than ± 1 can be used to adjust
the feed-back signal to obtain better performance. Many gene expression classification
problems are two class problems.
3.3.2 Ensemble of neural networks (boosting)
One of the main disadvantages of neural network for classification is that the training result
also depends on initial weights, which are generated randomly. Boosting can be used to
enhance the robustness of the neural network. The term Boosting refers to a machine learning
framework that combines a set of simple decision rules, which is generated by a set of
learners with different learning abilities, into a complex one that has higher accuracy and
lower variance. It is especially useful in handling real world problems that have the following
properties (Freund and Schapire, 1996): the samples have various degrees of hardness to
learn and the learner is sensitive to the change of training samples. The complexity of
learning hardness often occurs when applying machine learning method to tackle biological
problems. There are three boosting approaches (Haykin, 1999):
•
Boosting by filtering: If the number of training samples is large, the samples are either
discarded or kept during training.
•
Boosting by subsampling: With a fixed training sample set size, the probability to
include samples into training sample set for learning algorithms is adjusted.
3-31
•
Boosting by reweighting: This approach assumes that the training samples could be
weighted by the learning algorithms. The training errors are calculated by making use
of these weights.
AdaBoost (Freund and Schapire, 1996) is a simple and effective boosting algorithm through
subsampling. During learning process, the algorithm tries to make the learners focus on
different portions of the training samples by refining the sampling distribution. Figure 3.1
shows one of two versions of AdaBoost training algorithm.
The input of the algorithm are n training samples, {(b1 , y1 ),K, (b n , y n )} , where
bi
(i = 1K n) are the sample vectors, and yi ∈ Y
(i = 1K n) are their associated class
labels, which are also the desired outputs of the learning models. Y is the set of class labels.
The algorithm starts by setting the initial sampling distribution d 1,1Kn to a uniform one. It then
enters an iterative process. At the t th iteration, the algorithm calls the function
train_learner() with training samples b1..n and sampling distribution d t ,1Kn , to get the trained
learning model M t . By calling function get_hypothesis() with M t and b1..n , the algorithm
generates the hypothesized classes of training samples. The algorithm then calculates total
error ε t by adding up all the wrongly predicted samples’distribution. If the error is bigger
than
1
, the algorithm stops. Otherwise, it proceeds to calculate a factor β t . This factor is
2
used to reduce the portion of the correctly predicted training samples, in the sampling
distribution d t +1,1Kn of the next iteration. When the algorithm terminates, T learning models
M 1KT are trained with factors β1KT indicating the contribution of the learning models to the
combined hypothesis. The combined hypothesis for a test sample b can then be calculated as
3-32
1
h(b) = arg max
log
∑
y∈Y
t :get_hypothesis( M t ,b )= y β t
.
Eq. 3.3.2.1
When number of training samples is small, β t could be zero. We set a lower threshold to β t
when implementing the algorithm so that h(b) can be calculated.
function AdaBoost({(b 1 , y1 ),K, (b n , y n )})
d1,i =
1
n
(i = 1K n)
// Input: training samples
// Initial sampling distribution
for t = 1 to T do
M t = train_learner(b1..n , d t ,1Kn )
ht ,1..n = get_hypothesis( M t , b1Kn )
εt =
∑d
i :ht ,i ≠ y i
if ε t >
t ,i
1
then
2
T = t −1
terminate loop
// Terminate the algorithm
end
// Update sampling distribution to focus on the wrongly classified samples
ε
βt = t
1 − εt
β t d t ,i if ht ,i = yi
d t′,i =
(i = 1K n)
if ht ,i ≠ y i
d t ,i
d′
d t +1,i = n t ,i
(i = 1K n)
∑ d t′,i
i =1
end
return ( M 1KT , β 1KT )
end
Figure 3.1: AdaBoost algorithm
3-33
Theoretical study shows that if the hypothesis obtained by individual classifiers constantly
has error that is slightly better than random guess, the number of prediction errors of the final
hypothesis h drops to zero exponentially fast when T increases (Freund and Schapire,
1996).
Three layer neural networks can be the learning model integrated with AdaBoost. In this case,
train_learner() consists of initializing, training and optimizing neural network weights, and
get_hypothesis() corresponds to neural network decision making. The running of the
algorithm will construct T neural networks in total for making combined decision.
3.3.3 Neural network feature selector
This section first analyses the limitation of applying information gain measure for ranking of
features having continuous value, then describes the neural network feature selector method.
3.3.3.1 Disadvantage of information gain measure
Information gain can be applied in measuring both discrete and continuous feature values.
But in continuous case, it only takes the order of the values into consideration, which may not
be sufficient. For example, suppose the expression matrix is
3-34
6
7
8
9
10
B=
11
12
13
14
15
1
2
3
4
5
16
17
18
19
20
7
7.5
8
8 .5
9
12
12.5
13
13.5
14
and the corresponding class labels are y = [− 1,−1,−1,−1,−1,+1,+1,+1,+1,+1] . The three
T
features’values are plotted in Figure 3.2 at line y = 1 , y = 2 and y = 3 . Intuitively we can
see that the discrimination abilities of the feature at line y = 2 and y = 3 are better than the
one at line y = 1 . Compared with feature values at line y = 1 , the distance between feature
values of different classes at line y = 2 are longer, and the density of feature values within
different classes at line y = 3 are higher. But the information gain ratios of these three
features are the same because it only takes the order of the feature values into account, hence
achieving the same maximum gain ratio for all three features when split in the middle.
3-35
Figure 3.2: Plot of feature values with class labels, ‘o’for ‘-1’and ‘+’for ‘+1’
3.3.3.2 Neural network feature selector
Since feature selection is a preprocessing process for classification, classification accuracy
can also be a selection criterion. The advantage of integrating a classifier into the feature
selection process is that the feature set is optimized by the classification accuracy. Moreover,
the training of the classifier and the selection of the features use the same bias. The
consistency improves the classification performance. However, the computational cost of the
integration may be high. In (Liu and Motoda, 1998) this approach is called wrapper model.
This framework is shown in Figure 3.3.
Procedure wrapper accepts full feature set (F), training samples (D) and testing samples (T)
as input. It first generates a subset of features (S), and then performs cross validation using S
and D to get classification accuracy (A). These two steps are repeated until A is sufficiently
3-36
high under certain criterion. A classification model (M) is then obtained from D with selected
S. Finally M and S are used to perform the test to measure the performance. In the algorithm,
cross validation can be done by dividing D into a training set and a validating set; or by the
leave-one-out method (see more details below).
procedure wrapper (F, D, T)
do
S=feature_set_gen(F)
A=cross_validation(S,D)
until sufficient(A)
M=train(S,D)
test(M,S,T)
end
Figure 3.3: Wrapper model framework
Setiono and Liu, (1997) proposed a neural network feature selection method based on the
wrapper approach. Here in this thesis, we applied it to gene expression analysis, with some
modifications. The following is a detailed description of this method.
The algorithm starts with all features, and removes features that have minor contribution to
classification one by one. The algorithm first trains a neural network with a given feature set,
then disables each feature and estimates the classification performance of this neural network
with the remaining features. If the decrease of the estimated performance is within an
acceptable level, the algorithm constructs a neural network with the remaining features, and
calculates the actual classification performance. If the actual performance is also acceptable,
the feature is removed, and the algorithm continues searching for more features to be
removed. Otherwise, it keeps this feature and continues to test other features according to
3-37
their estimated performances. It is a greedy method, and usually achieves sub-optimal
solutions.
There are two ways to train and validate the neural network. Suppose there are n samples.
The first one is to separate these samples into two sets: training set and validating set. If n is
too small to produce sufficiently large training and validating sets, the leave-one-out
technique can be used to perform training and validation n times, and obtain the average
training and validation accuracy.
Neural network feature selector method requires the neural network training to force the
weights associated with an irrelevant input neuron to have small magnitude, in order to
reduce the effect of the corresponding feature’s removal on the classification performance.
This is implemented using weight decay on the weights between the input and the hidden
layer when applying the back-propagation training. After every element of the weights is
updated with ∆w (1) , according to Eq. 3.3.1.13,
′
wi(,1j) = wi(,1j) + ∆wi(,1j) ,
Eq. 3.3.3.1
the elements of the new weights are computed as follows
ε jη
(1)′
(1) ″
wi , j = 1 −
w ,
2 i, j
2
1 + ( w (1) )
∑i i, j
Eq. 3.3.3.2
where η is the learning rate parameter, and ε j is a penalty term associated with the j th
input neuron. Eq. 3.3.3.2 is similar to the weight decay method in (Hertz et al., 1991), but the
focus is different: the later can be used to eliminate hidden neurons.
3-38
In order to improve the convergence, we also tried the cross entropy error function instead of
Eq. 3.3.1.11,
n k2
F(w (1) , w ( 2 ) ) = − ∑∑ d kp log( y k( 2 ), p ) + (1 − d kp ) log(1 − y k( 2 ), p ) ,
p =1 k
[
]
Eq. 3.3.3.3
where the desired value d kp of p th sample of k th class is either 1 or 0 , which is different
from the one used in Eq. 3.3.1.11. The derivation of the back propagation with error function
in Eq. 3.3.3.3 is as follows:
∂ F n k2 ∂ F ∂y k( 2 ), p
= ∑∑
∂w p =1 k ∂yk( 2 ), p ∂w
,
Eq. 3.3.3.4
where
d kp
1 − d kp
∂F
=
−
−
y ( 2 ), p 1 − y ( 2 ), p .
∂yk( 2 ), p
k
k
Eq. 3.3.3.5
Because the three layer neural network model is
k1
k0
yk( 2 ), p = ϕ ( 2 ) ∑ϕ (1) ∑ xip w (ji1) wkj( 2 ) ,
i =1
j =1
Eq. 3.3.3.6
we have
∂yk( 2 ), p
′
′
= ϕ ( 2 ) vk( 2 ), p ⋅ ϕ (1) v (j1), p = ϕ ( 2 ) vk( 2 ), p ⋅ y (j1), p ,
(2)
∂wkj
(
)
(
)
(
)
Eq. 3.3.3.7
and
′
′
∂yk( 2 ), p
= ϕ ( 2 ) (vk( 2 ), p ) ⋅ ϕ (1) (v (j1), p ) ⋅ xip .
(1)
∂w ji
Eq. 3.3.3.8
Again according to the delta rule (Haykin, 1999) and the gradient descent rule (Hertz et al.,
1991), we obtain
∆wkj( 2 ) = −η
and
n d kp
1 − d kp
∂F
=
η
−
∑ ( 2 ), p
∂wkj( 2)
1 − yk( 2 ), p
p =1 yk
( 2 ) ′ ( 2), p (1), p
ϕ
vk
⋅ yj
(
)
Eq. 3.3.3.9
3-39
∆w (ji1) = −η
n k d kp
1 − d kp
∂F
=
−
η
∑∑ ( 2 ), p
∂w (ji1)
1 − y k( 2 ), p
p =1 k y k
2
( 2 ) ′ ( 2 ), p
′
ϕ
⋅ ϕ (1) v (j1), p ⋅ xip .
vk
(
)
(
)
Eq. 3.3.3.10
Because the implementation of the algorithm is on MATLAB, we transform the computation
d ij
1 − d ij
of weight into matrix form: Let F′ be a n × k2 matrix, the elements f ij = ( 2 ),i −
y
1 − y (j 2 ),i
j
which has the value opposite to that of
∂F
. Let X be n × k0 matrix; V (1) and Y (1) be
∂yk( 2), p
n × k1 matrix; V ( 2) and Y( 2 ) be n × k2 matrix. Then we have
′
∆W (2) = η (F ′ ⊗ ϕ ( 2) (V ( 2) )) T Y (1)
Eq. 3.3.3.11
and
T
∆W
(1)
′
′T
= η (F ′ ⋅ ϕ ( 2 ) (V ( 2) )) ⊗ I ⋅ ϕ (1) (V (1) ) X
Eq. 3.3.3.12
where I is a n × n identity matrix used to extract the diagonal elements out.
Figure 3.4 shows the algorithm: At line 1, the function neural_network_feature_selector()
accepts six parameters: a set of m feature vectors {f1 ,K, f m } , a set of penalty parameters
{ε 1 ,K, ε m } , a penalty parameter amplification factor f , penalty parameter thresholds
(ε min , ε max ) , a class label vector y , an allowable maximum decrease in validation accuracy
′ , rmin
′′ ) , and a linear combination factor rfactor for
∆r ′′ , training and validation thresholds (rmin
testing and training performances.
At lines 2 to 5, the function copies feature number, feature set, penalty parameter set into
′′ to a
three variables m′ , F and E , and initializes the maximum validating accuracy rmax
3-40
small positive value ε . It then enters an iterative process, at lines 6 to 40, to remove features
one by one until no more feature in F can be removed with sufficiently high performance. In
the removal process, at line 7, a neural network N is initialized according to m′ , by calling
function initialize () . Then at line 8, N is trained and validated with features in F and
penalty parameters in E by calling train_vali date() , which returns the trained network N ,
′′ . At line 10,
training accuracy r ′ and validating accuracy r ′′ . The algorithm then updates rmax
m′ neural networks N 1,K,m′ are initialized with same weights as N .
At lines 11 to 15, m′ feature sets F1,K,m′ and penalty parameter sets E1,K, m′ are constructed,
each of which has one feature omitted; and they are tested and validated with corresponding
neural network by calling simulate_validate() at line 14; the corresponding estimated training
accuracy r1′,K,m′ and validating accuracy r1′,′K,m′ are returned. At line 16, r1′,K,m′ and r1′,′K,m′ are
sorted according to their linear combination in descending order. The higher the linear
combination rfactor ri′ + ri′′ is, the more likely the corresponding i th feature to be eliminated.
The factor rfactor is usually set to be bigger than one. When the difference among r1′,K,m′ are
high, r1′,K,m′ will be the main contributor to the ranking. However, when the difference among
r1′,K,m′ are not so high, which often occurs for small of training set, ranking is mainly affected
by r1′,′K,m′ .
At lines 17 to 25, according to the sorted index s1K
, ,m′ , the corresponding neural networks
whose estimated training and validation accuracy rates, rs′i and rs′i′ are bigger than the
′ , rmin
′′ ) are retrained to get actual training accuracy rs′i and validating accuracy
threshold (rmin
rs′i′ . If the validating accuracy is sufficiently high compared with maximum validating
3-41
accuracy, as indicated in the conditions at lines 23 and 26, then the penalty parameters is
updated at lines 27 to 33. At lines 35 to 38, the selected feature f si and penalty parameter ε si
′′ is updated. At lines
are removed from F and E ; feature number m′ is decreased; and rmax
29 to 33, the way to update penalty parameter is as follows: for any r′j , if it is bigger than
average value r ′ , which means that the corresponding feature is likely to be removed, the
corresponding ε j is enlarged by the factor f ; otherwise it is reduced by f . A penalty
parameter’s lower and upper thresholds [ε min , ε max ] are set to prevent it to be too large or too
small. After the feature removal process at from lines 6 to 40, the selected feature set in F is
returned at line 41.
As mentioned earlier, neural network feature selector method is based on the wrapper model.
But its feature generation and cross validation part are not so distinct. In general, lines 7 to 16
and lines 26 to 39 relate to feature generation; lines 17 to 25 validate the feature set; and line
40 contains the stopping criterion.
1
2
3
function
3-42
′ , rmin
′′ ), rfactor )
neural_network_feature_selector({f1 ,K
, f m },{ε1 ,K, ε m}, f , (ε min , ε max ), y , ∆r ′′, (rmin
m′ = m
F = {f1 , , f m }
5
E = {ε 1 ,K , ε m }
′′ = ε
rmax
6
do
4
7
8
9
10
11
12
N = initialize (m ′)
( N , r ′, r ′′) = train_vali date( N , F , E , y )
′′ = max(rmax
′′ , r ′′)
rmax
( N1 ,K , N m′ ) = ( N ,K, N )
for i = 1 to m′
Fi = F − {N i }
Ei = E − {ε i }
(ri′, ri′′) = simulate_validate( N i , Fi , Ei , y )
13
14
15
end
((rs′1 ,K , rs′m′ ), (rs′1′ ,K, rs′′m′ )) = sort_descending( rfactor × (r1′,K , rm′ ′ ) + (r1′′, K, rm′′′ ))
i=0
16
17
18
do
19
20
i = i +1
′
if rs′ > rmin
i
and rs′′
i
N si = initialize (m′ − 1)
21
( N si , rs′i , rs′i′ ) = train_validate( N si , Fsi , E si , y )
′′ − rs′i′
rmax
δ =
′′
rmax
22
23
24
end
i ≥ m′
if δ ≤ ∆r ′′
25
until
26
or
δ ≤ ∆r ′′
then
1 m′
∑ ri′
m′ i =1
j = 1 to m′
if r j′ ≥ r ′
r′ =
27
28
for
29
and
ε j ∈ [ε min , ε max ]
ε j = fε j
30
31
else
εj =
32
33
34
εj
f
end
end
F = F − { f si }
35
E = E − {ε si }
m′ = m′ − 1
′′ = max(rmax
′′ , rs′′i )
rmax
36
37
38
39
end
until i ≥ m ′
return F
40
41
42
′′
> rmin
end
Figure 3.4: Neural network feature selector
then
3-43
The main modifications of the neural network feature selector in this thesis compared with
the one proposed by Setiono and Liu (1997) are summarized below.
1) Use of back propagation training method: Setiono and Liu employed BFGS (BroydenFletcher-Shanno-Goldfarb) method for training of neural networks, which is a variant of
quasi-Newton method that has been shown to be very effective. However, in the training
process, BFGS computes a Hessian matrix with dimension equal to the square of the number
of features. The size of the matrix is huge when the algorithm is applied to microarray dataset
consisting of a large number of features. So we replace the BFGS algorithm with back
propagation algorithm in our implementation.
2) Use of leave-one-out validation as an estimator: The neural network feature selector
algorithm proposed by Setiono and Liu uses a fixed partition of training and cross-validation
sample set as initial input. In a typical microarray dataset, when there are only a very limited
number of samples, a fixed training and cross-validation sample set may introduce large bias
in estimating the generalization performance of the trained neural networks. Instead of simply
splitting the sample set into two partitions, the neural network feature selector proposed in
this thesis employs leave-one-out method to reduce the bias in estimating the generalization
performance.
3) Use of different of penalty function: Penalty functions are employed in both versions of
neural network feature selectors to force small weights between the input layer and the
hidden layer to zero, in order to reduce the effect of irrelevant features to the classification.
The penalty function used in (Setiono and Liu, 1997) is
3-44
( ) ) + ε ( (w ) ) ,
P( w) = ε (∑∑
∑
1 + β (w )
β wi(,1j)
1
i
j
2
(1) 2
i, j
(1) 2
i, j
2
Eq. 3.3.3.13
i
where ε1 , ε 2 and β are parameters for deciding the detailed penalty effect. From Eq.
3.3.1.13 it can be easily seen that the derivative of the penalty function against a weight
involves only that weight itself:
2 βwi(,1j)
∂ P( w)
+ 2ε w (1) .
= ε1
2 i, j
2
2
∂wi(,1j)
(1)
1 + β wi , j
( )
Eq. 3.3.3.14
As a result, the update of the weights tries to force the individual small weights to zero. By
contrast, the penalty function used in this thesis forces all weights associated with an input
neuron to zero, if the summed square of these weights are small. The new penalty function is
implicitly implemented in Eq. 3.3.3.2. It helps reducing the effect of an input neuron on all
hidden neurons when this input neuron is removed.
3.3.4 Support vector machines
Support vector machine is a linear learning model that can also perform classification. It was
invented by Boser et al. (1992). The theoretical aspect of Support Vector Machines is based
on Statistical Learning Theory (Vapnik, 1998). Section 3.3.4.1 describes the basic SVM
learning model and its training; Section 3.3.4.2 illustrates the linearly non-separable case;
Section 3.3.4.3 describes SVM with nonlinear mapping; Section 3.3.4.4 introduces a
multivariate feature selection method based on SVM.
3-45
3.3.4.1 Linearly separable learning model and its training
From decision-making point of view, a linear classifier tries to obtain decision surfaces that
can discriminate samples of different classes. These decision surfaces are hyperplanes.
Suppose there are n training samples that belong to two classes, {(x1 , y1 ), K , (x n , yn )} ,
where x i
(i = 1K n) are sample vectors, and yi = ±1 (i = 1K n) are associated class labels.
If there exists a hyperplane w T x + b = 0 so that for all n samples x i ,
w T x i + b ≥ 0
T
w x i + b < 0
yi = + 1
yi = − 1
Eq. 3.3.4.1
hold, then these samples are linearly separable. The distance between the separating
hyperplane and its closest sample vector is called margin of separation, denoted as ρ . A
T
Support Vector Machine’s task is to find an optimal separating hyperplane w ∗ x + b ∗ = 0
that has the biggest margin of separation among all separating hyperplanes. Obviously, the
distances between this hyperplane and its nearest sample vectors on both sides are equal.
With w ∗ and b∗ properly scaled, we have
w ∗ T x i + b ∗ ≥ +1 yi = +1
∗T
w x i + b ∗ ≤ −1 yi = −1
Eq. 3.3.4.2
for all samples. The sample vectors that satisfy
yi ( w ∗ x i + b ∗ ) = 1
T
Eq. 3.3.4.3
are called support vectors. Let us denote a pair of support vectors on both sides of the
separating hyperplane as x + and x − ,
w ∗ T x + + b ∗ = +1
.
∗T −
w x + b ∗ = −1
We then have
Eq. 3.3.4.4
3-46
1 w ∗ x + w ∗ x −
ρ=
−
2 w∗
w ∗
T
=
=
T
(
T
T
1
w ∗ x+ − w∗ x−
∗
2w
)
Eq. 3.3.4.5
1
w∗
T
where w ∗ = w ∗ w ∗ .
So ρ is maximized when w ∗ is minimized. The training task can then be converted to an
optimization problem
minimize
w
subject to
yi (w x i + b) ≥ 1 i = 1,K, n
T
.
Eq. 3.3.4.6
This constrained optimization problem can be solved using the method of Lagrange
multipliers. First, the Lagrange function corresponding to problem in Eq. 3.3.4.6 is
constructed
J(w , b, a ) =
n
1
w − ∑α i [ yi (w T x i + b ) − 1]
2
i =1
Eq. 3.3.4.7
where the nonnegative variables a = [α1 ,K,α n ]T are called Lagrange multipliers. The
problem in Eq. 3.3.4.6 is equivalent to minimizing J(w, b, a ) with respect to w and b , or
maximizing J( w, b, a ) respect to a . The solution lies in the saddle point of J(w , b, a )
∂ J( w, b, a )
=0
∂w
.
∂ J( w, b, a )
=0
∂b
By solving Eq. 3.3.4.8, we have
Eq. 3.3.4.8
3-47
n
=
w
α i yi x i
∑
i =1
,
n
∑α i yi = 0
i =1
Eq. 3.3.4.9
and by substituting Eq. 3.3.4.9 into Eq. 3.3.4.7 we get
J( w , b , a ) =
n
∑ αi −
i =1
1
2
n
n
i =1
j =1
∑∑
y i y jα iα j x iT x j .
Eq. 3.3.4.10
The problem in Eq. 3.3.4.6 can then be transformed into the quadratic optimization problem
n
maximize W(a ) = ∑α i −
i =1
1 n n
∑∑ yi y jα iα j x iT x j
2 i =1 j =1
n
∑ yiα i = 0, α i ≥ 0, i = 1,K, n
subject to
.
Eq. 3.3.4.11
i =1
This quadratic optimization problem has a unique solution that can be expressed as a
weighted combination of the training samples.
Suppose the solution of the problem is a ∗ = [α1∗ ,K,α n∗ ]T , according to Eq. 3.3.4.9, we have
n
w ∗ = ∑α i y i x i .
Eq. 3.3.4.12
i =1
and using Eq. 3.3.4.4, we get
T
T
b ∗ = 1 − w ∗ x + = −1 − w ∗ x − ,
Eq. 3.3.4.13
where support vectors x + and x − could be determined using the Karush-Kuhn-Tucker
conditions (Fletcher, 1987; and Bertsekas, 1995). According to the Karush-Kuhn-Tucker
condition, following equation holds,
α i [ yi (w T x i + b) − 1] = 0 i = 1,K, n .
So the sample vectors with positive Lagrange multipliers are support vectors.
Eq. 3.3.4.14
3-48
3.3.4.2 Linearly non-separable learning model and its training
In the linearly non-separable case, there is no hyperplane that can separate all samples. In this
circumstance, a set of slack variables {ξ i } ξ i ≥ 0, i = 1,K, n is introduced, and the separation
hyperplane is redefined as
yi (w T x i + b ) ≥ 1 − ξ i
i = 1,K, n ,
Eq. 3.3.4.15
and the optimization problem becomes
n
minimize
w T w + C∑ξi
subject to
y i (w x i + b) ≥ 1 i = 1, K, n
i =1
.
Eq. 3.3.4.16
T
where the parameter C balances the generalization ability represented in the first term, and
the separation ability indicated in the second term. Problem in Eq. 3.3.4.16 can be converted
to its dual problem similar to that in the separable case, in which the slack variables are
omitted
n
maximize W(a ) = ∑α i −
i =1
n
subject to
∑yα
i =1
i
i
1 n n
∑∑ yi y jα iα j x Ti x j
2 i =1 j =1
,
Eq. 3.3.4.17
= 0, 0 ≤ α i ≤ C , i = 1,K, n
and the optimum solution becomes
ns
w ∗ = ∑ α si y si x si ,
Eq. 3.3.4.18
i =1
where n s is the number of support vectors, and si , i = 1K n s are indices corresponding to
those support vectors. To identify support vectors, the Karush-Kuhn-Tucker condition is
defined as
α i [ y i (w T x i + b) − 1 + ξ i ] = 0
i = 1,K , n .
ξ i (α i − C ) = 0
Eq. 3.3.4.19
3-49
According to this condition, all sample vectors with positive Lagrange multipliers are support
vectors; and the slack variable is non-zero only when its corresponding Lagrange multiplier
equals to C . The value of b ∗ can be determined by choosing any support vector x i with
Lagrange multiplier 0 < α i < C :
T
b∗ = 1 − w ∗ xi
if yi = +1
Eq. 3.3.4.20
or
T
b ∗ = w ∗ x i − 1 if yi = −1 .
Eq. 3.3.4.21
When the classification task has more than two class labels, there are two ways to transform
it to a binary classification problem. One way is to encode class labels using a binary
representation. Suppose there are l class labels in the task. log 2 (l ) support vector machines
are needed to perform classification task together. If a sample has the k th class label, a
vector y = [ y1 , K , y log 2 (l ) ]T ,
y1 , K , y log 2 (l ) ∈ {+1,−1} ,
+ 1 if ith bit of binary form of k is 1
where yi =
i = 1, K, log 2 (l ) ,
− 1 if ith bit of binary form of k is 0
is sufficient to represent corresponding desired outputs of the support vector machines.
Another way is the one against others method. In this method, l support vector machines are
needed. The corresponding desired output of a sample with k th class label
+ 1 if i = k
is y = [ y1 ,K , yl ]T , where yi =
i = 1,K, l .
− 1 otherwise
3.3.4.3 Nonlinear Support Vector Machines
Support vector machine is basically a linear model. It can be extended to handle non-linear
cases by introducing slack variables, as is shown in Section 3.3.4.2. In addition, nonlinear
3-50
mapping functions for transformation of input vectors can also be employed. This approach
makes the learning more flexible. Let t (x) = [t 0 (x), t 1 (x ),K , t l (x)]T be a vector of nonlinear
transform functions, where t 0 (x) = 1 . The optimal hyperplane is then defined as follows
w T t ( x) = 0 ,
Eq. 3.3.4.22
where the bias term is included implicitly in w . By adapting Eq. 3.3.4.12 we get
l
w = ∑ α i yi t (x i ) .
Eq. 3.3.4.23
i =1
Substituting Eq. 3.3.4.23 into Eq. 3.3.4.22, we obtain
l
∑α
i =1
i
y i t T ( x i )t ( x ) = 0 .
Eq. 3.3.4.24
Let the inner product kernel K(x i , x ) = t T (x i )t (x) be a symmetric function, Eq. 3.3.4.24
becomes
l
∑α
i =1
i
yi K(x i , x ) = 0 .
Eq. 3.3.4.25
Problem Eq. 3.3.4.17 is then reformulated as
n
maximize W(a ) = ∑ α i −
i =1
n
subject to
∑yα
i =1
i
i
1 n n
∑∑ y i y jα iα j K(x i , x j )
2 i =1 j =1
.
Eq. 3.3.4.26
= 0, 0 ≤ α i ≤ C , i = 1,K, n
The optimal decision hyperplane can be found by solving problem in Eq. 3.3.4.26 and
substituting the Lagrange multipliers into Eq. 3.3.4.25.
The complexity of the target function to be learned depends on the way it is represented. The
kernel approach provides a means to implicitly map input vectors into a feature space, i.e. the
kernel can be used without knowing its corresponding transforming function. The
3-51
introduction of kernel simplifies the design of a learner, and may improve generalization
ability. This approach can be used not only in support vector machine but also in other
learning models. Some of the examples of learning models that consist of kernel are listed
below (Haykin, 1999):
•
Polinomial learning machine
K(x, x i ) = (x T x i + 1) p .
•
Radial-basis function network
K(x, x i ) = e
•
Eq. 3.3.4.27
−
1
2σ 2
x − xi
.
Eq. 3.3.4.28
Three layer neural network
K(x, x i ) = tanh( β 0 x T x i + β 1 )
Eq. 3.3.4.29
where p , σ , β 0 and β1 are pre-specified parameters.
3.3.4.4 SVM recursive feature elimination (RFE)
The weights of a trained SVM can indicate the importance of the corresponding features to
the classification. Based on this idea, Guyon et al. (2002) proposed a recursive feature
elimination method. After training a linear kernel SVM, its weight can be obtained by Eq.
3.3.4.18. The algorithm iteratively trains SVM and eliminates the feature(s) with small
weights, until the feature set become empty. Figure 3.5 shows the algorithm:
3-52
RFE( S = {f1 ,K, f m } , y )
While S > 1
w = svm_training( S , y )
f = arg min( wi 2 ), i = 1,K, m
{ }
S =S− ff
end
Figure 3.5: Recursive feature elimination algorithm
3.3.5 Bayesian classifier
Keller et al. (2000) used a simple classification method based on naïve Bayes rule. Given an
expression vector x of m selected features, the classification of a sample is computed as
follows
class( x) = arg max(log P( M i | x)), i = a, b ,
i
Eq. 3.3.5.1
where P(M i | x) is a posteriori probability that M i is true given x . Applying the Bayes
rule once again, the class for vector x can be predicted as
class( x) = arg max(log P(x | M i ))
i
m
= arg max(∑ log P( x g | M i ))
i
g =1
m
= arg max(∑ − log δ ig −
i
g =1
Eq. 3.3.5.2
( x g − µ ig ) 2
), i = a, b ,
2(δ ig ) 2
where µig and δ ig are the mean and standard deviation of the feature values of the training
samples of class i . In a binary classification, we can be more confident about the
3-53
classification when the difference between log P(x | M a ) and log P(x | M b ) is bigger.
However, in order to obtain more information regarding the confidence of the classification,
we need to compute the following
class( x ) = log P(x | M a ) − log P(x | M b )
( x g − µ ag ) 2 m
( x g − µ bg ) 2
g
log
δ
−
−
−
= ∑ − log δ ag −
∑
b
.
2(δ ag ) 2 g =1
2(δ bg ) 2
g =1
m
Eq. 3.3.5.3
A positive difference means that the sample is predicted to be class a , and a negative
difference means that the sample is predicted to be class b . The larger the difference, the
more confident we are about the classification. We also make use of this difference when
computing another measure of accuracy, i.e. acceptance rate, which will be discussed in
Section 4.1.6.1.
3.3.6 Two discriminant methods for multivariate feature selection
We found that a number of multivariate feature ranking methods can be placed in a unified
framework. The methods attempt either to find a vector projection of the samples onto which
maximizes or to minimize certain objective function f(w ) . The function f(w ) is originally
used on individual features to measure the discrimination ability or diversity of the feature.
The magnitude of the elements in the vector then indicates the relative importance of the
features. When f(w ) is extremal margin, the method is equivalent to RFE. When f(w ) is set
to Fisher’s criterion (see Section 3.3.6.1), the method becomes Fisher’s Linear
Discrimination method. When the function f(w ) is substituted by Eq. 3.3.3.3, the method
resembles neural network feature selector in the sense that the optimized weights between
input and hidden layer can indicate the relative importance of the input neuron. When f(w ) is
3-54
the standard deviation of the projection of all samples, the method becomes PCA. Section
3.3.6.1 describes Fisher’s linear discriminator. In Section 3.3.6.2, we attempt to use
Likelihood ranking method as the objective function to rank relative discrimination
contribution of features.
3.3.6.1 Fisher’s linear discriminant
Let G = {g1 ,K, g m } be a set of features. By performing linear transform ya ,i = ∑ w g xag,i and
g∈G
yb ,i = ∑ w g xbg,i , that is projecting all samples from m dimension space to a unit vector w ,
g∈G
we can obtain the Fisher’s criterion of the projections alone w ,
F(w ) =
(µ a′ − µ b′ )2
2
2
δ a′ + δ b′
(w
u a − w Tub )
= T
w ( n a Σ a + nb Σ b ) w
2
T
,
Eq. 3.3.6.1
where µ ′ and δ ′ are the mean and standard deviation of two classes of projections
respectively, u a and u b are the mean of the original sample vectors of the two classes
respectively, and Σ a and Σ b are the covariance matrix of the samples from the two classes
respectively. Fisher’s linear discriminant tries to find the weight w that maximizes F(w ) ,
that is,
maximize F(w )
s.t.
wTw = 1
.
Eq. 3.3.6.2
The solution of the maximization problem is
w * = (n a Σ a + nb Σ b ) −1 (u a − u b ) .
Eq. 3.3.6.3
The value of w * can be an indicator of the contribution of features to discrimination.
3-55
3.3.6.2 Multivariate Likelihood feature ranking
We propose a multivariate likelihood feature selection method that is based on a similar idea
as that of Fisher’s linear discriminant. Suppose there are n = na + nb samples with m
features. Recall Keller’s Likelihood method for ranking for individual gene g , which are
expressed in Eq. 3.2.2.6 and Eq. 3.2.2.7 in Section 3.2.2. Let G = {g1 ,K, g m } be a set of
features. By performing linear transform ya , i =
∑ w g xag,i
g ∈G
and yb, i =
∑ w g xbg, i ,
that is
g ∈G
projecting all samples from m dimension space to a unit vector w , we can obtain the
likelihood of the projections of the samples of two classes on vector w ,
( y a ,i − µ a ) 2
( y a ,i − µ b ) 2
= ∑ − log(δ a ) −
+ log(δ b ) +
2(δ a ) 2
2(δ b ) 2
i =1
na
LIK a→b
Eq. 3.3.6.4
and
nb
( y − µb )2
( y b ,i − µ a ) 2
.
LIK b→a = ∑ − log(δ b ) − b ,i
+
+
log(
δ
)
a
2(δ b ) 2
2(δ a ) 2
i =1
Eq. 3.3.6.5
Where µ and δ are the mean and standard deviation of two classes of projections,
respectively.
The multivariate feature selection process becomes:
maximize f(w )
s.t.
wTw = 1
Eq. 3.3.6.6
where f(w ) = LIK a→b , f(w ) = LIK b→a or f(w ) = LIK a→b + LIK b→a . When the maximization
problem in Eq. 3.3.6.6 is solved, it becomes an indicator of the contribution of features to
3-56
discrimination. We use Sequential Quadratic Programming method (Fletcher, 1987) to find a
suitable w .
3.3.7 Combining univariate feature ranking method and multivariate
feature selection method
Information gain and likelihood method are univariate feature ranking methods in the sense
that they assume genes contribute to classification independently, and rank the genes
according to their individual contribution. This assumption has computational advantages.
But in real world, genes are often working together for a certain function, and the
combinatorial effect of these genes is not considered by univariate selection methods. On the
other hand, neural network feature selector, recursive feature elimination and multivariate
likelihood method consider the whole contribution of subset of features to the classification.
These three approaches have the potential to select smaller subsets of features with higher
classification performance. However, the selection process may be obscured when applied to
microarray datasets with high dimensionality and in the presence of large number of
irrelevant features. Take RFE as an example, the presence of large number of irrelevant
features hides the discriminative information from relevant features. This can be seen in the
formulation of the SVM dual problem where the coefficients of the quadratic terms in the
dual problem are computed as the scalar products of two inputs
m
x i T x j = ∑ xik x jk
k =1
Eq. 3.3.7.1
.
The gene elimination process is very sensitive to change in the feature set. SVM also has the
disadvantage that it is sensitive to outliers as discussed in Guyon et al. (2002). In microarray
data, the outliers may be introduced by: 1) noise in the expression data, or 2) incorrectly
3-57
identified or labeled samples in the training dataset. It is therefore more beneficial to apply
RFE on a dataset with a reduced number of features. A univariate feature selection algorithm
can be used to first efficiently reduce the large number of features originally present in the
dataset and a multivariate feature selection method such as RFE can then be applied to
remove more features. To summarize, we first identify and remove genes that are expected to
have low discrimination ability as indicated by LIK scores. Then, we apply RFE to remove
the size of the feature set further. With this integrated approach to feature selection, we are
able to achieve good classification performance with fewer genes than those reported by
Guyon et al. (2002) and Keller et al. (2000).
A multivariate method is usually very time-consuming when applied to a dataset with
thousands of genes. For RFE, in order to eliminate one or more genes, a new SVM has to be
trained, and the overall computational cost is Ω(m 2 n 2 ) . On the other hand, LIK ranks genes
independently, which makes the computational complexity of LIK, Ο(mn) . Using LIK first
to reject a large number of genes, and then using RFE to perform further selection will save
significant running time compared to just using RFE alone. This is especially important when
the improvement in microarray technology makes it possible to obtain gene expression values
from tens or even hundreds of thousands of genes.
4-58
4 Experimental results and discussion
The experimental results from and discussions on applying machine learning methods to
global gene expression analysis in our research will be described in this chapter. We start
with the experiments on the datasets of the second type classification problem, which consist
of large number of features and small number of samples. The high dimension nature of this
kind of problem makes it distinct from common classification problems. We then describe the
test result on a newly released dataset, which is of the first type of classification problem with
large number of samples and small number of features.
4.1 Second type - high dimension problems
The main work in this thesis focuses on the second type of classification problems, which
involves a small number of samples with a large number of features, and feature selection
problems associated with this type of problems. Our strategy is to first try various analysis
methods, most of which were described in Chapter 3, on a well known benchmark dataset,
human acute leukemia microarray dataset (Golub et al., 1999), select one that has the best
performance, then try that method on other datasets of same type, including small, round blue
cell tumors (SRBCTs) (Khan et al., 2001) dataset and artificial datasets.
The human acute leukemia microarray dataset consists of 72 microarray experiments with
expression values of 7129 clones from 6817 human genes. Here the term clone refers to the
fragment of a gene. Each of the genes has a short description; and each clone is represented
by an accession number. Each microarray is assigned with a class label, either Acute Myeloid
Leukemia (AML) or Acute Lymphoblastic Leukemia (ALL), according to the organism used
for the hybridization. A second type of classification problem arises from this dataset. We
4-59
used the clone id as feature id. Each of the hybridizations corresponds to a classification
sample. These samples were divided into two sets by Golub et al.: The first sample set, which
consists of 27 ALL samples and 11 AML samples, is for training the classifier. The second
sample set, consisting of 20 ALL samples and 14 AML samples, is for testing the classifier.
Figure 4.1: The combination of feature selection / integration methods and classification
methods.
Due to the time constraint, we only tested a subset of all possible combinations of these
methods. In our experiment described in the following sections, the combination of
Likelihood and Recursive Feature Selection method achieved the best feature selection
4-60
performance. With the optimal feature sets selected using the above combination, Bayesian
classification method achieved the best classification performance. The combinations that
used are summarized in Table 4.1. Characteristic of some of the combinations are described.
The term homogenous in the table refer to the methods that use same kind of criterion in
single dimension and multi-dimension. Those combinations that are not tested correspond to
blank cells in the table.
Information gain
Neural network feature
Fisher Linear
selector
Discriminator
Homogeneous , stopping
criterion, heavy
computation
Likelihood
Tested
Fisher Criterion
Homogeneous
Multivariate Likelihood
Selection Method
Recursive Feature Elimination
Best selection performance /
Homogeneous
Extensively studied
Two gene sets can be
obtained, each distinguish
one class from the other Tested
Second best combination we
found
Extermal Margin
Baseline
Tested
Table 4.1: List of combined univariate and multivariate feature selection methods tested.
4.1.1 Principle Component Analysis for feature integration
Our first effort was to study the relevance of features to the class labels. The study involves
both feature integration and feature selection. For feature integration, we chose principle
component analysis to see how much the most informative component extracted from the
features can contribute to the classification. Experiment was done using statistics toolbox and
neural network toolbox in MATLAB v6.1. Computation of the principle components form
large number of features consumes large amount of computer memory. Due to the limitation
of our computer system, the program was unable to process all 7129 features. We generated
72 components from randomly chosen 4500 features, which is the maximum number of
4-61
features that can be processed by the program, and then used all samples with these
components to perform leave-one-out training and validation using a three-layer feed-forward
neural network with different number of hidden units on all samples. In the training process,
batch mode was used. In each leave-one-out iteration, a test of the classification performance
on the 71 samples was done after every 10 epochs. If the accuracy on training data is 100%,
the training is stopped, and the remaining one sample is tested.
Table 4.2 shows the
performance. The result shows that although principle component extracted most informative
information in terms of standard variance from the features, this information is hardly
relevant to the classification.
Hidden neuron number
600
600
50
50
Training method
RPROP backpropagation
One Step Secant Algorithm
One Step Secant Algorithm
One Step Secant Algorithm
Performance
51.39%
51.39%
43.06%
48.61%
Table 4.2: Performance of neural network using principle components as input.
4.1.2 C4.5 for feature selection
We applied C4.5 algorithm on all 72 samples of the leukemia dataset. The constructed
decision tree was surprisingly simple, which only involves two genes. It could correctly
classify 71 samples. The tree is as follows:
M84526_at > 290 : -1
M84526_at <= 290 :
| X54489_rna1_at <= 91 : 1
| X54489_rna1_at > 91 : -1
4-62
Using only 38 training samples, a even simpler tree involving only one feature is generated,
this tree correctly classified all 38 training samples and 31 out of 34 testing samples
(accuracy: 91.2%).
X95735_at <= 938 : 1
X95735_at > 938 : -1
The decision tree construction algorithm C4.5 also supports constructing trees in iterative
mode. In this mode, the algorithm randomly selects an initial sample subset from training set
to construct a decision tree, and then iteratively add the samples that are misclassified by the
tree into the subset and reconstruct the tree until there is no misclassification among all
training samples. We tested C4.5 with iterative mode for 20 trials using all 72 samples. All
trials generated a two-layer tree with three nodes. These trees are listed in Table 4.3. For
simplification, we just list the gene accession numbers of the first layer and the second layer
to represent the tree.
Feature of first layer
Feature of second layer
Occurrence
M84526_at
M83652_s_at
4
M84526_at
D86967_at
2
M84526_at
X54489_rna1_at
2
U46499_at
M98833_at
1
U46499_at
X86401_s_at
1
U46499_at
D89289_at
1
L09209_s_at
D80003_at
1
D88422_at
M31166_at
5
D88422_at
U23070_at
2
D88422_at
M83652_s_at
1
Table 4.3: Decision trees constructed by iterative mode from all 72 samples.
4-63
We then tried leave-one-out test to construct decision trees from all samples. A total of 72
trees were constructed, most of which have two layers, but some had only one layer. Table
4.4 summarize these trees. When applying leave-one-out to 38 training samples, all the trees
constructed had one layer. They are summarized in Table 4.5.
Feature of first layer
Feature of second layer
Occurance
M23197_at
M20902_at
1
M23197_at
D80003_at
1
M23197_at
2
M27891_at
M31166_at
2
M27891_at
M55418_at
1
M27891_at
2
M84526_at
X06948_at
1
M84526_at
Y07604_at
1
M84526_at
M81883_at
1
M84526_at
D86967_at
1
M84526_at
X54489_rna1_at
52
U46499_at
M60527_at
1
U46499_at
M98833_at
1
U46499_at
U36922_at
1
U46499_at
X86401_s_at
1
X95735_at
HG2160-HT2230_at
1
X95735_at
M83652_s_at
1
M31211_s_at
1
Table 4.4: Decision trees constructed by leave-one-out mode from all 72 samples
Feature of first layer
Occurance
X95735_at
35
Classification accuracy
on test samples
0.912
M27891_at
1
0.941
M31166_at
1
0.706
M55150_at
1
0.794
Table 4.5: Decision trees constructed by leave-one-out mode from 38 training samples with
prediction accuracy on 34 test samples.
4-64
Table 4.3, Table 4.4 and Table 4.5 show that certain features occur very frequently in these
three experiments. It appears that the selectivity of C4.5 algorithm is high for the dataset. In
Table 4.3 and Table 4.4, feature M84526_at has the highest occurrence frequency, which
implies that this feature is important in deciding the classes when all 72 samples are taken
into account. But when only taking the 38 training samples into account, the algorithm
selected a very different set of features, which is shown in Table 4.5. Only feature X95735_at
appeared two times in leave-one-out mode for all 72 samples (see Table 4.4). The fact that
the trees generated on all training and test samples are very simple and the trees generated on
the training samples are even much simpler make us expect that some feature selection
method should that can generate a very small feature set when presented with a very limited
number of training samples. It may also be possible to obtain high classification accuracy on
test samples by certain classifiers constructed using this small feature set.
Rank
Feature
Training
All samples samples
Test
samples
1
M84526_at
0.652
0.408
0.689
2
M27891_at
0.652
0.685
0.689
3
D88422_at
0.651
0.578
0.584
4
M23197_at
0.648
0.581
0.684
5
X95735_at
0.647
0.844
0.522
6
U46499_at
0.634
0.565
0.692
7
M31523_at
0.590
0.511
0.851
8
L09209_s_at 0.589
0.562
0.577
9
M83652_s_at 0.550
0.578
0.420
10
M11722_at
0.542
0.332
0.683
22
M31166_at
0.405
0.689
0.182
26
M55150_at
0.398
0.671
0.198
Table 4.6: Leukemia features and their information gain of all samples, training samples and
test samples, sorted by gain of all samples.
4-65
We continued to investigate the information gain computed for the features of first level. In
Table 4.6, top ten ranked features are listed according to the information gain of all samples.
The features that are listed in Table 4.5 whose rank are higher than ten are also listed in Table
4.6.
Information gain is a measure of relevance of a feature to classification. It appeared that the
features with high gain in training samples are likely to have high gain in all samples and test
samples. As mentioned before, C4.5 can construct a decision tree that consists of only one
feature from training samples using information gain. The tree is very simple but it can only
correctly classify 31 of 34 test samples. The generalization ability of the classifiers
constructed using only information gain measure for continuous features is not high, as was
discussed in Chapter 3, so we tried to use information gain measure as a feature selection
method, and test neural networks based on the selected features to see whether there is any
improvement in performance.
4.1.3 Neural networks with features selected using information gain
We tested the neural network with different configurations and different numbers of features,
as is listed in the following tables, with the highest information gain obtained from training
samples. The test results are listed in Table 4.7. Two kinds of training methods were used:
trainoss (One Step Secant Algorithm) and traingd (Gradient descent backpropagation). The
training was done in batch mode. In the training process, a test of the classification error of
training samples was performed repeatedly after a certain number of epochs indicated in test
interval column in the table until the error converge to no greater than training error
tolerance. The reason why we choose different tolerance is to check how well the trained
neural network could generalize, under a given over-fitting limit. If the number of tests
4-66
exceeded the number indicated in number of unsuccessful trials column, then the training was
considered to be unsuccessful and the training result was rejected. Once the training was
successfully terminated, we tested the classification accuracy of the trained network on test
samples. For every configuration corresponding to each row in the table, we collected 100
successful trainings and then calculates the mean and standard deviation of the classification
accuracy on the test samples.
Feature
number
Number of Training
hidden units method
Test
interval Number
of Training
error Test accuracy
(number of epochs) unsuccessful tolerance (number of
trials
samples)
50
60
trainoss
20
N/A
0
0.810±0.090
20
60
10 (from all
samples)
20
trainoss
20
N/A
0
0.855±0.080
trainoss
20
20
0
0.952±0.024
5
20
trainoss
20
20
0
0.880±0.075
100
40
trainoss
20
20
0
0.824±0.101
50
50
trainoss
200
20
0
0.843±0.089
50
50
trainoss
20
20
0
0.848±0.090
50
10
traingd
100
5
2
0.936±0.036
50
10
traingd
100
5
2
0.931±0.041
20
10
traingd
100
5
4
10
10
traingd
100
3
4
0.942±0.025
(Repeats
can
hardly converge)
20
5
traingd
100
2
4
0.947±0.021
Table 4.7: Prediction performance of neural network using features selected by information
gain.
Note that the experiment in the third row of Table 4.7 used top ten features from information
gain of all features. The aim of this test is to see how high the accuracy could be when
information from test samples is used. We found that when the training error tolerance of the
training sample was as large as 4, with top 20 features, the test accuracy could approximate
4-67
the test in the third row. If the training error tolerance is smaller than 4, the test performance
may be affected by over-fitting.
The test in Table 4.7 gave us some indications on how to optimize the parameters to obtain
better generalization ability. We continued to test whether the test errors were located on a
few test samples or evenly distributed with fine tuned parameters. The method was the same
in the experiments in Table 4.7. We collected 100 successful trainings and counted the
number of times the test samples that were wrongly predicted as well. In Table 4.8 the
configuration and prediction accuracy are listed, and the wrong prediction frequency is
shown in Figure 4.2.
As can be seen from Figure 4.2, it seemed that for individual experiments of certain
configurations, the wrong predication frequency was very high in certain test samples. In
addition, most of trained neural networks were likely to make wrong prediction on the class
of test samples 28 and 29.
Experiment ID Number of Training Test
interval Number
of Feature
hidden
method (number
of unsuccessful number
units
epochs)
trials
Training
error Mean precision
tolerance (number
of samples)
Standard
deviation
14
5
traingd
100
1
15
5
0.893
0.044
15
3
traingd
50
1
50
2
0.932
0.036
16
3
traingd
50
1
50
4
0.914
0.050
17
3
traingd
50
1
20
4
0.948
0.019
18
3
traingd
50
1
100
4
0.924
0.030
19
3
traingd
50
1
100
2
0.930
0.021
Table 4.8: Test of neural network using features selected by information gain to identify
incorrectly predicted test samples.
4-68
Distribution of wrong predictions
Number of times of wrong prediction
100
90
80
Exp 14
Exp 15
Exp 16
Exp 17
Exp 18
Exp 19
70
60
50
40
30
20
10
0
1 3 5 7
9 11 13 15 17 19 21 23 25 27 29 31 33
Test sample ID
Figure 4.2: Wrong prediction times.
4.1.4 AdaBoost
AdaBoost has the mechanism to change the sampling distribution to focus on the training
samples that have high validation errors, and then AdaBoost may fit the training samples well
and keep good generalization ability as well. Hence we expect that AdaBoost framework
might further improve the overall classification performance. In each AdaBoost experiment, a
number of neural networks of the identical size were consecutively trained for 50 epochs, and
those neural networks with training error no higher than error tolerance were employed for
refining sampling distribution. The training process continued until 50 such neural networks
4-69
were obtained. The final hypothesis of the test samples could then be obtained by combining
hypothesis of individual neural networks and factors β1KT , according to Eq. 3.3.2.1. One of
disadvantages of AdaBoost is that the training process is slow, so we only conducted the test
no more than twice for different configurations. The test results are listed in Table 4.9. The
configurations in tests 1 and 4 were tested once for their training time is extremely long, but
the remaining configurations were tested twice.
As can be seen from Table 4.9, the test performance of tests using the top 20 to 50 features is
generally better than the test results using the same number of features in Table 4.7 even
when the error tolerance was as low as 2 validation errors. In AdaBoost tests, there are
generally one or two errors (97.1% or 94.1% accuracy) in the test samples. The trained
networks are expected to fit the training samples well, and the combined hypothesis can also
achieve high prediction performance on the test samples.
Test ID
1
2
3
4
5
6
7
8
9
10
Hidden
number
unit
Feature number
Training error tolerance
Test
accuracy
5
20
3
0.971
5
10
3
0.853
5
10
3
0.794
5
15
4
0.912
5
50
4
0.941
5
50
4
0.971
5
50
2
0.971
5
50
2
0.941
3
50
2
0.971
3
50
2
0.941
Table 4.9: AdaBoost test results.
4-70
4.1.5 Neural network feature selector
By seeing the test result in the previous section, we continued to investigate whether it is
possible to further reduce the number of relevant genes without losing too much prediction
accuracy. Neural network feature selector, being a wrapper feature selection approach, could
exploit the relevant information between features and classes in the trained neural network
models. Another advantage of the neural network feature selector is that it can decide the
optimum number of selected features. Experiments were carried out using neural network
feature selector. The neural network feature selector algorithm was implemented using
MATLAB. In the experiments, a set of features, which had highest information gain from the
training samples, was used as initial feature set whose size was to be reduced. Because the
number of training samples is very small, we used leave-one-out approach to get average
training accuracy and validation accuracy when calling function train_validate() and
simulate_validate() . Each neural network was trained for maximum 200 epochs in the
function train_validate() . The rfactor was set to 10. After the selection process, with the
selected features, 100 repetitions of training and testing were done, and the mean and
standard deviation of training and testing accuracy ranks were calculated. The results are
summarized in Table 4.10 and Table 4.11. The performance of both summed square and
cross entropy error functions are listed in the two tables respectively. In these two tables, the
column (ε min , ε max ) contains the thresholds of penalty parameter ε . Because we increased or
decreased the penalty parameter by a factor of 1.1 , the values reflect the minimum and
maximum number of times that the penalty parameter may increase or decrease
accumulatively.
4-71
Experiment
1
2
3
4
5
Feature
set size
′′
r′min rmin
∆r ′′ (ε min , ε max )
200
0.9
0.9
0.05
50
0.9
0.9
0.01
50
0.95
0.95
0.01
50
0.97
0.97
0.01
200
0.95
0.9
0.03
1.1±30
1.1±20
1.1±20
1.1±20
1.1±20
Number of Accuracy
on Accuracy on test
features
training samples
samples
selected
6
1.00±0.00
0.67±0.01
3
1.00±0.00
0.79±0.00
4
1.00±0.00
0.88±0.03
4
1.00±0.00
0.88±0.03
6
0.98±0.01
0.71±0.00
Table 4.10: Experiment result of neural network feature selector with summed square error
function.
′
rmin
′′
rmin
∆r ′′ (ε min , ε max )
200
0.9
0.9
0.05
50
0.9
0.9
0.01
50
0.95
0.95
0.01
50
0.97
0.97
0.01
200
0.95
0.9
0.03
Experiment Feature
set size
1
2
3
4
5
1.1±30
1.1±20
1.1±20
1.1±20
1.1±20
Number of Accuracy
on Accuracy on
features
training samples test samples
selected out
4
1.00±0.00
0.71±0.02
4
1.00±0.00
0.72±0.03
6
1.00±0.00
0.68±0.04
36
1.00±0.00
0.80±0.08
191
1.00±0.01
0.79±0.09
Table 4.11: Experiment result of neural network feature selector with cross entropy error
function.
From Table 4.10, we can see that under certain settings the neural network feature selector
can select a small set of four out of 50 genes and the prediction accuracy on the test samples
could be as high as 88%. In comparison, the experiments in Table 4.11 show that the
selection performances are generally worse in terms of the number of the features selected
and the prediction accuracy. We noticed that back-propagation training of neural networks
was much faster when using cross entropy error function than when using summed error
function.
4-72
4.1.6 Hybrid Likelihood and Recursive Feature Elimination method
We tried the combination of likelihood and recursive feature elimination method. Very good
feature selection performance was found using this hybrid method on Leukemia dataset,
which encouraged us to continue to test the method systematically, comparing it with feature
selection methods using LIK and RFE alone. We computed acceptance rate, which is a
performance measure that is stricter than accuracy.
Suppose there are n samples with predicted values of o1 , K , o n , and their corresponding
class labels are y1 , K , y n . Each of the class labels takes value of either + 1 or − 1 , and the
values of outputs are real numbers. If the prediction output of a classifier for a sample has the
same sign as that of its true class, we consider this sample to be correctly classified. The
performance measure of accuracy, i.e. the number of correctly classified samples over the
total number of test samples, is defined as
accuracy =
{i
o i y i > 0, i = 1, K , n }
n
,
Eq. 4.1.6.1
where S denotes the cardinality of the set S . In contrast, the acceptance rate is computed as
follows
i oi yi > − min( o j y j ), i = 1,K, n
j =1,K, n
acceptance rate =
.
n
Eq. 4.1.6.2
The strength of correct prediction of a sample can be obtained by multiplying the output and
class label of a sample together oi yi , the larger the value of the product, the better the
prediction made by the classifier. When the value of the product is negative, the classifier
makes a wrong prediction of the sample. To calculate the acceptance rate, we first select the
4-73
worst prediction out of all test samples. The worst prediction corresponds to the sample
whose o j y j is minimum. This minimum values is multiplied by − 1 and is used as a
threshold. All the predictions that have the o j y j value bigger than this threshold are
considered as being accepted. The acceptance rate is 1 when all the test samples are correctly
predicted. Otherwise, it will not be greater than the accuracy, because the prediction of the
classifier on some test samples may have small confidence, as indicated by the output-class
label products that are lower than the threshold. These test samples are correctly predicted
and are counted in the computation of accuracy, but they will not be counted in the
computation of acceptance rate.
In the tables and figures in this section, we will denote accuracy and acceptance rate by acu
and acp, respectively. Obviously, the acceptance rate cannot be higher than accuracy.
Because the number of samples was small, we used the leave-one-out method for validating
the classifier on training samples as well as on all samples. When there are n samples, leaveone-out is a technique to iteratively choose each sample for testing, and the remaining
samples for training. A total of n classifiers were trained, and n predictions were made. The
accuracy and acceptance rates were computed from the predictions and labels of the
corresponding test samples.
We ran our experiments on a Pentium 4 1.4GHz computer with 512-megabyte memory. We
wrote and ran our program using MATLAB 6.1. The support vector machine was constructed
with
the
Support
Vector
Machine
Toolbox
from
http://theoval.sys.uea.ac.uk/~gcc/svm/toolbox which was developed by Gavin Cawley. For
the SVM, we set C = 100.0 and used the linear kernel, the same as those used by Guyon et al.
(2002).
4-74
4.1.6.1 Leukemia dataset
Figure 4.3 shows the sorted LIK scores. We chose equal numbers of genes with the highest
LIK ALL → AML
and LIK AML→ ALL score as the initial gene sets for RFE. We plotted in this figure the
scores of the top (2 x 80) genes. The top i th gene according to LIK ALL → AML always has a
higher LIK ALL → AML value than the corresponding top i th gene’s LIK AML→ ALL score. Genes with
low LIK scores are not expected to be good discriminators. We decided to pick the top genes
to check their discriminating ability. In particular, we ran experiments using the top (2 x 10),
(2 x 20), (2 x 30) genes. We found the best performance was obtained when 2 x 20 top
ranking genes were selected. The performance was measured by computing the prediction
accuracy and acceptance rate of SVM and Bayesian classifiers built using the selected genes
on the test samples.
Figure 4.4 shows the accuracy and acceptance rate using two different experimental settings:
leave-one-out and train-test split. For the leave-one-out (LOO) setting, we computed the
performance measures using only the 38 training samples as well as on the entire dataset
consisting of 72 samples. For the train-test split, the measures shown were computed on the
34 test samples, while the measures on the 38 training samples are not reported in the table. A
series of experiments were conducted to find the smallest number of genes that would give
good performance measure. The experiments started with all 40 (= 2 x 20) genes selected by
the LIK feature selection. One gene at a time was eliminated using RFE. RFE feature
selection was conducted until there was only one gene left. For a selected subset of genes, the
performance measures were computed under all experimental settings and using both SVM
and Bayesian classifiers.
4-75
1800
1600
1400
LIK score
1200
1000
800
600
400
200
0
20
40
60
Gene rank
80
100
120
Figure 4.3: Sorted LIK score of a subset of genes in the leukemia dataset. Dots indicate
LIK ALL→ AML scores and circles indicate LIK AML→ ALL . The top 28 genes according to their
LIK ALL→ AML
values have scores between 92014 and 1978.3; and the top 15 genes according to
their LIK ALL → AML values have scores between 22852 and 2148.7; they are not shown in this
figure.
As can be seen from the figure, the SVM classifier achieved almost perfect accuracy and
acceptance rate when there were three to 14 genes used to find the separating hyperplane. On
the other hand, when the Naïve Bayesian method was used for classification, almost perfect
performance was achieved with as many as 40 genes in the model. Elimination of the genes
by RFE one by one showed that the results could be maintained as long as there are at least
4-76
three genes in the model. This stability in performance indicates the robustness of the RFE
feature selection method when given a pre-selected small subset of relevant genes, as
identified by the LIK method. It is worth noting that the acceptance rate on the test samples
was almost constant with at least three genes, both when the SVM classifier and the Naïve
Bayesian classifier were used for prediction. We emphasize here that the hybrid LIK+RFE
feature selection was run using the 38 training samples; the classifiers were also built using
the same set of training samples without the use of any information from the data in the test
set.
A set of three genes was discovered to give perfect accuracy and acceptance rate regardless of
the experimental settings and the classifiers used. These genes are listed in Table 4.12. They
have also been identified as relevant genes in this dataset by several researchers. Golub et al.
(1999) identified U05259_rna1_at and M27891_at as relevant, while Keller et al. (2000)
identified the gene X03934_at as relevant. On the other hand, Guyon et al. (2002) identified a
completely different set consisting of four genes. Among these four genes, only M27891_at
occurs in the previous C4.5 result.
4-77
LIK+RFE classification performance using SVM
1.00
0.80
Training sample LOO acu
Performance
Training sample LOO acp
0.60
Test sample acu
Test sample acp
0.40
All sample LOO acu
All sample LOO acp
0.20
0.00
40 37 34 31 28 25 22 19 16 13 10 7
4
1
Number of genes
LIK+RFE classification performance using Bayesian method
1.00
0.80
Training sample LOO acu
Performance
Training sample LOO acp
0.60
Test sample acu
Test sample acp
0.40
All sample LOO acu
All sample LOO acp
0.20
0.00
40 37 34 31 28 25 22 19 16 13 10 7
4
1
Number of genes
Figure 4.4: Classification performance of genes selected using the hybrid LIK+RFE.
4-78
Gene accession number
Description
U05259_rna1_at
MB-1 gene
M27891_at
CST3 Cystatin C (amyloid angiopathy and cerebral hemorrhage)
X03934_at
GB DEF = T-cell antigen receptor gene T3-delta
Table 4.12: The smallest gene set found that achieves prefect classification performance.
ALL T-cell Train
ALL B-cell Train
AML Train
ALL T-cell Test
ALL B-cell Test
AML Test
4
U05259__rna1__at
3
2
1
0
-1
-2
-2
-1
0
0
2
1
4
2
6
3
X03934__at
M27891__at
Figure 4.5: Plot of the leukemia data samples according to the expression values of the three
genes selected by the hybrid LIK+RFE.
Since there were only three genes selected by the hybrid LIK+RFE method, we are able to
visualize the distribution of both the training and test samples in a three-dimensional space.
Figure 4.5 shows the plot of the samples. In this figure, we differentiate between acute
4-79
myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL) samples. There were
actually two different types of ALL samples. These were B-cells or T-cells as determined by
whether they arose from a B or a T cell lineage (Keller et al., 2000). From the figure, we can
see that all except one B-cell sample had almost constant expression values for two genes,
namely M27891_at and X02934_at. Training sample number 17 was the one ALL B-cell
that was an outlier. On the other hand, all T-cell samples had almost constant expression
values for genes U05259_mal_at and M27891_at, while all AML samples had similar
expression values for U05259_mal_at and X03934_at. The plot shows that the three selected
genes were also useful in differentiating ALL B-cell and T-cell samples.
For comparison purposes, the classification performance of SVM and Naï
ve Bayesian
classifiers built using genes selected according to their LIK scores only is shown in Figure
4.6. For the results shown in this figure, we started with the same set of 40 genes and
removed one gene at a time according to their LIK scores. As can be seen from the figure, the
results were not as good as those shown in Figure 4.4. In particular, using SVM classifiers,
the accuracy and the acceptance rate were more than 80 percent when there were still more
than 20 genes in the model. The acceptance rate drops drastically when there are fewer genes.
Naï
ve Bayesian classifiers performed well when there were more than 21 genes. Further
removal of more genes according to their LIK scores caused the acceptance rate to drop
considerably. When there were fewer than five genes, the accuracy and the acceptance rate of
the classifiers were low.
4-80
LIK classification performance using SVM
1.00
0.80
Training sample LOO acu
Performance
Training sample LOO acp
0.60
Test sample acu
Test sample acp
0.40
All sample LOO acu
All sample LOO acp
0.20
0.00
40 37 34 31 28 25 22 19 16 13 10 7
4
1
Number of genes
LIK classification performance using Bayesian method
1.00
0.80
Training sample LOO acu
Performance
Training sample LOO acp
0.60
Test sample acu
Test sample acp
0.40
All sample LOO acu
All sample LOO acp
0.20
0.00
40 37 34 31 28 25 22 19 16 13 10 7
4
1
Number of genes
Figure 4.6: Performance of SVM and Naï
ve Bayesian classifiers built using genes selected
according to LIK scores.
The performance of SVM and Naï
ve Bayesian classifiers built using the genes selected by
RFE is depicted in Figure 4.7. We started with all 7129 genes in the feature set. We built an
4-81
SVM using the training samples with expression values of all the genes and measured its
performance on the test samples. We also built a Naï
ve Bayesian classifier and measured its
performance as well. The gene that had the smallest absolute weight in the SVM-constructed
hyperplane was removed, and the process of training and testing was repeated with one fewer
gene. This process was continued until there were no more genes to be removed.
RFE classification performance
1.00
Performance
0.80
0.60
SVM acu
SVM acp
Bayesian acu
0.40
Bayesian acp
169
517
865
1213
1561
1909
2257
2605
2953
3301
3649
3997
4345
4693
5041
5389
5737
6085
6433
6781
0.00
7129
0.20
Number of genes
Figure 4.7: Classification performance of purely using RFE. Classification performance of
SVM and Naï
ve Bayesian classifiers using genes selected by RFE starting from 7129 genes
down to only one gene. The experimental setting was training test split and the performance
measures were shown on the 34 test samples.
4-82
An interesting point to note from the results depicted in Figure 4.7 is the sharp improvement
in the acceptance rate of the Bayesian classifiers when the number of genes was reduced from
2773 to 2772. The gene that was eliminated at this stage was M26602_at. The acceptance
rates stayed at 100 percent when there were 2772 to 1437 genes. Further removal of genes
caused the rate to deteriorate gradually. On the other hand, the performance of SVM was
more stable. With more than 519 genes, both the accuracy and the acceptance rate were at
least 90 percent.
We also experimented with choosing the top genes according to their LIK scores. We
selected genes with LIK scores that were higher than a certain threshold. The threshold
values tested were 1500, 2000, and 2500. Note that there were always more genes selected
because of their high LIK ALL→ AML than genes selected because of their LIK AML→ ALL values. The
best performance was obtained when the threshold was set to 1500. A total of 62 genes met
this threshold value and were used to form the initial gene set for RFE. After applying RFE,
we obtained a set of four genes that achieves perfect accuracy and acceptance rate on the
training and test samples under all three experimental settings. The set of four selected genes
is shown in Table 4.13. Two out of the four genes were the same ones as those selected using
the (2 x 20) top initial genes listed in Table 4.12. These genes were U0529_rnal_at and
M27891_at. The genes M16336_s_at was also found by Keller at al. (2000) to be an
important gene for classification.
The performance of the SVM classifiers with genes selected using just the RFE approach was
slightly different from that reported by Guyon et al. (2002). The reason for this could be the
variation in the implementation of the quadratic programming solvers. The Matlab toolbox
uses Sequential Minimal Optimisation algorithm (Platt, 1999), while Guyon et al. used a
4-83
variant of the soft-margin algorithm for SVM training (Cortes, 1995). Our hybrid LIK+RFE
method achieved better performance than other methods reported in the literature. To achieve
perfect performance, the RFE implementation of Guyon et al. needed eight genes. When the
number of genes was reduced to four, the leave-one-out results on the training samples using
SVM achieved only 97 percent accuracy and 97 percent acceptance rate. SVMs trained on 38
training samples with the four selected genes achieved only 91 percent accuracy and 82
percent acceptance rate on the test samples.
Gene assection number
Description
U05259_rna1_at
MB-1 gene
M16336_s_at
CD2 CD2 antigen (p50), sheep red blood cell receptor
M27891_at
CST3 Cystatin C (amyloid angiopathy and cerebral hemorrhage)
X58072_at
GATA3 GATA-binding protein 3
Table 4.13: The genes selected by the hybrid LIK+RFE method. The genes that have LIK
scores of at least 1500 were selected initially. RFE was then applied to select these four genes
that achieved perfect performance.
Using the genes selected according to their LIK scores and applying the Bayesian method,
Keller et al. (2000) achieved 100 percent prediction with more than 150 genes. Hellem and
Jonassen (2002) required 20 to 30 genes to obtain accurate prediction by ranking pair-wise
contribution of genes to the classification. The classification of the samples was obtained by
applying k-nearest neighbours, diagonal linear discriminant and Fisher’s linear discriminant
methods. Guyon et al. also mentioned the performance of other works on this dataset
(Mukherjee et al., 2000; Chapelle et al., 2000; Weston et al., 2001). None of these works
reported performance results that are as good as ours.
4-84
Besides LIK, we also tried to combine other univariate feature ranking methods with RFE.
They are baseline criterion proposed by Golub et al. (1999), Fisher’s criterion, and Extremal
margin (Guyon et al., 2002). In the way that is similar to that in LIK+RFE, we choose the top
40 genes ranked by these three methods, then let RFE do further feature set reduction. Figure
4.8 shows the accuracy on test samples when running RFE starting from the initial feature set
selected by the three univariate methods.
It can be seen from the Figure 4.8, the prediction accuracy of baseline and Fisher’s criteria on
test samples are similar. In the elimination process, the accuracy were kept above 80% for
SVM prediction and 90% for Bayesian prediction, as long as there were more than 5 genes
remaining in the set. The prediction accuracy dropped drastically when there were less than 5
genes remaining in the set. The RFE starting from genes selected by Extremal margin ranking
performed better than the other two ranking methods especially in Bayesian prediction
accuracy, which remained no less than 97% until there was one gene left in the gene set.
Particularly, the perfect prediction was achieved when there are 5 genes left in the dataset.
4-85
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5
10
15
20
25
30
Baseline
Fisher
Extremal margin
35
40
Accuracy
Bayesian prediction
Number of genes
SVM prediction
1
Accuracy
0.8
Baseline
Fisher
Extremal margin
0.6
0.4
0.2
5
10
15
20
25
30
35
40
0
Number of genes
Figure 4.8: SVM and Bayesian prediction accuracy on test samples using features selected by
RFE combined with baseline criterion, Fisher’s criterion and Extremal margin method.
4-86
4.1.6.2 Small, round blue cell tumors dataset
The second dataset to test LIK+RFE is from small, round blue cell tumors (SRBCTs) (Khan
et al., 2001). There are 88 samples altogether with 2308 genes, divided into 4 classes:
neuroblastoma (NB), rhabdomyosarcoma (RMS), Burkitt lymphomas (BL) and the Ewing
family of tumors (EWS). Because our focus is on binary classification, we decomposed the
problem into 4 one-against-rest binary second type classification problems. In (Khan et al.,
2001) there are 63 training samples, which consist of 23 EWS, 8 BL, 12 NB, and 20 RMS
samples. There are 25 test samples, which consist of 6 EWS, 3 BL, 6 NB, 5 RMS samples,
and non-SRBCTs samples. After testing the classification performance of individual
problems, we then performed test on the prediction combining all the classifiers together to
predict the samples that is not belong to these four classes.
We obtained the expression ratio data of Khan et al. (2001). Before we conducted our
experiments, the expression values were transformed by computing their logarithmic values.
Base 2 log transformation was used, as this is the usual practice employed by researchers
analyzing micraoarray data. In Figure 4.9, the plot of the gene ranking according to their LIK
scores is shown. The LIK scores were computed for differentiating EWS samples from nonEWS samples. The set of top 20 genes according to their LIK EWS → Non− EWS ranking contained
eight genes that were also in the set of top 20 genes according to their LIK Non− EWS → EWS ranking.
Hence, when RFE was applied to further eliminate genes from the feature set, it started with
32 unique genes.
4-87
90
80
70
60
LIK score
50
40
30
20
10
0
-10
0
100
200
300
Gene rank
400
500
600
Figure 4.9: Sorted LIK scores of genes in the SRBCT dataset. Dots indicate LIK EWS → Non− EWS
scores and circles indicate LIK Non− EWS → EWS scores.
For the other three classification problems, the plots would look very similar to Figure 4.9
and are not shown in this paper. For each of the four problems, LIK selected the top (2 x 20)
genes. The number of unique genes selected by LIK and the results of the experiments from
solving four binary classification problems are summarized in Table 4.14. The numbers of
unique genes selected by LIK and the smallest numbers of genes required to achieve near
perfect performance during the gene elimination process by RFE are shown in the second
column of the table. For the three classification problems to identify EWS, BL and NB, the
accuracy and the acceptance rates were at least 98 percent for all experimental settings. Those
perfect performance results are highlighted in the table. For the fourth classification problem
4-88
to differentiate between RMS and non-RMS samples, the accuracy rates were at least 92
percent. However, the acceptance rate on the test samples dropped to eight percent for SVM
classifier and 16 percent for Naï
ve Bayesian classifier, respectively.
SVM
Leave-one-out
on
training
samples
Prediction
on
test
samples
Leave-oneout on all
samples
Bayesian
Leave-oneout
on
training
samples
Initial/final
number of
genes
32/5
Acu
1.00
Acp
1.00
Acu
1.00
Acp
1.00
Acu
0.99
Acp
0.99
Acu
1.00
Acp
1.00
Acu
1.00
Acp
1.00
Acu
1.00
Acp
1.00
BL vs non-BL
37/3
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
NB vs non-NB
34/3
1.00
1.00
1.00
1.00
1.00
1.00
0.98
0.98
1.00
1.00
1.00
1.00
RMS vs nonRMS
34/4
1.00
1.00
0.92
0.08
0.99
0.88
1.00
1.00
0.92
0.16
0.97
0.35
Classification
problem
EWS vs nonEWS
Prediction
on
test
samples
Leave-one-out
on all samples
Table 4.14: Experimental results for the SRBCT dataset using the hybrid LIK+RFE.
The poor acceptance rate obtained when predicting RMS test samples suggests that the
differences in the output of the classifiers and the actual target values were high for the
incorrectly predicted samples. In order to verify the predictions, we plotted the distribution of
the samples according to the expression values of three genes, ImageID784224 (fibroblast
growth factor receptor 4), ImageID796258 (sarcoglycan, alpha), and ImageID1409509
(troponin T1). The three were selected because their corresponding SVM weights were the
largest. The plot is shown in Figure 4.10. We can clearly see that the two incorrectly
classified non-RMS samples were outliers with large values for ImageID1409509 (troponin
T1). These two outliers were Sk. Muscle samples TEST-9 and TEST-13, which were
misclassified as RMS samples. It should be noted that there were no Sk. Muscle samples in
the training dataset.
4-89
RMS Train
Non-RMS Train
RMS Test
Non-RMS Test
3
2.5
2
troponin T1
1.5
1
0.5
0
-0.5
-1
-1.5
-2
-3
3
-2
2
-1
1
0
0
1
-1
2
-2
3
-3
sarcoglycan, alpha
fibroblast growth factor receptor 4
Figure 4.10: Plot of RMS and non-RMS samples. Plot of all 88 RMS and non-RMS samples
according to the expression values of three of the four selected genes.
The genes selected by the hybrid LIK+RFE for each of the four classification problems are
listed in Table 4.15. For the problem of differentiating EWS from non-EWS samples, our
method selected five genes, all of which were also selected by Khan et al. (2001). On the
other hand, to differentiate between NB and non-NB samples, only three genes were needed
and none was selected by Khan et al. All together, the hybrid LIK+RFE identified 15
important genes. This number compares favorably with the total of 96 genes selected by the
PCA (Principle Component Analysis) approach of Khan et al.
4-90
Classification
problem
EWS vs non-EWS
Reported
by Khan et
al. (2001)
Image ID
Description
Y
377461
caveolin 1, caveolae protein, 22kD
Y
295985
ESTs
Y
80338
selenium binding protein 1
Y
52076
olfactomedinrelated ER localized protein
Y
814260
follicular lymphoma variant translocation 1
Y
204545
ESTs
897164
catenin (cadherin-associated protein), alpha 1 (102kD)
241412
E74-like factor 1 (ets domain transcription factor)
45632
glycogen synthase 1 (muscle)
768246
glucose-6-phosphate dehydrogenase
810057
cold shock domain protein A
897177
phosphoglycerate mutase 1 (brain)
Y
784224
fibroblast growth factor receptor 4
Y
796258
sarcoglycan, alpha (50kD dystrophin-associated glycoprotein)
Y
1409509
troponin T1, skeletal, slow
BL vs non-BL
Y
NB vs non-NB
RMS
non_RMS
vs
Table 4.15: The genes selected by the hybrid LIK+RFE for the four binary classification
problems.
We also tested the classification performance of SVM and Naï
ve Bayesian classifiers on
genes selected based purely on their LIK scores. For comparison purpose, for each of the four
problems, the number of genes was set to be the same as the corresponding final number
selected by the hybrid LIK+RFE shown in Table 4.14. Table 4.16 summarizes the results. For
three of the four classification problems, the performance of the classifiers was not as good as
the results reported in Table 4.14. The accuracy and acceptance rates dropped to as low as 52
percent. The most unexpected results came from the fourth problem to differentiate between
RMS and non-RMS samples. The SVM classifier achieved perfect accuracy and acceptance
rate using four genes, while the Naï
ve Bayesian classifier managed to obtain at least 92
percent accuracy and acceptance rate. The four genes were ImageID461425 (MLY4),
ImageID784224 (fibroblast growth factor receptor 4), ImageID296448 (insulin-like growth
4-91
factor 2), and ImageID207274 (Human DNA for insulin-like growth factor II). All these
genes were among the 96 genes identified by Khan et al. (2001). Of these four, only one was
selected by LIK+RFE, that is, ImageID784224.
Classification
problem
EWS vs nonEWS
BL vs nonBL
NB vs nonNB
RMS vs nonRMS
Number
of
genes
SVM
Leave-one-out
on
training
samples
Prediction on
test samples
Leave-oneout on all
samples
Bayesian
Leave-one-out
on
training
samples
Prediction on
test samples
Leave-one-out
on all samples
Acu
1.00
Acp
1.00
Acu
0.92
Acp
0.88
Acu
0.95
Acp
0.88
Acu
0.98
Acp
0.97
Acu
0.84
Acp
0.84
Acu
0.95
Acp
0.86
0.95
0.92
0.88
0.88
0.97
0.83
0.98
0.98
0.88
0.76
0.93
0.88
0.95
0.92
0.84
0.76
0.97
0.86
0.97
0.97
0.80
0.52
0.95
0.92
1.00
1.00
1.00
1.00
1.00
1.00
0.97
0.95
0.92
0.92
0.97
0.95
5
3
3
4
Table 4.16: The performance of SVM and Naï
ve Bayesian classifiers built using the top
genes selected according to their LIK scores.
In comparison, Khan et al. (2001) used neural networks for multiple classifications to achieve
93 percent EWS, 96 percent RMS, 100 percent BL and 100 percent NB diagnostic
classification performance on the 88 training and test samples. Since there were four classes
of training data samples, each neural network had four output units. The target outputs were
binary encoded, for example, for an EWS sample the target was (EWS=1, RMS=NB=BL=0).
A total of 3750 neural networks calibrated with 96 genes were required. The highest average
output value from all neural networks determined the predicted class of a new sample. The
Euclidean distance between the average values and the target values was computed for all
samples in order to derive the probability distribution of the distances. A test sample would
be diagnosed as a member of one of the four classes based on the highest average value given
by the neural networks. This was provided that the distance value falls within the 95th
4-92
percentile of the corresponding distance probability of the predicted class. Otherwise, the
diagnosis would be rejected and the sample would be classified as a non-SRBCT sample. Of
the 88 samples in the training and test datasets, eight were rejected. Five of these were nonSRBCT samples in the test set, while the other three actually belonged to the correct class but
their distances lay outside the threshold of the 95th percentile.
In order to visualize the distribution of the samples based on the expression values of the
selected genes, we performed clustering of the genes using the EPCLUST program
(http://ep.ebi.ac.uk/EP/EPCLUST). The default setting of the program was adopted; the
average linkage clustering and uncentered correlation distance measure were used. Figure
4.11 shows the clusters. It can be seen clearly from this figure that there existed four distinct
clusters corresponding to the four classes in the data. Most of the samples of a class fell into
their own corresponding clusters. The five non-SRBCT samples lay between clusters. We
conjecture that samples between clusters might not belong to any classes found in the training
dataset. Two between-cluster samples, RMS-T7 and TEST-20 were exceptions. RMS-T7,
which was nearer to the two Sk. Muscle samples TEST-9 and TEST-13 was actually an RMS
sample. TEST-20, which was nearer to Prostate sample TEST-11 than to EWS cluster was
actually an EWS sample. These exceptions were consistent with the neural network
prediction results of Khan et al. (2001) as the neural networks predicted TEST-9 and TEST13 to be RMS class, and they predicted TEST-20 and TEST-11 to be EWS class. Both
predictions, however, did not meet the 95th percentile distance criterion and were therefore
rejected. This indicated that these samples were also difficult to differentiate by the neural
networks. Different results from our clustering and the neural network classification can be
seen for test sample TEST-3, a non-SRBCT sample. The clustering placed TEST-3 between
4-93
BL and NB clusters. But the neural networks predicted this sample as an RMS sample
without meeting the 95 th percentile distance criterion.
Most of the genes selected from Leukemia and SRBCT datasets by our hybrid LIK + RFE
method have some relevance to cancer according to literature search in PubMed, a document
retrieval service of the National Library of Medicine of United States. However, biological
experiments need to be done for further validation of the role of these genes. The
performance of the method is also data dependent, as demonstrated in the significant
difference in the acceptance rate of the classifiers for the first three binary classification
problems and the fourth problem in the SRBCT dataset. Overall, we observe that the
classification performance on the test set generally does not change much with the
consecutive elimination of a few genes. The removal of one gene would not normally cause a
drastic change in the performance of the classifier. Significant drops in accuracy and/or the
acceptance rate are observed most frequently when a gene is removed from the optimal set.
4-94
Figure 4.11: Hierarchical clustering of SRBCT samples with selected 15 genes.
4-95
4.1.6.3 Artificial datasets
In order to study how expression values of irrelevant genes affect the selection performance
on RFE, we generated three types of artificial datasets. Each of the datasets consists of 40
training samples (20 positive class + 20 negative class) and 40 test samples (also 20 positive
class + 20 negative class). All datasets consist of features that are relevant and irrelevant to
the classification. The values of the irrelevant features are sampled from a normal distribution
with standard deviation 1 and mean 0 regardless of the class labels of the samples. The three
types of datasets are different in the way the relevant features are constructed. For the first
type, the relevant features contribute independently to classification. Their values are
normally distributed with standard deviation 1, and mean x or − x according to the class
labels. For different relevant features, x is randomly chosen from a uniform distribution on
the interval (0,1) . If x is big enough, univariate feature selection method is likely to be able
to select the relevant features. The relevant features of the second type dataset is constructed
by considering the joint effect of the features: suppose there are k relevant features
X 1 ,K , X k , where X 1 , K X k −1 are sampled from normal distribution in the same way as that
k −1
from irrelevant features. We set the remaining relevant feature X k = ∑ X i ± α , where α is
i =1
randomly sampled from a normal distribution of standard deviation 0 and mean 1, and α is
added or subtracted from the sum depending on the class labels of samples. It is expected to
be harder for a univariate method to select the first k − 1 relevant features. The third type of
datasets is constructed by setting the first k − 1 relevant features the same way as that in first
type datasets. But the k th feature is set in the same way as that in the second type datasets.
This type of datasets models the expression of the genes that related to cancer better, where
the genes have certain degrees of individual contributions to cancer from observation, but
they interact with other genes together to cause the disease.
4-96
We tried LIK, RFE and LIK+RFE on the datasets of all the three types above, each contained
a total of 100 features, 4 of which are relevant features. In each experiment, we let the
selection method select the features based on the training samples, then test the classification
performance on the testing samples. For testing RFE, one feature is eliminated each time. For
testing LIK, we include the same number of top ranking features of both LIK ag→b and
LIKbg→a ranks. For testing the combination of LIK+RFE, we choose the top 5+5 and the top
10+10 LIK ranked features to input into RFE. We use SVM to test the classification
performance. The setting of SVM for RFE and classification is same as those in Section
4.1.6.1. Before applying the methods, the datasets were normalized in the same way the
Leukemia dataset. For each setting, 100 experiments were conducted, each on a different
dataset generated, and the mean and standard deviation of the test result are shown in Table
4.17. The table shows the type of the datasets; the number of LIK ranked features to be used
as input for RFE (lik_num), within which the number of distinct ones (num_lik_rfe); the
accuracy of SVM prediction on datasets using all and relevant features (acc_all, acc_rev
respectively), where acc_rev, is used to find out what is the ideal prediction performance; the
corresponding number of support vectors of trained SVMs (num_sv_all and num_sv_rev
respectively); for running LIK, RFE, and LIK+RFE, the highest accuracy obtained
(max_rfe_acc, max_lik_acc, and max_lik_rfe_acc respectively), the smallest number of
features to achieve highest accuracy (num_rfe_acc, num_lik_acc, and num_lik_rfe_acc
respectively); the number of relevant features existing in selected features (num_rfe_acc_rev,
num_lik_acc_rev, and num_lik_rfe_acc_rev respectively); and the percentage of times of the
last
relevant
feature
existing
in
these
features
(contain_rfe_acc_last_rev,
contain_lik_acc_last_rev, and contain_lik_rfe_acc_last_rev respectively).
4-97
lik_num
10
dataset_type
1
2
3
20
1
2
3
num_lik_rfe
8.57 ± 0.97
9.01 ± 1.00
8.40 ± 0.83
14.66 ± 1.34
14.97 ± 1.47
14.86 ± 1.20
num_lik_sel
2.71 ± 0.87
1.05 ± 0.59
2.91 ± 0.71
2.79 ± 0.95
1.24 ± 0.62
3.22 ± 0.66
acc_all
0.66 ± 0.09
0.55 ± 0.08
0.71 ± 0.09
0.65 ± 0.11
0.55 ± 0.08
0.71 ± 0.08
num_sv_all
35.25 ± 1.56
36.07 ± 1.61
34.39 ± 1.61
35.14 ± 1.92
36.16 ± 1.64
34.13 ± 1.94
acc_rev
0.83 ± 0.08
0.81 ± 0.07
0.88 ± 0.06
0.82 ± 0.09
0.81 ± 0.06
0.89 ± 0.06
num_sv_rev
13.15 ± 6.45
14.87 ± 4.47
8.76 ± 4.21
12.79 ± 6.66
15.08 ± 4.68
8.29 ± 3.75
max_rfe_acc
0.82 ± 0.08
0.69 ± 0.07
0.89 ± 0.07
0.82 ± 0.09
0.69 ± 0.07
0.89 ± 0.07
num_rfe_acc
9.79 ± 17.15
13.38 ± 17.62
3.51 ± 7.06
7.20 ± 11.85
11.75 ± 17.55
3.98 ± 8.70
num_rfe_acc_rev
2.19 ± 1.01
1.39 ± 0.75
1.29 ± 0.62
1.90 ± 0.89
1.39 ± 0.82
1.29 ± 0.61
contain_rfe_acc_last_rev
0.53 ± 0.50
0.90 ± 0.30
0.97 ± 0.17
0.50 ± 0.50
0.90 ± 0.30
0.94 ± 0.24
max_lik_acc
0.84 ± 0.08
0.70 ± 0.07
0.91 ± 0.06
0.84 ± 0.08
0.70 ± 0.07
0.90 ± 0.06
num_lik_acc
10.20 ± 17.07
24.26 ± 27.35
5.20 ± 11.70
6.97 ± 11.14
18.02 ± 20.83
3.76 ± 5.94
num_lik_acc_rev
2.45 ± 1.00
1.66 ± 1.06
1.81 ± 0.95
2.08 ± 0.97
1.40 ± 0.90
1.81 ± 0.97
contain_lik_acc_last_rev
0.63 ± 0.49
0.94 ± 0.24
1.00 ± 0.00
0.52 ± 0.50
0.92 ± 0.27
1.00 ± 0.00
max_lik_rfe_acc
0.81 ± 0.09
0.66 ± 0.09
0.89 ± 0.07
0.82 ± 0.09
0.66 ± 0.09
0.88 ± 0.07
num_lik_rfe_acc
3.17 ± 2.06
3.14 ± 2.56
1.83 ± 1.48
3.92 ± 3.20
4.24 ± 3.94
2.41 ± 3.15
num_lik_rfe_acc_rev
1.96 ± 0.97
0.84 ± 0.55
1.25 ± 0.58
1.91 ± 0.91
0.85 ± 0.61
1.38 ± 0.76
contain_lik_rfe_acc_last_rev
0.51 ± 0.50
0.72 ± 0.45
0.96 ± 0.20
0.41 ± 0.49
0.70 ± 0.46
0.92 ± 0.27
Table 4.17: Test result of LIK, RFE and LIK+RFE on artificial datasets
It can be seen from Table 4.17, as expected, in terms of the number of feature selected and
the accuracy (rows max_*_acc and num_*_acc), for datasets of type 1 and 3, RFE is similar
to LIK; for datasets of type 2, RFE is significantly better than LIK. But the testing results
from all three types of datasets indicate that LIK+RFE selected significantly more accurate
feature sets, in terms of the number of relevant features over the number of selected features,
without significant loss of prediction performance (rows max_*_acc, num_*_acc and
num_*_acc_rev).
It can also be seen from the table that when the number of irrelevant features is large
compared with the number of relevant features, the classification of SVM becomes bad;
4-98
almost all training samples became support vectors (row num_sv_all compared with row
num_sv_rev). This phenomenon also occurs when SVM is applied on the second type
datasets like Leukemia and SRBCT dataset. From our experience, the number of support
vectors reduces significantly only when the number of features for training is near or less
than the number of training samples. We suspect the phenomenon is related to the learning
capacity of SVM but we have not found theoretical basis for this.
The assumption of single or double normal distribution on irrelevant and relevant features are
simple and may not accurately reflect the real situation in microarray datasets, especially, the
combinatorial effect of features are not modeled. However, as can be seen from the test result
from these simple datasets, it is quite likely that the weights of trained SVM on a mixture of
many irrelevant features and few relevant features are unable to truly measure the
contribution of the features to the classification. Some relevant features are incorrectly
eliminated by RFE starting from all features. By comparison, LIK ranking is able to keep
most of the relevant features for type 1 and 3 datasets, which enables RFE to perform further
selection more effectively.
4.1.7 The combination of Likelihood method and Fisher’s method
We also compared LIK+RFE with the combination of univariate and multivariate versions of
Likelihood method and Fisher’s method. In our experiment, a set of features was first
selected by a univariate method, in the same fashion as in the experiment for LIK+RFE. A
multivariate method was then used to eliminate the features recursively from that feature set.
We tested the method on the Leukemia dataset and set the size of the initial set to 30 as
Fisher’s linear discriminant encounters matrix inversion problems if the initial gene set size is
4-99
bigger than the number of training samples. The algorithm was implemented and run on
Matlab 6.1.
Figure 4.12 shows that when using the combination of F_F, L_L and F_L, the accuracy of
Bayesian classification on test samples was generally better than that of SVM prediction,
based on same set of genes. But the SVM prediction of L_F is better that that of Bayesian
method. In both SVM and Bayesian test results, L_L outperformed the other three
combinations when there are more than 15 genes remaining in the gene set. However, its
performance dropped drastically when there are less than four genes remaining in the gene
set. Under this situation, the combination of F_F is the best. However, none of these
combinations of these four methods outperformed the combination of LIK+RFE on our test
on Leukemia dataset in terms of classification accuracy based on the same number of genes.
4-100
SVM prediction
1
0.9
0.8
Accuracy
0.7
L_F
0.6
F_F
0.5
L_L
F_L
0.4
0.3
0.2
0.1
3
6
9
12
15
18
21
24
27
30
0
Number of genes
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
3
6
9
12
15
18
21
24
L_F
F_F
L_L
F_L
27
30
Accuracy
Bayesian prediction
Number of genes
Figure 4.12: SVM and Bayesian prediction accuracy when running combination of univariate
and multivariate feature selection methods. L_F: Likelihood + Fisher’s linear discriminator;
F_F: Fisher’s criterion + Fisher’s linear discriminator; L_L: Likelihood + Multivariate
Likelihood Method; F_L: Fisher’s criterion + Multivariate Likelihood Method.
4-101
4.2 First type - low dimension problems
The dataset we used that consists of the low dimension analysis problem is a Zebra fish
developmental microarray dataset obtained from Lab of Functional Genomics of Institute of
Molecular and Cell Biology, Singapore (Lo et al., 2003). In recent years, Zebra fish has been
adopted as a model system for the studies of vertebrate development owing to some of its
unique characteristics favorable for genetic studies compared to other vertebrate systems.
These characteristics include reasonably short lifetime, large number of progenies, external
fertilization and embryonic development, and translucent embryos (Talbot and Hopkins,
2000). In the microarray experiment, there were altogether 11,552 Expression Sequence Tag
(EST) clones representing 3100 genes printed onto the microarray glass slides. According to
BLAST (Basic Local Alignment Search Tool) search, 4519 of the 11,552 clones have
matches to 728 distinct publicly deposited protein sequences. That is, the functions of these
4519 clones are known, and the functions of the remaining clones are unknown. The relative
expression of the 11,552 clones in Zebra fishes’six developmental stages, including cleavage
(E2), gastrula (E3), blastula (E4), segmentation (E5), pharyngula (E6) and hatching (E7),
based on their developmental morphology, was monitored using microarray experiments, in
comparison to the expression of these clones in stage unfertilized eggs (E0). A first type
classification problem was constructed, which included 11,449 samples from the 11,552
clones with 6 features. A total 3887 of the 11,449 samples corresponding to the known clones
were labeled according to whether they are muscle genes or not. Within these 3887 clones,
248 were clones from 17 muscle genes. We did the classification by employing SVM. The
labeled clones were randomly split into two sets, 2500 for training and remaining 1387 for
testing. There were 157 and 91 positive samples in the training and testing sets respectively.
The remaining 7562 unlabelled samples were then used to perform prediction.
4-102
log 10 δ
-3.0
10
Number
Vectors
312
34
False
positive
9
-2.5
10
291
42
9
1287
49
100
-2.0
10
271
44
10
1286
47
110
-1.5
10
263
41
10
1286
50
109
-1.0
10
271
42
10
1286
49
104
-0.5
10
379
42
14
1282
49
113
0.0
10
631
31
19
1277
60
131
-3.0
20
300
41
9
1287
50
101
-2.5
20
282
42
9
1287
49
107
-2.0
20
267
42
10
1286
49
112
-1.5
20
251
40
10
1286
51
111
-1.0
20
270
43
12
1284
48
101
-0.5
20
362
39
14
1282
52
128
0.0
20
611
33
26
1270
58
176
-3.0
50
289
42
10
1286
49
102
-2.5
50
275
43
10
1286
48
108
-2.0
50
260
43
10
1286
48
112
-1.5
50
248
41
9
1287
50
103
-1.0
50
257
42
12
1284
49
110
-0.5
50
322
36
18
1278
55
159
0.0
50
567
32
29
1267
59
238
C
of
Support
True positive
True
negative
1287
False
negative
57
Predicted
positive
88
Table 4.18: Test of SVM with RBF kernel using different parameters.
Table 4.18 shows the test result of the SVM using different parameters with radial basis
function kernel. It can be seen that best test performance was obtained when setting C = 10
and σ = 0.01 . Using this parameter, the number of correctly predicted positive test samples
(true positive) reached 44, which was the highest among all configurations we tried; the
number of incorrectly predicted negative (false negative) samples was as low as 10,
compared favorably with the lowest false negative we obtained, which is 9. The trained
model under this setting is also not complex, which can be seen from the number of support
vectors was as low as 271. We provided a list of the 110 positively predicted unknown clones
to the biological researchers in the Lab of Functional Genomics of Institute of Molecular and
Cell Biology. Ten clones were selected from these 110 clones to do further biological
4-103
validation (re-sequencing of these clones) was done. Eight of these ten clones are proven to
be from muscle genes. The remaining two were found to be known genes after repeated
sequencing. Although these two are not muscle genes, they are functionally related to muscle
genes, therefore showed similar expression patterns to that of the other eight clones. Besides
these ten clones, another positively predicted clone was re-sequenced, which is a putative
novel gene. In situ hybridization on this clone showed that the corresponding gene truly had
muscle function (Lo et al., 2003).
5-104
5 Conclusion and future work
5.1 Biological knowledge discovery process
Biological research can be treated as a knowledge discovery process, which has been greatly
facilitated by the emergence of the field of Bioinformatics. Take microarray data as an
example, in the discovery process, biological information is extracted out by biological
studies, stored in DNA sequence trace files, scanned microarray images, descriptions of
samples and descriptions of experimental conditions. The information can then be quantified
or symbolized for biological data with certain structure. The biological data include
sequential representations of nucleotides / proteins, gene expression matrices. Computational
and statistical methods are then applied to extract biological knowledge from the data. The
knowledge is in essence relationships, which could be relationship between genes from their
sequential and expression similarity; relationship between gene expression and sample
property, experimental condition, intermediate product, or cellular process. Machine learning
plays an important role in discovering these relationships. However, the amount of
knowledge that can be discovered depends on two factors: the amount and quality of the
biological data available, and the suitability of the machine learning methods for various
biological problems.
There are more and more researchers who work to accelerate the discovery process. There are
currently two main journals and several main conferences in the bioinformatics area: Journal
of Bioinformatics, Journal of Computational Biology, Pacific Symposium on Biocomputing
(PSB), International Conference on Intelligent Systems for Molecular Biology (ISMB), and
Annual International Conference on Research in Computational Molecular Biology
5-105
(RECOMB). There is also a conference specifically focus on microarray data analysis, which
is Critical Assessment of Microarray Data Analysis (CAMDA).
Although much effort have been made for multidisciplinary collaboration in the discovery
process, researchers from non-biology discipline still need to gain more insight of the nature
of the biological problems, together with biologists. The real good design of machine
learning methods lie in full incorporation of biological knowledge rather than simply abstract
the biological problem to fit to well developed models. This criterion dictates the future
direction of applying machine learning methods for biological problems.
5.2 Contribution, limitation and future work
This thesis focuses on the classification and feature selection problems for gene expression
analysis. In the research work, we have reviewed current work in the literature and identified
the classification problems. We then applied seven feature selection methods, feature
extraction methods and some of their combinations for gene selection, and employed five
classification methods for prediction of cancer tissue type and gene function. We improved
neural network feature selector to make it more suitable to the gene selection problem for the
datasets having high dimension but few samples. We also developed a multivariate version of
likelihood feature selection method. We found the hybridization of Likelihood and Recursive
Feature Elimination achieved significantly better gene selection performance on the
benchmark Leukemia dataset than other methods. The hybridization of LIK+RFE on SRBCT
dataset also significantly outperformed a neural network method proposed by other
researchers.
5-106
The thesis shows a process of understanding of the nature of the problem and choosing
suitable methods. We first tested whether the most informative components may contribute to
the classification by applying principle component analysis and neural networks on Leukemia
dataset. Result showed that those components had little discrimination ability. We then
applied decision tree to find sets of rules that were able to classify all the samples. The
simplicity of the rules implied the possibility to find small set of genes that have high
classification ability. Due to the discrete nature of C4.5 algorithm, the small decision tree
generated from training samples, with continuous expression values, did not have good
generalization ability; subsequently the prediction accuracy of the tree on test samples was
low.
The possibly simple underlying classification model and the deficiency of decision tree
method inspired us to use information gain as a gene ranking method, but use neural network
as classifier. The classification was improved. After studying of the distribution of neural
network prediction errors, we found that, by employing AdaBoost technique to summarize
ensemble of neural network classifiers, the classification performance could be improved. We
then moved on into looking for methods that could reduce the number of genes used for
classification without significant loss of prediction performance. By properly tuning the
parameters, the combination of information gain and neural network feature selection
achieved that goal.
We then tested other combinations of univariate selection method and multivariate selection
method, including the methods that are based on extremal margin, likelihood, and Fisher’s
criterion. Within these combinations the hybrid of Likelihood method and Recursive Feature
Elimination method (LIK+RFE) selected the most compact gene set that had prefect
5-107
prediction performance. We did systematic test on this hybrid method using other datasets of
the high dimension classification problem. Test results were very promising. Applying the
classification methods on the Zebra fish dataset with large number of samples was
straightforward. But instead of purely validating the classification performance on known
genes, the real prediction of the function of some unknown genes were confirmed by
biological experiments.
Our experiments of the hybrid of LIK+RFE on SRBCT dataset showed that the feature
selection and classification methods for gene expression analysis are data dependent. Our
experiments also showed that, for microarray datasets of high dimension classification
problem, the choice of feature selection methods are more important than the choice of
classification methods. It is possible to design better selection methods or better combinations
of selection methods for Leukemia dataset; and although some method used in this thesis did
not achieve high selection performance on Leukemia dataset, they may do well on other
datasets.
The study on linear separability (Cover, 1965) suggests that when the number of samples is
small compared with the number of features, it is possible to find a number of subsets of
features that can perfectly distinguish all samples. Our experiments on the leukemia dataset
also support this hypothesis: we found two different gene sets consisting of just three or four
genes, which can achieve perfect classification performance. Biological study shows that
although many genes do not have direct relevance to the cancer under study, their expression
may have subtle and systematic difference in different classes of tissues (Alon et al., 1999).
Hence, a new challenge for cancer classification arises: to find as many as possible small
subsets of genes that can achieve high classification performance. Using only microarray data
5-108
with these subsets of genes, we can build different classifiers and look for those that have
desirable properties such as extremal margin, i.e. wide difference between the smallest output
of the positive class samples and the largest output of the negative class samples. Another
property could be median margin, which is the difference between the median output of the
positive class samples and the median output of the negative class samples. Exhaustively
enumerating and evaluating all the gene combinations is computationally NP-hard (nondeterministic Polynomial-time hard) and is feasible only when the number of relevant genes
is relatively very small.
Due to its cost, microarray experiments conducted for identifying the genes that are crucial
for cancer diagnosis are still scarce. The measurements obtained from the experiments are
noisy. These facts make the selection of different sets of relevant genes vital. Moreover,
cancer is a complex disease. It is not caused by only a few genes, but also by many other
factors (Kiberstis and Roberts, 2002). So even the best selected subsets may not actually be
the most crucial ones to the cancer under study. They can, however, be important candidates
for a further focused study on the gene interactions within individual subsets, and the
relationship between these interactions and the disease. There has been work done on the
second order selection. For example, Goyun et al. (2002) found a gene pair that could have
zero leave-one-out error on the training samples, but achieved poor performance on test
samples. Hellem and Jonassen (2002) also evaluated the contribution of pairs of genes to the
classification for the ranking of genes, but they still have to combine multiple pairs of genes
to perform classification. We plan to work on finding better ways to develop methods for
high order feature selection that would allow the classifiers to achieve high performance with
different small sets of genes.
109
Reference
•
Akutsu, T., Miyano, S., and Kuhara, S., (1999). Identification of genetic networks
from a small number of gene expression patterns under the Boolean network model.
Pacific Symposium on Biocomputing, 4, 17-28.
•
Aleksander, I., and Morton, H., (1990). An Introduction to Neural Computing.
Chapman and Hall, London.
•
Alon, U., Barkai, N., Notterman, D. A., Gish, K., Ybarra, S., Mack, D. and Levine, A.
J. (1999). Broad patterns of gene expression revealed by clustering analysis of tumor
and normal colon tissues probed by oligonucleotide arrays. Proceedings of the
National Academy of Sciences, 96, 6745-6750.
•
Alter, O., Brown, P. O., and Botstein, D., (2000). Singular value decomposition for
genome-wide expression data processing and modeling. Proceedings of the National
Academy of Sciences, 97, 10101-10106.
•
Becker, S., (1991). Unsupervised learning procedures for neural networks.
International Journal of Neural Systems, 2, 17-33.
•
Bersekas, D. P., (1995). Nonlinear Programming. Athenas Scientific.
•
Bishop, C.M., (1995). Neural Networks for Pattern Recognition. Clarendon Press,
Oxford.
•
Boser, B., Guyon, I., and Vapnik, V. N., 1992. A training algorithm for optimal
margin classifiers. Fifth Annual Workshop on Computational Learning Theory. 144152.
•
Brazma, A., and Vilo, J., (2000). Gene expression data analysis. FEBS Letters,
480,17-24.
110
•
Brown, M. P. S., Grundy, W. N., Lin, D., Sugnet, C., Ares, M., and Haussler, D.,
(2000). Knowledge-based analysis of microarray gene expression data by using
support vector machines. Proceedings of the National Academy of Sciences, 97, 262267.
•
Butte, A. J., and Kohane, I. S., (2000). Mutual information relevance networks:
Functional genomic clustering using pairwise entropy measurements. Pacific
Symposium on Biocomputing, 5, 415-426.
•
Cai, J., Dayanik, A., Yu, H., Hasan, N., Terauchi, T., and Grundy, W. N., (2000).
Classification of Cancer Tissue Types by Support Vector Machines Using Microarray
Gene Expression Data. International Conference on Intelligent Systems for Molecular
Biology.
•
Chapelle, O., Vapnik, V., Bousquet, O. and Mukherjee, S., (2000). Choosing kernel
parameters for support vector machines. AT&T Labs Technical Report.
•
Chen, T., Filkov, V., and Skiena, S. S., (1999). Identifying gene regulatory networks
from experimental data. Annual International Conference on Computational Biology.
•
Cortes, C., and Vapnik, V., (1995). Support vector networks. Machine Learning, 20,
273-297.
•
Cover, T., (1965). Geometrical and Statistical Properties of Systems of Linear
Inequalities with Applications in Pattern Recognition. IEEE Transaction on
Electronic Computer, 14, 326-334.
•
Datta, S., (2001). Exploring relationships in gene expressions: A partial least squares
approach. Gene Expression, 9, 257-264.
•
Dewey, T. G., and Bhan, A., (2001). A linear systems analysis of expression time
series. Methods of Microarray Data Analysis, Kluwer Academic.
111
•
D'haeseleer, P., (2000). Reconstructing Gene Networks from Large Scale Gene
Expression Data. Ph.D. thesis, University of New Meixco.
•
D'haeseleer, P., Wen, X., Fuhrman, S., and Somogyi, R., (1997). Mining the gene
expression matrix: inferring gene relationships from large scale gene expression data.
Information processing in cells and tissues, Paton, R. C., and Holcombe, M., Eds.,
Plenum Press, 203-212.
•
Ewing, R. M., Kahla, A. B., Poirot, O., Lopez, F., Audic, S., and Claverie, J. M.,
(1999). Large-scale statistical analyses of rice ESTs reveal correlated patterns of gene
expression. Genome Research, 9, 950-959.
•
Filkov, V, Skiena, S, and Zhi, J., (2001). Analysis techniques for microarray timeseries data. Annual International Conference on Computational Biology. 124-131.
•
Fitch, W. M., and Margoliash, E., (1967). Construction of phylogenetic trees. Science,
155, 279-284.
•
Fletcher, R., (1987). Practical Methods of Optimization, 2nd edition, Wiley, New
York.
•
Freund, Y., and Schapire, R. E., (1996). Experiments with a new boosting algorithm.
Machine learning: Proceedings of the Thirteenth International Conference, 148-156.
•
Friedman, N., Linial, M., Nachman, I., and Pe'er, D., (2000). Using Bayesian
networks to analyze expression data. Journal of Computational Biology, 7, 601-620.
•
Fuhrman, S., Cunningham, M. J., Wen, X., Zweiger, G., Seihamer, J. J., and
Somogyi, R., (2000). The application of Shannon entropy in the identification of
putative drug targets. Biosystems, 55, 5-14.
•
Furey, T. S., Cristiannini, N., Duffy, N., Bednarski, D. W., Schummer, M., and
Haussler, D., (2000). Support vector machine classification and validation of cancer
tissue samples using microarray expression data, Bioinformatics, 16, 906-914.
112
•
Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gassenbeek, M., Mesirov, J. P.,
Coller, H., Loh, M. L., Downing, J. R., Caligiuri, M. A., Bloomfield, C. D., and
Lander, E. S., (1999). Molecular Classification of Cancer: Class Discovery and Class
Prediction by Gene Expression Monitoring. Science, 286, 531-537.
•
Guyon, I., Weston, J., Barnhill, S., and Vapnik, V., (2002). Gene selection for cancer
classification using support vector machines. Machine Learning, 46, 389-422.
•
Haykin, S., (1999). Neural networks: A comprehensive foundation. Prentice Hall.
•
Hellem, T. and Jonassen, I., (2002). New feature subset selection procedures for
classification of expression profiles. Genome Biology, 3(4), research0017.1-0017.11
•
Hertz, J., Krogh, A., and Pla,er, R. G., (1991). Introduction to the Theory of Neural
Computation. Addison-Wesley.
•
Hieter, P., and Boguski, M., (1997). Functional genomics: it's all how you read it.
Science, 278, 601-602.
•
Holter, N. S., Maritan, A., Cieplak, M., Fedoroff, N. V., and Banavar, J. R., (2001).
Dynamic modeling of gene expression data. Proceedings of the National Academy of
Sciences, 98, 1693-1698.
•
Huang, S., (1999). Gene expression profiling, genetic networks, and cellular states: an
integrating concept for tumorigenesis and drug discovery. Journal of Molecular
Medicine, 77, 469-480.
•
Hwang, K. B., Cho, D. Y., Park, S. W., Kim, S. D., and Zhang, B. T., (2001).
Applying machine learning techniques to analysis of gene expression data: cancer
diagnosis. Methods of Microarray Data Analysis. Kluwer Academic, 167-182.
•
Ideker, T. E., Thorsson, V. and Karp, R. M., (2000). Discovery of regulatory
interactions through pertubation: inference and experimental design. Pacific
Symposium on Biocomputing, 5, 302-313.
113
•
Kauffman, S. A., (1969). Metabolic stability and epigenesis in randomly connected
nets. Journal of Theoretical Biology, 22, 437-467.
•
Keller, A. D., Schummer, M., Ruzzo, W. L., and Hood, L., (2000). Bayesian
classification of DNA array expression data. Technical Report, University of
Washington - Computer Science Engineering - 2000-08-01.
•
Khan, J., Wei, J. S., Ringner, M., Saal, L. H., Ladanyi, M, Westermann, F, Berthold,
F., Schwab, M., Antonescu, C. R., Peterson, C., and Meltzer, P. S., (2001).
Classification and diagnostic prediction of cancers using gene expression profiling
and artificial neural networks. Nature Medicine, 7, 673-679.
•
Kiberstis, P., and Roberts, L., (2002). It's Not Just the Genes. Science, 296, 685.
•
Kitano, H., (2002). Systems biology: a brief overview. Science, 295, 1662-1664.
•
Klevecz, R. R., (2000). Dynamic architecture of the yeast cell cycle uncovered by
wavelet decomposition of expression microarray data. Functional and Integrative
Genomics, 1, 186-192.
•
Li, W., (2002). Zipf's law in importance of genes for cancer classification using
microarray Data. Journal of Theoretical Biology, 219, 539-51.
•
Liang, S., Fuhrman, S., and Somogyi, R., (1998). REVEAL, A general reverse
engineering algorithm for inference of genetic network architectures, Pacific
Symposium on Biocomputing.
•
Liu, H., and Motoda, H., (1998). Feature selection for knowledge discovery and data
mining. Kluwer Academic Publishers, 1998.
•
Lo, J., Lee, S., Xu, M., Liu, F., Ruan, H., Eun, A., He, Y., Ma, W., Wang, W., Wen,
Z., and Peng, J., (2003). 15,000 Unique Zebrafish EST Clusters and Their Use in
Microarray for Profiling Gene Expression Patterns During Embryogenesis. Genome
Research, 13, 455-466.
114
•
Maki, Y., Tominaga, D., Okamoto, M., Watanabe, S., and Eguchi, Y., (2001).
Development of a System for the Inference of Large Scale Genetic Networks. Pacific
Symposium on Biocomputing, 6, 446-458.
•
Michaels, G. S., Carr, D. B., Askenazi, M., Fuhrman, S., Wen, X., and Somogyi, R.,
(1998). Cluster analysis and data visualization of large-scale gene expression data.
Pacific Symposium on Biocomputing, 3, 42-53.
•
Michigan, (1999). Challenges and Opportunities in Understanding the Complexity of
Living Systems. University of Michigan, report into initiatives of Life Sciences
Commission within context of biocomplexity.
•
Muirhead, R. J., (1982). Aspects of Multivariate Statistical Theory. Wiley series in
probability and mathematical statistics. Wiley, New York.
•
Mukherjee, S., Tamayo, P., Slonim, D., Verri, A., Golub, T., Messirov, J. P., and
Poggio, T., (2000). Support vector machine classification of microarray data. AI
memo, CBCL paper 182. MIT.
•
Pe'er, D., Regev, A., Elidan, G., and Friedman, N., (2001). Inferring subnetworks
from perturbed expression profiles. International Conference on Intelligent Systems
for Molecular Biology.
•
Pineda, F. J., (1987). Generalization of back-propagation to recurrent neural networks.
Physical Review Letters, 59, 2229-2232.
•
Platt, J., (1999). Fast training of SVMs using sequential minimal optimisation.
Advances in Kernel Methods: Support Vector Learning. MIT press, Cambridge, MA,
185-208.
•
Quinlan, J. R., (1993). C4.5: Programming for Machine Learning. Morgan Kaufmann
Publishers.
115
•
Raychaudhuri, S., Stuart, J. M., and Altman, R. B., (2000). Principal components
analysis to summarize microarray experiments: application to sporulation time series.
Pacific Symposium on Biocomputing, 5, 452-463.
•
Samsonova, M. G., and Serov, V. N., (1999). NetWork: An interactive interface to the
tools for analysis of genetic network structure and dynamics. Pacific Symposium on
Biocomputing.
•
Savageau, M. A., (1976). Biochemical Systems analysis: a study of function and
design in molecular biology. Addison-Wesley, Reading.
•
Setiono, R., and Liu, H., (1997). Neural-network feature selector. IEEE Transactions
on Neural Networks, 8, 654-662.
•
Slonim, D., Tamayo, P., Mesirov, J., Golub, T. R., and Lander, E., (2000). Class
prediction and discovery using gene expression data. Annual International
Conference on Computational Biology.
•
Someren, E. P. V., Wessels, L. F. A. and Reinders, M. J. T., (2001). Genetic Network
Models: A Comparative Study. Proceedings of SPIE, Micro-arrays: Optical
Technologies and Informatics.
•
Someren, E. V., Wessels, L. F. A., and Reinders, M. J. T., (2000). Linear modeling of
genetic networks from experimental data. International Conference on Intelligent
Systems for Molecular Biology.
•
Somogyi, R., Fuhrman, S., Askenazi, M. and Wuensche, A., (1996). The gene
expression matrix: towards the extraction of genetic network architectures.
Proceedings of Second World Congress of Nonlinear Analysis, 30, 1815-1824.
•
Spellman, P. T., Sherlock, G., Zhang, M. Q., Iyer, V. R., Anders, K., Eisen, M. B.,
Brown, P. O., Botstein, D., and Futcher, B., (1998). Comprehensive identification of
116
cell cycle-regulated genes of the yeast sacccharomyces cerevisiae by microarray
hybridization. Molecular Biology of the Cell, 9, 3272-3297.
•
Stone, M., and Brooks, R. J., (1992). Continuum regression: cross-validated
sequentially constructed prediction embracing ordinary least squares, partial least
squares and principal component regression. Journal of the Royal Statistical Society
B, 52, 237-269: 1990. Corrigendum 54, 906-907: 1992.
•
Szallasi, Z., (1999). Genetic network analysis in light of massively parallel biological
data acquisition. Pacific Symposium on Biocomputing.
•
Talbot, W.S., and Hopkins, N., (2000). Zebra fish mutations and functional analysis
of the vertebrate genome. Genes and Development, 14, 755-762.
•
Tibshirani, R., Hastie, T., Eisen, M., Ross, D., Botstein, D., Brown, P., (1999).
Clustering methods for the analysis of DNA microarray data. Technical Report,
Stanford University.
•
Vapnik, V., (1998). Statistical Learning Theory. Wiley, New York.
•
Vohradsky, J., (2001). Neural network model of gene expression. FASEB Journal, 15,
846-854.
•
Weigend, A. S., Rumelhart, D. E., and Huberman, B. A., (1991). Generalization by
weight-elimination with application to forecasting. Advances in Neural Information
Processing Systems, Lippmann, R. P., Moody, J., and Touretzky, D. S., Eds., Morgan
Kaufmann, 3, 875-882.
•
Wen, X., Fuhrman, S., Michaels, G. S., Carr, D. B., Smith, S., Barker, J. L., and
Somogyi, R., (1998). Large-scale temporal gene expression mapping of CNS
development. Proceedings of the National Academy of Sciences, 95, 334-339.
•
Werbos, P. J., (1990). Backpropagation through time: what it does and how to do it.
Proceedings of the IEEE, 78, 1550-1560.
117
•
Wessels, L. F. A., Someren, E. P. V., and Reinders, M. J. T., (2001). A Comparison of
Genetic Network Models. Pacific Symposium on Biocomputing, 508-519, Hawai.
•
Weston, J., Mukherjee, S., Chapelle, O., Pontil, M., Poggio, T. and Vapnik, V.,
(2001). Feature Selection for SVMs. Advances in Neural Information Processing
Systems, 13, 668-674.
•
Yang, Y. H., Dudoit, S., Luu, P., and Speed, T. P., (2001). Normalization for cDNA
Microarray Data. Microarrays: Optical Technologies and Informatics, Proceedings of
SPIE.
•
Yeung, K. Y., Haynor, D. R., Ruzzo, W. L., (2001). Validating clustering for gene
expression data. Bioinformatics, 17, 309-318.
•
Zhang, B. T., Ohm, P., and Muhlenbein, H., (1997). Evolutionary Induction of Sparse
Neural Trees. Evolutionary Computation, 5, 213-236.
•
Zipf, G. F., (1965) [1935]. Psycho-Biology of Languages. Mass. MIT Press.
| 5cs.CE
|
ON THE RATIONAL HOMOLOGY AND ASSEMBLY MAPS
OF GENERALISED THOMPSON GROUPS
arXiv:1605.00840v1 [math.GR] 3 May 2016
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
Abstract. Let Vr (Σ) be the generalised Thompson group defined as
the automorphism group of a valid, bounded, and complete Cantor algebra. We show that the rational homology of Vr (Σ) vanishes above a
certain degree depending only on s, the number of descending operations
used in the definition of the Cantor algebra. In particular this applies
to Brin-Thompson groups sV . We explicitly compute the rational homology for 2V and show that it vanishes everywhere except in degrees
0 and 3, where it is isomorphic to Q. We also determine the number
of conjugacy classes of finite cyclic subgroups of a given order m in
Brin-Thompson groups. We apply our computations to the rationalised
Farrell-Jones assembly map in algebraic K-theory.
1. Introduction
There are many well-known generalisations of Thompson’s group V due
to, for example, Higman [10], Stein [16], and Brin [2], which share many of
the properties of V . For example, they all contain infinite torsion and are
of type F∞ [4, 8]. Furthermore, the Higman-Thompson groups Vn,r and the
Brin-Thompson groups sV are simple [10, 3]. It turns out that all these are
examples of automorphism groups Vr (Σ) of valid, bounded, and complete
Cantor algebras. For notation and definitions we refer to [13, 14], where it
was shown that these Vr (Σ) and the centralisers of their finite subgroups are
of type F∞ , and an explicit description of these centralisers was given.
Brown showed in [5] that Thompson’s group V is rationally acyclic; he also
conjectured that V is acyclic, and this conjecture was recently proved in [17].
In this paper we show that the rational homology of Vr (Σ) vanishes above
a certain degree depending only on the number s of descending operations
used in the definition of the Cantor algebra. In particular:
Theorem 1.1. Let Ur (Σ) be a complete, valid, and bounded Cantor algebra,
s ≥ 2 and k ≥ s2s−1 . Then Hk (Vr (Σ); Q) = 0.
This shows in particular that all Higman-Thompson groups Vn,r are rationally acyclic.
The method of the proof gives a very explicit construction to actually
compute the homology groups, which we do for Brin’s group Vr (Σ) = 2V. It
turns out that its rational homology vanishes everywhere except in degrees
0 and 3, where it is isomorphic to Q; compare Theorem 2.11.
Date: May 4, 2016.
2010 Mathematics Subject Classification. 20J05, 19D55.
The first named author was supported by Gobierno de Aragón, European Regional
Development Funds and MTM2015-67781-P.
1
2
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
In [13] it was shown that, for any finite subgroup K ≤ Vr (Σ), there are
only finitely many conjugacy classes of subgroups isomorphic to K. Using
the method employed there, in Proposition 3.1 we explicitly calculate the
number of conjugacy classes of cyclic subgroups of a given order in the Brinlike groups sVn,r , where n denotes the arity of the descending operations.
Note that sV = sV2,1 .
In Section 4 we discuss an application of our results to the algebraic Ktheory of the integral group rings of these generalized Thompson groups. We
first review the Farrell-Jones Conjecture, and then explain how our computations, combined with the results from [12, 13, 14], imply the following
theorem. This generalizes the analysis that was carried out for Thompson’s
group T in [?]. All terms and notation are explained in Section 4.
Theorem 1.2. Consider a Brin-Thompson group G = sV .
If the rationalised Farrell-Jones Conjecture in algebraic K-theory holds
for sV , then Kn (Z[sV ]) ⊗Z Q is isomorphic to
(1.3)
M
M
M
i
i
]) ⊗ Q ,
Kq (Z[Cm
); Q)
⊗
ΘCm
Hp (ZG (Cm
i
1≤m 1≤i≤2φ(m) −1
p+q=n
0≤p≤φ(m)s2s−1
−1≤q
i )]
Q[WG (Cm
Z
i denote the representatives of the conjugacy classes of finite cyclic
where Cm
groups of order m in sV .
If the Leopoldt-Schneider Conjecture holds for all cyclotomic fields, then
Kn (Z[sV ]) ⊗Z Q contains as a direct summand the subspace of (1.3) indexed
by q ≥ 0.
The results used in the proof of this Theorem also imply that the homoli ) can be computed in terms of the ordinary homology
ogy groups Hp (ZG (Cm
Hp (sV ) of sV . Finally, we remark that using [12, 14] one also deduces
that for any valid, bounded, and complete Cantor algebra, the rationalised
Whitehead group W h(Vr (Σ))⊗Z Q is an infinite dimensional Q-vector space;
compare Corollary 4.5.
2. The rational homology of Vr (Σ)
Throughout this section we will use the notation used in [14]. Let Ur (Σ)
denote a valid, bounded, and complete Cantor algebra, and let G = Vr (Σ)
be its automorphism group. The examples to keep in mind are those where
r = 1, and G = V1 (Σ) is either the Higman-Thompson-group G = Vn,1 or
the Brin-Thompson group sV . For each basis A ∈ Ur (Σ) we have a Morse
function t(A) = |A|. We denote the Stein complex by X, which was denoted
by Sr (Σ) in [14].
Remark 2.1. The Stein complex is known to be contractible [5, 8, 14]. By
the same argument as in the remark at the end of Section 4 of [5], nothing
changes if one fixes an integer q and replaces the Stein complex by the Stein
complex obtained by only considering bases of cardinality ≥ q.
The truncated Stein complex. As in [5] we will consider a truncated
Stein complex. For 1 ≤ p ≤ q let Xp,q denote the full subcomplex of X
generated by bases A with p ≤ |A| ≤ q. Note that dim Xp,q ≤ q − p.
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
3
Recall from [14, Lemma 3.4] that each basis has a unique maximal elementary descendant, denoted E(A), with n1 · · · ns |A| leaves obtained by applying all descending operations exactly once to each element of A. Hence
dim Xp,q = q − p if q ≤ N p, where N = n1 · · · ns .
Remark 2.2. For V we have N = 2; for Vn we have N = n; and for sV we
have N = 2s .
Proposition 2.3. Suppose that s ≥ 2 and let Ur (Σ) be a complete, valid,
and bounded Cantor algebra. For all n ≥ 1 there is an integer p0 , depending
on n, such that Xp,q is n-connected for p ≥ p0 and q ≥ p + (N − 1)n.
Proof. Let t = N − 1. Since the complex given by the union of
Xp,p+tn ⊂ Xp,p+tn+1 ⊂ · · ·
is contractible, see Remark 2.1, it suffices to show that there is some p0 such
that for p ≥ p0 and q ≥ p + tn, the pair is (Xp,q+1 , Xp,q ) is n-connected, i.e.,
the inclusion Xp,q ֒→ Xp,q+1 induces an isomorphism between the homotopy
groups πi for 0 ≤ i ≤ n. Exactly as in [5, Theorem 2], this will be satisfied
if for any A with |A| = q + 1, then
Lp (A) = |{B ∈ Xp,q | B < A}|
is n-connected. We follow the lines of the argument of [8], and begin by
showing that
Lp0 (A) = |{B ∈ Xp,q | B < A very elementary}|
is n-connected. Consider the complex Kq , which was denoted Kn in the proof
of [14, Lemma 3.8]. This complex has as vertices the labelled subsets of A
where the possible labels are precisely the set of colours and the cardinality
of a subset with label i is the arity ni . The simplices are given by pairwise
disjoints set of such subsets. As in [14, Lemma 3.8], the argument of [4,
Lemma 4.20] shows that Kq is n-connected if q is big enough. Now, q−p−1 ≥
tn − 1 ≥ n + 1 and therefore the (q − p − 1)-skeleton of Kq is also nconnected. Finally observe that dim Lp (A) = q − p − 1 and in fact, Lp (A) is
the barycentric subdivision of that (q − p − 1)-skeleton.
Now let B ∈ Lp (A) r Lp0 (A) and consider its descending link lkp ↓ (B)
with respect to the height function h̄, but now truncating above p. We claim
that this link is (n − 1)-connected. Using Morse theory (see [8, Lemma 3.1])
this will yield the result. We have
lkp ↓ (B) = downlkp ↓ (B) ∗ uplkp ↓ (B),
where
downlkp ↓ (B) = |{C | C < B, h̄(C) < h̄(B), p ≤ |C|}|
is the truncated downlink, and
uplkp ↓ (B) = uplk ↓ (B) = |{C | B < C, h̄(C) < h̄(B)}|
is the uplink (truncation does not affect the uplink).
As in [8], we may assume that no element of A is obtained by a simple
expansion of an element in B, otherwise the uplink is contractible and the
claim holds true. This means that each element of B yields either some
number l with min{ni | i ∈ S} + 1 = m + 1 ≤ l ≤ N of elements in A, or
4
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
remains invariant in A. Let kb,l be the number of elements of B yielding
exactly l elements
P in A and let ks be the number of the invariant elements
of B. Put k = N
l=m kb,l . Then
(2.4)
|B| = k + ks ,
and
(2.5)
|A| =
N
X
lkb,l + ks = q + 1 ≥ p + tn + 1.
l=m
Set |B| = x. Now, downlkp ↓ (B) is equivalent to the complex associated
to
Lx−p
(A1 ),
0
where A1 has exactly ks elements. Note that this complex has dimension
x − p. Therefore, choosing ks big enough, we may assume that the downlink
is (x − p − 2)-connected.
On the other hand, the argument of [8] implies that the uplink is (k − 2)connected. Therefore, lkp ↓ (B) is (x − p − 2 + k)-connected.
Now, from (2.4) and (2.5) we get
tk + x ≥
N
X
(l − 1)kb,l + x ≥ p + tn + 1.
l=m
Thus
x − p − 1 − t + tk ≥ tn − t.
Assume first that x − p 6= 0. Then t(x − p − 2) ≥ x − p − 1 − t, and hence
t(x − p − 2 + k) ≥ x − p − 1 − t + tk ≥ tn − t.
Finally, x − p − 2 + k ≥ n − 1. In the case when p − x = 0 we have
tk − 1 ≥ tn.
Thus k − 1/t ≥ n, and since both k and n are integers, this implies that
k − 1 ≥ n, and therefore
k − 2 ≥ n − 1.
This means that in both cases lkp ↓ (B) is (n−1)-connected, as required.
From now on fix a basis A with |A| = p, and let Yp,q = Xp,q /G. Consider the
complex Zbp+1,q with m-simplices of the form σ : B0 < B1 < · · · < Bm with
|Bm | ≤ p + n, A < B0 and A ≤ Bm elementary. We let the (finite) group
H = {g ∈ G | gA = A}
bp+1,q , and set
act on Z
bp+1,q /H.
Zp+1,q = Z
Lemma 2.6. Yp+1,q is a subcomplex of Yp,q , and there is an isomorphism
of chain complexes
C∗ (Yp,q , Yp+1,q ) ∼
= C∗ (CZp+1,q , Zp+1,q )
(where C denotes the cone) and therefore a long exact sequence in homology
· · · → H (Yp+1,q ) → H (Yp,q ) → H (CZp+1,q , Zp+1,q ) ∼
= H̃ (Zp+1,q ) → · · · .
j
j
j
j−1
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
5
Proof. Note first that CZp+1,q can be seen as the complex with m-simplices
of the form H(B0 < B1 < · · · < Bm ) with A ≤ B0 . The fact that Yp+1,q lies
inside Yp,q is obvious. Moreover, if Gσ is an orbit with Gσ ∈ Yp,q r Yp+1,q ,
then σ has the form σ : A0 < A1 < · · · < Am with |A0 | = p. We may choose
some g ∈ G with gA0 = A, and map
ϕ : Gσ̄ 7→ H(ν̄),
where ν = A < gA1 < · · · < gAn , and ¯ denotes passing to the quotient
chain complex. Note that if we choose some other g ′ 6= g with g ′ A0 = A,
then g = hg ′ for some h ∈ H, thus
H(A < gA1 < · · · < gAn ) = H(A < g′ A1 < · · · < g ′ An ).
In particular this map does not depend on the choice of g. It is obvious that
ϕ induces an isomorphism at each degree of the chain complexes, we only
have to show that this is in fact a chain map. But this follows from the fact
that, if δ is the boundary map, then
δ(Gσ̄) =
m
X
(−1)i G(τ̄i )
i=1
with τi = A0 < · · · < Ai−1 < Ai+1 < · · · < Am .
The main advantage of this result is that Zp+1,q is much easier to understand than Yp,q . For example, it is a simplicial complex, as opposed to Yp,q .
It is in fact it is the simplicial realisation |Zp+1,q | of the poset (Zp+1,q , ≤)
with
Zp+1,q = {HB | A < B elementary expansion},
where ≤ is given by
HB0 ≤ HB1 ⇐⇒ there is some h ∈ H with B0 ≤ hB1 .
We let
b=
N
s
Y
(ni + 1).
i=1
Proposition 2.7. Let Ur (Σ) be a complete, valid, and bounded Cantor alb − 2s , and p ≥ 2s , then Zp+1,q is contractible.
gebra. If q − p ≥ N
Proof. Recall that we may view Zp+1,q as the geometrical realisation of the
poset (Zp+1,q , ≤).
We say that A < B is a full extension if the following property holds: for
each a ∈ A, the set of colours in the path from a to each of its descendants
in B is the same, i.e., all such paths have the same length. For example, if
we had three colours denoted by bold, dashed, and dotted lines, then:
6
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
a
a
is a full extension;
is not a full extension.
We also say that HB ∈ Zp+1,q is full if A < B is a full extension. Note
that this does not depend on the chosen representative in the coset HB.
f
These elements form a subposet of Zp+1,q which we denote by Zp+1,q
. Now,
f
we claim that for any HB ∈ Zp+1,q there is an A0 with HA0 ∈ Zp+1,q
, such
f
that for any other HC < HB with HC ∈ Zp+1,q
we have HC ≤ HA0 < HB.
To see this, consider the expansion A < B and for each a ∈ A let Ωa
be the largest set of colours which appear in the path from a to each of its
descendants in B. Note that unless a is unchanged in B, the set Ωa contains
at least one element. Let A0 be obtained from A by applying to every a ∈ A
precisely the operations Ωa in such a way that A < A0 is complete. Then
f
obviously A < A0 < B and A 6= A0 ; thus HA0 ∈ Zp+1,q
. Now if HC < HB
is full, by changing C into hC for some h ∈ H if necessary, we may assume
that A < C < B and that A < C is full. But the way A0 is defined together
with the fact that A < B is elementary (i.e., no repetitions of colours are
allowed in any path), imply that A0 ≥ C.
Now using Quillen’s Poset Lemma (see [1, Lemma 6.5.2]) we deduce that
f
the poset inclusion Zp+1,q
֒→ Zp+1,q induces a homotopy equivalence
f
|Zp+1,q
| ≃ |Zp+1,q | = Zp+1,q .
f
. Observe first that each full
We want to further reduce the poset Zp+1,q
extension A < B is uniquely determined by the p-tuple ωB = (Ωa )a∈A of
subsets Ωa ⊆ S needed to get from a its descendants in B. As H permutes
f
are H-orbits of tuples of the
the elements of A, the elements of Zp+1,q
fn
f
previous form. Let Zp+1,q
be the subposet of Zp+1,q
consisting of orbits for
which Ωa = Ωb for a 6= b both in A implies Ωa = Ωb = ∅. Obviously this
property does not depend on the chosen representative. As before we have
fn
f
the poset inclusion Zp+1,q
֒→ Zp+1,q
, and for any full extension A < B there
fn
is some A < A0 ≤ B with HA0 ∈ Zp+1,q
and such that, for any A < C < B
fn
with HC ∈ Zp+1,q
, we have HC ≤ HA0 . To see it note that it suffices
to let A0 be the expansion corresponding to the p-tuple obtained from ωB
by changing all repeated instances of Ωa ’s except one to ∅. Also note that
it does not matter which element a associated to the particular instance
that is unchanged, as with any other choice we would end up in another
representative of the same orbit.
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
7
Now, using Quillen’s Poset Lemma again, we get a homotopy equivalence
fn
f
|Zp+1,q
| ≃ |Zp+1,q
| ≃ |Zp+1,q | = Zp+1,q .
fn
Now, the elements of the smaller poset Zp+1,q
can be seen as sets of the
form
α = {Ω1 , . . . , Ωm | m ≤ p, ∅ 6= Ωi ⊆ S}
fn
Recall that repetitions are not allowed. Therefore, Zp+1,q
is a subposet of
the poset A whose elements are the α’s above, and given two elements
α = {Ω1 , . . . , Ωm | m ≤ p, ∅ 6= Ωi ⊆ S},
α′ = {Ω′1 , . . . , Ω′m | m′ ≤ p, ∅ 6= Ω′i ⊆ S},
we have α′ ≤ α if and only if m′ ≤ m and there is some reordering of the
components of α′ such that Ω′i ⊆ Ωi for i = 1, . . . , m′ . Obviously, A is
contractible as
β = {Ω | ∅ 6= Ω ⊆ S}
is un upper bound for any other element α ∈ A (the fact that β ∈ A
is guaranteed by the hypothesis p ≥ 2s ). So we only have to check that
fn
the hypothesis on q − p implies β ∈ Zp+1,q
. To do that, observe that the
s
cardinality of β is 2 − 1 ≤ p, so all the expansions associated to β can
be performed in A. For each element a ∈ A, performing the descending
operations encoded by Ω = {si1 , ..., sid } (i1 , ..., id ∈ S) yields exactly NΩ =
ni1 · · · nid elements. Therefore, to be able to perform the expansions encoded
fn
by β inside Zp+1,q
, we need
X
X
b − 2s .
q−p≥
{NΩ − 1 | Ω ∈ β} =
{NΩ | Ω ∈ β} − 2s = N
As this holds true by hypothesis, we get the result.
As seen in the proof of Proposition 2.7, to compute the reduced homology
fn
of Zp+1,q we need to consider only Zp+1,q
, which is the geometric realisation
fn
of the poset Zp+1,q
. This poset has bounded dimension, with this bound
only depending on the number of colours involved. To begin with, suppose
b − 2s . Then Zp+1,q is contractible and Z f n has a maximal
q−p ≥ N
p+1,q
fn
fn
∪ 0̂, where 0̂ denotes a
element β. Consider the poset Z p+1,q = Zp+1,q
fn
unique minimal element. Then Z p+1,q is equal to the interval [0̂, β], and
we can find the length of a maximal chain as follows. We give the colours
the natural ordering and begin by including the one-element sets, i.e 0̂ <
{{1}} < {{1}, {2}} < · · · < {{1}, {2}, . . . , {s}}. Given any 2 ≤ k ≤ s,
suppose we have α ∈ [0̂, β] an element of our maximal chain constructed in
such a way that, for all l < k, all l-element sets Ωl ∈ α. We now add a kelement subset Ωk ∈
/ α and get an interval [α, α ∪ Ωk ] of lengths k as follows.
There is a (k − 1)-element subset Ωk−1 ∈ α such that Ωk = Ωk−1 ∪ {i} for
some i ∈ S. Then α < {α\Ωk−1 , (Ωk−1 ∪ {i})} = α1 . Now there is a (k − 2)element subset Ωk−2 ∈ α1 such that Ωk−1 = Ωk−2 ∪ {j} for some i 6= j ∈ S,
and α < α1 < {α1 \Ωk−2 , (Ωk−2 ∪{j})} = α2 . Continue this process to arrive
at
α < α1 < · · · < αk = α ∪ Ωk .
8
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
Continue this process until finally we construct the only s-element subset
Ωs = S. Hence the length of a maximal chain in [0̂, β] is
s
X
s
i
= s2s−1 .
D(s) =
i
i=1
fn
This implies that the dimension of Zp+1,q
is equal to D(s)−1. Now removing
the condition on q − p, we only decrease the dimension and hence
fn
dim(Zp+1,q
) ≤ D(s) − 1,
fn
fn
and Zp+1,q
is contractible if dim(Zp+1,q
) = D(s) − 1.
We are now ready to prove Theorem 1.1: if Ur (Σ) is a complete, valid,
and bounded Cantor algebra, s ≥ 2 and k ≥ D(s), then Hk (Vr (Σ), Q) = 0.
Proof of Theorem 1.1. Let p be as in Proposition 2.3 and as in Proposition
2.7 and set q = p + tk with t = N − 1. As Xp,q is k-connected, Hk (Vr (Σ)) =
Hk (Yp,q ). Consider for j ≥ 0 the long exact sequence of Lemma 2.6:
. . . → Hk (Yp+j+1,q ) → Hk (Yp+j,q ) → H̃k−1 (Zp+j+1,q ) → . . .
When j = (t − 1)k, note that dim Yp+j+1,q = p + tk − (p + tk − k + 1) = k − 1
thus
. . . → 0 = Hk (Yp+(t−1)k+1,q ) → Hk (Yp+(t−1)k,q ) → H̃k−1 (Zp+(t−1)k+1,q ) → . . .
fn
Proposition 2.7 implies that H̃k−1 (Zp+j+1,q ) ∼
= H̃k−1 (Zp+j+1,q ). The arfn
either has dimension strictly less that
gument above implies that Zp+j+1,q
k − 1 or is contractible. Hence H̃k−1 (Zp+j+1,q ) = 0 for all 0 ≤ j ≤ (k − 1)t .
This means that
0 = Hk (Yp+(t−1)k,q ) ∼
= Hk (Yp+(t−1)k−1,q ) ∼
= ...
∼
. . . = Hk (Yp+1,q ) ∼
= Hk (Yp,q ) ∼
= Hk (Vr (Σ)).
Corollary 2.8. The Higman-Thompson groups Vn,r are rationally acyclic.
Proof. Use [5, Theorem 2] instead of Proposition 2.3. The computations for
D(s) also work for s = 1 and the argument of Theorem 1.1 goes through
unchanged.
Corollary 2.9. Let sV be Brin-Thompson group. Then for all k ≥ s2s−1
and k = 1, we have Hk (sV ) = 0.
Proof. For k ≥ D(s) this is Theorem 1.1 and for k = 1 this follows from the
fact that sV is simple [3].
Remark 2.10. We do not know whether the bound in Theorem 1.1 is sharp,
yet the computations in Theorem 2.11 below show that it is sharp for 2V .
It is the same for all groups Vr (Σ) for a given s, yet it is not clear how much
fn
arities influence the homology of the complex Zp+j+1,q
. The proof for 2V
below uses the fact that all n1 = n2 = 2 when considering explicity the
fn
complexes Zp+j+1,q
.
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
9
Theorem 2.11. Let G = 2V . Then
H3 (2V ) ∼
= Q,
and for all k 6= 3, 0
Hk (2V ) = 0.
Proof. In light of Corollary 2.9, noting that D(2) = 4, we only need to
verify the claim for k = 2, 3. We begin by computing the reduced homology
of Zp+1,q for all q − p ≤ 4. Using the description at the end of the proof of
Proposition 2.7 we get the following cases:
• q−p = 1 : Here Zp+1,q consists of two points and hence H̃0 (Zp+1,q ) ∼
=
Q and H̃i (Zp+1,q ) = 0 for all i ≥ 1.
• q − p = 2 : Here Zp+1,q ≃ ∗ and H̃i (Zp+1,q ) = 0 for all i ≥ 0.
• q − p = 3 : Zp+1,q ≃ S 1 and H̃1 (Zp+1,q ) ∼
= Q and H̃i (Zp+1,q ) = 0 for
all i 6= 1.
• q − p = 4 : Zp+1,q ≃ S 2 and H̃2 (Zp+1,q ) ∼
= Q and H̃i (Zp+1,q ) = 0 for
all i 6= 2.
We compute H2 (2V ) : Here q = p + 6 and for j ≥ 4 the dimension of
Yp+j+1,q is less or equal to 1 and hence the second homology vanishes.
For j = 4, we have H2 (Yp+4+1,q ) = 0 and H̄1 (Zp+4+1,q ) = 0 because
Zp+4+1,q is contractible. Thus the long exact sequence yields H2 (Yp+3+1,q ) =
0.
For j = 3 we have that q − (p + j) = 3 and hence H1 (Zp+3+1,q ) ∼
= Q and
we obtain the following long exact sequence:
0 → H2 (Yp+3,q ) → Q → H1 (Yp+3+1,q ) → H1 (Yp+3,q ).
(∗)
∼ H1 (2V ) = 0.
Proposition 2.3 and Corollary 2.9 imply that H1 (Yp+3,q ) =
Computing H1 (Yp+3+1,q ) is the same as computing H1 (Yp′ +1,p′ +3 ). Since
Yp′ +3,p′ +3 is zero-dimensional, H1 (Yp′ +3,p′ +3 ) = 0 and H̃0 (Zp′ +2+1,p′ +3 ) ∼
=
Q. Furthermore, since Yp′ +3,p′ +3 is connected, we have H̃0 (Yp′ +3,p′ +3 ) = 0.
Hence the long exact sequence of Lemma 2.6 yields that H1 (Yp′ +2,p′ +3 ) ∼
= Q.
We now have the long exact sequence:
H̃1 (Zp′ +1+1,p′ +3 ) → H1 (Yp′ +2,p′ +3 ) → H1 (Yp′ +1,p′ +3 ) → H̃0 (Zp′ +1+1,p′ +3 ).
Since Zp′ +1+1,p′ +3 is contractible, the two outside terms are zero and hence
H1 (Yp′ +1,p′ +3 ) ∼
= Q. This means that (∗) reduces to
0 → H2 (Yp+3,q ) → Q → Q → 0.
Since the map Q → Q is onto, it is an isomorphism and this implies that
H2 (Yp+3,q ) = 0.
For j = 2 we get that H1 (Zp+2+1,q ) = 0. Hence the long-exact sequence
of Lemma 2.6 yields
H2 (Yp+2,q ) = 0.
Now, for j ≤ 1, Zp+j+1,q is contractible and as above we have that
H2 (2V ) ∼
= H2 (Yp,q ) = 0.
We finish by computing H3 (2V ), where q = p + 9. For j ≥ 6 we have
dim(Yp+j+1,q ) ≤ 2 and hence H3 (Yp+6+1,q ) = 0. Now we repeatedly use the
long exact sequence of Lemma 2.6.
10
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
For j = 6 we have that q − (p + j) = 3 and hence H3 (Zp+6+1,q ) =
H2 (Zp+6+1,q ) = 0 and
H3 (Yp+6+1 ) = H3 (Yp+6,q ) = 0.
∼ Q. Furthermore
Let j = 5. Then H3 (Zp+5+1,q ) = 0 and H2 (Zp+5+1,q ) =
H2 (Yp+5+1,q ) = 0, which follows from the calculations for H2 (2V ), where we
showed that H2 (Yp′ ,p′ +3 ) = 0. Hence the long exact sequence now gives
H3 (Yp+5,q ) ∼
= H2 (Zp+5+1,q ) ∼
= Q.
Now, for j ≤ 4, we have that Zp+j+1,q is contractible, hence the second and
third homology vanishes, and hence H3 (Yp+j+1,q ) ∼
= H3 (Yp+j,q ) ∼
= Q, which
yields
H3 (2V ) ∼
= Q.
When computing H2 (2V ) above, we used the fact that 2V was simple.
One can also get the result without referring to the simplicity of 2V by explicitly computing the connecting homomorphism in (∗). One could then
continue by using the long exact sequence of Lemma 2.6 to show that
H1 (2V ) = 0.
3. Conjugacy classes of finite cyclic subgroups in sVn,r
Proposition 3.1. For any m > 0 such that m and n − 1 are coprime, the
number of conjugacy classes of cyclic groups of order m in sVn,r is
nφ(m) − 1,
where φ(m) is the Euler function. In particular, for any m > 0 the number
of conjugacy classes of cyclic groups of order m in sV2,r is
2φ(m) − 1.
Proof. We claim first that the number of conjugacy classes of cyclic subgroups of order m equals the cardinality of the set C(n, r, m) of the possible
{ks | s|m, 0 ≤ ks ≤ n − 1}
P
such that km 6= 0 and s|m sks ≡ r mod n − 1. The claim follows essentially from [13, Proposition 4.2 and Theorem 4.3], but we briefly recall the
argument here. The main idea goes back to Higman (see [10] and also [7]):
one can prove that any finite subgroup H ≤ sVn,r acts on a certain n-ary
r-rooted forest by permuting the set of leaves L. That set of leaves can be
split into transitive subsets, each corresponding to a particular type as a
permutation representation, and at least one of the orbits must be faithful.
The number of orbits of a certain type can be modified modulo n − 1 just
by passing to subtrees. This yields the same group, yet note that if some
type does not appear it can not be created. Any other copy H1 ≤ sVn,r of
the same group is conjugate to H in sVn,r if and only if, in both groups for
each type, the set of orbits of that precise type is either zero in both, or non
zero in both and congruent modulo n − 1.
In the particular case of a cyclic group of order m, the types of permutation representations correspond to the possible divisors of m, more precisely,
to each s|m we may associate the transitive permutation representation of
length s. The number ks above refers to the number of orbits of length s.
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
11
The only faithful orbit has length m, hence km 6=P0. From the fact that our
group is defined using an r-rooted forest we get s|m sks ≡ r mod n − 1.
Now, assume that m and n − 1 are coprime. Then for each choice of ks ,
s|m, s 6= 1, solving the equation
X m
ks mod n − 1
mx ≡ r −
s
s|m,s6=1
yields exactly one possible value of k1 , i.e., a single element in the set
C(n, r, m) above. Note here that we are using the fact that k1 can not
be zero, otherwise there could in some cases be two valid choices for k1 ,
namely 0 and n − 1. This means that
|C(n, r, m)| = nφ(m) − 1.
For any m > 0 the number of conjugacy classes of cyclic groups of order
m in sVn,r can be computed as follows: Consider the set
X
Ω = {r −
α(s)s | 0 ≤ α(s) ≤ n} ⊆ Z,
s|m,0<s6=m
observe that it consist of nφ(m)−1 elements. Let d = gcd(m, n − 1), the
desired number is then
d · {a ∈ Ω | d divides a} .
More detailed formulas for the special case when m = pa for a prime p can
be found in [10] or [7].
4. Assembly maps and algebraic K-theory
In this final section we explain an application of our computations to
algebraic K-theory. We begin by recalling the rationalised version of the
Farrell-Jones Conjecture. For more details and background information we
refer to the introduction of [12] and to the references cited there.
Let G be a group. Denote by (FC) the set of conjugacy classes of finite
cyclic subgroups of G. The centraliser of a subgroup C ≤ G is denoted
ZG (C) and the Weyl group WG (C) is defined as the quotient WG (C) =
NG (C)/ZG (C) of the normaliser modulo the centraliser. Notice that, when
C is finite, then WG (C) is finite, too.
For any group G and any n ∈ Z there is a natural homomorphism
(4.1)
M
M
Hp (ZG (C); Q)
⊗
ΘC Kq (Z[C]) ⊗ Q → Kn (Z[G]) ⊗ Q ,
Z
Q[WG (C)]
(C)∈(F C) p+q=n
p≥0
q≥−1
Z
called rationalised Farrell-Jones assembly map. The rationalised FarrellJones Conjecture asserts that (4.1) is an isomorphism for every G and every n ∈ Z.
Here ΘC is an idempotent endomorphism of Kq (Z[C]) ⊗Z Q, whose image
is a direct summand of Kq (Z[C]) ⊗Z Q isomorphic to
!
M
Kq (Z[D]) ⊗ Q → Kq (Z[C]) ⊗ Q .
(4.2)
coker
DC
Z
Z
12
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
The Weyl group acts via conjugation on C and hence on ΘC (Kq (Z[C]) ⊗Z Q).
The Weyl group action on the homology comes from the fact that the space
ENG (C)/ZG (C) is a model for BZG (C). The dimensions of the Q-vector
spaces in (4.2) are explicitly computed in [15]*Theorem on page 9 for any q
and any finite cyclic group C.
The following injectivity result about the Farrell-Jones Conjecture is proven
in [12].
Theorem 4.3 ([12]*Theorem 1.11). Suppose that the following conditions
are satisfied for each finite cyclic subgroup C of G.
[A] For every p ≥ 1, Hp (ZG (C); Z) is a finitely generated abelian group.
[B] For every q ≥ 0, the natural homomorphism
Y
Kq (Z[ζc ]) ⊗ Q →
Kt (Zℓ ⊗Z Z[ζc ], Zℓ ) ⊗ Q
Z
Z
ℓ prime
is injective, where c is the order of C and ζc is any primitive c-th
root of unity.
Then the restriction of the rationalised Farrell-Jones assembly map (4.1) to
the summands with q ≥ 0 is injective for each n.
We remark that condition [B] is conjecturally always true; more precisely,
it is true if the Leopoldt-Schneider Conjecture in algebraic number theory
holds for all cyclotomic fields. For more details we refer to [12, Section 2].
We can now prove Theorem 1.2.
Proof of Theorem 1.2. First, we need to show that the source of the rationalised Farrell-Jones assembly map (4.1) is isomorphic to (1.3):
M
M
M
i
i
Kq (Z[Cm
]) ⊗ Q .
Hp (ZG (Cm
); Q)
⊗
ΘCm
i
1≤m 1≤i≤2φ(m) −1
p+q=n
0≤p≤φ(m)s2s−1
−1≤q
i )]
Q[WG (Cm
Z
Using Proposition 3.1 we obtain the number of conjugacy classes of finite
cyclic subgroups of order m. It was shown in [13, Theorem 4.4] that each
centraliser is decomposed into a direct product ZsV (Cm ) = Z1 × · · · × Zl ,
where l is the number of transitive permutation representations of Cm , and
where each Zi fits into a short exact sequence of groups
0 → Ki → Zi → sV → 0
with Ki a locally finite group. As already observed in the proof of Proposition 3.1, l = φ(m). Now an easy spectral sequence argument shows that for
all p ≥ 0
Hp (Zi ; Q) ∼
= Hp (sV ; Q).
The Künneth Theorem together with Corollary 2.9 yields that, if p ≥
φ(m)s2s−1 , then Hp (ZsV (Cm ); Q) vanishes. This establishes the first statement of the theorem.
The last statement is then a consequence of Theorem 4.3, because [14,
Theorems 3.1 and 4.9] imply that sV and all centralisers of finite subgroups
are of type F∞ , and therefore condition [A] is satisfied.
ON THE RATIONAL HOMOLOGY OF GENERALISED THOMPSON GROUPS
13
Remark 4.4. The Weyl groups WsV (C) for finite subgroups of sV were described in [14, Theorem 5.1]. With the notation used in the proof of Proposition 3.1, for each finite
P cyclic group of order m there is a set of leaves Y of
a fixed order |Y | = s|m sks , where 0 ≤ ks ≤ 1 and km = 1. Note that here
i of the conjugacy classes of cyclic
n = 2. Hence, for each representative Cm
P
subgroups of order m, there is a Yi with m ≤ |Yi | ≤ s|m s and
i
i
i
WsV (Cm
) = NS(Yi ) (Cm
)/ZS(Yi ) (Cm
),
where S(Yi ) denotes the symmetric group on |Yi | letters.
Notice that, in the case of Thompson’s group V , the formula (1.3) drastically simplifies, since Hp (V ; Q) vanishes for all p 6= 0, and so (1.3) reduces
to
M
M
i
Kn (Z[Cm
]) ⊗ Q .
Q
⊗
ΘCm
i
1≤m 1≤i≤2φ(m) −1
i )]
Q[WG (Cm
Z
Also for Brin-Thompson’s group 2V , using Theorem 2.11, one can describe
(1.3) in a little more detail.
Finally, we remark that [12, Theorem 1.16] applies to general automorphism groups of Cantor algebras because of [14, Theorems 3.1 and 4.9], and
immediately gives the following result about their Whitehead groups.
Corollary 4.5. Let G = Vr (Σ) be the automorphism group of a valid,
bounded, and complete Cantor algebra. Then there is an injective homomorphism
lim W h(H) ⊗Z Q → W h(G) ⊗Z Q,
−→
OF G
where the colimit is taken over the orbit category of finite subgroups of G.
In particular, W h(G) ⊗Z Q is an infinite dimensional Q-vector space.
References
[1] D. J. Benson. Representations and cohomology II: Cohomology of groups and modules. Cambridge Un. Press, Cambridge, 1991.
[2] Matthew G. Brin. Higher dimensional Thompson groups. Geom. Dedicata, 108:163–
192, 2004.
[3] M. Brin, On the Bakers map and the Simplicity of the Higher Dimensional Thompson
Groups nV , Publ. Mat. 54 (2010), no. 2, 433439.
[4] Kenneth S. Brown. Finiteness properties of groups. Journal of Pure and Applied
Algebra, 44 (1987), 45–75.
[5] K. S. Brown. The geometry of finitely presented infinite simple groups. Algorithms
and classification in combinatorial group theory (Berkeley, CA, 1989), 121-136, Math.
Sci. Res. Inst. Publ., 23, Springer, New York, 1992.
[6] K. S. Brown. Cohomology of groups, Graduate Texts in Mathematics 87. SpringerVerlag, New York, 1994.
[7] W. Dicks, C. Martı́nez-Pérez. Isomorphisms Brin-Higman-Thompson groups. Israel
J. Math. 199 (2014), no. 1, 189218.
[8] M.G.Fluch, M. Marschler, S. Witzel and M.C.B. Zaremsky. The Brin-Thompson
groups sV are of type F∞ , Pacific J. Math. 266 (2013), no. 2, 283295.
[9] R. Geoghegan, and M. Varisco On Thompson’s group T and algebraic K-theory.
Preprint 2014 /http://arxiv.org/pdf/1401.0357
[10] G. Higman, Finitely presented infinite simple groups, Notes on Pure Mathematics 8,
Australian National University, Canberra 1974
14
CONCHITA MARTÍNEZ-PÉREZ, BRITA NUCINKIS, AND MARCO VARISCO
[11] D. Kochloukova, C. Martı́nez-Pérez and B. E. A. Nucinkis. Cohomological finiteness
properties of the Brin-Thompson-Higman groups 2V and 3V, Proceedings of the Edinburgh Mathematical Society (2) 56 no. 3, 777-804 (2013).
[12] W. Lück, H. Reich, J. Rognes and M. Varisco. Algebraic K-theory of group rings and
the cyclotomic trace map. preprint, arXiv:1504.03674, (2015).
[13] C. Martı́nez-Pérez, and B. E. A. Nucinkis. Bredon cohomological finiteness conditions
for generalisations of Thompson’s groups, Groups Geom. Dyn. 7 (2013), 931–959,
[14] C. Martı́nez-Pérez, F. Matucci and B. E. A. Nucinkis. Cohomological finiteness conditions and centralisers in generalisations of Thompson’s group V. preprint (2013),
to appear Forum Math.
[15] D. Patronas, The Artin Defect in Algebraic K-Theory, Ph.D Thesis, Freie Universität
Berlin (2014).
[16] M. Stein. Groups of piecewise linear homeomorphisms. Trans. Amer. Math. Soc.,
332(2):477–514, 1992.
[17] M. Szymik. Homology and the stability problem in the Thompson group family.
Preprint (2014), arXiv:1411.5035.
Conchita Martı́nez-Pérez, Departamento de Matemáticas, Universidad de
Zaragoza, 50009 Zaragoza, Spain
E-mail address: [email protected]
Brita E. A. Nucinkis, Department of Mathematics, Royal Holloway, University of London, Egham, TW20 0EX, UK.
E-mail address: [email protected]
Marco Varisco, Department of Mathematics and Statistics, University at
Albany, SUNY, USA
E-mail address: [email protected]
| 4math.GR
|
Towards Energy Consumption Verification via Static
Analysis
∗
P. Lopez-Garcia †
[email protected]
R. Haemmerlé
[email protected]
U. Liqat †
arXiv:1512.09369v1 [cs.PL] 31 Dec 2015
[email protected]
ABSTRACT
In this paper we leverage an existing general framework
for resource usage verification and specialize it for verifying
energy consumption specifications of embedded programs.
Such specifications can include both lower and upper bounds
on energy usage, and they can express intervals within which
energy usage is to be certified to be within such bounds.
The bounds of the intervals can be given in general as functions on input data sizes. Our verification system can prove
whether such energy usage specifications are met or not.
It can also infer the particular conditions under which the
specifications hold. To this end, these conditions are also
expressed as intervals of functions of input data sizes, such
that a given specification can be proved for some intervals
but disproved for others. The specifications themselves can
also include preconditions expressing intervals for input data
sizes. We report on a prototype implementation of our approach within the CiaoPP system for the XC language and
XS1-L architecture, and illustrate with an example how embedded software developers can use this tool, and in particular for determining values for program parameters that
ensure meeting a given energy budget while minimizing the
loss in quality of service.
Keywords: Energy consumption analysis and verification,
resource usage analysis and verification, static analysis, verification.
1.
†
INTRODUCTION
In an increasing number of applications, particularly those
running on devices with limited resources, it is very important and sometimes essential to ensure conformance with respect to specifications expressing non-functional global properties such as energy consumption, maximum execution time,
memory usage, or user-defined resources. For example, in a
real-time application, a program completing an action later
than required is as erroneous as a program not computing
the correct answer. The same applies to an embedded application in a battery-operated device (e.g., a portable or
implantable medical device, an autonomous space vehicle,
or even a mobile phone) if the application makes the device
∗Spanish Council for Scientific Research (CSIC).
†IMDEA Software Institute, Madrid, Spain.
‡Universidad Politécnica de Madrid (UPM).
M. Klemen †
[email protected]
M. Hermenegildo †
‡
[email protected]
run out of batteries earlier than required, making the whole
system useless in practice.
In general, high performance embedded systems must control, react to, and survive in a given environment, and this
in turn establishes constraints about the system’s performance parameters including energy consumption and reaction times. Therefore, a mechanism is necessary in these
systems in order to prove correctness with respect to specifications about such non-functional global properties.
To address this problem we leverage an existing general
framework for resource usage analysis and verification [22,
23], and specialize it for verifying energy consumption specifications of embedded programs. As a case study, we focus on the energy verification of embedded programs written in the XC language [37] and running on the XMOS
XS1-L architecture (XC is a high-level C-based programming language that includes extensions for communication,
input/output operations, real-time behavior, and concurrency). However, the approach presented here can also be
applied to the analysis of other programming languages and
architectures. We will illustrate with an example how embedded software developers can use this tool, and in particular for determining values for program parameters that
ensure meeting a given energy budget while minimizing the
loss in quality of service.
2.
OVERVIEW OF THE ENERGY VERIFICATION TOOL
In this section we give an overview of the prototype tool
for energy consumption verification of XC programs running on the XMOS XS1-L architecture, which we have implemented within the CiaoPP system [13]. As in previous
work [20, 25], we differentiate between the input language,
which can be XC source, LLVM IR [17], or Instruction Set
Architecture (ISA) code, and the intermediate semantic program representation that the CiaoPP core components (e.g.,
the analyzer) take as input. The latter is a series of connected code blocks, represented by Horn Clauses, that we
will refer to as “HC IR” from now on. We perform a transformation from each input language into the HC IR and
pass it to the corresponding CiaoPP component. The main
reason for choosing Horn Clauses as the intermediate representation is that it offers a good number of features that
make it very convenient for the analysis [25]. For instance,
it supports naturally Static Single Assignment (SSA) and
recursive forms, as will be explained later. In fact, there is
a current trend favoring the use of Horn Clause programs
Energy Model
Energy Consumption Analysis & Verification Tool
HC IR
Translator
Program
Static
Analysis
#pragma true
Inferred
#pragma checked
Proved
Assertions
# pragma check
# pragma trust
...
XC Code
XC
Compiler
int f(int arg){
...
Static
Comparator
#pragma false
Disproved
#pragma check
Unproved
Figure 1: Energy consumption verification tool using CiaoPP.
.
as intermediate representations in analysis and verification
tools [2].
Figure 1 shows an overview diagram of the architecture of
the prototype tool we have developed. Hexagons represent
different tool components and arrows indicate the communication paths among them.
The tool takes as input an XC source program (left part of
Figure 1) that can optionally contain assertions in a C-style
syntax. As we will see later, such assertions are translated
into Ciao assertions, the internal representation used in the
Ciao/CiaoPP system.
The energy specifications that the tool will try to prove
or disprove are expressed by means of assertions with check
status. These specifications can include both lower and upper bounds on energy usage, and they can express intervals
within which energy usage is to be certified to be within such
bounds. The bounds of the intervals can be given in general
as functions on input data sizes. Our tool can prove whether
such energy usage specifications are met or not. It can also
infer the particular conditions under which the specifications
hold. To this end, these conditions are also expressed as intervals of functions of input data sizes, such that a given
specification can be proved for some intervals but disproved
for others.
In addition, assertions can also express trusted information such as the energy usage of procedures that are not developed yet, or useful hints and information to the tool. In
general, assertions with status trust can be used to provide
information about the program and its constituent parts
(e.g., individual instructions or whole procedures or functions) to be trusted by the analysis system, i.e., they provide base information assumed to be true by the inference
mechanism of the analysis in order to propagate it throughout the program and obtain information for the rest of its
constituent parts.
In our tool the user can choose between performing the
analysis at the ISA or LLVM IR levels (or both). We refer
the reader to [19] for an experimental study that sheds light
on the trade-offs implied by performing the analysis at each
of these two levels, which can help the user to choose the
level that fits the problem best.
The associated ISA and/or LLVM IR representations of
the XC program are generated using the xcc compiler. Such
representations include useful metadata. The HC IR translator component (described in Section 4) produces the internal representation used by the tool, HC IR, which includes
the program and possibly specifications and/or trusted information (expressed in the Ciao assertion language [32,
15]).
The tool performs several tasks:
1. Transforming the ISA and/or LLVM IR into HC IR.
Such transformation preserves the resource consumption semantics, in the sense that the resource usage
information inferred by the tool is applicable to the
original XC program.
2. Transforming specifications (and trusted information)
written as C-like assertions into the Ciao assertion language.
3. Transforming the energy model at the ISA level [16],
expressed in JSON format, into the Ciao assertion language. Such assertions express the energy consumed
by individual ISA instruction representations, information which is required by the analyzer in order to
propagate it during the static analysis of a program
through code segments, conditionals, loops, recursions,
etc., in order to infer analysis information (energy consumption functions) for higher-level entities such as
procedures, functions, or loops in the program.
4. In the case that the analysis is performed at the LLVM
IR level, the HC IR translator component produces a
set of Ciao assertions expressing the energy consumption corresponding to LLVM IR block representations
in HC IR. Such information is produced from a mapping of LLVM IR instructions with sequences of ISA
instructions and the ISA-level energy model. The mapping information is produced by the mapping tool that
was first outlined in [3] (Section 2 and Attachments
D3.2.4 and D3.2.5) and is described in detail in [11].
Then, following the approach described in [20], the CiaoPP
parametric static resource usage analyzer [30, 28, 34] takes
the HC IR, together with the assertions which express the
energy consumed by LLVM IR blocks and/or individual ISA
instructions, and possibly some additional (trusted) information, and processes them, producing the analysis results,
which are expressed also using Ciao assertions. Such results include energy usage functions (which depend on input
data sizes) for each block in the HC IR (i.e., for the whole
program and for all the procedures and functions in it.).
The analysis can infer different types of energy functions
(e.g., polynomial, exponential, or logarithmic). The procedural interpretation of the HC IR programs, coupled with
the resource-related information contained in the (Ciao) assertions, together allow the resource analysis to infer static
bounds on the energy consumption of the HC IR programs
that are applicable to the original LLVM IR and, hence, to
their corresponding XC programs. Analysis results are given
using the assertion language, to ensure interoperability and
make them understandable by the programmer.
The verification of energy specifications is performed by
a specialized component which compares the energy specifications with the (safe) approximated information inferred
by the static resource analysis. Such component is based on
our previous work on general resource usage verification presented in [22, 23], where we extended the criteria of correctness as the conformance of a program to a specification expressing non-functional global properties, such as upper and
lower bounds on execution time, memory, energy, or user defined resources, given as functions on input data sizes. We
also defined an abstract semantics for resource usage properties and operations to compare the (approximated) intended
semantics of a program (i.e., the specification) with approximated semantics inferred by static analysis. These operations include the comparison of arithmetic functions (e.g.,
polynomial, exponential, or logarithmic functions) that may
come from the specifications or from the analysis results. As
a possible result of the comparison in the output of the tool,
either:
1. The original (specification) assertion (i.e., with status
check) is included with status checked (resp. false),
meaning that the assertion is correct (resp. incorrect)
for all input data meeting the precondition of the assertion,
2. the assertion is “split” into two or three assertions with
different status (checked, false, or check) whose preconditions include a conjunct expressing that the size
of the input data belongs to the interval(s) for which
the assertion is correct (status checked), incorrect (status false), or the tool is not able to determine whether
the assertion is correct or incorrect (status check), or
3. in the worst case, the assertion is included with status
check, meaning that the tool is not able to prove nor
to disprove (any part of) it.
If all assertions are checked then the program is verified.
Otherwise, for assertions (or parts of them) that get false
status, a compile-time error is reported. Even if a program
contains no assertions, it can be checked against the assertions contained in the libraries used by the program, potentially catching bugs at compile time. Finally, and most
importantly, for assertion (or parts of them) left with status
check, the tool can optionally produce a verification warn-
ing (also referred to as an “alarm”). In addition, optional
run-time checks can also be generated.
3.
THE ASSERTION LANGUAGE
Two aspects of the assertion language are described here:
the front-end language in which assertions are written and
included in the XC programs to be verified, and the internal
language in which such assertions are translated into and
passed, together with the HC IR program representation, to
the core analysis and verification tools, the Ciao assertion
language.
3.1
The Ciao Assertion Language
We describe here the subset of the Ciao assertion language which allows expressing global “computational” properties and, in particular, resource usage. We refer the reader
to [32, 13, 15] and their references for a full description of
this assertion language.
For brevity, we only introduce here the class of pred assertions, which describes a particular predicate and, in general, follows the schema:
:- pred Pred [: Precond ] [=> Postcond ] [+ Comp-Props].
where Pred is a predicate symbol applied to distinct free
variables while Precond and Postcond are logic formulae
about execution states. An execution state is defined by
variable/value bindings in a given execution step. The assertion indicates that in any call to Pred, if Precond holds in
the calling state and the computation of the call succeeds,
then Postcond also holds in the success state. Finally, the
Comp-Props field is used to describe properties of the whole
computation for calls to predicate Pred that meet Precond.
In our application Comp-Props are precisely the resource
usage properties.
For example, the following assertion for a typical append/3
predicate:
: - pred append (A ,B , C )
: ( list (A , num ) , list (B , num ) , var ( C ) )
= > ( list (C , num ) ,
rsize (A , list ( ALb , AUb , num ( ANl , ANu ) ) ) ,
rsize (B , list ( BLb , BUb , num ( BNl , BNu ) ) ) ,
rsize (C ,
list ( ALb + BLb , AUb + BUb ,
num ( min ( ANl , BNl ) , max ( ANu , BNu ) ) ) ) )
+ resource ( steps , ALb +1 , AUb +1) .
states that for any call to predicate append/3 with the first
and second arguments bound to lists of numbers, and the
third one unbound, if the call succeeds, then the third argument will also be bound to a list of numbers. It also states
that an upper bound on the number of resolution steps required to execute any of such calls is AU b + 1, a function
on the length of list A. The rsize terms are the sized types
derived from the regular types, containing variables that represent explicitly lower and upper bounds on the size of terms
and subterms appearing in arguments. See Section 5 for an
overview of the general resource analysis framework and how
sized types are used.
The global non-functional property resource/3 (appearing in the “+” field), is used for expressing resource usages
and follows the schema:
resource(Res Name, Low Arith Expr, Upp Arith Expr)
where Res Name is a user-provided identifier for the resource
the assertion refers to, Low Arith Expr and Upp Arith Expr
are arithmetic functions that map input data sizes to re-
source usage, representing respectively lower and upper bounds
on the resource consumption.
Each assertion can be in a particular status, marked with
the following prefixes, placed just before the pred keyword:
check (indicating the assertion needs to be checked), checked
(it has been checked and proved correct by the system),
false (it has been checked and proved incorrect by the system; a compile-time error is reported in this case), trust
(it provides information coming from the programmer and
needs to be trusted), or true (it is the result of static analysis and thus correct, i.e., safely approximated). The default
status (i.e., if no status appears before pred) is check.
3.2
The XC Assertion Language
The assertions within XC files use instead a different syntax that is closer to standard C notation and friendlier for
C developers. These assertions are transparently translated
into Ciao assertions when XC files are loaded into the tool.
The Ciao assertions output by the analysis are also translated back into XC assertions and added inline to a copy of
the original XC file.
More concretely, the syntax of the XC assertions accepted
by our tool is given by the following grammar, where the
non-terminal hidentifier i stands for a standard C identifier,
hinteger i stands for a standard C integer, and the nonterminal hground-expr i for a ground expression, i.e., an expression of type hexpr i that does not contain any C identifiers that appear in the assertion scope (the non-terminal
hscopei).
hassertioni ::= ‘#pragma’ hstatusi hscopei ‘:’ hbodyi
hstatusi ::= ‘check’ | ‘trust’ | ‘true’ | ‘checked’ | ‘false’
hunary-expr i := hidentifier i
| hinteger i
| ‘sum’ ‘(’ hidentifier i ‘,’ hexpr i ‘,’ hexpr i ‘,’ hexpr i ‘)’
| ‘prod’ ‘(’ hidentifier i ‘,’ hexpr i ‘,’ hexpr i‘,’ hexpr i ‘)’
| ‘power’ ‘(’ hexpr i ‘,’ hexpr i ‘)’
| ‘log’ ‘(’ hexpr i ‘,’ hexpr i ‘)’
| ‘(’ hexpr i ‘)’
| ‘+’ hunary-expr i
| ‘-’ hunary-expr i
| ‘min’ ‘(’ hidentifier i ‘)’
| ‘max’ ‘(’ hidentifier i ‘)’
XC assertions are directives starting with the token #pragma
followed by the assertion status, the assertion scope, and the
assertion body. The assertion status can take several values,
including check, checked, false, trust or true, with the
same meaning as in the Ciao assertions. Again, the default
status is check.
The assertion scope identifies the function the assertion is
referring to, and provides the local names for the arguments
of the function to be used in the body of the assertion. For
instance, the scope biquadCascade(state, xn, N) refers to
the function biquadCascade and binds the arguments within
the body of the assertion to the respective identifiers state,
xn, N. While the arguments do not need to be named in a
consistent way w.r.t. the function definition, it is highly recommended for the sake of clarity. The body of the assertion
expresses bounds on the energy consumed by the function
and optionally contains preconditions (the left hand side of
the ==> arrow) that constrain the argument sizes.
Within the body, expressions of type hexpr i are built from
standard integer arithmetic functions (i.e., +, -, *, /) plus
the following extra functions:
• power(base, exp) is the exponentiation of base by
exp;
• log(base, expr) is the logarithm of expr in base base;
• sum(id, lower, upper, expr) is the summation of
the sequence of the values of expr for id ranging from
lower to upper;
• prod(id, lower, upper, expr) is the product of the
sequence of the values of expr for id ranging from
lower to upper;
• min(arr) is the minimal value of the array arr;
• max(arr) is the maximal value of the array arr.
hscopei ::= hidentifier i ‘(’ ‘)’
| hidentifier i ‘(’ hargumentsi ‘)’
hargumentsi ::= hidentifier i | hargumentsi ‘,’ hidentifier i
hbodyi ::= hprecond i ‘==>’ hcost-boundsi | hcost-boundsi
hprecond i ::= hupper-cond i | hlower-cond i
| hlower-cond i ‘&&’ hupper-cond i
hlower-cond i ::= hground-expr i ‘<=’ hidentifier i
Note that the argument of min and max must be an identifier
appearing in the assertion scope that corresponds to an array
of integers (of arbitrary dimension).
hupper-cond i ::= hidentifier i ‘<=’ hground-expr i
hcost-boundsi ::= hlower-bound i | hupper-bound i
| hlower-bound i ‘&&’ hupper-bound i
hlower-bound i := hexpr i ‘<=’ ‘energy’
hupper-bound i := ‘energy’ ‘<=’ hexpr i
hexpr i := hexpr i ‘+’ hmult-expr i
| hexpr i ‘-’ hmult-expr i
hmult-expr i := hmult-expr i ‘*’ hunary-expr i
| hmult-expr i ‘/’ hunary-expr i
4.
ISA/LLVM IR TO HC IR TRANSFORMATION
In this section we describe briefly the HC IR representation and the transformations into it that we developed in
order to achieve the verification tool presented in Section 2
and depicted in Figure 1. The transformation of ISA code
into HC IR was described in [21]. We provide herein an
overview of the LLVM IR to HC IR transformation.
The HC IR representation consists of a sequence of blocks
where each block is represented as a Horn clause:
< block id > (< params >) : − S1 , . . . , Sn .
Each block has an entry point, that we call the head of
the block (to the left of the : − symbol), with a number
of parameters < params >, and a sequence of steps (the
body, to the right of the : − symbol). Each of these Si
steps (or literals) is either (the representation of) an LLVM
IR instruction, or a call to another (or the same) block.
The analyzer deals with the HC IR always in the same way,
independent of its origin.
LLVM IR programs are expressed using typed assemblylike instructions. Each function is in SSA form, represented
as a sequence of basic blocks. Each basic block is a sequence
of LLVM IR instructions that are guaranteed to be executed
in the same order. Each block ends in either a branching or
a return instruction. In order to represent each of the basic
blocks of the LLVM IR in the HC IR, we follow a similar
approach as in the ISA-level transformation [21]. However,
the LLVM IR includes an additional type transformation as
well as better memory modelling. It is explained in detail
in Appendix 5 of [3]. The main aspects of this process, are
the following:
1. Infer input/output parameters to each block.
2. Transform LLVM IR types into HC IR types.
3. Represent each LLVM IR block as an HC IR block and
each instruction in the LLVM IR block as a literal (Si ).
4. Resolve branching to multiple blocks by creating clauses
with the same signature (i.e., the same name and arguments in the head), where each clause denotes one
of the blocks the branch may jump to.
The translator component is also in charge of translating
the XC assertions to Ciao assertions and back. Assuming
the Ciao type of the input and output of the function is
known, the translation of assertions from Ciao to XC (and
back) is relatively straightforward. The Pred field of the
Ciao assertion is obtained from the scope of the XC assertion to which an extra argument is added representing the
output of the function. The Precond fields are produced directly from the type of the input arguments: to each input
variable, its regular type and its regular type size are added
to the precondition, while the added output argument is declared as a free variable. Finally the Comp-Props field is set
to the usage of the resource energy, i.e., a literal of the form
resource(energy, Lower, Upper) where Lower and Upper
are the lower and upper bounds from the energy consumption specification.
5.
ENERGY CONSUMPTION ANALYSIS
As already mentioned in Section 2, we use an existing
static analysis to infer the energy consumption of XC programs [21]. It is a specialization of the generic resource analysis presented in [35] that uses the instruction-level models
described in [16]. Such generic resource analysis is fully
based on abstract interpretation [9], defining the resource
analysis itself as an abstract domain that is integrated into
the PLAI abstract interpretation framework [27, 33] of CiaoPP,
obtaining features such as multivariance, efficient fixpoints,
and assertion-based verification and user interaction.
In the rest of this section we give an overview of the general resource analysis, using the following append/3 predicate as a running example:
append ([] ,
S, S).
append ([ E | R ] , S , [ E | T ]) : - append (R ,S , T ) .
The first step consists of obtaining the regular type of the
arguments for each predicate. To this end, we use one of
the type analyses present in the CiaoPP system [36]. In our
example, the system infers that for any call to the predicate
append(X, Y, Z) with X and Y bound to lists of numbers
and Z a free variable, if the call succeeds, then Z also gets
bound to a list of numbers. The regular type for representing
“list of numbers” is defined as follows:
listnum := [] | [ num | listnum ].
From this type definition, sized type schemas are derived,
which incorporate variables representing explicitly lower and
upper bounds on the size of terms and subterms. For example, in the following sized type schema (named listnum-s):
listnum-s → listnum(α,β) (num(γ,δ) )
α and β represent lower and upper bounds on the length
of the list, respectively, while γ and δ represent lower and
upper bounds of the numbers in the list, respectively.
In a subsequent phase, these sized type schemas are put
into relation, producing a system of recurrence equations
where output argument sizes are expressed as functions of
input argument sizes.
The resource analysis is in fact an extension of the sized
type analysis that adds recurrence equations for each resource. As the HC IR representation is a logic program,
it is necessary to consider that a predicate can fail or have
more than one solution, so we need an auxiliary cardinality
analysis to get more precise results.
We develop the append example for the simple case of the
resource being the number of resolution steps performed by
a call to append/3 and we will only focus on upper bounds,
rU . For the first clause, we know that only one resolution
step is needed, so:
rU ln(0,0) (n(γX ,δX ) ), ln(αY ,βY ) (n(γY ,δY ) ) ≤ 1
The second clause performs one resolution step plus all the
resolution steps performed by all possible backtrackings over
the call in the body of the clause. This number can be
bounded as a function of the number of solutions. After
setting up and solving these equations we infer that an upper
bound on the number of resolution steps is the (upper bound
on) the length of the input list X plus one. This is expressed
as:
rU ln(αX ,βX ) (n(γX ,δX ) ), ln(αY ,βY ) (n(γY ,δY ) ) ≤ βX + 1
We refer the reader to [35] for a full description of this
analysis and tool.
6.
THE GENERAL RESOURCE USAGE VERIFICATION FRAMEWORK
In this section we describe the general framework for (static)
resource usage verification [22, 24] that we have specialized
in this paper for verifying energy consumption specifications
of XC programs.
The framework, that we introduced in [22], extends the
criteria of correctness as the conformance of a program to
a specification expressing non-functional global properties,
such as upper and lower bounds on execution time, memory,
energy, or user defined resources, given as functions on input
data sizes.
Both program verification and debugging compare the actual semantics [[P ]] of a program P with an intended semantics for the same program, which we will denote by I. This
intended semantics embodies the user’s requirements, i.e.,
it is an expression of the user’s expectations. In the framework, both semantics are given in the form of (safe) approximations. The abstract (safe) approximation [[P ]]α of the
concrete semantics [[P ]] of the program is actually computed
by (abstract interpretation-based) static analyses, and compared directly to the (also approximate) specification, which
is safely assumed to be also given as an abstract value I α .
Such approximated specification is expressed by assertions
in the program. Program verification is then performed by
comparing I α and [[P ]]α .
In this paper, we assume that the program P is in HC
IR form (i.e., a logic program), which is the result of the
transformation of the ISA or LLVM IR code corresponding
to an XC program. As already said, such transformation
preserves the resource consumption semantics, in the sense
that the resource usage information inferred by the static
analysis (and hence the result of the verification process) is
applicable to the original XC program.
Resource usage semantics.
Given a program p, let Cp be the set of all calls to p.
The concrete resource usage semantics of a program p, for a
particular resource of interest, [[P ]], is a set of pairs (p(t̄), r)
such that t̄ is a tuple of data (either simple data such as
numbers, or compound data structures), p(t̄) ∈ Cp is a call
to procedure1 p with actual parameters t̄, and r is a number
expressing the amount of resource usage of the computation
of the call p(t̄). The concrete resource usage semantics can
be defined as a function [[P ]] : Cp 7→ R where R is the set of
real numbers (note that depending on the type of resource
we can take other set of numbers, e.g., the set of natural
numbers).
The abstract resource usage semantics is a set of 4-tuples:
(p(v̄) : c(v̄), Φ, inputp , sizep )
where p(v̄) : c(v̄) is an abstraction of a set of calls. v̄ is a
tuple of variables and c(v̄) is an abstraction representing a
set of tuples of data which are instances of v̄. c(v̄) is an
element of some abstract domain expressing instantiation
states. Φ is an abstraction of the resource usage of the calls
represented by p(v̄) : c(v̄). We refer to it as a resource usage
interval function for p, defined as follows:
• A resource usage bound function for p is a monotonic
arithmetic function, Ψ : S 7→ R∞ , for a given subset
S ⊆ Rk , where R is the set of real numbers, k is
the number of input arguments to procedure p and
R∞ is the set of real numbers augmented with the
special symbols ∞ and −∞. We use such functions to
express lower and upper bounds on the resource usage
of procedure p depending on input data sizes.
• A resource usage interval function for p is an arithmetic function, Φ : S 7→ RI, where S is defined as
before and RI is the set of intervals of real numbers,
such that Φ(n̄) = [Φl (n̄), Φu (n̄)] for all n̄ ∈ S, where
Φl (n̄) and Φu (n̄) are resource usage bound functions
1
Also called predicate in the HC IR.
that denote the lower and upper endpoints of the interval Φ(n̄) respectively for the tuple of input data sizes
n̄. Although n̄ is typically a tuple of natural numbers,
we do not want to restrict our framework. We require
that Φ be well defined so that ∀n̄ (Φl (n̄) ≤ Φu (n̄)).
inputp is a function that takes a tuple of data t̄ and returns
a tuple with the input arguments to p. This function can be
inferred by using the existing mode analysis or be given by
the user by means of assertions. sizep (t̄) is a function that
takes a tuple of terms t̄ and returns a tuple with the sizes of
those data under the size metric described in Section 5.
In order to make the presentation simpler, we will omit
the inputp and sizep functions in abstract tuples, with the
understanding that they are present in all such tuples.
Intended meaning.
The intended approximated meaning I α of a program is
an abstract semantic object with the same kind of tuples:
(p(v̄) : c(v̄), Φ, inputp , sizep ), which is represented by using
Ciao assertions (which are part of the HC IR) of the form:
:- check Pred [: Precond ] + ResUsage.
where p(v̄) : c(v̄) is defined by Pred and Precond, and Φ is defined by ResUsage. The information about inputp and sizep
is implicit in Precond and ResUsage. The concretization of
I α , γ(I α ), is the set of all pairs (p(t̄), r) such that t̄ is a tuple
of terms and p(t̄) is an instance of Pred that meets precondition Precond, and r is a number that meets the condition
expressed by ResUsage (i.e., r lies in the interval defined by
ResUsage) for some assertion.
Example 6.1. Consider the following HC IR program that
computes the factorial of an integer.
fact (N , Fact ) : - N = <0 , Fact =1.
fact (N , Fact ) : - N >0 , N1 is N -1 ,
fact ( N1 , Fact1 ) , Fact is N * Fact1 .
One could use the assertion:
: - check pred fact (N , F )
: (num( N ) , var ( F ) )
= > (num( N ) , num( F ) ,
r s i z e (N , num( Nmin , Nmax ) ) ,
r s i z e (F , num( Fmin , Fmax ) ) )
+ resource ( steps , Nmin +1 , Nmax +1) .
to express that for any call to fact(N,F) with the first argument bound to a number and the second one a free variable,
the number of resolution (execution) steps performed by the
computation is always between Nmin + 1 and Nmax + 1, where
Nmin and Nmax respectively stand for a lower and an upper
bound of N. In this concrete example, the lower and upper
bounds are the same, i.e., the number of resolution steps is
exactly N + 1, but note that they could be different. 2
Example 6.2. The assertion in Example 6.1 captures the
following concrete semantic tuples:
( fact(0, Y), 1 )
( fact(8, Y), 9 )
but it does not capture the following ones:
( fact(N, Y), 1 )
( fact(1, Y), 35 )
the left one in the first line above because it is outside
the scope of the assertion (i.e., N being a variable, it does
not meet the precondition Precond), and the right one because it violates the assertion (i.e., it meets the precondition Precond, but does not meet the condition expressed by
ResUsage). 2
Partial correctness: comparing to the abstract semantics.
Given a program p and an intended resource usage semantics I, where I : Cp 7→ R, we say that p is partially correct
w.r.t. I if for all p(t̄) ∈ Cp we have that (p(t̄), r) ∈ I, where
r is precisely the amount of resource usage of the computation of the call p(t̄). We say that p is partially correct with
respect to a tuple of the form (p(v̄) : cI (v̄), ΦI ) if for all
p(t̄) ∈ Cp such that r is the amount of resource usage of the
computation of the call p(t̄), it holds that: if p(t̄) ∈ γ(p(v̄) :
cI (v̄)) then r ∈ ΦI (s̄), where s̄ = sizep (inputp (t̄)). Finally,
we say that p is partially correct with respect to I α if:
• For all p(t̄) ∈ Cp , there is a tuple (p(v̄) : cI (v̄), ΦI ) in
I α such that p(t̄) ∈ γ(p(v̄) : cI (v̄)), and
• p is partially correct with respect to every tuple in I α .
Let (p(v̄) : c(v̄), Φ) and (p(v̄) : cI (v̄), ΦI ) be tuples expressing an abstract semantics [[P ]]α inferred by analysis
and an intended abstract semantics I α , respectively, such
that cI (v̄) v c(v̄),2 and for all n̄ ∈ S (S ⊆ Rk ), Φ(n̄) =
[Φl (n̄), Φu (n̄)] and ΦI (n̄) = [ΦlI (n̄), ΦuI (n̄)]. We have that:
(1) If for all n̄ ∈ S, ΦlI (n̄) ≤ Φl (n̄) and Φu (n̄) ≤ ΦuI (n̄),
then p is partially correct with respect to (p(v̄) : cI (v̄), ΦI ).
(2) If for all n̄ ∈ S Φu (n̄) < ΦlI (n̄) or ΦuI (n̄) < Φl (n̄), then
p is incorrect with respect to (p(v̄) : cI (v̄), ΦI ).
Checking the two conditions above requires the comparison of resource usage bound functions.
Resource Usage Bound Function Comparison.
Since the resource analysis we use is able to infer different
types of functions (e.g., polynomial, exponential, and logarithmic), it is also desirable to be able to compare all of
these functions.
For simplicity of exposition, consider first the case where
resource usage bound functions depend on one argument.
Given two resource usage bound functions (one of them
inferred by the static analysis and the other one given in
an assertion/specification present in the program), Ψ1 (n)
and Ψ2 (n), n ∈ R the objective of the comparison operation is to determine intervals for n in which Ψ1 (n) >
Ψ2 (n), Ψ1 (n) = Ψ2 (n), or Ψ1 (n) < Ψ2 (n). For this, we
define f (n) = Ψ1 (n) − Ψ2 (n) and find the roots of the equation f (n) = 0. Assume that the equation has m roots,
n1 , . . . , nm . These roots are intersection points of Ψ1 (n) and
Ψ2 (n). We consider the intervals S1 = [0, n1 ), S2 = (n1 , n2 ),
Sm = . . . (nm−1 , nm ), Sm+1 = (nm , ∞). For each interval
Si , 1 ≤ i ≤ m, we select a value vi in the interval. If
f (vi ) > 0 (respectively f (vi ) < 0), then Ψ1 (n) > Ψ2 (n)
(respectively Ψ1 (n) < Ψ2 (n)) for all n ∈ Si .
There exist powerful algorithms for obtaining roots of
polynomial functions. In our implementation we have used
2
Note that the condition cI (v̄) v c(v̄) can be checked using
the CiaoPP capabilities for comparing program state properties such as types.
the GNU Scientific Library [10], which offers a specific polynomial function library that uses analytical methods for
finding roots of polynomials up to order four, and uses numerical methods for higher order polynomials.
We approximate exponential and logarithmic resource usage functions using Taylor series. In particular, for exponential functions we use the following formulae:
ex ≈ Σ∞
n=0
xn
x2
x3
=1+x+
+
+ ...
n!
2!
3!
f or all x
(x ln a)3
(x ln a)2
+
+ ...
2!
3!
In our implementation these series are limited up to order 8.
This decision has been taken based on experiments we have
carried out that show that higher orders do not bring a significant difference in practice. Also, in our implementation,
the computation of the factorials is done separately and the
results are kept in a table in order to reuse them.
Dealing with logarithmic functions is more complex, as
Taylor series for such functions can only be defined for the
interval (−1, 1).
For resource usage functions depending on more than one
variable, the comparison is performed using constraint solving techniques.
ax = ex
ln a
≈ 1 + x ln a +
Safety of the Approximations.
When the roots obtained for function comparison are approximations of the actual roots, we must guarantee that
their values are safe, i.e., that they can be used for verification purposes, in particular, for safely checking the conditions presented above. In other words, we should guarantee
that the error falls on the safe side when comparing the corresponding resource usage bound functions. For this purpose we developed an algorithm for detecting whether the
approximated root falls on the safe side or not, and in the
case it does not fall on the safe side, performing an iterative
process to increment (or decrement) it by a small value until
the approximated root falls on the safe side.
7.
USING THE TOOL: EXAMPLE
As an illustrative example of a scenario where the embedded software developer has to decide values for program
parameters that meet an energy budget, we consider the
development of an equaliser (XC) program using a biquad
filter. In Figure 2 we can see what the graphical user interface of our prototype looks like, with the code of this biquad
example ready to be verified. The purpose of an equaliser
is to take a signal, and to attenuate / amplify different frequency bands. For example, in the case of an audio signal,
this can be used to correct for a speaker or microphone frequency response. The energy consumed by such a program
directly depends on several parameters, such as the sample
rate of the signal, and the number of banks (typically between 3 and 30 for an audio equaliser). A higher number of
banks enables the designer to create more precise frequency
response curves.
Assume that the developer has to decide how many banks
to use in order to meet an energy budget while maximizing
the precision of frequency response curves at the same time.
In this example, the developer writes an XC program where
the number of banks is a variable, say N. Assume also that
the energy constraint to be met is that an application of the
Figure 2: Graphical User Interface of the prototype with the XC biquad program.
biquad program should consume less than 125 millijoules
(i.e., 125000000 nanojoules). This constraint is expressed
by the following check assertion (specification):
#pragma check biquadCascade(state,xn,N) :
(1 <= N) ==> (energy <= 125000000)
where the precondition 1 <= N in the assertion (left hand
side of ==>) expresses that the number of banks should be
at least 1.
Then, the developer makes use of the tool, by selecting the
following menu options, as shown in the right hand side of
Figure 2: check_assertions, for Action Group, res_plai,
for Resource Analysis, mathematica, for Solver, llvm, for
Analysis Level (which will tell the analysis to take the
LLVM IR option by compiling the source code into LLVM IR
and transform into HC IR for analysis) and finally source,
for Output Language (the language in which the analysis /
verification results are shown). After clicking on the Apply
button below the menu options, the analysis is performed,
which infers a lower and an upper bound function for the
consumption of the program. Concretely those bounds are
represented by the following assertion, which is included in
the output of the tool:
#pragma true biquadCascade(state,xn,N) :
(16502087*N + 5445103 <= energy &&
energy <= 16502087*N + 5445103)
In this particular case, both bounds are identical. In other
words, the energy consumed by the program is exactly characterized by the following function, depending on N only:
Ebiquad (N) = 16502087 × N + 5445103 nJ
Then, the verification of the specification (check assertion) is performed by comparing the energy bound functions
above with the upper bound expressed in the specification,
i.e., 125000000, a constant value in this case. As a result,
the two following assertions are produced (and included in
the output file of the tool):
#pragma checked biquadCascade(state,xn,N) :
(1 <= N && N <= 7)
==> (energy <= 125000000)
#pragma false biquadCascade(state,xn,N) :
(8 <= N)
==> (energy <= 125000000)
The first one expresses that the original assertion holds
subject to a precondition on the parameter N, i.e., in order to meet the energy budget of 125 millijoules, the number of banks N should be a natural number in the interval
[1, 7] (precondition 1 <= N && N <= 7). The second one
expresses that the original specification is not met (status
false) if the number of banks is greater or equal to 8.
Since the goal is to maximize the precision of frequency
response curves and to meet the energy budget at the same
time, the number of banks should be set to 7. The developer
could also be interested in meeting an energy budget but this
time ensuring a lower bound on the precision of frequency
response curves. For example by ensuring that N ≥ 3, the
acceptable values for N would be in the range [3, 7].
In the more general case where the energy function inferred by the tool depends on more than one parameter, the
determination of the values for such parameters is reduced
to a constraint solving problem. The advantage of this approach is that the parameters can be determined analytically
at the program development phase, without the need of determining them experimentally by measuring the energy of
expensive program runs with different input parameters.
8.
RELATED WORK
As mentioned before, this work adds verification capabilities to our previous work on energy consumption analysis for
XC/XS1-L [21], which builds on of our general framework
for resource usage analysis [31, 28, 35, 14, 26] and its support for resource verification [22, 24], and the energy models
of [16].
Regarding the support for verification of properties expressed as functions, the closest related work we are aware
of presents a method for comparison of cost functions inferred by the COSTA system for Java bytecode [1]. The
method proves whether a cost function is smaller than another one for all the values of a given initial set of input
data sizes. The result of this comparison is a Boolean value.
However, as mentioned before, in our approach [22, 24] the
result is in general a set of subsets (intervals) in which the
initial set of input data sizes is partitioned, so that the result of the comparison is different for each subset. Also,
[1] differs in that comparison is syntactic, using a method
similar to what was already being done in the CiaoPP system: performing a function normalization and then using
some syntactic comparison rules. Our technique goes beyond these syntactic comparison rules. Moreover, [1] only
covers (generic) cost function comparisons while we have addressed the whole process for the case of energy consumption
verification. Note also that, although we have presented our
work applied to XC programs, the CiaoPP system can also
deal with other high- and low-level languages, including, e.g.,
Java bytecode [29, 26].
In a more general context, using abstract interpretation in
debugging and/or verification tasks has now become well established. To cite some early work, abstractions were used
in the context of algorithmic debugging in [18]. Abstract
interpretation has been applied by Bourdoncle [4] to debugging of imperative programs and by Comini et al. to the
algorithmic debugging of logic programs [7] (making use of
partial specifications in [6]), and by P. Cousot [8] to verification, among others. The CiaoPP framework [5, 12, 14]
was pioneering in many aspects, offering an integrated approach combining abstraction-based verification, debugging,
and run-time checking with an assertion language.
9.
CONCLUSIONS
We have specialized an existing general framework for resource usage verification for verifying energy consumption
specifications of embedded programs. These specifications
can include both lower and upper bounds on energy usage, expressed as intervals within which the energy usage
is supposed to be included, the bounds (end points of the
intervals) being expressed as functions on input data sizes.
Our tool can deal with different types of energy functions
(e.g., polynomial, exponential or logarithmic functions), in
the sense that the analysis can infer them, and the specifications can involve them. We have shown through an example, and using the prototype implementation of our approach within the Ciao/CiaoPP system and for the XC language and XS1-L architecture, how our verification system
can prove whether such energy usage specifications are met
or not, or infer particular conditions under which the specifications hold. These conditions are expressed as intervals of
input data sizes such that a given specification can be proved
for some intervals but disproved for others. The specifica-
tions themselves can also include preconditions expressing
intervals for input data sizes. We have illustrated through
this example how embedded software developers can use this
tool, and in particular for determining values for program
parameters that ensure meeting a given energy budget while
minimizing the loss in quality of service.
10.
ACKNOWLEDGEMENTS
The research leading to these results has received funding
from the European Union 7th Framework Programme under
grant agreement 318337, ENTRA - Whole-Systems Energy
Transparency, Spanish MINECO TIN’12-39391 StrongSoft
and TIN’08-05624 DOVES projects, and Madrid TIC-1465
PROMETIDOS-CM project. We also thank all the participants of the ENTRA project team, and in particular John
P. Gallagher, Henk Muller, Kyriakos Georgiou, Steve Kerrison, and Kerstin Eder for useful and fruitful discussions.
Henk Muller (XMOS Ltd.) also provided benchmarks (e.g.,
the biquad program) that we used to test our tool.
11.
ADDITIONAL AUTHORS
Additional authors: John Smith (The Thørväld Group,
email: [email protected]) and Julius P. Kumquat
(The Kumquat Consortium, email: [email protected]).
12.
REFERENCES
[1] E. Albert, P. Arenas, S. Genaim, I. Herraiz, and
G. Puebla. Comparing cost functions in resource
analysis. In 1st International Workshop on
Foundational and Practical Aspects of Resource
Analysis (FOPARA’09), volume 6234 of Lecture Notes
in Computer Science, pages 1–17. Springer, 2010.
[2] N. Bjørner, F. Fioravanti, A. Rybalchenko, and
V. Senni, editors. Workshop on Horn Clauses for
Verification and Synthesis, July 2014. To appear in
Electronic Proceedings in Theoretical Computer
Science.
[3] N. Bohr, K. Eder, J. P. Gallagher, K. Georgiou,
R. Haemmerlé, M. V. Hermenegildo, B. Kafle,
S. Kerrison, M. Kirkeby, X. Li, U. Liqat,
P. Lopez-Garcia, H. Muller, M. Rhiger, and
M. Rosendahl. Initial Energy Consumption Analysis.
Technical report, FET 318337 ENTRA Project, April
2014. Deliverable 3.2, http://entraproject.eu.
[4] F. Bourdoncle. Abstract debugging of higher-order
imperative languages. In Programming Languages
Design and Implementation’93, pages 46–55, 1993.
[5] F. Bueno, P. Deransart, W. Drabent, G. Ferrand,
M. Hermenegildo, J. Maluszynski, and G. Puebla. On
the Role of Semantic Approximations in Validation
and Diagnosis of Constraint Logic Programs. In Proc.
of the 3rd. Int’l Workshop on Automated
Debugging–AADEBUG’97, pages 155–170, Linköping,
Sweden, May 1997. U. of Linköping Press.
[6] M. Comini, G. Levi, M. C. Meo, and G. Vitiello.
Abstract diagnosis. Journal of Logic Programming,
39(1–3):43–93, 1999.
[7] M. Comini, G. Levi, and G. Vitiello. Declarative
diagnosis revisited. In 1995 International Logic
Programming Symposium, pages 275–287, Portland,
Oregon, December 1995. MIT Press, Cambridge, MA.
[8] P. Cousot. Automatic Verification by Abstract
Interpretation, Invited Tutorial. In Fourth
International Conference on Verification, Model
Checking and Abstract Interpretation (VMCAI),
number 2575 in LNCS, pages 20–24. Springer, January
2003.
[9] P. Cousot and R. Cousot. Abstract Interpretation: a
Unified Lattice Model for Static Analysis of Programs
by Construction or Approximation of Fixpoints. In
ACM Symposium on Principles of Programming
Languages (POPL’77). ACM Press, 1977.
[10] M. Galassi, J. Davies, J. Theiler, B. Gough,
G. Jungman, P. Alken, M. Booth, and F. Rossi. GNU
Scientific Library Reference Manual. Network Theory
Ltd, 2009. Available at
http://www.gnu.org/software/gsl/.
[11] K. Georgiou, S. Kerrison, and K. Eder. A Multi-level
Worst Case Energy Consumption Static Analysis for
Single and Multi-threaded Embedded Programs.
Technical Report CSTR-14-003, University of Bristol,
December 2014.
[12] M. Hermenegildo, G. Puebla, and F. Bueno. Using
Global Analysis, Partial Specifications, and an
Extensible Assertion Language for Program
Validation and Debugging. In K. R. Apt, V. Marek,
M. Truszczynski, and D. S. Warren, editors, The Logic
Programming Paradigm: a 25–Year Perspective, pages
161–192. Springer-Verlag, July 1999.
[13] M. Hermenegildo, G. Puebla, F. Bueno, and P. L.
Garcı́a. Integrated Program Debugging, Verification,
and Optimization Using Abstract Interpretation (and
The Ciao System Preprocessor). Science of Computer
Programming, 58(1–2), 2005.
[14] M. Hermenegildo, G. Puebla, F. Bueno, and
P. Lopez-Garcia. Integrated Program Debugging,
Verification, and Optimization Using Abstract
Interpretation (and The Ciao System Preprocessor).
Science of Computer Programming, 58(1–2):115–140,
October 2005.
[15] M. V. Hermenegildo, F. Bueno, M. Carro, P. López,
E. Mera, J. Morales, and G. Puebla. An Overview of
Ciao and its Design Philosophy. TPLP,
12(1–2):219–252, 2012.
http://arxiv.org/abs/1102.5497.
[16] S. Kerrison and K. Eder. Energy modelling of software
for a hardware multi-threaded embedded
microprocessor. ACM Transactions on Embedded
Computing Systems (TECS), 2015. To appear.
[17] C. Lattner and V. Adve. LLVM: A compilation
framework for lifelong program analysis and
transformation. In Proc. of the 2004 International
Symposium on Code Generation and Optimization
(CGO), pages 75–88. IEEE Computer Society, March
2004.
[18] Y. Lichtenstein and E. Y. Shapiro. Abstract
algorithmic debugging. In R. A. Kowalski and K. A.
Bowen, editors, Fifth International Conference and
Symposium on Logic Programming, pages 512–531,
Seattle, Washington, August 1988. MIT.
[19] U. Liqat, K. Georgiou, S. Kerrison, P. Lopez-Garcia,
M. V. Hermenegildo, J. P. Gallagher, and K. Eder.
Inferring Energy Consumption at Different Software
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
Levels: ISA vs. LLVM IR. Technical report, ENTRA
Project, April 2014. Appendix D3.2.4 of Deliverable
D3.2. Available at http://entraproject.eu.
U. Liqat, S. Kerrison, A. Serrano, K. Georgiou,
P. Lopez-Garcia, N. Grech, M. Hermenegildo, and
K. Eder. Energy Consumption Analysis of Programs
based on XMOS ISA-level Models. In Proceedings of
LOPSTR’13, 2014.
U. Liqat, S. Kerrison, A. Serrano, K. Georgiou,
P. Lopez-Garcia, N. Grech, M. Hermenegildo, and
K. Eder. Energy Consumption Analysis of Programs
based on XMOS ISA-level Models. In Proceedings of
the 23rd International Symposium on Logic-Based
Program Synthesis and Transformation (LOPSTR’13),
2014.
P. López-Garcı́a, L. Darmawan, and F. Bueno. A
Framework for Verification and Debugging of
Resource Usage Properties. In Technical
Communications of ICLP, volume 7 of LIPIcs, pages
104–113. Schloss Dagstuhl, July 2010.
P. Lopez-Garcia, L. Darmawan, F. Bueno, and
M. Hermenegildo. Interval-Based Resource Usage
Verification: Formalization and Prototype. In R. P.
na, M. Eekelen, and O. Shkaravska, editors,
Foundational and Practical Aspects of Resource
Analysis, volume 7177 of LNCS, pages 54–71.
Springer-Verlag, 2012.
P. Lopez-Garcia, L. Darmawan, F. Bueno, and
M. Hermenegildo. Interval-Based Resource Usage
Verification: Formalization and Prototype. In R. P.
na, M. Eekelen, and O. Shkaravska, editors,
Foundational and Practical Aspects of Resource
Analysis. Second Iternational Workshop FOPARA
2011, Revised Selected Papers, volume 7177 of Lecture
Notes in Computer Science, pages 54–71.
Springer-Verlag, 2012.
M. Méndez-Lojo, J. Navas, and M. Hermenegildo. A
Flexible (C)LP-Based Approach to the Analysis of
Object-Oriented Programs. In LOPSTR 2007, number
4915 in LNCS, pages 154–168. Springer-Verlag,
August 2007.
M. Méndez-Lojo, J. Navas, and M. Hermenegildo. A
Flexible (C)LP-Based Approach to the Analysis of
Object-Oriented Programs. In 17th International
Symposium on Logic-based Program Synthesis and
Transformation (LOPSTR 2007), number 4915 in
LNCS, pages 154–168. Springer-Verlag, August 2007.
K. Muthukumar and M. Hermenegildo. Compile-time
Derivation of Variable Dependency Using Abstract
Interpretation. Journal of Logic Programming,
13(2/3):315–347, July 1992.
J. Navas, M. Méndez-Lojo, and M. Hermenegildo. Safe
Upper-bounds Inference of Energy Consumption for
Java Bytecode Applications. In The Sixth NASA
Langley Formal Methods Workshop (LFM 08), April
2008. Extended Abstract.
J. Navas, M. Méndez-Lojo, and M. Hermenegildo.
User-Definable Resource Usage Bounds Analysis for
Java Bytecode. In BYTECODE’09, volume 253 of
ENTCS, pages 6–86. Elsevier, March 2009.
J. Navas, E. Mera, P. López-Garcı́a, and
M. Hermenegildo. User-Definable Resource Bounds
[31]
[32]
[33]
[34]
[35]
[36]
[37]
Analysis for Logic Programs. In Proc. of ICLP’07,
volume 4670 of LNCS, pages 348–363. Springer, 2007.
J. Navas, E. Mera, P. López-Garcı́a, and
M. Hermenegildo. User-Definable Resource Bounds
Analysis for Logic Programs. In 23rd International
Conference on Logic Programming (ICLP’07), volume
4670 of Lecture Notes in Computer Science. Springer,
2007.
G. Puebla, F. Bueno, and M. Hermenegildo. An
Assertion Language for Constraint Logic Programs. In
Analysis and Visualization Tools for Constraint
Programming, number 1870 in LNCS, pages 23–61.
Springer-Verlag, 2000.
G. Puebla and M. Hermenegildo. Optimized
Algorithms for the Incremental Analysis of Logic
Programs. In International Static Analysis Symposium
(SAS 1996), number 1145 in LNCS, pages 270–284.
Springer-Verlag, September 1996.
A. Serrano, P. Lopez-Garcia, and M. Hermenegildo.
Resource Usage Analysis of Logic Programs via
Abstract Interpretation Using Sized Types. TPLP,
ICLP’14 Special Issue, 14(4-5):739–754, 2014.
A. Serrano, P. Lopez-Garcia, and M. Hermenegildo.
Resource Usage Analysis of Logic Programs via
Abstract Interpretation Using Sized Types. Theory
and Practice of Logic Programming, 30th Int’l.
Conference on Logic Programming (ICLP’14) Special
Issue, 14(4-5):739–754, 2014.
C. Vaucheret and F. Bueno. More Precise yet Efficient
Type Inference for Logic Programs. In International
Static Analysis Symposium, volume 2477 of Lecture
Notes in Computer Science, pages 102–116.
Springer-Verlag, September 2002.
D. Watt. Programming XC on XMOS Devices. XMOS
Limited, 2009.
| 6cs.PL
|
arXiv:1402.0240v4 [cs.DS] 26 Mar 2016
Graph Cuts with Interacting Edge Costs –
Examples, Approximations, and Algorithms
Stefanie Jegelka and Jeff Bilmes
Abstract
We study an extension of the classical graph cut problem, wherein we
replace the modular (sum of edge weights) cost function by a submodular
set function defined over graph edges. Special cases of this problem have
appeared in different applications in signal processing, machine learning,
and computer vision. In this paper, we connect these applications via
the generic formulation of “cooperative graph cuts”, for which we study
complexity, algorithms, and connections to polymatroidal network flows.
Finally, we compare the proposed algorithms empirically.
1
Introduction
Graphs have been a ubiquitous modeling tool in areas as diverse as operations research, signal processing and machine learning. Graphical representations reveal
structure in the problem that is often the key to obtaining efficient algorithms
for real-world data analysis problems. As a prominent example, the Minimum
(s, t)-Cut problem underlies important problems in low-level computer vision
[Boykov and Veksler, 2006] (e.g., image segmentation and regularization), probabilistic inference in graphical models [Greig et al., 1989, Ramalingam et al.,
2008], and for representing pseudo-boolean functions in computer vision and
constraint satisfaction problems [Kolmogorov and Boykov, 2005, Ramalingam
et al., 2011, Z̆ivný et al., 2009]. The reduction to cuts has had a tremendous
practical impact.
The algorithmic efficiency of cuts comes at the price of reduced modeling
power: graph cuts model problems that correspond to a special class of functions
(a sub-class of submodular functions defined on the nodes of the graph [Z̆ivný
et al., 2009]). Section 2 lists applications that do not fall into this category.
Motivated by these limitations, this paper studies a non-additive generalization
of Minimum (s, t)-Cut, where the cut cost function is a submodular set function
over graph edges.
A set function f : 2E → R defined on subsets of a ground set E is submodular
if it satisfies diminishing marginal costs: for all sets A ⊂ B ⊆ E and e ∈ E \ B,
it holds that
f (A ∪ {e}) − f (A) ≥ f (B ∪ {e}) − f (B).
1
(1)
This generalization – we refer to it as Cooperative Cut – introduces dependencies
between edges, and expresses a wider set of functions than graph cuts.
1.1
Minimum Cut and Minimum Cooperative Cut
The Minimum (s, t)-Cut problem is defined as follows.
Problem 1 (Minimum (s, t)-Cut). Given a weighted graph G = (V,
PE, w) with
terminal nodes s, t ∈ V, find a cut C ⊆ E of minimum cost w(C) = e∈C w(e).
A cut is a set of edges whose removal disconnects all paths between s and t.
We assume throughout that w ≥ 0. Many very efficient algorithms are
known to solve Minimum (s, t)-Cut; the reader is referred to [Ahuja et al.,
1993, Schrijver, 2004] for an overview.
P
In graph cuts, the cost of any given cut C ⊆ E is a sum w(C) = e∈C w(e)
of edge weights. This function is said to be modular or, equivalently, additive on
the edge set E. It implies two important modeling characteristics for graph cuts:
First, the nonnegativity of the weights can only penalize two nodes for being
separated — this introduces a form of positive correlation between nodes, also
hence this is sometimes referred to as attractive potentials in the computer vision
community. Second, modular edge weights express interactions ofPonly two
nodes at a time. That is, the additive contribution w(e) to the cost e∈C w(e)
of a cut C by a given edge e ∈ C is the same regardless of the cut in which the
edge e is considered.
Several applications, however, are more accurately modeled when allowing
non-additive interactions between edge weights. We survey some examples and
applications in Section 2. These examples are captured when replacing the
modular cost function w by a nonnegative and nondecreasing submodular set
function over graph edges. The definition (1) implies that with submodular
edge weights, the cost of an additional edge depends on which other edges are
contained in the cut. This non-additivity allows specific long-range dependencies
between multiple (pairs of) nodes simultaneously that cannot be expressed by
graph cuts. These observations motivate the definition of cooperative graph cuts.
Problem 2 (Minimum cooperative cut (MinCoopCut)). Given a graph G =
(V, E, f ) with terminal nodes s, t ∈ V and a nonnegative, monotone nondecreasing submodular function f : 2E → R+ defined on subsets of edges, find an
(s, t)-cut C ⊆ E of minimum cost f (C).
A set function f is nondecreasing if A ⊆ B ⊆ E implies that f (A) ≤ f (B).
MinCoopCut is a constrained submodular minimization problem:
minimize f (C)
subject to C ⊆ E is an (s, t)-cut in G.
(2)
As cooperative cuts employ the same graph structures as standard graph cuts,
they easily integrate into and extend many of the applications of graph cuts.
We note, however, that the graph G need not have any relationship to the submodular function f other than the fact that the edges of the graph G constitute
the ground set of f .
2
1.2
Relation to the literature
Cooperative graph cuts relate to the recent literature in two aspects. First, a
number of models in signal processing, computer vision, security, and machine
learning are special cases of cooperative cuts, as is discussed in Section 2.
Second, recent interest has emerged in the literature regarding the implications of extending classical combinatorial problems (such as Shortest Path,
Minimum Spanning Tree, or Set Cover) from a sum-of-weights to submodular cost functions [Svitkina and Fleischer, 2008, Iwata and Nagano, 2009, Goel
et al., 2009, 2010, Jegelka and Bilmes, 2011a,b, Koufogiannakis and Young,
2009, Hassin et al., 2007, Zhang et al., 2011, Baumann et al., 2013]. None of
this work, however, has addressed cuts. In this work, we provide lower and
upper bounds on the approximation factor of MinCoopCut.
One approach to solving MinCoopCut is via relaxations. For Minimum
(s, t)-Cut, a celebrated result of [Ford and Fulkerson, 1956, Dantzig and Fulkerson, 1955] states that the dual of the linear programming relaxation is a Maximum Flow problem, and that their optimal values coincide. We refer to the
ratio between the maximum flow value (i.e., the optimal value of the relaxation),
and the optimal value of the discrete cut, as the flow-cut gap. For Minimum
(s, t)-Cut, this ratio is one. In Section 4, we formulate a convex relaxation of
MinCoopCut whose dual problem is a generalized flow problem, where submodular capacity constraints are placed not only on individual edges but on
arbitrary sets of edges simultaneously. The flow-cut gap for this problem can be
on the order of n, the number of nodes. In contrast, the related polymatroidal
maximum flow problem [Lawler and Martel, 1982, Hassin, 1982] (defined in Section 5.1.3) still has a flow-cut gap of one. Polymatroidal flows are equivalent to
submodular flows, and have recently gained attention for modeling information
flow in wireless networks [Kannan et al., 2011, Kannan and Viswanath, 2011,
Chekuri et al., 2012]. Their dual problem is a minimum cut problem where the
edge weights are defined by a convolution of local submodular functions [Lovász,
1983]. Such convolutions are generally not submodular (see Equation (28)).
1.3
Summary of main contributions
In this paper, we survey diverse examples of cooperative cuts in different applications, and provide a detailed theoretical analysis:
√
• We show a lower bound of Ω( n) on the approximation factor for the
general MinCoopCut problem.
• We analyze two families of approximation algorithms. The first relies
on substituting the submodular cost function by a tractable approximation. The second family consists of rounding algorithms that build on the
relaxation of the problem. Both families contain algorithms that use a
partitioning of edges into node incidence sets, but in different ways.
• We provide a lower bound of n − 1 on the flow-cut gap, and relate it to
different families of submodular functions.
3
In Section 5.1.3 we draw connections to polymatroidal flows [Lawler and Martel,
1982, Hassin, 1982]. The non-additive cut cost function used in the resulting
approximation is solvable exactly and may itself be interesting to consider as
an exactly solvable class of e.g. higher-order potentials in computer vision. The
paper concludes with a discussion of open problems. In particular, the results
of this paper motivate a wider and more refined study of the complexity and
expressive power of non-additive graph cut problems.
The paper is structured as follows. In Section 2 we discuss various instances
of cooperative cuts and their properties. The complexity of MinCoopCut is
addressed in Section 3, convex relaxations in Section 4, and algorithmic approaches in Section 5.
1.4
Notation and technical preliminaries
Throughout this paper, we are given a directed graph1 G = (V, E) with n nodes
and m edges, and terminal nodes s, t ∈ V. The function f : 2E → R+ is submodular and monotone nondecreasing, where by 2E we denote the power set of
E. We assume f to be normalized, i.e., f (∅) = 0. Equivalently to Definition (1),
the function f is submodular if for all A, B ⊆ E, it holds that
f (A) + f (B) ≥ f (A ∪ B) + f (A ∩ B).
(3)
The function f generalizes commonly used modular edge weight functions w :
E → R+ that satisfy Equation 3 with equality. We denote the marginal cost of
an element e ∈ E with respect to a set A ⊆ E by f (e | A) , f (A∪{e})−f (A). A
function f (A) = g(w(A)) for a nonnegative modular function w and a concave
scalar function g is always submodular.
For any node v ∈ V, let δ + (v) = {(v, u) ∈ E} be the set of edges directed
out of v, and δ − (v) = {(u, v) ∈ E} be the set of edges into v. Together, these
two directed sets form the (undirected) incidence set δ(v) = δ + (v) ∪ δ − (v).
These definitions directly extend to sets of nodes: for a set S ⊆ V of nodes,
δ + (S) = {(v, u) ∈ E : v ∈ S, u ∈
/ S} is the set of edges leaving S. Without loss
of generality, we assume all graphs are simple.
The Lovász extension f˜ : [0, 1]m → R of the submodular function f is
its lower convex envelope and is defined as follows [Lovász, 1983]. Given a
vectorPx ∈ [0, 1]m , we can uniquely decompose x into its level sets {Bj }j as
x =
j λj χBj where B1 ⊂ B2 ⊂ . . . are distinct subsets. Here and in the
following, χB ∈ [0, 1]m is the characteristic vectorPof the set B, with χB (e) = 1 if
e ∈ B, and χB (e) = 0 otherwise. Then f˜(x) = j λj f (Bj ). This construction
illustrates that f˜(χB ) = f (B) for any set B. The Lovász extension can be
computed by sorting the entries of the argument x in O(m log m) time and
calling f m times.
A separator of a submodular function f is a set S ⊆ E with the property
that f (S) + f (E \ S) = f (E), implying that for any B ⊆ E, f (B) = f (B ∩
S) + f (B ∩ (E \ S)). If S is a minimal separator, then we say that f couples all
1 Undirected
graphs can be reduced to bidirected graphs.
4
0.1
x1
s
9.9
0.1
9.9
pP
Let f (A) =
e∈A w(e), so
qP
h(X) =
e∈δ + (X) w(e).
x2
0.1
Then h is not submodular:
t
h({s, x1 }) + h({s, x2 }) =
√
19.9 +
√
0.2
<
√
2 10 = h({s}) + h({s, x1 , x2 })
Figure 1: The node function induced by a cooperative cut is in general not
submodular. The above h violates inequality (3) for A = {s, x1 }, B = {s, x2 }
but satisfies it (strictly) for A = {t, x1 }, B = {t, x2 }.
edges in S. For the edges within a minimal separator, f is strictly subadditive:
P
e∈S f (e) > f (S) for |S| > 1. That means, the joint cost of this set of edges is
smaller than the sum of their individual costs.
1.5
Node functions induced by submodular edge weights
Every cost function f on graph edges defines a cut function h : 2V → R+ on
sets X ⊆ V of nodes:
h(X) , f (δ + (X)).
(4)
It is well known that if f is a (modular) sum of nonnegative edge weights, then
h is submodular [Schrijver, 2004]. In fact, the following is true:
Proposition 1. The function f is a non-negative modular function on the edge
set if and only if h(X) = f (δ + (X)) is a submodular function on the nodes for
all edge-induced sub-graphs of G = (V, V × V).
If, however, f is an arbitrary nondecreasing submodular function, then this is
not always the case, as Figure 1 illustrates. Proposition 2, proven in Appendix B,
summarizes some key properties of h.
Proposition 2. Let h : 2V → R, h(X) , f (δ + (X)) be the node function induced
by a cooperative cut with nonnegative nondecreasing submodular cost function
f . Then:
1. h is not always submodular, and
2. h is subadditive, i.e., h(A) + h(B) ≥ h(A ∪ B) for any A, B ⊆ V.
The non-submodularity of h shows that cooperative cuts are strictly more
general than (modular-weight) graph cuts. In some cases, the function h is submodular. One obvious sufficient condition is that f is nonnegative and modular,
but this condition is not necessary as shown in the following.
5
Proposition 3. Let f be monotone submodular and permutation symmetric in
the sense that f (A) = f (σ(A)) for any permutation σ of the set E. If G is a
complete graph, then h is submodular.
Proof. Symmetry implies that f is of the form f (A) = g(|A|) for a scalar function g. Submodularity of f implies that there is always a function g 0 that
interpolates g on R \ {0, 1, . . . , m}, i.e., f (A) = g 0 (|A|) = g(|A|), and g 0 is a
piecewise linear concave function. Let EXY be the edges between sets X and
Y . The submodularity of h is identical to the condition that for all X ⊆ V,
x, y ∈
/ X, it holds that
h(X ∪ x) + h(X ∪ y) ≥ h(X) + h(X ∪ x ∪ y).
(5)
Let R = V \ (X ∪ x ∪ y). By the concavity and monotonicity of g 0 we have
h(X) + h(X ∪ x ∪ y)
= g 0 (|EXR | + |EXx | + |EXy |) + g 0 (|EXR | + |ERx | + |ERy |)
= g 0 (|X||R| + 2|X|) + g 0 (|X||R| + 2|R|)
≤ 2g 0 (|X||R| + |X| + |R|)
≤ g 0 (|EXR | + |EXy | + |ERx | + |Exy |) + g 0 (|EXR | + |EXx | + |ERy | + |Exy |)
= h(X ∪ x) + h(X ∪ y),
and hence submodularity of h follows.
Note that if G is not complete, then h might no longer be submodular. An
exact (possibly algebraic or graph-theoretic) characterization of the conditions
on G and f that imply submodularity of h is currently an open problem.
2
Motivation and special cases
We begin by surveying special cases of cooperative cuts from applications in
signal processing, machine learning, and computer vision. Notably, some of
these applications lead to submodular node functions h as defined in (4) and are
hence polynomial-time solvable, while for others h is not submodular. We first
discuss the latter case which motivated this paper, and then special submodular
cases. Additional special cases are discussed in Section 6.
2.1
General, non-submodular examples
Image segmentation. The classical task of segmenting an image into a foreground object and its background is commonly formulated as a maximum a
posteriori (MAP) inference problem in a Markov Random Field (MRF) or Conditional Random Field (CRF). If the potential functions of the random field are
submodular functions (of the node variables), then the MAP solution can be
computed efficiently via the minimum cut in an auxiliary graph [Greig et al.,
1989, Boykov and Jolly, 2001, Kolmogorov and Zabih, 2004].
6
While these graph cut models have been very successful in computer vision,
they suffer from known shortcomings. For example, the cut model implicitly
penalizes the length of the object boundary, or equivalently the length of a corresponding graph cut around the object. As a result, MAP solutions (minimum
cuts) tend to truncate fine parts of an object (such as branches of trees, or
animal hair), and to neglect carving out holes (such as the mesh grid of a fan,
or written letters on paper). This tendency is aggravated if the image has regions of low contrast, where local information is insufficient to determine correct
object boundaries.
A solution to both of these problems is proposed in [Jegelka and Bilmes,
2011a]. It relies on the continuation of “obvious” object boundaries: one aims
to reduce the cut cost if the cut consists of edges (pairs of pixels) with similar
appearance. This aim is impossible to model with a modular-cost graph cut
where edges are independent. Hence, Jegelka and Bilmes [2011a] replace the
graph cut by a cooperative cut that lowers the cost for sets of similar edges.
Specifically, the edges in the image graph are partitioned into groups Si of
similar edges, where similarity is defined via the adjacent pixels (nodes), and
the cut cost is
k
X
f (C) =
gi (w(C ∩ Si )),
(6)
i=1
P
where the gi are increasing, strictly concave functions, and w(C) = e∈C w(e)
is a sum of nonnegative weights.
From the viewpoint of graphical models, the function h induced by (6) is
a higher-order potential, i.e., a polynomial of degree much larger than 2. The
model (6) also applies to multi-label (scene labeling) problems and other computer vision tasks [Kohli et al., 2013, Silberman et al., 2014, Taniai et al., 2015].
An alternative cooperative cut function has been studied to improve image
segmentation results:
f (C) = max w(e).
(7)
e∈C
Contrary to the cost function (6), the function (7) couples all edges in the grid
graph uniformly, without any similarity constraints or grouping. As a result,
the cost of any long cut is discounted. Sinop and Grady [2007] and Allène
et al. [2009] derive this function as the `∞ norm of the (gradient) vector of pixel
differences; this vector is the edge indicator vector y in the relaxation we define
in Section 4. Conversely, the relaxation of the cooperative cut problem leads to
new, possibly non-uniform and group-structured regularization terms [Jegelka,
2012].
Label Cuts. In computer security, attack graphs are state graphs modeling
the steps of an intrusion. Each transition edge is labeled by an atomic action a,
and blocking an action a blocks the set of all associated edges Sa ⊆ E that carry
label a. To prevent an intrusion, one must separate the initial state s from the
goal state t by blocking (cutting) appropriate edges. The cost of cutting a set
of edges is the cost of blocking the associated actions (labels), and paying for
7
one action a accounts for all edges in Sa . If each action has a cost c(a), then a
minimum label cut that minimizes the submodular cost function
X
f (C) =
c(a) min{1, |C ∩ Sa |}
(8)
a
indicates the lowest-cost prevention of an intrusion [Jha et al., 2002].
Sparse separators of Dynamic Bayesian Networks. A graphical model
G = (V, E) defines a family of probability distributions. It has a node vi for
each random variable xi , and any represented distribution
p(x) must factor
Q
with respect to the edges of the graph as p(x) ∝ (vi ,vj )∈E ψij (xi , xj ). A dynamic graphical model (DGM) [Bilmes, 2010] consists of three template parts:
a prologue Gp = (Vp , Ep ), a chunk Gc = (Vc , Ec ) and an epilogue Ge = (Ve , Ee ).
Given a length τ , an unrolling of the template is a model that begins with Gp
on the left, followed by τ + 1 repetitions of the “chunk” part Gc and ending in
the epilogue Ge .
To perform inference efficiently, a periodic section of the partially unrolled
model is identified on which an effective inference strategy (e.g., a graph triangulation, an elimination order, or an approximate inference method) is developed
and then repeatedly used for the complete duration of the model unrolled to
any length. This periodic section has boundaries corresponding to separators
in the original model [Bilmes, 2010] which are called the interface separators.
Importantly, the efficiency of any inference algorithm derived within the periodic section depends critically on properties of the interface, since the variables
within must become a clique.
In general, the computational cost of inference is lower bounded, and the
memory cost of inference is exactly given, by the size of the joint state space of
the interface variables. A “small” separator corresponds to a minimum vertex
cut in the graphical model, where the cost function measures the size of the
joint state space. Vertex cuts can be rephrased as standard edge cuts. Often,
a modular cost function suffices for good results. Sometimes, however, a more
general cost function is needed. In Bilmes and Bartels [2003], for example
(which is our original motivation for studying MinCoopCut), a state space
function that considers deterministic relationships between variables is found to
significantly decrease inference costs.
An example of a function that respects determinism is the following. In
a Bayesian network possessing determinism, let D be the subset of fully deterministic nodes. That means any xi ∈ D is a deterministic function of the
variables corresponding to its parent nodes par(i) meaning p(xi |xpar(i) ) = 1[xi =
g(xpar(i) )] for some deterministic function g. Let Di be the state space of variable
xi . Furthermore, given a set A of variables, let AD = {xi ∈ A ∩ D | par(i) ⊆ A}
be its subset of fully determined variables. If the state space of a deterministic
variable is not restricted by fixing a subset of its parents,
Q then the function
measuring the state space of a set of variables A is f (A) = xi ∈A\AD |Di |. The
logarithm of this function is a submodular function, and therefore the problem
of finding a good separator is a cooperative (vertex) cut problem. In fact, this
8
function is a lower bound on the computational complexity of inference, and
corresponds exactly to the memory complexity since memory need be retained
only at the boundaries between repeated sections in a DGM.
More generally, a similar slicing mechanism applies for partitioning a graph
for use on a parallel computer — we may seek separators that require little
information to be transferred from one processor to another. A reasonable
proxy for such “compressibility” is the entropy of a set of random variables, a
well-known submodular function. The resulting optimization problem of finding
a minimum-entropy separator is a cooperative cut that is different from any
known special cases.
Robust optimization. Assume we are given a graph where the weight of
each edge e ∈ E is noisy and distributed as N (µ(e), σ 2 (e)) for nonnegative
mean weights µ(e). The noise on different edges is independent, and the cost
of a cut is the sum of edge weights of an unknown draw from that distribution.
In such a case, we might want to not only minimize the expected cost, but also
take the variance into consideration. This is the aim in mean-risk minimization
(which is equivalent to a probability tail model or value-at-risk model), where
we aim to minimize
sX
X
σ 2 (e).
(9)
f (C) =
µ(e) + λ
e∈C
e∈C
This is a cooperative cut, and this special case admits an FPTAS [Nikolova,
2010].
2.2
Special cases that lead to submodular functions h
Curiously, some instances of cooperative cuts provably yield submodular node
functions h and are hence easier. In the first two examples below, f is defined
over edges in a complete graph and is symmetric. Here, symmetry is meant in
the sense of [Vondrák, 2013] and Proposition 3, the function is indifferent to
permutations of the arguments.
Higher-order potentials in computer vision. A number of higher-order
potentials (pseudo-boolean functions) from computer vision, i.e., potential functions that introduce dependencies between more than two variables at a time,
can be reformulated as cooperative cuts. As an example, P n Potts functions
[Kohli et al., 2009a] and robust P n potentials [Kohli et al., 2009b] bias image
labelings to assign the same label to larger patches of pixels (of uniform appearance). The potential is low if all nodes in a given patch take the same label,
and high if a large fraction deviates from the majority label. These potential
functions correspond to a complete graph with a cooperative cut cost function
f (C) = g(|C|),
9
(10)
for a concave function g. The larger the fraction of deviating nodes, the more
edges are cut between labels, leading to a higher penalty. The function g makes
this robust by capping the maximum penalty.
Regularization and Total Variation. A popular regularization term in signal processing, and in particular for image denoising, has been the Total Variation (TV) and its discretization [Rudin et al., 1992]. The setup commonly
includes a pixel variable (say xj or xij ) corresponding to each pixel or node
in the image graph G, and an objective that consists of a loss term and the
regularization. The discrete TV for variables xij corresponding to pixels vij in
an M × M grid with coordinates i, j is given as
TV1 (x) =
M q
X
i,j=1
(xi+1,j − xij )2 + (xi,j+1 − xij )2 .
(11)
If x is constrained to be a {0, 1}-valued vector, then this is an instance of
cooperative cut — the pixels valued at unity correspond to the selected
elements
P p
X ⊆ V, and the edge submodular function corresponds to f (C) = ij C ∩ Sij
for C ⊆ E where Sij = {(vi+1,j vij ), (vi,j+1 , vij )} ⊆ E ranges over all relevant
neighborhood pairs of edges. Discrete versions of other variants of total variation
are also cooperative cuts. Examples include the combinatorial total variation
of Couprie et al. [2011]:
Xs X
TV2 (x) =
νi2 (xi − xj )2 ,
(12)
i
(vi ,vj )∈E
and the submodular oscillations in [Chambolle and Darbon, 2009], for instance,
X
TV3 (x) =
max{xi,j , xi+1,j , xi,j+1 , xi+1,j+1 }
1≤i,j≤M
=
X
1≤i,j≤M
− min{xi,j , xi+1,j , xi,j+1 , xi+1,j+1 }
max
`,r∈Uij ×Uij
|x` − xk |,
(13)
(14)
where for notational convenience we used Uij = {(i, j), (i + 1, j), (i, j + 1), (i +
1, j + 1)}. The latter term (14), like P n potentials, corresponds to a symmetric submodular function on a complete graph, and both (10) and (14) lead to
submodular node functions h(X).
Approximate submodular minimization. Graph cuts have been useful
optimization tools but cannot represent any arbitrary set function, not even all
submodular functions [Z̆ivný et al., 2009]. But, using a decomposition theorem
by Cunningham [1983], any submodular function can be phrased as a cooperative graph cut. As a result, any fast algorithm that computes an approximate
minimum cooperative cut can be used for (faster) approximate minimization of
certain submodular functions [Jegelka et al., 2011].
10
2.3
General Cooperative Cut
The above examples indicate that certain special cases reduce MinCoopCut to
a submodular minimization problem, or result in a simpler optimization problem than the general form of MinCoopCut with an arbitrary non-negative
submodular cost function f and an arbitrary graph G. We will discuss such
examples in further detail in Section 6.
Yet, there are many reasons for studying the optimization landscape of general MinCoopCut. First, not all of the examples in Section 2.1 fall into one
of the “easier” classes. Second, applications often benefit from learning the
submodular function rather than specifying it a priori. While learning a submodular function is hard in general [Goemans et al., 2009, Balcan and Harvey,
2012], learning can be practically viable for sub-classes of submodular functions
[Lin and Bilmes, 2012, Fix et al., 2013, Stobbe and Krause, 2012, Tschiatschek
et al., 2014]. Applications such as those in computer vision [Jegelka and Bilmes,
2011a] would likely benefit from learning too, but the resulting cooperative cut
problem would not necessarily fall into an easy sub-class. Moreover, applications
often benefit from combining different objective terms. In computer vision, this
may be a cooperative cut potential for encouraging object boundaries of homogeneous appearance, combined with terms that express the data likelihood for
different object classes, terms that encourage the coherence of uniform patches
of pixels, e.g. the potentials in [Kohli et al., 2009b], and possibly others. All
of these terms are cooperative cuts, but together they quickly exceed special
sub-classes of the problem.
In fact, empirical results enabled by general algorithms may hint at the
existence of further, easier special cases that help map the complexity landscape.
The empirical results in Section 7, for example, are based on the results in this
paper and open up questions for further study. Hence, this paper deliberately
takes a general viewpoint to connect the many examples from a spectrum of
areas to a common optimization problem.
3
Complexity and lower bounds
In this section, we address the hardness of the general MinCoopCut problem.
Assuming
that the cost function is given as an oracle, we show a lower bound
√
of Ω( n) on the approximation factor. In addition, we include a proof of NPhardness. NP-hardness holds even if the cost function is completely known and
polynomially computable and representable.
Our results complement known lower bounds for related combinatorial problems having submodular cost functions. Table 1 provides an overview of known
results from the literature. In addition, Zhang et al. [2011] show a lower bound
for the special case of Minimum Label Cut via a reduction from Minimum
1−(log log m̄)−c
Label Cover. Their lower bound is 2(log m̄)
for c < 0.5, where m̄
is the input length of the instance. Their proof is based on the PCP theorem.
In contrast, the proof of the lower bound in Theorem 1 is information-theoretic.
11
Problem
Set Cover
Minimum Spanning Tree
Shortest Path
Perfect Matching
Minimum Cut
Lower Bound
Ω(ln |U|)
Ω(n)
Ω(n2/3 )
Ω(n)
√
Ω( n)
Reference
Iwata and Nagano [2009]
Goel et al. [2009]
Goel et al. [2009]
Goel et al. [2009]
Theorem 1
Table 1: Hardness results for combinatorial problems with submodular costs,
where n is the number of nodes, and U the universe to cover. These results
assume oracle access to the cost function.
Theorem 1. No polynomial-time
algorithm can solve MinCoopCut with an
p
approximation factor of o( |V|/ log |V|).
The proof relies on constructing two submodular cost functions f , h that are
almost indistinguishable, except that they have quite differently valued minima.
In fact, with high probability they cannot be distinguished with a polynomial
number of function queries. If the optima of h and f differ by a factor larger than
α, then any solution for f within a factor α of the optimum would be enough
evidence to discriminate between f and h. As a result, a polynomial-time algorithm that guarantees an approximation factor α would lead to a contradiction.
The proof technique follows along the lines of the proofs in [Goemans et al.,
2009, Svitkina and Fleischer, 2008].
One of the functions, f , depends on a hidden random set R ⊆ E that will
be its optimal cut. We will use the following lemma that assumes f will depend
on a random set R.
Lemma 1 (Svitkina and Fleischer [2008], Lemma 2.1). If for any fixed set
Q ⊆ E, chosen without knowledge of R, the probability of f (Q) 6= h(Q) over
the random choice of R is m−ω(1) , then any algorithm that makes a polynomial number of oracle queries has probability at most m−ω(1) of distinguishing
between f and h.
Consequently, the two functions f and h in Lemma 1 cannot be distinguished
with high probability within a polynomial number of queries, i.e., within polynomial time. Hence, it suffices to construct two functions for which Lemma 1
holds.
Theorem 1. We will prove the bound in terms of the number m = |E| of edges
in the graph. The graph we construct has n = m − ` + 2 nodes, and therefore
the proof also shows the lower bound in terms of nodes.
Construct a graph G = (V, E) with ` parallel disjoint paths from s to t,
where each path has k edges. The random set R ⊂ E is always a cut consisting
of |R| = ` edges, and contains one edge from each path uniformly at random.
12
k
ss
S
k
t
...
`t
t
`
Figure 2:
forfor
the the
proof
of Theorem
1.
Figure
2: Graph
Graph
proof
of Theorem
1.
Figure
2: Graph
for the
ofrandom
Theorem
One
the
functions,
ff, ,depends
on on
aproof
hidden
set R1.✓set
E that
One
of of
the
functions,
depends
a hidden
random
R ✓will
E that will
be its
optimal
willuse
use
the
following
Lemma
that assumes
f to depend
be its
optimalcut.
cut. We
We will
the
following
Lemma
that assumes
f to depend
on aonrandom
a randomset
setR.
R.
WeLemma
define β 1= (Svitkina
8`/k < ` (for
k > 8), and, for any Q ⊆ E,
and Fleischer [2008], Lemma
If for
Lemma 1 (Svitkina and Fleischer [2008], Lemma 2.1). If2.1).
for any
fixedany
set fixed set
Q ✓
E
,
chosen
without
knowledge
of
R,
the
probability
of
f
(Q)
=
6
Q ✓ E, chosen without knowledge!(1)
of R, the probability of f (Q) 6= h(Q) overh(Q) over
h(Q)of
=R
min{|Q|,
the random choice
is m `} , then any algorithm that makes a(15)
polynothe random choice of R is m !(1) , then any algorithm that makes
!(1)a polynomial number of oracle queries has probability at most
of distinguishing
!(1)m
f
(Q)
=
min{|Q
\
R|
+
min{|Q
∩
R|,
β},
`}.
(16)
mial number of oracle queries has probability at most m
of distinguishing
between f and h.
between f and h.
Consequently,
f andfew
h insets
Lemma
1 cannot
be distinguished
The functions
differ the
onlytwo
forfunctions
the relatively
Q with
|Q ∩ R|
> β and
Consequently,
the two
functions
f and h in Lemma
1 cannot
be distinguished
high
probability
within
a polynomial
number
of queries,
i.e.,
within
|Qwith
\ R|
<
`
−
β,
with
min
h(A)
=
h(C)
=
`,
min
f
(A)
=
f
(R)
= β,polyA∈C
A∈C
with time.
high probability
within
a polynomial
numbertwo
of queries,
i.e., within
poly- Lemma 1
nomial
Hence, it
suffices
to construct
functions
for which
where
C is the
set Hence,
of cuts,
and C to
is construct
any√cut. two
Wefunctions
must have
k` = m,
so define
nomial
time.
it suffices
Lemma
1
holds.
√for which
2
such that
holds. = 8/7 log m, and set k = 8 m/ and ` = m/8.
Proof (Theorem 1). We will prove the bound in terms of the number m = |E |
We compute
the probability
that we
f and
h differhas
forn a=given
query
set Q.
of edges
in the graph.
graph
construct
m
`m+=2|E|
nodes, and
Proof (Theorem
1). WeThe
will prove
the bound
in terms of the
number
Probabilities
are
over
the
random
unknown
R.
Since
f
≤
h,
the
probability
therefore
the
proof
also
shows
the
lower
bound
in
terms
of
nodes.
of edges in the graph. The graph we construct has n = m ` + 2 nodes, and of a
Construct
graph
G shows
= (VIf,the
E
) lower
with
`then
parallel
disjoint
paths
to∩t,R|,
where
difference
is Pr(f
<also
h(Q)).
|Q|
≤ `,bound
fin(Q)
< of
h(Q)
only iffrom
β <s|Q
therefore
thea(Q)
proof
terms
nodes.
each
has ka graph
edges.
Our
random
set
R ⇢
E is>paths
always
bes a
consisting
and
thepath
probability
Pr(f G(Q)
< h(Q))
Pr(|Q
∩ R|
β) increases
Q
grows. of
Construct
=
(V,
E) with =
` parallel
disjoint
from
to cut
t,aswhere
|R| = ` edges, and contains one edge from each path uniformly at random. We
path has
k edges.
Our randomsince
set Rh(Q)
⇢ E is=always
be a cut consisting of
If, define
on each
the other
hand,
` the
= 8`/k
< `|Q|
(for≥k`,>then
8), and, for
any Q
✓ E probability
,
|R| = ` edges, and contains one edge from each path uniformly at random. We
8`/k <
`Pr(|Q
(for=k min{|Q|,
>
8), +
and,
for any∩QR|,
✓ E,β} < `) = Pr(|Q ∩ R| > β)
h(Q)
`}
Pr(fdefine
(Q) < =h(Q))
=
\ R|
min{|Q
f (Q) = min{|Q \ R| + min{|Q \ R|,
h(Q) = min{|Q|, `}
}, `}.
(12)
(13)
(12)
decreases as Q grows. Hence, the probability of a difference is largest when
The functions di↵er
only
for the
Q with |Q \(13)
R| >
and
f (Q)
= min{|Q
\ R|relatively
+ min{|Q \few
R|, sets
}, `}.
|Q||Q=\ R|
`. < `
, with minA2C h(A) = h(C) = `, minA2C f (A) = f (R) = , C
So The
let
= `. If
Q spreads
b ≤We
k edges
a path
P ,|Q
then
the
set |Q|
of
cuts,
and
Conly
is any
cut.
must
have
=
m,
define
✏ such that
functions
di↵er
forover
the
relatively
few of
sets
with
\so
R|the
> probability
and
p Qk`
p
✏2 Q
=|Qincludes
8/7
and
setin
k P=
m/✏
`==expected
m/8.
that
∩8R
is b/k.
The
between
\ R|log
< m,
` the
,edge
with
min
h(A)
= and
h(C)
`,✏min
(A) = f (R)
= , CQ and
A2C
A2C foverlap
We
compute
the
probability
that
and
h|Q|/k
given
query
set Q.
the sum
set
ofofcuts,
is paths:
any p
cut. E[
We
must
have
=di↵er
m, so
define
✏ such the
that
R is the
hitsand
onCall
|Q
∩fR|
=k`
=for
`/k.a Since
edges
p] R.
2
Probabilities
are
oversetthe
random
unknown
Since f h, the probability of a
✏
=
8/7
log
m,
and
k
=
8
m/✏
and
`
=
✏
m/8.
in di↵erence
R are independent
across
different
paths,
may
boundh(Q)
the probability
is P (f (Q)
<
h(Q)).
Ifthat
|Q|
`, we
then
f (Q)
only set
if Q.< of
|Qa\ R|,
We compute
probability
fand
h di↵er
for a<given
query
large
by the
a Chernoff
bound
(with
= 7\ in
andintersection
the probability
P (f (Q) <
h(Q))
= Pδ(|Q
R|[56]):
> ) increases as Q grows.
Probabilities are over the random unknown R. Since f h, the probability of a
If, on the other hand,
|Q|
`, then since h(Q)
` the probability
<=h(Q)
di↵erence is P (f (Q)< h(Q)). If |Q| `, then f (Q)
only if < |Q \ R|,
Pr
f
(Q)
=
6
h(Q)
≤
Pr
|Q
∩
R|
≥
8`/k
and the probability P (f (Q) < h(Q)) = P (|Q \ R| > ) increases as Q grows. (17)
P (f (Q) < h(Q)) = P (|Q \ R| + min{|Q \ R|,
} < `) = P (|Q \ R| >
2
/8 = ` −ω(log
m)
If, on the other hand, ≤
|Q|2−7`/k
`, then=since
2−7h(Q)
= 2 the probability
= m−ω(1) .
P (f (Q) < h(Q)) = P (|Q \ R| + min{|Q \ R|,
} < `) = P (|Q \ R| > )
)
(18)
11
With this result, Lemma 1 applies. No polynomial-time
algorithm can guarantee
to be able to distinguish f and h with high probability. A polynomial algorithm
with approximation factor better than11the ratio of optima h(R)/f (R) would
discriminate the two functions and thus lead to a contradiction. As a result, the
lower bound is determined by the ratio of optima of h and f . The optimum of
f is f (R) = β, and h has
cost ` for all minimal cuts. Hence, the ratio
√ uniformp
is h(R)/f (R) = `/β = m/ = o( m/ log m).
13
Building on the construction in the above proof with ` = n1/3 and a different
cut cost function, Balcan and Harvey [2012] proved that if the data structure
used by an algorithm (even with an arbitrary number of queries) has polynomial size, then this data structure cannot represent the minimizers of their
cooperative cut problem to an approximation factor of o(n1/3 / log n).
In addition, we mention that a reduction from Graph Bisection serves to
prove that MinCoopCut is NP-hard. We defer the proof to Appendix C, but
point out that in the reduction, the cost function is fully accessible and given
as a polynomial-time computable formula.
Theorem 2. Minimum Cooperative (s, t)-Cut is NP-hard.
4
Relaxation and the flow dual
As a first step towards approximation algorithms, we formulate a relaxation of
MinCoopCut and analyze the flow-cut gap. The minimum cooperative cut
problem can be relaxed to a continuous convex optimization problem using the
convex Lovász extension f˜ of f :
min
y∈R|E| , x∈R|V|
s.t.
f˜(y)
− x(u) + x(v) + y(e) ≥ 0
x(s) − x(t) ≥ 1
(19)
for all e = (u, v) ∈ E
y≥0
The dual of this problem can be derived by writing the Lovász extension as a
maximum f˜(y) = maxz∈P(f ) z > y of linear functions. The maximum is taken
over the submodular polyhedron
X
P(f ) = {y |
y(e) ≤ f (A) ∀A ⊆ E}.
(20)
e∈A
The resulting dual problem is a flow problem with non-local capacity constraints:
max
X
ν∈R,ϕ∈R|E|
s.t.
ϕ(A) ,
e∈δ + u
ϕ(e) −
X
e0 ∈δ − u
(21)
ϕ(e) ≤ f (A)
for all A ⊆ E
ϕ(e0 ) = d(u)ν
for all u ∈ V
e∈A
X
ν
(22)
ϕ ≥ 0,
where d(u) = 1 if u = s, d(u) = −1 if u = t, and d(u) = 0 otherwise. Constraint (22) demands that ϕ must, in addition to satisfying the common flow
conservation, reside within the submodular polyhedron P(f ). This more restrictive constraint replaces the edge-wise capacity constraints that occur when
f is a sum of weights.
14
As an alternative to (19), the constraints can be stated in terms of paths: a
set of edges is a cut if it intersects all (s, t)-paths in the graph.
min f˜(y)
X
s.t.
(23)
e∈P
y(e) ≥ 1 for all (s, t)-paths P ⊆ E
y ∈ [0, 1]E .
We will use this form in Section 5.2.1, and the relaxation (19) in Section 5.2.2.
4.1
Flow-cut gap
The relaxation (19) of the discrete problem (2) is not tight. This becomes
evident when analyzing the ratio f (C ∗ )/f˜(y ∗ ) between the optimal value of the
discrete problem and the relaxation (19) (i.e., the integrality gap). This ratio
is, by strong duality between Problems (19) and (21), also the flow-cut gap
f (C ∗ )/ν ∗ of the optimal cut and maximal flow values.
Lemma 2. Let P be the set of all (s, t)-paths in the graph. The flow-cut gap
f (C ∗ )/ν ∗ can be upper and lower bounded as follows:
f (C ∗ )
P
P ∈P
minP 0 ⊆P
f (P 0 )
|P 0 |
≤
f (C ∗ )
f (C ∗ )
≤
ν∗
maxP ∈P minP 0 ⊆P
f (P 0 )
|P 0 |
.
Proof. The Lemma straightforwardly follows from bounding the optimal flow
ν ∗ . The flow through a single path P ∈ P, if all other edges e ∈
/ P are empty, is
restricted by the minimum average capacity for any subset of edges within the
(P 0 )
path, i.e., minP 0 ⊆P f|P
0 | . Moreover, we obtain a family of feasible solutions as
those that send nonzero flow only along one path and remain within that path’s
capacity. Hence, the maximum flow must be at least as big as the flow for any
of those single-path solutions. This observation yields the upper bound on the
ratio.
A similar argumentation shows the lower
P bound: the total joint capacity
constraint is upper bounded by fˆ(A) =
P ∈P f (A ∩ P ) ≥ f (A). Hence,
P
f (P 0 )
minP 0 ⊆P
is the value of the maximum flow with capacity fˆ if each
0
P ∈P
|P |
edge is only contained in one path, and is an upper bound on the flow otherwise.
Corollary 1. The flow-cut gap for MinCoopCut can be as large as n − 1.
Proof. Corollary 1 can be shown via an example where the upper and lower
bound of Lemma 2 coincide. The worst-case example for the flow-cut gap is a
simple graph that consists of one single path from s to t with n − 1 edges. For
this graph one of the capacity constraints is that
X
ϕ(E) =
ϕ(e) ≤ f (E).
(24)
e∈E
15
approximating f
√
generic (§5.1.1)
O( m log m)
|C ∗ |
semigradient (§5.1.2)
(|C ∗ |−1)(1−κf )+1
polymatroidal flow (§5.1.3)
min{∆s , ∆t }
relaxation
randomized (§5.2.1)
rounding I (§5.2.2)
rounding II (§5.2.2)
|Pmax |
|Pmax |
|V| − 1
Table 2: Overview of the algorithms and their approximation factors.
Constraint (24) is the only relevant capacity constraint if the capacity (and cut
cost) function is f (A) = maxe∈A w(e) with weights w(e) = γ for all e ∈ E and
some constant γ > 0 and, consequently, f (E) = γ. By Constraint (24), the
γ
. The optimum cooperative cut C ∗ , by contrast,
maximum flow is ν ∗ = n−1
consists of any single edge and has cost f (C ∗ ) = γ.
Single path graphs as used in the previous proof can provide worst-case
examples for rounding methods too: if f is such that f (e) ≥ f (E)/|E| for all
edges e in the path, then the solution to the relaxed cut problem is maximally
(E)
uninformative: all entries of the vector y are y(e) = fn−1
.
5
Approximation algorithms
We next address approximation algorithms whereby we consider two complementary approaches. The first approach is to substitute the submodular cost
function f by a simpler function fˆ. Appropriate candidate functions fˆ that admit an exact cut optimization are the approximation by Goemans et al. [2009]
(Section 5.1.1), semi-gradient based approximations (Section 5.1.2), or approximations by making f separable across local neighborhoods (Section 5.1.3).
The second approach is to solve the relaxations from Section 4 and round the
resulting optimal fractional solution (Section 5.2.2). Conceptually very close to
the relaxation approach, we offer another algorithm that solves the mathematical program (23) via a randomized greedy algorithm (Section 5.2.1).
The relaxations approaches are affected by the flow-cut gap, or, equivalently,
the length of the longest path in the graph. The approximations that use a
surrogate cost function are complementary and not affected by the “length”,
but by a notion of the “width” of the graph.
5.1
Approximating the cost function
We begin with algorithms that use a suitable approximation fˆ to f , for which
the problem
minimize fˆ(C) s.t. C ⊆ E is a cut
(25)
is solvable exactly in polynomial time. The following lemma will be the basis
for the approximation bounds.
16
Lemma 3. Let Sb = argminS∈S fˆ(S). If for all S ⊆ E, it holds that f (S) ≤
fˆ(S), and if for the optimal solution S ∗ to Problem (2), it holds that fˆ(S ∗ ) ≤
αf (S ∗ ), then Sb is an α-approximate solution to Problem (2):
b ≤ αf (S ∗ ).
f (S)
b ≤ fˆ(S ∗ ), it follows that f (S)
b ≤ fˆ(S)
b ≤ fˆ(S ∗ ) ≤ αf (S ∗ ).
Proof. Since fˆ(S)
5.1.1
A generic approximation
Goemans et al. [2009] define a generic approximation of a monotone submodular
pP
function2 that has the functional form fˆea (A) =
e∈A wf (e). The weights
ˆ
wf (e) depend on f . When using fea , we compute a minimum cut for the
2
, which is a modular sum of weights and hence results in a standard
cost fˆea
Minimum (s, t)-Cut problem. In practice, the bottleneck lies in computing
the weights wf . Goemans et al. [2009] show how to compute weights such
√
that f (A) ≤ √fˆ(A) ≤ αf (A) with α = O( m) for a matroid rank function,
and α = O( m log m) otherwise. We add that for an integer polymatroid
rank function bounded by M = max√
e∈E f (e), the logarithmic factor can be
replaced by a constant to yield α = O( mM ) (if one approximates the matroid
expansion3 of the polymatroid instead of f directly). Together with Lemma 3,
this yields the following approximation bounds.
b = argminC∈C fˆea (C) be the minimum cut for cost fˆea , and
Lemma 4. Let C
∗
b = O(√m log m)f (C ∗ ). If f is integer-valued
C = argminC∈C f (C). Then f (C)
√
b = O( mM )f (C ∗ ), where
and we approximate its matroid expansion, then f (C)
M ≤ maxe f (e).
The lower bound in Theorem 1 suggests that for sparse graphs, the bound
in Lemma 4 is tight up to logarithmic factors.
5.1.2
Approximations via semigradients
For any monotone submodular function f and any set A, there is a simple way
to compute a modular upper bound fˆs to f that agrees with f at A. In other
words, fˆs is a discrete supergradient of f at A. We define fˆs as [Jegelka and
Bilmes, 2011a, Iyer et al., 2013a]
X
X
fˆs (B; A) = f (A) +
f (e | A) −
f (e | E \ e).
(26)
e∈B\A
e∈A\B
2 We will also call it the ellipsoidal approximation since it is based on approximating a
symmetrized version of the submodular polyhedron by an ellipsoid.
3 The expansion is described in Section 10.3 in [Narayanan, 1997]. In short, we replace each
element e by a set ê of f (e) parallel elements. Thereby we extend f to a submodular function
S
fˆ on subsets of i êiS
. The desired rank function is now the convolution r(·) = fˆ(·) ∗ | · | and
it satisfies f (S) = r( e∈S ê).
17
v4
v1
S
v2
v5
v3
v6
fˆpf (C) =f ({(v1 , v4 ), (v2 , v4 )})
t
+ f ({(v3 , v4 ), (v3 , v5 )})
+ f ({(v3 , v6 )})
Figure 3: Approximation of a cut cost. Red edges are in CvΠ4 (head), blue dashed
edges in CvΠ3 (tail), and the green dash-dotted edge in CvΠ6 (head).
b ∈ argminC∈C fˆs (C; ∅). Then
Lemma 5. Let C
b ≤
f (C)
where κf = maxe 1 −
(|C ∗ |
f (e|E\e)
f (e)
|C ∗ |
f (C ∗ ),
− 1)(1 − κf ) + 1
is the curvature of f .
Lemma 5 was shown in [Iyer et al., 2013a]. As m (and correspondingly
|C ∗ |) gets large, the bound eventually no longer depends on m and instead only
on the curvature of f . In practice, results are best when the supergradient
is used in an iterative algorithm: starting with C0 = ∅, one computes Ct ∈
argminC∈C fˆs (C; Ct−1 ) until the solution no longer changes between iterations.
The minimum cut for the cost function fˆs (C; A) can be computed as a minimum
cut with edge weights
(
f (e | E \ e) if e ∈ A
(27)
w(e) =
f (e | A)
if e ∈
/ A.
Consequently, the semigradient approximation yields a very easy and practical
algorithm that iteratively uses standard minimum cut as a subroutine. This
algorithm was used e.g. in [Jegelka and Bilmes, 2011a], and the visual results in
[Kohli et al., 2013] show that it typically yields very good solutions in practice
on certain problem instances where the optimum solution can be computed
exactly.
5.1.3
Approximations by introducing separation
The approximations in Section 5.1.1 and 5.1.2 are indifferent to the structure of
the graph, while following approximation is not. One may say that Problem (2)
is hard because f introduces non-local dependencies between edges that might
be anywhere in the graph. Indeed, the problem is easier if dependencies between
edges are restricted to local neighborhoods defined by the graph, for example,
edges that might be incident to the same vertex.
Hence, we define an approximation fˆpf that is globally separable but locally
exact. To measure the cost of an edge set C ⊆ E, we partition C into groups
Π(C) = {CvΠ }v∈V , where the edges in set CvΠ must be incident to node v (CvΠ
may be empty). That is, we assign each edge either to its head or to its tail
18
node in any partition, as illustrated in Figure 3. Let PC be the family of all
such partitions (which vary over the head or tail assignment of each edge). We
define an approximation
X
fˆpf (C) = min
f (CvΠ )
(28)
Π(C)∈PC
v∈V
that (once the partition is fixed) decomposes across different node incidence
edge sets, but is accurate within a group CvΠ . Thanks to the subadditivity of
f , the function fˆpf is an upper bound on f . It is a convolution of submodular
functions and always is the tightest approximation that is a direct sum over
any partition in PC . Perhaps surprisingly, even though the approximation (28)
looks difficult to compute and is in general not even a submodular function (an
example is in Appendix D), it is possible to solve a minimum cut with cost
fˆpf exactly. To do so, we exploit its duality to a generalized maximum flow
problem, namely polymatroidal network flows.
Polymatroidal network flows. Polymatroidal network flows [Lawler and
Martel, 1982, Hassin, 1982] generalize the capacity constraint of traditional flow
problems. They retain the constraint of flow conservation (a function ϕ : E →
R+ is a flow if the inflow at each node v ∈ V \ {s, t} equals the outflow). The
edge-wise capacity constraint ϕ(e) ≤ cap(e) for all e ∈ E, given a capacity
function cap : E → R+ is replaced by local submodular capacities over sets of
out
for outgoing
edges incident at each node v: capin
v for incoming edges, and capv
edges. The capacity constraints at each v ∈ V are
ϕ(A) ≤ capin
v (A)
ϕ(A) ≤ capout
v (A)
for all A ⊆ δ − (v) (incoming edges), and
for all A ⊆ δ + (v) (outgoing edges).
Each edge (u, v) belongs to two incidence sets, δ + u and δ − v. A maximum flow
with such constraints can be found in time O(m4 τ ) by the layered augmenting
path algorithm by Tardos et al. [1986], where τ is the time to minimize a submodular function on any set δ + v, δ − v for any v. Hence, the incidence sets are
in general much smaller than E.
A special polymatroidal maximum flow turns out to be dual to the cut
problem we are interested in. To see this, we will use the restriction f A of the
function f to a subset A. For ease of reading we drop the explicit restriction
notation later. We assume throughout that the desired cut is minimal4 , since
additional edges can only increase its cost.
Lemma 6. Minimum (s, t)-cut with cost function fˆpf is dual to a polymatroidal
out
network flow with capacities capin
= f δ+ v at each node
v = f δ − v and capv
v ∈ V.
The proof is provided in Appendix E. It uses, with some additional considerations, the dual problem to a polymatroidal maxflow, which can be stated
4A
cut C ⊆ E is minimal if no proper subset B ⊂ C is a cut.
19
as follows. P
Let capin : 2E → R+ be the joint incoming capacity function, i.e.,
in
−
out
cap (C) = v∈V capin
be the correspondv (C ∩δ v), and let equivalently cap
ing joint outgoing capacity. The dual of the polymatroidal maximum flow is a
minimum cut problem whose cost is a convolution of edge capacities [Lovász,
1983]:
h
i
cap(C) = (capin ∗ capout )(C) , min capin (A) + capout (C \ A) .
(29)
A⊆C
This convolution is in general not a submodular function. Lemma 6 implies that
we can solve the approximate MinCoopCut via its dual flow problem. The
primal cut solution will be given by a set of full edges, i.e., edges whose joint
flow equals their joint capacity.
We can now state the resulting approximation bound for MinCoopCut.
Let C ∗ be the optimal cut for cost f . We define ∆s to be the tail nodes of the
edges in C ∗ : ∆s = {v | ∃(v, u) ∈ C ∗ }, and similarly, ∆t = {v | ∃(u, v) ∈ C ∗ }.
The sets ∆s , ∆t provide a measure of the “width” of the graph.
b be the minimum cut for cost fˆpf , and C ∗ the optimal cut
Theorem 3. Let C
for cost f . Then
b ≤ min{|∆s |, |∆t |} f (C ∗ ) ≤ |V| f (C ∗ ).
f (C)
2
Proof. To apply Lemma 3, we need to show that f (C) ≤ fˆpf (C) for all C ⊆ E,
and find an α such that fˆpf (C ∗ ) ≤ αf (C ∗ ). The first condition follows from the
subadditivity of f .
To bound α, we use Lemma 6 and Equation 29:
fˆpf (C ∗ ) = (capin ∗ capout )(C ∗ )
in
∗
(30)
out
∗
≤ min{cap (C ), cap (C )}
nX
o
X
≤ min
f (C ∗ ∩ δ + v),
f (C ∗ ∩ δ − v)
v∈∆s
v∈∆t
o
n
∗
+
≤ min |∆s | max f (C ∩ δ v), |∆t | max f (C ∗ ∩ δ − v)
v∈∆t
v∈∆s
∗
≤ min |∆s |, |∆t | f (C ).
Thus, Lemma 3 implies an approximation bound α ≤ min |∆s |, |∆t |
|V|/2.
(31)
(32)
(33)
(34)
≤
Iyer et al. [2013b] show that the bound in Theorem 3 can be tightened to
|V|
2+(|V|−2)(1−κf ) by taking into account the curvature κf of f .
5.2
Relaxations
An alternative approach to approximating the edge weight function f is to
relax the cut constraints via the formulations (23) and (19). We analyze two
20
Algorithm 1 Greedy randomized path cover
Input: graph G = (V, E), terminal nodes s, t ∈ V, cost function f : 2E → R+
C = ∅, P
y=0
while e∈Pmin y(e) < 1 for the current shortest path Pmin do
choose β within the interval β ∈ (0, mine∈Pmin f (e|C)]
for e in Pmin do
with probability β/f (e|C), set C = C ∪ {e}, y(e) = 1.
end for
end while
prune C to C 0 and return C 0
algorithms: the first, a randomized algorithm, maintains a discrete solution,
while the second is a simple rounding method. Both cases remove the constraint
that the cut must be minimal: any set B is feasible that has a subset C ⊆ B that
is a cut. Relaxing the minimality constraint makes the feasible set up-monotone
(equivalently up-closed). This is not major problem, however, since any superset
of a cut can easily be pruned to a minimal cut while only, if anything, improving
the solution due to the monotonicity of f .
5.2.1
Randomized greedy covering
The constraints in the path-based relaxation (23) suggest that a minimum (s, t)cut problem is also a min-cost cover problem: a cut must intersect or “cover”
each (s, t)-path in the graph. The covering formulation of the constraints in (23)
clearly show the relaxation of the minimality constraint. Algorithm 1 solves a
discrete variant of the formulation (23) and maintains a discrete y ∈ {0, 1}, i.e.,
y is eventually the indicator vector of a cut.
Since a graph can have exponentially many (s, t)-paths, there can be exponentially many constraints. But all that is needed in the algorithm is to find a
violated constraint, and this is possible by computing the shortest path Pmin ,
using y as the (additive) edge lengths. If Pmin ’s weight is at least one, then y is
feasible. Otherwise, Pmin defines a violated constraint in formulation (23).
Owing to the form of the constraints, we can adapt a randomized greedy
cover algorithm [Koufogiannakis and Young, 2009] to Problem (23) and obtain
Algorithm 1. In each step, we compute the shortest path with weights y to find
a possibly uncovered path. Ties are resolved arbitrarily. To cover the path, we
randomly pick edges from Pmin . The probability of picking edge e is inversely
proportional to the marginal cost f (e|C) of adding e to the current selection
of edges5 . We must also specify an appropriate β. With the maximal allowed
β = mine∈Pmin f (e|C), the cheapest edges are selected deterministically, and
others randomly. In that case, C grows by at least one edge in each iteration,
5 If min
e∈Pmin f (e|C) = 0, then we greedily pick all such edges with zero marginal cost,
because they do not increase the cost. Otherwise we sample as indicated in the algorithm.
21
and the algorithm terminates after at most m iterations.
If the algorithm returns a set C that is feasible but not a minimal cut, it is
easy to prune it to a minimal cut C 0 ⊆ C without any additional approximation
error, since monotonicity of f implies that f (C 0 ) ≤ f (C). Such pruning can for
example be done via breadth-first search. Let Vs be the set of nodes reachable
from s after the edges in C have been removed. Then we set C 0 = δ + (Vs ). The
set C 0 must be a subset of C, since if there was an edge (u, v) ∈ C 0 \ C, then v
would also be in Vs , and then (u, v) cannot be in C 0 , a contradiction.
The approximation bound for Algorithm 1 is the length of the longest path,
like that of the rounding methods in Section 5.2.2. This is not a coincidence,
since both algorithms essentially use the same relaxation.
Lemma 7. In expectation (over the probability of sampling edges), Algorithm 1
b 0 with E[f (C
b 0 )] ≤ |Pmax |f (C ∗ ), where Pmax is the longest
returns a solution C
simple (s, t)-path in G.
b be the cut before pruning. Since f is nondecreasing, it holds
Proof. Let C
0
b
b
that f (C ) ≤ f (C).
By Theorem 7 in [Koufogiannakis and Young, 2009],
a greedy randomized procedure like Algorithm 1 yields in expectation an αapproximation for a cover, where α is the maximum number of variables in
any constraint. Here, α is the maximum number of edges in any simple path,
b 0 )] ≤ E[f (C)]
b ≤
i.e., the length of the longest path. This implies that E[f (C
∗
|Pmax |f (C ).
Indeed, randomization is important. Consider a deterministic algorithm that
always picks the edge with minimum marginal cost in the next path to cover.
bd returned by this algorithm can be much worse. As an example,
The solution C
consider a graph consisting of a clique V of n nodes, with nodes s and t. Let
S ⊆ V be a set of size n/2. Node s is connected to all nodes in S, and node t is
0
connected to the clique only by a distinct node v 0 ∈
PV \ S via edge (v , t). Let
the cost function be a sum of edge weights, f (C) = e∈C w(e). Edge (v 0 , t) has
weight γ > 0, all edges in δ + (S) have weight γ(1 − ) for a small > 0, and
all remaining edges have weight γ(1 − /2). The deterministic algorithm will
bd = δ + (S) as the solution, with cost n2 γ (1 − ), which is by a factor
return C
4
bd |(1 − ) = n2 (1 − ) worse than the optimal cut, f ({(v 0 , t)}) = γ. Hence,
of |C
4
for the deterministic variant of Algorithm 1, we can only show the following
approximation bound:
bd returned by the greedy deterministic heuristic,
Lemma 8. For the solution C
∗
b
b
it holds that f (Cd ) ≤ |Cd |f (C ). This approximation factor cannot be improved
in general.
bd assign the path P (e) which it was chosen to cover.
Proof. To each edge e ∈ C
By the nature of the algorithm, it must hold that f (e) ≤ f (C ∗ ∩ P (e)), because
otherwise an edge in C ∗ ∩ P (e) would have been chosen. Since C ∗ is a cut, the
22
Algorithm 2 Rounding procedure given y ∗
order E such that y ∗ (e1 ) ≥ y ∗ (e2 ) ≥ . . . ≥ y ∗ (em )
for i = 1, . . . , m do
let Ci = {ej | y ∗ (ej ) ≥ y ∗ (ei )}
if Ci is a cut then
b and return C
b
prune Ci to C
end if
end for
set C ∗ ∩ P (e) must be non-empty. These observations imply that
X
X
bd ) ≤
bd | max f (C ∗ ∩ P (e)) ≤ |C
bd |f (C ∗ ).
f (e) ≤
f (C ∗ ∩ P (e)) ≤ |C
f (C
bd
e∈C
bd
e∈C
b
e∈C
Tightness follows from the worst-case example described above.
5.2.2
Rounding
Our last approach is to solve the convex program (19) and round the continuous
to a discrete solution. We describe two types of rounding, each of which achieves
a worst-case approximation factor of n − 1. This factor equals the general flowcut gap in Lemma 1. Let x∗ , y ∗ be the optimal solution to the relaxation (19)
(equivalently, to (23)). We assume w.l.o.g. that x∗ ∈ [0, 1]n , y ∗ ∈ [0, 1]m .
Rounding by thresholding edge lengths. The first technique uses the edge
weights y ∗ . We pick a threshold θ and include all edges e whose entry y ∗ (e)
is larger than θ. Algorithm 2 shows how to select θ, namely the largest edge
length that when treated as a threshold yields a cut.
b be the rounded solution returned by Algorithm 2, θ the threshLemma 9. Let C
old at the last iteration i, and C ∗ the optimal cut. Then
b ≤ 1 f (C ∗ ) ≤ |Pmax |f (C ∗ ) ≤ (n − 1)f (C ∗ ),
f (C)
θ
where Pmax is the longest simple path in the graph.
Proof. The proof is analogous to that for covering problems [Iwata and Nagano,
2009]. In the worst case, y ∗ is uniformly distributed along the longest path, i.e.,
y ∗ (e) = |Pmax |−1 for all e ∈ Pmax as y ∗ must sum to at least one along each
path. Then θ must be at least |Pmax |−1 to include at least one of the edges in
Pmax . Since f˜ is nondecreasing like f and also positively homogeneous, it holds
that
b ≤ f (Ci ) = f˜(χC ) ≤ f˜(θ−1 y ∗ ) = θ−1 f˜(y ∗ ) ≤ θ−1 f˜(χC ∗ ) = θ−1 f (C ∗ ).
f (C)
i
b ⊆
The first inequality follows from monotonicity of f and the fact that C
−1 ∗
˜
˜
Ci . Similarly, the relation between f (χCi ) and f (θ y ) holds because f˜ is
23
nondecreasing: by construction, y ∗ (e) ≥ θχCi (e) for all e ∈ E, and hence
χCi (e) ≤ θ−1 y ∗ (e). Finally, we use the optimality of y ∗ to relate the cost
to f (C ∗ ); the vector χC ∗ is also feasible, but y ∗ optimal. The lemma follows
since θ−1 ≤ |Pmax |.
Rounding by node distances. Alternatively, we can use x∗ to obtain a
discrete solution. We pick a threshold θ uniformly at random from [0, 1] (or
find the best one), and choose all nodes u with x∗ (u) ≥ θ (call this Vθ ). This
induces the cut Cθ = δ(Vθ ). Since the node labels x∗ can also be considered as
distances from s, we refer to this rounding methods as distance rounding.
Lemma 10. The worst-case approximation factor for a solution Cθ obtained
with distance rounding is Eθ [f (Cθ )] ≤ (n − 1)f˜(y ∗ ) ≤ (n − 1)f (C ∗ ).
Proof. To upper bound the quantity Eθ [f (Cθ )], we partition the set of edges
into (n − 1) sets δ + (v), that is, each set corresponds to the outgoing edges
of a node v ∈ V. We sort the edges in each δ + (v) in nondecreasing order
by their values y ∗ (e). Consider one specific incidence set δ + (u) with edges
eu,1 , . . . , eu,h and y ∗ (eu,1 ) ≤ y ∗ (eu,2 ) ≤ . . . ≤ y ∗ (eu,h ). Edge eu,i is in the cut if
θ ∈ [x∗ (u), x∗ (u) + y ∗ (eu,i )). Therefore, it holds for each node u that
Z 1
Eθ [f (Cθ ∩ δ + (u))] =
f (Cθ ∩ δ + (u))dθ
(35)
0
h
X
=
(y ∗ (eu,j ) − y ∗ (eu,j−1 ))f ({eu,j , . . . eu,h })
(36)
= f˜(y ∗ (δ + (u))),
(37)
j=1
where we define y ∗ (eu,0 ) = 0 for convenience, and assume that f (∅) = 0. This
implies that
X
Eθ [f (Cθ )] ≤ Eθ [
f (Cθ ∩ δ + (v))]
(38)
v∈V
=
X
v∈V
f˜(y ∗ (δ + (v))) ≤ (n − 1)f˜(y ∗ ) ≤ (n − 1)f (C ∗ ).
A more precise approximation factor is
6
P
v
(39)
f˜(y ∗ (δ + (v)))
.
f (y ∗ )
Special cases
The complexity of MinCoopCut is not always as bad as the worst-case bound
in Theorem 1. While it is useful to consider this most general case (see Section 2.3), we next discuss properties of the submodular cost function and the
graph structure that lead to better approximation factors. Our discussion is
not specific to cooperative cuts; it is rather a survey of properties that make a
number of submodular optimization problems easier.
24
generic (§5.1.1)
O(maxi
p
|Bi | log |Bi |)
∗
∩Bi |
maxi (|C ∗ ∩B|C
i |−1)(1−κf )+1
semigradient (§5.1.2)
i
maxi min{∆s ∩ Bi , ∆t ∩ Bi }
bd ∩ Bi |
maxi |C
polymatroidal flow (§5.1.3)
deterministic greedy (§5.2.1)
Table 3: Improved approximation bounds for functions of the form f (A) =
Pk
i=1 fi (A ∩ Bi ). The bounds are now determined by the largest support
maxi |Bi |, but not by k.
6.1
Separability and sums with bounded support
An important factor for tractability and approximations is the separability of
the cost function, that is, whether there are separators of f whose structure
aligns with the graph.
Definition 1 (Separator of f ). A set S ⊆ E is called a separator of f : 2E → R
if for all B ⊆ E, it holds that f (B) = f (B ∩ S) + f (B \ S). The set of separators
of f is closed under union and intersection.
The structure of the separators strongly affects the complexity of MinCoopCut. First and obviously, the extreme case that f is a modular function (and
each e ∈ E is a separator)
S can beSsolved exactly. Second, if the separators of
f form a partition E = v Ev+ ∪ v Ev− that aligns with node neighborhoods
such that Ev+ ⊆ δ + (v), and Ev− ⊆ δ − (v), then both fˆpf and distance rounding solve the problem exactly. No change in the algorithm is needed, i.e., the
exact partition need not be known. In that case, the flow-cut
gap is zero, as
P
∗
becomes obvious from the proof of Lemma 10, since ( v f˜(yE
))/f˜(y ∗ ) = 1.
v
These separators respect the graph structure and rule out any non-local edge
interactions.
Sums of functions with bounded support. P
A generalization of the case
of separators are functions that are a sum f (A) = i fi (A ∩ Bi ) of functions fi ,
each of which has bounded support Bi . The Bi can be overlapping. In this case,
the approximation bounds improve for many of the algorithms in Section 5.1
that rely on a surrogate function, and for the greedy approximation in Lemma 8.
Those bounds, summarized in Table 3, can be shown by approximating each fi
separately by fˆi with approximation factor αi that now depends on |Bi |, and
P
using f (A) ≤ maxj αj i fˆi (A). This separate approximation is implicit in
all those algorithms except the approximation from Section 5.1.1. In those
implicit cases, no changes need to be made in the implementation and the
partition need not be known. For the generic approximation in Section 5.1.1,
one can approximate each fi explicitly
and separately, if the partition is known.
P
Optimizing the resulting sum i fˆi or its square is however no longer a minimum
cut problem. It admits an FPTAS [Nikolova, 2010, Kohli et al., 2013].
25
For the relaxations, it is not immediately clear that the decomposition always
leads to improvements. Consider for example a function f (A) = f1 (A ∩ B1 ) +
f2 (A ∩ B2 ), where f1 (B1 ) = f2 (B2 ), P , Pmax = B1 ∪ B2 and |B1 | = |B2 | =
|Pmax /2|. Then f˜( |P1 | χP ) = f˜( |B11 | χB1 ). In that case, the proof of Lemma 9
may still require θ−1 = |Pmax |.
6.2
Symmetry and “unstructured” functions
Going one step further, one may consider sums of that do not necessarily have
bounded support P
but are of a simpler form. An important such class are functions fi (A) = g( e∈A wi (e)) = g(wi (A)) for nonnegative weights wi (e) and
a nondecreasing concave function g. We refer to the submodular functions
g(w(A)) as unstructured, because they only consider a sum of weights, but otherwise do not make any distinction between edges (unlike, e.g., graphic matroid
rank functions). One may classify such functions into a hierarchy, where F(k)
Pk
contains all functions f (A) = j=1 gj (wj (A)) with at most k such components.
The functions F(k) are special cases of low-rank quasi-concave functions, where
k is the rank of the function.
If k = 1, then it suffices to minimize w1 (C) directly and the problem reduces
to Minimum (s, t)-Cut. For k > 1, several combinatorial problems admit an
FPTAS with running time exponential in k [Goyal and Ravi, 2008, Mittal and
Schulz, 2012]. This holds for cooperative cuts too [Kohli et al.,
p 2013]. A special
case for k = 2 is the mean-risk objective f (A) = w1 (A) + w2 (A) [Nikolova,
2010]. Goel et al. [2010] show that these functions can yield better bounds in
combinatorial multi-agent problems than general polymatroid rank functions, if
each agent has a cost function in F(1).
Even for general, unconstrained submodular minimization6 , the class F(k)
admits specialized improved optimization algorithms [Kohli et al., 2009a, Stobbe and Krause, 2010, Kolmogorov, 2012, Jegelka et al., 2013]. The complexity
of those faster specialized algorithms depends on the rank k as well. An interesting question arising from the above observations is whether F(k) contains
all submodular functions if k is large enough? The answer is no: even if k is
allowed to be exponentially large in the ground set size, this class is a strict
sub-class of all submodular functions. If the addition of auxiliary variables is
allowed, this class coincides with the class of graph-representable functions in
the sense of [Z̆ivný et al., 2009]: any graph cut function h : 2V → R+ is in
F(|E|), and any function in F(k) can be represented as a graph cut function in
an extended auxiliary graph [Jegelka et al., 2011]. However, not all submodular
functions can be represented in this way [Z̆ivný et al., 2009].
The parameter k is a measure of complexity. If k is not fixed, then MinCoopCut is NP-hard; for example, the reduction in Section C uses such functions. Even more, unrestricted k may induce large lower
Pk bounds, as has been
proved for label cost functions of the form f (A) = j=1 wj min{1, |A ∩ Bj |}
6 For unconstrained submodular function minimization we drop the constraint that the
functions gj are nondecreasing.
26
[Zhang et al., 2011].
A subclass of unstructured submodular functions are the aforementioned
permutation symmetric submodular functions7 that are indifferent to any permutation of the ground set: f (A) = f (σ(A)) for all permutations σ (possibly
within a group). This symmetry makes submodular optimization problems easier, as shown in Proposition 3 and work on submodular maximization [Vondrák,
2013] and partitioning problems [Ene et al., 2013].
6.3
Symmetry and graph structure
Proposition 3 shows that symmetry and the graph structure can work together
to make the cut problem easier, in fact, a submodular minimization problem on
graph nodes. Section 2 outlines some examples that come from applications.
6.4
Curvature
The curvature κf ∈ [0, 1] of a submodular function f is defined as
κf = max 1 −
e∈E
f (e | E \ e)
,
f (e)
(40)
and characterizes the deviation from being a modular function. Curvature is
known to affect the approximation bounds for submodular maximization [Conforti and Cornuéjols, 1984, Vondrák, 2008], and also for submodular minimization problems, approximating and learning submodular functions [Iyer et al.,
2013b]. The lower the curvature, the better the approximation factors. For
MinCoopCut and many other combinatorial minimization problems with submodular costs, the approximation factor is affected as follows. If αn is the
worst-case factor (e.g., for the semigradient approximation), then the tightened
αn
factor is (αn −1)(1−κ
. The lower bounds can be tightened accordingly.
f )+1
6.5
Flow-cut gaps revisited
The above properties that facilitate MinCoopCut reduce the flow-cut gaps
in some cases. The proof of Lemma 1 illustrates that the flow-cut gap is intricately linked to the edge cooperation (non-separability) along paths in the
graph. Therefore, the separability described in Section 6.1 affects the flow-cut
gap if it breaks up cooperation along paths: the gap depends only on the longest
cooperating path within any separator of f , and this can be much smaller than
n. If, however, an instance of MinCoopCut is better solvable because the cost
function is a member of F(`) for small constant `, then the gap may still be as
large as in Lemma 1. In fact, the example in Lemma 1 belongs to F(1): it is
equivalent to the function f (A) = γ min{1, |A|}.
Two variants of a final example may serve to better understand the flowcut (and integrality) gap. The first has a large gap, but the rounding methods
7 These are distinct from the other previously-used notion of symmetric submodular functions Queyranne [1998] where, for all A ⊆ E, f (A) = f (E \ A).
27
still find an optimal solution. The second has a gap of one, but the rounding
methods may return solutions with a large approximation factor. Consider a
graph with m edges consisting of m/k disjoint paths of length k each (as in
Figure 2), with a cost function f (C) = maxe∈C w(e). The edges are partitioned
into a cut B ⊂ E with |B| = m/k and the remaining edges E \ B. Let w(e) = γ
for e ∈
/ B and w(e) = β for e ∈ B.
For the first variant, let β = γ; so that for k = 1, we obtain the graph
in Lemma 1. With β = γ (for any k), any minimal cut is optimal, and all
rounding methods find an optimal solution. The maximum flow in Problem (21)
is ν ∗ = γ/k (γ/k flow on one path or γ/m flow on each edge in m/k paths in
parallel). Hence, the flow-cut gap is γ/(γ/k) = k despite the optimality of the
rounded (and pruned) solutions.
For the second variant, let β = γ/k. The maximum flow remains ν ∗ = γ/k,
and the optimal cut is B with f (B) = γ/k, so f (C ∗ ) = ν ∗ . An optimal solution
y ∗ to Program (19) is the uniform vector y ∗ = (γ/m)1m . Despite the zero gap,
for such y ∗ the rounding methods return an arbitrary cut, which can be by a
factor k worse than the optimal solution B. In contrast, the approximation
algorithms in Sections 5.1.2, 5.1.3 based on substitute cost functions do return
an optimal solution.
7
Experiments
We provide a summary of benchmark experiments that compare the proposed
algorithms empirically. We use two types of data sets. The first is a collection of
average-case submodular cost functions on two types of graph structures, clustered graphs and regular grids. The second consists of a few difficult examples
that show the limits of some of the proposed methods.
The task is to find a minimum cooperative cut in an undirected graph8 . This
problem can be solved directly or via n − 1 minimum (s, t)-cuts. Most of the
algorithms solve the (s, t) version. The above approximation bounds still apply,
as the minimum cut is the minimum (s, t)-cut for at least one pair of source
and sink. We observe that, in general, the algorithms perform well, typically
much better than their theoretical worst-case bounds. Which algorithm is best
depends on the cost function and graph at hand.
Algorithms and baselines. Apart from the algorithms discussed in this article, we test some baseline heuristics. First, to test the benefit of the more
sophisticated approximations fˆea and fˆpf we define the simple approximation
X
fˆadd (C) =
f (e).
(41)
e∈C
8 An undirected graph can easily be turned into a directed one by replacing each edge by
two opposing directed ones that have the same cost. A cut will always only include one of
those edges
28
The first baseline (MC) simply returns the minimum cut with respect to fˆadd .
The second baseline (MB) computes the minimum cut basis C = {C1 , . . . , Cn−1 }
b = argminC∈C f (C). The minimum cut
with respect to fˆadd and then selects C
basis can be computed via a Gomory-Hu tree [Bunke et al., 2007]. As a last
baseline, we apply an algorithm (QU) by Queyranne [1998] to h(X) = f (δ(X)).
This algorithm minimizes symmetric submodular functions in O(n3 ) time. However, h not always submodular (e.g., see Props. 1, 2, and 3), and therefore this
algorithm cannot provide any approximation guarantees in general. In fact, we
will see in Section 7.2 that it can perform arbitrarily poorly.
Of the algorithms described in this article, EA denotes the generic (ellipsoidbased) approximation of Section 5.1.1. The iterative semigradient approximation from Section 5.1.2 is initialized with a random cut basis (RI) or a minimumweight cut basis (MI). PF is the approximation via polymatroidal network flows
(Section 5.1.3). These three approaches approximate the cost functions. In addition, we use algorithms that solve relaxations of Problems (23) and (19): CR
solves the convex relaxation using Matlab’s fmincon, and applies Algorithm 2
for rounding. DB implements the distance rounding by thresholding x∗ . Finally,
we test the randomized greedy algorithm from Section 5.2.1 with the maximum
possible β = βmax (GM) and an almost maximal β = 0.9βmax (GA). GH denotes
the deterministic greedy heuristic. All algorithms were implemented in Matlab,
with the help of a graph cut toolbox [Bagon, 2006, Boykov and Kolmogorov,
2004] and the SFM toolbox [Krause, 2010].
7.1
Average-case
The average-case benchmark data has two components: graphs and cost functions. We first describe the graphs, then the functions.
Grid graphs. The benchmark contains three variants of regular grid graphs
of degree four or six. Type I is a plane grid with horizontal and vertical edges
displayed as solid edges in Figure 4(a). Type II is similar, but has additional
diagonal edges (dashed in Figure 4(a)). Type III is a cube with plane square
grids on four faces (sparing the top and bottom faces). Different from Type I,
the nodes in the top row are connected to their counterparts on the opposite
side of the cube. The connections of the bottom nodes are analogous.
Clustered graphs. The clustered graphs consist of a number of cliques that
are connected to each other by few edges, as depicted in Figure 4(b).
Cost functions. The benchmark includes four families of functions. The first
group (Matrix rank I,II, Labels I, II ) consists of matroid rank functions or sums
of three such functions. The functions used here are either based on matrix
rank or ranks of partition matroids. We summarize those functions as rank-like
costs.
29
(a) Grids I and II
(b) Clustered graph
Figure 4: Examples of the test graph structures. The grid (a) was used with and
without the dashed diagonal edges, and also with a variation of the connections
in the first and last row. The clustered graphs were similar to the example
shown in (b).
The second group (Unstructured I, II ) contains two variants of unstructured
functions g(w(C)), where g is either a logarithm or a square root. These functions are designed to favor a certain random optimal cut. The construction
ensures that the minimum cut will not be one that separates out a single node,
but one that cuts several edges.
The third family (Bestcut I, II ) is constructed to make a cut optimal that
has many edges and that is therefore different from the cut that uses fewest
edges. For such a cut, we expect fˆadd to yield relatively poor solutions.
The fourth set of functions (Truncated rank ) is inspired by the difficult truncated functions that can be used to establish lower bounds on approximation
factors. These functions “hide” an optimal set, and interactions are only visible
when guessing a large enough part of this hidden set. The following is a detailed
description of all cost functions:
Matrix rank I. Each element e ∈ E indexes a column in matrix X ∈ Rd×m .
The cost of A ⊆ E is the rank of the sub-matrix XA of the columns indexed
by the e ∈ A: fmrI (A) = rank(XA ). The matrix X is of the form√[ I0 R ],
where R ∈ {0, 1}d×(m−d) is a random binary matrix with d = 0.9 m, and
I0 is a column-wise permutation of the identity matrix.
P3
(i)
Matrix rank II. The function fmrII (A) = 0.33 i=1 fmrI (A) sums up three
(i)
functions fmrI of type matrix rank I with different random matrices X.
S
Labels I. This class consists of functions of the form f`I (A) = √
| e∈A `(e)|.
Each element e is assigned a random label `(e) from a set of 0.8 m possible
labels. The cost counts the number of labels in A.
P3
(i)
Labels II. These functions f`II (A) = 0.33 i=1 f`I (A) are the sum of three
functions of type labels I with different random labels.
P
Unstructured I. These are functions fdpI (A) = log e∈A w(e), where weights
w(e) are chosen randomly as follows. Sample a set X ⊂ V with |X| = 0.4n,
30
and set w(e) = 1.001 for all e ∈ δX. Then randomly assign some “heavy”
weights in [n/2, n2 /4] to some edges not in δX, so that each node is
incident to one or two heavy edges. The remaining edges get random
(mostly integer) weights between 1.001 and n2 /4 − n + 1.
pP
Unstructured II. These are functions fdpII (A) =
e∈A w(e) with weights
assigned as for unstructured function II.
∗
Bestcut I. We randomly pick a connected subset
P X ⊆ V of size 0.4n and
∗
define the cost fbcI (A) = 1[|A ∩ δX | ≥ 1] + e∈A\δX ∗ w(e). The edges in
E \ δX ∗ are assigned random weights w(e) ∈ [1.5, 2]. If there is still a cut
C 6= δX ∗ with cost one or lower, we correct w by increasing the weight of
one e ∈ C to w(e) = 2. The optimal cut is then δX ∗ , but it is usually not
the one with fewest edges.
Bestcut II. Similar to bestcut I (δX ∗ is again optimal), but with submodularity on all edges: E is partitionedP
into three sets, E = (δX ∗ ) ∪ B ∪ C.
∗
Then fbcII (A) = 1[|A ∩ δX | ≥ 1] + e∈A∩(B∪C) w(e) + maxe∈A∩B w(e) +
maxe∈A∩C w(e). The weights of two edges in B and two edges in C are
set to w(e) ∈ (2.1, 2.2).
Truncated rank. This function is similar to the truncated rank in the proof
of the lower bound (Theorem 1). Sample a connected X ⊆ V with |X| =
0.3|V| and set R = δX.
p The cost is ftr (A) = min{|A ∩ R| + min{|A ∩
R|, λ1 }, λ2 } for λ1 = |R| and λ2 = 2|R|. Here, R is not necessarily the
optimal cut.
To estimate the approximation factor on one problem instance (one graph
and one cost function), we divide by the cost of the best solution found by any of
the algorithms, unless the optimal solution is known (this is the case for Bestcut
I and II ).
7.1.1
Results
Figure 5 shows average empirical approximation factors and also the worst observed factors. The first observation is that all algorithms remain well below
their theoretical approximation bounds9 . That means the theoretical bounds
are really worst-case results. For several instances we obtain optimal solutions.
The general performance depends much on the actual problem instance; the
truncated rank functions with hidden structure are, as may be expected, the
most difficult. The simple benchmarks relying on fˆadd perform worse than the
more sophisticated algorithms. Queyranne’s algorithm performs surprisingly
well here.
9 Most of the bounds proved above are absolute, and not asymptotic. The only exception
is fˆea . For simplicity, it is here treated as an absolute bound.
31
grid graphs
clustered graphs
rank-like cost functions (average over 61 (left), 80 (right) instances)
3
3
approx. factor
approx. factor
4
2
1
0
QU MC MB RI
2
1
0
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
unstructured functions (average over 30 (left), 40 (right) instances)
1.5
approx. factor
approx. factor
2
1.5
1
0.5
0
QU MC MB RI
1
0.5
0
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
bestcut functions (average over 15 (left), 20 (right) instances)
approx. factor
approx. factor
4
4
3
2
1
0
QU MC MB RI
3
2
1
0
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
approx. factor
approx. factor
4
4
3
2
1
0
QU MC MB RI
3
2
1
0
MI EA PF CR DB GM GA GH
truncated rank (average over 15 (left), 20 (right) instances)
2.5
1.5
approx. factor
approx. factor
2
1
0.5
0
QU MC MB RI
2
1.5
1
0.5
0
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
Figure 5: Results for average-case experiments. The bars show the mean empirical approximation factors, and red crosses mark the maximum observed
empirical approximation factor. The left column refers to grid graphs, the right
column to clustered graphs. The first three algorithms (bars) are baselines, the
next four approximate f , the next four solve a relaxation, and the last is the
deterministic greedy heuristic.
32
7.2
Difficult instances
Lastly, we show two difficult instances. More examples may be found in [Jegelka,
2012, Ch. 4]. The example demonstrates the drawbacks of using approximations
like fˆadd and Queyranne’s algorithm.
Our instance is a graph with n = 10 modes, shown in Figure 6. The graph
edges are partitioned into n/2 sets, indicated by colors. The black set Ek makes
up the cut with the maximum number of edges. The remaining edge sets are
constructed as
Ei = (vi , vj ) ∈ E | i < j ≤ n/2 ∪ (vn/2+i , vj ) ∈ E | n/2 + i < j ≤ n
(42)
for each 1 ≤ i < n/2. In Figure 6, set E1 is red, set E2 is blue, and so on. The
cost function is
n/2−1
X
b · 1 |A ∩ Ei | ≥ 1 + |A ∩ Ek |,
fa (A) = 1 |A ∩ Ek | ≥ 1 +
(43)
i=1
with b = n/2. The function 1[·] denotes the indicator function. The cost of the
2
optimal solution is f (C ∗ ) = f (Ek ) = 1 + n4 ≈ 1. The second-best solution is
2
the cut δ(v1 ) with cost f (δv1 ) = 1 + n4 + b ≈ 1 + n2 = 6, i.e., it is by a factor of
almost b = n/2 worse than the optimal solution. Finally, MC finds the solution
2
2
δ(vn ) with f (δvn ) = 1 + n4 + b( n2 − 1) ≈ n4 = 21.
Variant (b) uses the cost function
n/2−1
fb (A) = 1[|A ∩ Ek | ≥ 1] +
X
i=1
b · 1[|A ∩ Ei | ≥ 1]
(44)
with a large constant b = n2 = 100. For any b > n/2, any solution other than
C ∗ is more than n2 /4 = |C ∗ | > n times worse than the optimal solution. Hence,
thanks to the upper bounds on their approximation factors, all algorithms except
for QU find the optimal solution. The result of the latter depends on how it
selects a minimizer of f (B ∪ e) − f (e) in the search for a pendent pair; this
quantity often has several minimizers here. Variant (b) uses a specific adversarial
permutation of node labels, for which QU always returns the same solution δv1
with cost b + 1, no matter how large b is: its solution can become arbitrarily
poor.
8
Discussion
In this work, we have defined and analyzed the MinCoopCut problem, that
is, a minimum (s, t)-cut problem with a submodular cost function on graph
edges. This problem unifies a number of non-additive graph cut problems in
the literature that have arisen in different application areas.√
We showed an information-theoretic lower bound of Ω( n) for the general
MinCoopCut problem if the function is given as an oracle, and NP-hardness
33
vn/2+1
v1
vn
vn/2
(a)
(b)
20
15
10
5
0
100
approx. factor
approx. factor
25
QU MC MB RI
80
60
40
20
0
MI EA PF CR DB GM GA GH
QU MC MB RI
MI EA PF CR DB GM GA GH
Figure 6: Difficult instance and empirical approximation factors with n = 10
nodes. White bars illustrate theoretical approximation bounds where applicable. In (b),
√ the second-best cut δv1 has cost fb (δv1 ) = b + 1 = 101
max{|C ∗ |, n, m log m}.
even if the cost function is fully known and polynomially representable. We
propose and compare complementary approximation algorithms that either rely
on representing the cost function by a simpler function, or on solving a relaxation
of the mathematical program. The latter are closely tied to the longest path
of cooperating edges in the graph, as is the flow-cut gap. We also show that
the flow-cut gap may be as large as n − 1, and therefore larger than the best
approximation factor possible.
The lower bound and analysis of the integrality gap use a particular graph
structure, a graph with parallel disjoint paths of equal length. Taken all proposed algorithms together, all instances of MinCoopCut on graphs with parallel paths
of the same length can be solved within an approximation bound at
√
most n. This leaves the question
whether there is an instance that makes all
√
approximations worse than n.
Section 6 outlined properties of submodular functions that facilitate submodular minimization under combinatorial constraints, and also submodular
minimization in general. Apart from separability, we defined the hierarchy of
function classes F(k). The F(k) are related to graph-representability and might
therefore build a bridge between recent results about limitations of representing submodular functions as graph cuts [Z̆ivný et al., 2009] (and, even stricter,
the limitations of polynomial representability) and the results discussed in Section 6.2 that provide improved algorithms whose complexity depends on k.
8.1
Open problems
This paper is part of a growing collection of work that studies submodular cost
functions in combinatorial optimization problems over cuts, trees, matchings,
and so on. Such problems are not only of theoretical interest: they occur in a
34
spectrum of problems in computer vision [Jegelka and Bilmes, 2011a, Shelhamer
et al., 2014, Heng et al., 2015, Taniai et al., 2015] and machine learning [Iyer
et al., 2013a, Iyer and Bilmes, 2013, Khalil et al., 2014]. In several cases, the
functions used do not directly fall into any of the “easier” sub-classes (e.g., the
entropy cuts outlined in Section 2, and also see the discussion in Section 2.3).
At the same time, the empirical results in this paper and others [Iyer et al.,
2013a] suggest that in many cases, the results of approximate algorithms can
still be good, even though in the worst case they are not. Section 6 outlines
beneficial properties. Is there a more precise quantification of the complexity of
these problems? Do there exist other properties that lead to better algorithms?
One direction that is less explored is the interaction of the graph structure with
the cost function.
Specific to this work, cut functions induce a function on nodes. Propositions 2 and 3 imply that the node function can be submodular, but in very
many cases it is not. Yet, the results for Queyranne’s algorithm in Section 7
suggest that often the function h may remain close to submodular. This could be
the case, for example, if the graph is almost complete and f symmetric, or if the
symmetry of f is more restricted. A deeper study of the functions h induced by
cooperative cuts could reveal insights about a refined complexity of the problem,
and explain the good empirical results. One particular interesting example was
the polymatroidal flow case where the function f defined in Eqn. (28) was not
necessarily submodular (see Proposition 5), but where the resulting h (also not
necessarily submodular) could be optimized exactly in polynomial time. This
suggests an interesting future direction, namely to fully characterize a class of
functions f for which polytime algorithms can be obtained to solve minimum
“interacting cut” problems (i.e., cut problems where the edges may interact
but not necessarily in a purely submodular or supermodular fashion). The dual
polymatroidal flow case shows one instance of interacting cut that can be solved
exactly.
Finally, finding optimal bounds and algorithms for related cut problems with
submodular edge weights is an open problem. Appendix A outlines some initial
results for cooperative multi-way and sparsest cut.
References
R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network Flows. Prentice Hall,
1993.
C. Allène, J.-Y. Audibert, M. Couprie, and R. Keriven. Some links between
extremum spanning forests, watersheds, and min-cuts. Image and Vision
Computing, 2009.
S. Bagon. Matlab wrapper for graph cut, December 2006. http://www.wisdom.
weizmann.ac.il/~bagon.
35
N. Balcan and N. Harvey. Submodular functions: Learnability, structure, and
optimization. arXiv:0486478, 2012.
F. Baumann, S. Berckey, and C. Buchheim. Facets of Combinatorial Optimization — Festschrift for Martin Grötschel, chapter Exact Algorithms for
Combinatorial Optimization Problems with Submodular Objective Functions,
pages 271–294. Springer, 2013.
J. Bilmes. Dynamic graphical models – an overview. IEEE Signal Processing
Magazine, 27(6):29–42, 2010.
J. Bilmes and C. Bartels. On triangulating dynamic graphical models. In
Uncertainty in Artificial Intelligence, pages 47–56, Acapulco, Mexico, 2003.
Morgan Kaufmann Publishers.
Y. Boykov and M.-P. Jolly. Interactive graph cuts for optimal boundary and
region segmentation of objects in n-d images. In Int. Conf. on Computer
Vision (ICCV), 2001.
Y. Boykov and V. Kolmogorov. An experimental comparison of min-cut/maxflow algorithms for energy minimization in vision. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 26(9):1124–1137, 2004.
Y. Boykov and O. Veksler. Handbook of Mathematical Models in Computer Vision, chapter Graph Cuts in Vision and Graphics: Theories and Applications.
Springer, 2006.
F. Bunke, H. W. Hamacher, F. Maffioli, and A. Schwahn. Minimum cut bases
in undirected networks. Report in Wirtschaftsmathematik (WIMA Report)
108, Universität Kaiserslautern, 2007.
A. Chambolle and J. Darbon. On total variation minimization and surface
evolution using parametric maximum flows. Int. Journal of Computer Vision,
84(3), 2009.
C. Chekuri, S. Kannan, A. Raja, and P. Viswanath. Multicommodity flows
and cuts in polymatroidal networks. In Innovations in Theoretical Computer
Science (ITCS), 2012.
M. Conforti and G. Cornuéjols. Submodular set functions, matroids and the
greedy algorithm: tight worst-case bounds and some generalizations of the
Rado-Edmonds theorem. Discrete Applied Mathematics, 7(3):251–274, 1984.
C. Couprie, L. Grady, H. Talbot, and L. Najman. Combinatorial continuous
maximum flow. SIAM Journal on Imaging, pages 905–930, 2011.
W. H. Cunningham. Decomposition of submodular functions. Combinatorica,
3(1):53–68, 1983.
G. Dantzig and D. Fulkerson. On the max flow min cut theorem of networks.
Technical Report P-826, The RAND Corporation, 1955.
36
A. Ene, J. Vondrák, and Y. Wu. Local distribution and the symmetry gap:
Approximability of multiway partitioning problems. In Proc. SIAM-ACM
Symp. on Discrete Algorithms (SODA), 2013.
A. Fix, T. Joachims, S. M. Park, and R. Zabih. Structured learning of sum-ofsubmodular higher order energy functions. In Int. Conf. on Computer Vision
(ICCV), 2013.
L. Ford and D. Fulkerson. Maximal flow through a network. Canadian Journal
of Mathematics, 8:399–404, 1956.
M. R. Garey, D. S. Johnson, and L. Stockmeyer. Some simplified NP-complete
graph problems. Theoretical Computer Science, 1(3):237–267, 1976.
G. Goel, C. Karande, P. Tripati, and L. Wang. Approximability of combinatorial
problems with multi-agent submodular cost functions. In Proc. IEEE Symp.
on Foundations of Computer Science (FOCS), 2009.
G. Goel, P. Tripathi, and L. Wang. Combinatorial problems with discounted
price functions in multi-agent systems. In Foundations of Software Technology
and Theoretical Computer Science (FSTTCS), 2010.
M. Goemans, N. J. A. Harvey, S. Iwata, and V. S. Mirrokni. Approximating
submodular functions everywhere. In Proc. SIAM-ACM Symp. on Discrete
Algorithms (SODA), 2009.
M. X. Goemans, N. J. A. Harvey, R. Kleinberg, and V. S. Mirrokni. On learning
submodular functions – a preliminary draft. Unpublished Manuscript.
V. Goyal and R. Ravi. An FPTAS for minimizing a class of low-rank quasiconcave functions over a convex domain. Technical Report 366, Tepper School
of Business, Carnegie Mellon University, 2008.
D. M. Greig, B. T. Porteous, and A. H. Seheult. Exact maximum a posteriori
estimation for binary images. Journal of the Royal Statistical Society, 51(2),
1989.
R. Hassin. Minimum cost flow with set constraints. Networks, 12:1–21, 1982.
R. Hassin, J. Monnot, and D. Segev. Approximation algorithms and hardness
results for labeled connectivity problems. J. Comb. Optim., 14(4):437–453,
2007.
L. Heng, A. Gotovos, A. Krause, and M. Pollefeys. Efficient visual exploration
and coverage with a micro aerial vehicle in unknown environments. In IEEE
Int. Conf. on Robotics and Automation (ICRA), 2015.
S. Iwata and K. Nagano. Submodular function minimization under covering constraints. In Proc. IEEE Symp. on Foundations of Computer Science (FOCS),
2009.
37
R. Iyer and J. Bilmes. Submodular optimization with submodular cover and
submodular knapsack constraints. In Neural Information Processing Society
(NIPS), 2013.
R. Iyer, S. Jegelka, and J. Bilmes. Fast semidifferential-based submodular function optimization. In Proc. Int. Conf. on Machine Learning (ICML), 2013a.
R. Iyer, S. Jegelka, and J. Bilmes. Curvature and optimal algorithms for learning and minimizing submodular functions. In Neural Information Processing
Society (NIPS), 2013b.
S. Jegelka. Combinatorial Problems with submodular coupling in machine learning and computer vision. PhD thesis, ETH Zurich, 2012.
S. Jegelka and J. Bilmes. Submodularity beyond submodular energies: coupling edges in graph cuts. In IEEE Conf. on Computer Vision and Pattern
Recognition (CVPR), 2011a.
S. Jegelka and J. Bilmes. Approximation bounds for inference using cooperative
cuts. In Proc. Int. Conf. on Machine Learning (ICML), 2011b.
S. Jegelka, H. Lin, and J. Bilmes. On fast approximate submodular minimization. In Neural Information Processing Society (NIPS), 2011.
S. Jegelka, F. Bach, and S. Sra. Reflection methods for user-friendly submodular
optimization. In Neural Information Processing Society (NIPS), 2013.
S. Jha, O. Sheyner, and J. Wing. Two formal analyses of attack graphs. In Proc.
of the 15th Computer Security Foundations Workshop, pages 49–63, 2002.
S. Kannan and P. Viswanath. Multiple-unicast in fading wireless networks: A
separation scheme is approximately optimal. In IEEE Int. Symposium on
Information Theory (ISIT), 2011.
S. Kannan, A. Raja, and P. Viswanath. Local phy + global flow: A layering
principle for wireless networks. In IEEE Int. Symposium on Information
Theory (ISIT), 2011.
E. B. Khalil, B. Dilkina, and L. Song. Scalable diffusion-aware optimization of
network topology. In Proceedings of the 20th ACM SIGKDD international
conference on Knowledge discovery and data mining, pages 1226–1235. ACM,
2014.
P. Kohli, M. Kumar, and P. Torr. P3 & beyond: Move making algorithms for
solving higher order functions. IEEE Trans. on Pattern Analysis and Machine
Intelligence, pages 1645–1656, 2009a.
P. Kohli, L. Ladický, and P. Torr. Robust higher order potentials for enforcing
label consistency. International Journal of Computer Vision, 82(3):302–324,
2009b.
38
P. Kohli, A. Osokin, and S. Jegelka. A principled deep random field for image
segmentation. In IEEE Conf. on Computer Vision and Pattern Recognition
(CVPR), 2013.
V. Kolmogorov. Minimizing a sum of submodular functions. Discrete Applied
Mathematics, 160(15), 2012.
V. Kolmogorov and Y. Boykov. What metrics can be approximated by geo-cuts,
or global optimization of length/area and flux. In Int. Conf. on Computer
Vision (ICCV), 2005.
V. Kolmogorov and R. Zabih. What energy functions can be minimized via
graph cuts? IEEE Trans. on Pattern Analysis and Machine Intelligence, 26
(2):147–159, 2004.
C. Koufogiannakis and N. E. Young. Greedy ∆-approximation algorithm for
covering with arbitrary constraints and submodular costs. In Int. Colloquium
on Automata, Languages and Programming (ICALP), 2009.
A. Krause. SFO: A toolbox for submodular function optimization. Journal of
Machine Learning Research, 11:1141–1144, 2010.
E. L. Lawler and C. U. Martel. Computing maximal “Polymatroidal” network
flows. Mathematics of Operations Research, 7(3):334–347, 1982.
H. Lin and J. Bilmes. Learning mixtures of submodular shells with application
to document summarization. In Uncertainty in Artificial Intelligence (UAI),
2012.
L. Lovász. Mathematical programming – The State of the Art, chapter Submodular Functions and Convexity, pages 235–257. Springer, 1983.
S. Mittal and A. Schulz. An FPTAS for optimizing a class of low-rank functions
over a polytope. Mathematical Programming, 2012.
M. Mitzenmacher and E. Upfal. Probability and Computing: Randomized Algorithms and Probabilistic Analysis. Cambridge University Press, 2005.
H. Narayanan. Submodular Functions and Electrical Networks. Elsevier Science,
1997.
E. Nikolova. Approximation algorithms for reliable stochastic combinatorial
optimization. In APPROX, 2010.
M. Queyranne. Minimizing symmetric submodular functions. Mathematical
Programming, 82:3–12, 1998.
S. Ramalingam, P. Kohli, K. Alahari, and P. Torr. Exact inference in multilabel CRFs with higher order cliques. In IEEE Conf. on Computer Vision
and Pattern Recognition (CVPR), 2008.
39
S. Ramalingam, C. Russell, L. Ladicky, and P. H. S. Torr. Efficient minimization of higher order submodular functions using monotonic boolean functions.
ArXiv 1109.2304, 2011.
L. I. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise
removal algorithms. Physica D, (60), 1992.
A. Schrijver. Combinatorial Optimization. Springer, 2004.
E. Shelhamer, S. Jegelka, and T. Darrell. Communal cuts: Sharing cuts across
images. In NIPS workshop on Discrete Optimization in Machine Learning,
2014.
N. Silberman, L. Shapira, R. Gal, and P. Kohli. A contour completion model
for augmenting surface reconstructions. In Europ. Conf. on Computer Vision
(ECCV), 2014.
A. Sinop and L. Grady. A seeded image segmentation framework unifying graph
cuts and random walker which yields a new algorithm. In Int. Conf. on
Computer Vision (ICCV), 2007.
R. P. Stanley. Enumerative Combinatorics, volume I of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, 1997.
P. Stobbe and A. Krause. Efficient minimization of decomposable submodular
functions. In Neural Information Processing Society (NIPS), 2010.
P. Stobbe and A. Krause. Learning Fourier sparse set functions. In Conference
on Artificial Intelligence and Statistics (AISTATS), 2012.
Z. Svitkina and L. Fleischer. Submodular approximation: Sampling-based algorithms and lower bounds. In Proc. IEEE Symp. on Foundations of Computer
Science (FOCS), 2008.
T. Taniai, Y. Matsushita, and T. Naemura. Superdifferential cuts for binary energies. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR),
2015.
E. Tardos, C. A. Tovey, and M. A. Trick. Layered augmenting path algorithms.
Mathematics of Operations Research, 11(2), 1986.
S. Tschiatschek, R. Iyer, H. Wei, and J. Bilmes. Learning mixtures of submodular functions for image collection summarization. In Neural Information
Processing Society (NIPS), 2014.
J. Vondrák. Submodularity and curvature: the optimal algorithm.
Kôkyûroku Bessatsu, 2008.
RIMS
J. Vondrák. Symmetry and approximability of submodular maximization problems. SIAM J on Computing, 42(1):265–304, 2013.
40
P. Zhang, C. J.-Y, L.-Q. Tang, and W.-B. Zhao. Approximation and hardness
results for label cut and related problems. Journal of Combinatorial Optimization, 2011.
S. Z̆ivný, D. A. Cohen, and P. G. Jeavons. The expressive power of binary submodular functions. Discrete Applied Mathematics, 157(15):3347–3358, 2009.
A
Cooperative multi-cut and sparsest cut
An extension of MinCoopCut is the problem of cooperative multi-way cut and
sparsest cut. Using the approximation fˆea from Section 5.1.3, we can transform
any multi-way or sparsest cut problem with a submodular cost function on
edges (instead of a sum of edge weights) into a cut problem whose cut cost is a
convolution of local submodular functions. The relaxation of this cut problem
is dual to the polymatroidal flow problems considered by Chekuri et al. [2012].
Combining their results with ours, we get the following Lemma.
Lemma 11. Let α be the approximation factor for solving a sparsest cut /
multi-way cut in a polymatroidal network. If we solve a cooperative sparsest cut
/ multi-way cut by first approximating the cut cost f by a function fˆea and, on
this instance, using the method with factor α, we get an O(αn)-approximation
for cooperative sparsest cut / multi-way cut.
Using Theorems 6 and 8 in [Chekuri et al., 2012], we obtain the following
bounds:
Corollary 2. There is an O(n log k) approximation for cooperative sparsest cut
in undirected graphs that is dual to a maximum multicommodity flow problem
with k pairs, and an O(n log k) approximation for cooperative multi-way cut.
We leave it as an open problem whether these bounds are optimal.
B
Proof of Proposition 2
The first part of Proposition 2 is proven by Figure 1. Here, we show the second
part that the function h(X) = f (δ + (X)) is subadditive if f is nondecreasing
and submodular. Let X, Y ⊆ V. Then it holds that
h(X) + h(Y ) = f (δ + (X)) + f (δ + (Y ))
(45)
+
+
+
+
+
+
≥ f (δ (X) ∪ δ (Y ))
(47)
+
≥ f (δ (X ∪ Y ))
(48)
= h(X ∪ Y ).
(49)
≥ f (δ (X) ∪ δ (Y )) + f (δ (X) ∩ δ (Y ))
(46)
In Inequality (46), we used that f is submodular, and in Inequality (47), we
used that f is nonnegative.
41
(s, v1 ) (s, v2 ) (s, v3 ) (s, v4 ) (s, v5 ) (s, v6 )
s
t
(v1 , t) (v2 , t) (v3 , t) (v4 , t) (v5 , t) (v6 , t)
(a) graph G with Es (blue), Et (red) and
GB (black)
(b) graph Hσ
Cs
(s, v1 ) (s, v2 ) (s, v3 ) (s, v4 ) (s, v5 ) (s, v6 )
Cs (s, v1) (s, v2) (s, v3) (s, v4) (s, v5) (s, v6)
Ct
(v1 , t) (v2 , t) (v3 , t) (v4 , t) (v5 , t) (v6 , t)
Ct (v1, t) (v2, t) (v3, t) (v4, t) (v5, t) (v6, t)
(c) hσ (φ(C)) = 5 connected components
(d) balanced cut C: hσ (φ(C)) = 3 connected components
Figure 7: Graph for the reduction and examples for the definition of fbal
via ranks hσ , with nB = 6. In (c), Cs = {(s, v1 ), (s, v2 )} and Ct =
{(v3 , t), (v4 , t), (v5 , t), (v6 , t)}; in (d), Cs = {(s, v1 ), (s, v4 ), (s, v5 )} and Ct =
{(v2 , t), (v3 , t), (v6 , t)}. Connected components are indicated by dashed lines.
C
Reduction from Graph Bisection to MinCoopCut
In this section, we prove Theorem 2 via a reduction from Graph Bisection,
which is known to be NP-hard [Garey et al., 1976].
Definition 2 (Graph Bisection). Given a graph GB = (VB , EB ) with edge
˙ 2 = VB with |V1 | = |V2 | = |VB |/2
weights wB : EB → R+ , find a partition V1 ∪V
with minimum cut weight w(δ(V1 )).
Proof. To reduce Graph Bisection to MinCoopCut, we construct an auxiliary graph G = (VB ∪ {s, t}, EB ∪ Es ∪ Et ) that contains an unchanged copy
of GB and two additional terminal nodes s, t. The submodular weights on
the edges adjacent to the terminal nodes will express the balance constraint
|V1 | = |V2 | = |VB |/2. In G, we retain the modular costs w on EB and connect
s, t to every vertex in GB with corresponding new edge sets Es and Et , as illustrated in Figure 7(a). The cost of a cut in G is measured by the submodular
function
X
f (C) =
w(e) + βfbal (C ∩ (Es ∪ Et )),
(50)
e∈C∩EB
where β is an appropriately large constant, and fbal will be defined later. Obviously, any minimal (s, t)-cut C must include nB = |VB | edges from Es ∪ Et , and
partitions VB . Moreover, the cardinality of Cs = C ∩ Es is the number of nodes
in VB assigned to t. Hence, in an equipartition, |Cs | = |Ct | = nB /2, where
Ct = C ∩ Et .
42
It remains to define fbal as a nondecreasing submodular function that implements the equipartition constraint. The function will be an expectation over
matroid rank functions hσ . Let Hσ = (Es , Et , Fσ ) be a bipartite graph with
nodes Es ∪ Et whose edges Fσ form a derangement between Es and Et , as illustrated in Fig. 7(b).
Let φ(Cs ∪ Ct ) be the image of Cs ∪ Ct in the set of nodes of Hσ . The
function hσ : 2φ(Es ∪Et ) → N0 counts the number of connected components in
the subgraph induced by the nodes φ(Cs ∪ Ct ), and is the rank of a partition
matroid. Figure 7 shows some examples.
Let S be the set of all derangements σ of nB elements, i.e., all possible
edge configurations in Hσ . We define fbal to be the expectation (under uniform
draws of σ)
X
fbal (C) = Eσ [hσ (φ(C))] = |S|−1
hσ (φ(C)).
(51)
σ∈S
For a fixed derangement σ 0 and a fixed size |Cs ∪Ct | = nB , the value hσ0 (Cs ∪Ct )
is minimal if σ 0 (Cs ) = Ct and |Cs | = |Ct |. For a fixed σ, the rank hσ (C) is
|φ(Cs ∪ Ct )| = |Cs | + |Ct | minus the matching nodes. Denoting (s, vi ) in Hσ by
xi and (vi , t) by yi , the rank is
hσ (φ(Cs ) ∪ φ(Ct )) = |Cs | + |Ct | −
(xi , yσ(i) )
n
i=1
Hence, the sum in (51) becomes
X
X
(xi , yσ(i) )
hσ (C) = |S| |Cs | + |Ct | −
σ∈S
n
i=1
σ∈S
= |S| |Cs | + |Ct | −
X
X
xi ∈φ(Cs ) σ∈S
∩ (φ(Cs ) × φ(Ct )) . (52)
∩ (φ(Cs ) × φ(Ct )) (53)
(xi , yσ(i) ) ∩ ({xi } × φ(Ct )) (54)
To compute the sum over σ in the second term, let Cs∩t , {(s, v) | {(s, v), (v, t)} ⊆
C} be the set of s-edges whose counterpart on the t side is also contained in C.
Let further D0 (nB − 1) denote the number of permutations of nB − 1 elements
(pair (xi , yk ), i.e., σ(i) = k, is fixed), where one specific element xk can be
mapped to any other of the nB − 1 elements, and the remaining elements must
not be mapped to their counterparts (σ(j) 6= j). Then there are D0 (nB − 1)
derangements σ realizing a specific mapping σ(i) = k. Denoting the number of
derangements of n elements by D(n), the sum above becomes
D(nB )fbal (C) = (|Cs | + |Ct |)D(nB )
X
X
−
D0 (nB − 1) −
xi ∈Cs \Cs∩t yk ∈Ct
(55)
X
X
xi ∈Cs∩t yk ∈Ct ,k6=i
D0 (nB − 1)
= (|Cs | + |Ct |)D(nB )
(56)
− |Cs | − |Cs∩t | |Ct |D0 (nB − 1) − |Cs∩t |(|Ct | − 1)D0 (nB − 1)
= (|Cs | + |Ct |)D(nB ) − (|Cs ||Ct | − |Cs∩t |)D0 (nB − 1),
43
(57)
Pn
Pn−1
with D(n) = |S| = n! k=0 (−1)k /k! [Stanley, 1997], and D0 (n−1) = k=0 (n−
2)!(n − 1 − k)!(−1)k by Proposition 4 below.
Given that |Cs | + |Ct | must cut at least nB edges and that fbal is increasing,
fbal is minimized if |Cs | = |Ct | = nB /2. As a result, if β is large enough such
that fbal dominates the cost, then a minimum cooperative cut in G bisects the
GB subgraph of G optimally.
Proposition 4. Let D0 (n) be the number of permutations of n elements where
for one fixed element i0 we allow σ(i0 ) ∈ {1, . . . , n}, but σ(i) 6= i for all i 6= i0 .
Pn
k
Then D0 (n) = k=0 (n−1)!
k! (n − k)!(−1) .
Proof. D0 (n) can be derived by the method of the forbidden board [Stanley,
1997, pp. 71-73]. Let, without loss of generality, i0 = n, so the forbidden board
is B = {(1, 1), (2, 2), . . . , (n − 1, n − 1)}. Let Nj be the number of permutations
σ for which {(i, σ(i)}ni=1 ∩ B = j, and let rk be the number of k-subsets of B
such that no two elements have a coordinate in common. Here, rk = n−1
k .
Then D0 (n) = N0 = Nn (0) for
Nn (x) =
X
Nj xj =
j
n
X
k=0
rk (n − k)!(x − 1)k =
n
X
(n − 1)!
k=0
k!
(n − k)(x − 1)k ,
(58)
and hence D0 (n) =
D
Pn
k=0
(n−1)!
k! (n
− k)!(−1)k .
Convolutions of submodular functions are not
always submodular
The non-submodularity of convolutions was mentioned already in [Lovász, 1983].
For completeness, we show an explicit example that illustrates that non-submodularity
also holds for the special case of polymatroidal flows.
Proposition 5. The convolution of two submodular functions (f ∗ g)(A) =
minB⊆A f (B) + g(A \ B) is not in general submodular. In particular, this also
holds for the cut cost functions occurring in the dual problems of polymatroidal
maximum flows.
To show Proposition 5, consider the graph in Figure 5 with a submodular
edge cost function f (A) = maxe∈A w(e). The two submodular functions that
are convolved in the corresponding polymatroidal flow are the decompositions
X
capout (A) =
f (A ∩ δ + (v))
(59)
v∈V
in
cap (A) =
X
v∈V
44
f (A ∩ δ − (v)).
(60)
e1
1
t
e2
e5
e4
s
e3
Let f (A) = maxe∈A w(e) and
w(e1 ) = w(e2 ) = a,
w(e3 ) = b,
w(e4 ) = w(e5 ) = .
2
Figure 8: Example showing that the convolution of submodular functions is not
always submodular, e.g., for a = 1.5, b = 2 and = 0.001.
Both capout and capin are submodular functions from 2E to R+ . Their convolution h is
h(A) = (capout ∗ capin )(A) = min capout (B) + capin (A \ B) = fˆpf (A).
B⊆A
(61)
For h to be submodular, it must satisfy the condition of diminishing marginal
costs, i.e., for any e and A ⊆ B ⊆ E \ e, it must hold that h(e | A) ≥ h(e | B).
Now, let A = {e2 } and B = {e1 , e2 }. The convolution here means to pair e3
either with e1 or e2 . Then, if a < b,
h(e3 | A) = min{a + b, b} − a = b − a
h(e3 | B) = a + b − min{a + a, a} = b.
(62)
(63)
Hence, h(e3 | A) < h(e3 | B), disproving submodularity of h.
E
Cooperative Cuts and Polymatroidal Networks
We next prove Lemma 6 that relates the approximation fˆpf to maxflow problems
in polymatroidal networks.
Proof. (Lemma 6) The first step is the dual of the polymatroidal
flow. Let
P
capin : 2E → R+ be the joint incoming capacity, capin (C) = v∈V capin
v (C ∩
δ − v), and let equivalently capout be the joint outgoing capacity. The dual of
the polymatroidal maximum flow is a minimum cut problem whose cost is a
convolution of edge capacities [Lovász, 1983]:
cap(C) = (capin ∗ capout )(C) , min capin (A) + capout (C \ A) .
(64)
A⊆C
We will relate this dual to the approximation fˆpf . Given a minimal (s, t)-cut
C, let Π(C) be a partition of C, and Cvin = CvΠ ∩ δv− and Cvout = CvΠ ∩ δv+ . The
cut C partitions the nodes into two sets Vs containing s and Vt containing t.
Since C is a minimal directed cut, it contains only edges from the s side Vs to
the t side Vt of the graph. In consequence, Cvin = ∅ if v is on the s side, and
Cvout = ∅ otherwise. Hence, Cvin ∪ Cvout is equal to either Cvin or Cvout , and since
45
f (∅) = 0, it holds that f (Cvin ∪ Cvout ) = f (Cvin ) + f (Cvout ). Then, starting
the definition of fˆpf ,
X
fˆpf (C) = min
f (CvΠ )
v∈V
Π(C)∈PC
X
= min
f (Cvin ∪ Cvout )
v∈V
Π(C)∈PC
X
= min
f (Cvin ) + f (Cvout )
v∈V
Π(C)∈PC
X
in in
out
= min
capv (Cv ) + capout
v (Cv )
v∈V
Π(C)∈PC
= min capin (C in ) + capout (C out )
in
out
C ,C
= min capin (C in ) + capout (C \ C in )
C in ⊆C
= (capin ∗ capout )(C).
with
(65)
(66)
(67)
(68)
(69)
(70)
(71)
The minimum in Equation (67) is taken over all feasible partitions Π(C) and
their resulting
with the sets δ + v, δ − v. Then we use the notaS intersections
in
out
in
tion
=
S C out= v∈V Cv for all edges assigned to their head nodes, and C
v∈V Cv . The minima in Equations (69) and (70) are again taken over all
partitions in PC . The final equality follows from the above definition of a convolution of submodular functions.
46
| 8cs.DS
|
arXiv:1505.06881v3 [math.GR] 21 Feb 2016
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY
INVARIABLE GENERATION
T. GELANDER AND C. MEIRI
Abstract. It was suggested in [KLS14] that for arithmetic groups Invariable Generation
is equivalent to the Congruence Subgroup Property. In this paper we dismiss this conjecture by proving that certain arithmetic groups which possess the later property do not
possess the first one.
1. introduction
In an attempt to give a purely algebraic characterisation for the Congruence Subgroup
Property (CSP), it was asked in [KLS14] whether for arithmetic groups the CSP is equivalent to Invariable Generation (IG). This was partly motivated by the result of [KLS14] that
the pro-finite completion of an arithmetic group with the CSP is topologically finitely invariably generated. Another result in the positive direction was obtained in [G15], namely
that discrete subgroups, and in particular arithmetic subgroups, of rank one simple Lie
groups are not IG. Recall the famous Serre conjecture that an irreducible arithmetic lattice
Γ ≤ G in a semisimple Lie group has the CSP iff rankR (G) ≥ 2. In the current paper
we give a negative answer to the KLS question by producing various counterexamples of
arithmetic groups in higher rank which are not IG. Some of our examples are known to
possess the CSP.
1.1. Invariable generation. Recall that a subset S of a group Γ invariably generates Γ if
Γ = hsg(s) |s ∈ Si for every choice of g(s) ∈ Γ, s ∈ S. One says that a group Γ is invariably
generated, or shortly IG if one of the following equivalent conditions holds:
(1) There exists S ⊆ Γ which invariably generates Γ.
(2) The set Γ invariably generates Γ.
(3) Every transitive permutation representation on a non-singleton set admits a fixedpoint-free element.
(4) Γ does not have a proper subgroup which meets every conjugacy class.
The main results of this paper are that certain lattices are not IG.
A matrix g ∈ GLn (C) is net if the multiplicative group A(g) generated by the eigenvalues
of g does not contain any nontrivial root of unity. A linear group Γ ≤ GLn (C) is net if all
its nontrivial elements are net. It is well known that every finitely generated linear group
Date: September 11, 2017.
1
2
T. GELANDER AND C. MEIRI
admits a finite index net subgroup (see [R72, Theorem 6.11]). The assumption that Γ is
net is not crucial but very convenient for our purpose.
Theorem 1.1. Let Q be a rational quadratic form of signature (2, 2), let G = SO(Q) and
let Γ be a finite index net subgroup in G(Z). Then Γ is not invariably generated.
The group Γ in Theorem 1.1 is an irreducible lattice in the higher rank semisimple group
G = SO(2, 2) ∼
= PSL(2, R) × PSL(2, R), and as such is super-rigid, possesses the normal
subgroup property and the strong approximation theorem. By Kneser’s theorem [KM79]
Γ has the CSP (when the form is isotropic this also follows from Raghunathan’s theorem
[R76, R86]). In particular the groups obtained in this way are counterexamples to the
question of [KLS14].
We also prove a general theorem (see Theorem 2.2) that under a certain condition on
G every regular lattice in G (see Section 2.1 for the definition) is not IG. This produces a
large family of examples, for instance:
Example 1.2. (See [WM14, Proposition 6.61]) Let L be a cubic Galois extension of Q.
Let D be a central division algebra of degree 3 over Q which contains L as a subfield. Let
OD be an order in D. Then SL(1, OD ) is isomorphic to an anisotropic arithmetic lattice in
SL(3, R).
Being a lattice in SL(3, R), the group SL(1, OD ) possesses many rigidity properties such
as those mentioned after Theorem 1.1 and even more; for instance it has Kazhdan’s property (T) and IRS-rigidity (the Stuck–Zimmer theorem). The corresponding lattice is cocompact. By Serre’s conjecture, SL(1, OD ) is expected to possess the CSP, but this is
currently unknown.
Theorem 1.3. The arithmetic group SL1 (OD ) is not IG.
Theorem 2.2 is more general and applies to many other example.
1.2. The motivation. Apart from trying to better understand the notion of invariable
generation in infinite groups1, our motivation arose from an attempt to understand better
‘thin subgroups’ of arithmetic groups. For instance an important question in this theory is
whether every element belongs to some thin subgroup.
Definition 1.4. Let Γ be an arithmetic group. Let us say that an element γ ∈ Γ is fat if
every Zariski dense subgroup containing it is automatically of finite index, i.e. γ does not
belong to any thin subgroup of Γ.
The existence and classification problems of fat elements are extremely interesting and
may shed light on many problems concerning thin subgroups. Being fat is clearly invariant under conjugacy, hence the existence of such elements is closely related to invariable
generation. Indeed, if Γ admits an infinite index Zariski dense subgroup which meets every
conjugacy class then there are no fat elements in Γ. Therefore our results imply that in
certain arithmetic groups all elements are ‘thin’:
1The
origin of this notion is from Galois theory, and it is well known that finite groups are always IG.
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY INVARIABLE GENERATION
3
Corollary 1.5. The arithmetic groups from Theorem 1.1 and Theorem 2.2 do not admit
fat elements.
This applies for instance to certain lattices in SL(3, R) (cf. Example 1.2). However, our
method does not apply to SL(3, Z). New tools are required in order to study property IG
or existence of fat elements for this group.
Question 1.6. Is SL(3, Z) invariably generated? What about SL(n, Z) for n ≥ 4?
1.3. About the proofs. As in [W76] and [G15] the idea is to construct an independent
set consisting of one representative of each non-trivial conjugacy class. We use of course
the dynamics on the associated projective space P, where we construct such an infinite
family of ping-pong partners. In the proof of Theorem 2.2 we use the standard action on
points in P. The dynamical picture in the proof of Theorem 1.1 is more involved and we
are led to consider the action on lines in P rather than points. Yet the spirit of the proof
is similar.
The main technical novelty of this paper is a method that allows us to conjugate certain
elements, in particular R-regular elements, such that their dynamics will simulate unipotent
dynamics — in the sense that all non-trivial powers contract the complement of some small
open set into that set. We first allow ourselves to use elements from the enveloping group
G and then apply Poincaré recurrence theorem in order to approximate the solution in Γ.
1.4. Acknowledgment. The first author was partially supported by ISF-Moked grant
2095/15. The second author was partially supported by ISF grant 662/15.
2. R-regular lattices
Suppose that G = G(R) is a Zariski connected linear real algebraic group with a faithful
irreducible representation in a d-dimensional real vector space, with associated projective
space P ∼
= Pd−1 (R). We identify G with its image in GLd (R). An element g ∈ G is regular
if it is either a regular unipotent or all its eigenvalues are of distinct absolute value (in
particular each has multiplicity one). A subgroup Γ ≤ G is regular if every nontrivial
element of Γ is regular. For an element g ∈ G we denote by vg and Hg the attracting point
and repelling hyperplane of g in P, assuming they are unique. Let us say that a point or a
hyperplane is G-defined if it is of the form vg (resp. Hg ) for some g ∈ G. In addition let us
say that an incident pair (v, H) is G-defined if there is a regular element g ∈ G for which
v = vg and H = Hg−1 . Let us also say that an ordered pair (U, V ) of open neighborhoods
in P is G-applicable if there is a G-defined incident pair (v, H) with v ∈ U and H entirely
contained in the interior of the complement of V . We shall further suppose the following:
Assumption 2.1. The set of G-defined incident pairs is not contained in a proper subvariety of the set of incident pairs.
Theorem 2.2. Let G be a Zariski connected linear real algebraic group with a faithful
irreducible representation on a d-dimensional vector space which satisfies Assumption 2.1
4
T. GELANDER AND C. MEIRI
and let Γ ≤ G be a regular lattice. Then Γ admits an infinite rank free subgroup which
intersects every conjugacy class. In particular, Γ is not invariably generated.
Remark 2.3. For G = SLd (R) with its standard d-dimensional representation Assumption
2.1 is obviously satisfied. Moreover it is easy to check that the lattices in Example 1.2 are
regular, by construction. Thus Theorem 1.3 is a special case of Theorem 2.2.
2.1. The proof of Theorem 2.2.
Proposition 2.4 (Main Proposition). Let C be a nontrivial conjugacy class in Γ, and let
(U, V ) be a G-applicable pair of open sets in P. Then there is γ ∈ C such that γ n · V, n ∈
Z \ {0} are all disjoint and contained in U.
Lemma 2.5. Let γ ∈ Γ \ {1}, and let (U, V ) be a G-applicable pair of open sets in P. Then
there is δ ∈ Γ such that
(1) Hδ−1 ∩ V = ∅,
/ Hδ ,
(2) vγ , vγ −1 ∈
/ Hγ ∪ Hγ −1 ,
(3) vδ−1 ∈
(4) vδ ∈ U, and
/ Hδ for all n 6= 0.
(5) γ n · vδ−1 ∈
Proof. As a first step, observe that there exists a regular g ∈ G such that all the five
conditions hold with g instead of δ. Indeed, as (U, V ) is G-applicable there is g for which
(1) and (4) are satisfied. Moreover, in view of Assumption 2.1, for every n ∈ N \ {0} the set
of regular g ∈ G such that γ n · vg−1 ∈ Hg is nowhere-dense. Hence by the Baire category
theorem and the fact that G admits an analytic structure, we may slightly deform g in order
to guarantee that condition (5) holds as well. Since G is Zariski connected and irreducible,
the set of h ∈ G such that either (2) or (3) do not hold for g h is nowhere-dense. Thus, up
to replacing g by a conjugate g h with h sufficiently close to 1, all the five conditions are
satisfied.
Next note that γ n · vg−1 tends to vγ when n → ∞ and to vγ −1 when n → −∞. Thus if
we pick a sufficiently small symmetric identity neighborhood Ω ⊂ G and ǫ > 0 sufficiently
small, then we have:
(1) Ω · (Hg−1 )ǫ ∩ V = ∅,
/ Ω · (Hg )ǫ ,
(2) vγ , vγ −1 ∈
(3) Ω · (vg−1 )ǫ ∩ (Hγ ∪ Hγ −1 ) = ∅,
(4) Ω · (vg )ǫ ⊂ U,
(5) γ n Ω · (vg−1 )ǫ ∩ Ω · (Hg )ǫ = ∅ for all n 6= 0, and
(6) Ω · (vg )ǫ ∩ Ω · (Hg )ǫ = ∅ = Ω · (vg−1 )ǫ ∩ Ω · (Hg−1 )ǫ .
Finally we wish to replace g by an element δ ∈ Γ with similar dynamical properties.
Since Γ is a lattice, G/Γ carries a G-invariant probability measure. Let π : G → G/Γ
denote the quotient map. By the Poincaré recurrence theorem we have that for arbitrarily
large m, g m · π(Ω) ∩ π(Ω) 6= ∅, which translates to Ωg m Ω ∩ Γ 6= ∅, since Ω was chosen to
be symmetric.
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY INVARIABLE GENERATION
5
Note that replacing g by a power g m , m > 0 does not change the attracting points and
repelling hyperplanes, while for larger and larger m the element g m is becoming more and
more proximal. In particular, if m is sufficiently large then
g m · (P \ (Hg )ǫ ) ⊂ (vg )ǫ , and g −m · (P \ (Hg−1 )ǫ ) ⊂ (vg−1 )ǫ .
In view of Item (6), this implies that any element of the form h = w1 g m w2 with w1 , w2 ∈ Ω
is very proximal and satisfies
vh ∈ Ω · (vg )ǫ , vh−1 ∈ Ω · (vg−1 )ǫ , Hh ⊂ Ω · (Hg )ǫ , Hh−1 ⊂ Ω · (Hg−1 )ǫ .
In particular, we obtain the desired δ by choosing any element from the nonempty set
Γ ∩ Ωg m Ω.
Proof of Proposition 2.4. Pick γ ′ in C and let δ be the element provided by Lemma 2.5.
Items (2,3,5) of the lemma together with the contraction property of high powers of γ ′
guarantee that if O is a sufficiently small neighborhood vδ−1 then its hγ ′ i orbit stays away
from Hδ , i.e. there is a neighborhood W of Hδ disjoint from all γ ′n · O, n ∈ Z \ {0}.
In particular the hγ ′ i translates of O are all disjoint, i.e. O is hγ ′ i-wondering (in the
terminology of [G15]). If k is sufficiently large then by Item (1) of the lemma δ −k · V ⊂ O
while by Item (4), δ k · (P \ W ) ⊂ U. It follows that γ = δ k γ ′ δ −k is our desired element.
We shall also require the following:
Lemma S
2.6. There is a countable family of disjoint open sets Uk , k ∈ N such that if we
let Vk = i6=k Ui for any k ∈ N then the pairs (Uk , Vk ), k ∈ N are all G-applicable.
Proof. Let g ∈ G be a regular element and set (v, H) = (vg , Hg ). Pick h ∈ G near the
identity such that h · v ∈
/ H and v ∈
/ h · H. Let c : [0, 1] → G be an analytic curve
with c(0) = g and c(1) = h. Since the curve c is analytic, the boundary conditions allow
us to chose a sequence tk → 1 such that if we let (vk , Hk ) = (c(tk ) · v, c(tk ) · H) and
(v∞ , H∞ ) = (h · v, h · H), then for all i 6= j ≤ ∞ we have vi ∈
/ Hj . In addition, the set of
pairs {(vk , Hk ) : k < ∞} is discrete and accumulates to (v∞ , H∞ ). Thus, we may choose
neighborhoods Uk ⊃ vk sufficiently small so that Ui ∩ Hj = ∅, ∀i 6= j and so that the
only point in ∪Uk \ ∪Uk is h · v. It is evident that the sets Uk , k ∈ N satisfy the required
property.
The strategy of the proof of Theorem 2.2 is as follows. Let (Uk , Vk ), k ∈ N be as in
Lemma 2.6. Let Ck be an enumeration of the non-trivial conjugacy classes in Γ. In view
of Proposition 2.4, we may pick γk ∈ Ck so that
γkn · Vk ⊂ Uk , ∀n ∈ Z \ {0}.
In order to complete the proof of Theorem 2.2 we may address the following variant of
the ping-pong lemma (cf. [G15, Proposition 3.4]):
6
T. GELANDER AND C. MEIRI
Proposition 2.7. Let Γ be a group and X a compact Γ-space. Let γk ∈ Γ, k ∈ N be
elements of Γ, and suppose there are disjoint subsets Uk ⊂ X such that for all k =
6 j and
n
all n ∈ Z \ {0} we have γk · Uj ⊂ Uk . Then
∆ = hγk , k = 1, 2, 3, . . .i = ∗k hγk i,
is the free product of the infinite cyclic groups hγk i, k = 1, 2, 3, . . ..
Furthermore, the limit set L(∆) is contained in ∪k Uk .
3. Arithmetic groups of type SO(2, 2)
Assumption 3.1. Let q(x) be an integral quadratic form of signature (2, 2) and let (·, ·)
be the corresponding bilinear form. Let Γ ≤ SO(q, Z) be a finite index net subgroup.
Lemma 3.2. Let f be the characteristic polynomial of some g ∈ Γ. Then f is self reciprocal
and one of the following two conditions holds:
• All the roots of f are real.
• All the roots of f are non-real complex numbers of absolute values different than 1.
Proof. We first show that f is reciprocal. The free coefficient of f is 1 and f is a monic
polynomial of degree 4. Thus, in order to show that f is reciprocal it is enough to prove
that if α is a root of f then so is 1/α. Let v1 ∈ C4 be an eigenvector of g with eigenvalue
α. Denote V := {x ∈ C4 | (v1 , x) = 0} and let v2 ∈ C4 be such that (v1 , v2 ) = 1. Then,
gv2 = (1/α)v2 + v3 for some v3 ∈ V . Moreover, g preserves V and acts on C4 /V as
multiplication by 1/α. This implies that 1/α is an eigenvalue of g.
The next step is to show that it is not possible for f to have a real root α with absolute
values different than 1 as well as a non-real root reiθ . Assume otherwise, since f is real
and reciprocal r = 1 and f (x) = (x − α)(x − α1 )(x − eiθ )(x − e−iθ ) where α and θ are
real numbers, |α| =
6 1 and θ is not an integral multiple of π. Choose v1 ∈ R4 such that
gv1 = αv1 . Define U := {x ∈ R4 | (g − eiθ )(g − e−iθ )x = 0}, V := {x ∈ R4 | (v1 , x) = 0}
and W := {x ∈ R4 | (g − α)(g − α1 )x = 0}. The non-trivial subspace U ∩ V of U is invariant
under g so that the assumption on θ implies that it must be equal to U. In particular, every
element of U is orthogonal to v1 . Choose v2 ∈ W to be an eigenvector with eigenvalue α1 .
Let x ∈ U be orthogonal to v2 , then
1
(1)
0 = (v2 , x) = (gv2 , gx) = (v2 , gx).
α
Thus, the assumption on θ implies that (v2 , x) = 0 for every x ∈ U. It follows that R4 is an
orthogonal sum of W and U and either W and U both have signature (1, 1) or one of them is
positive definite and the other is negative definite. The second case is not possible since an
orthogonal transformation of a definite space cannot have an eigenvalue with absolute value
strictly greater than 1. The first case cannot happen since any orthogonal transformation
of a quadratic space of signature (1, 1) is diagonalisable over R.
If z ∈ C \ R is a root of f ∈ Z[x] then so is z̄. Thus if f (x) has a non-real root z = reiθ
with absolute value greater than 1 then re−iθ , 1r eiθ and 1r e−iθ are the other roots of f .
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY INVARIABLE GENERATION
7
Finally, f (x) cannot have a non-real root with absolute value 1 since then the previous
paragraph implies that all the roots of f (x) must have absolute value 1. Thus they are
roots of unity which contradicts the assumption about the eigenvalues of elements in Γ.
Lemma 3.3. Let γ ∈ Γ be a non-trivial element and denote its eigenvalues by α, β, β1 , α1
where |α| ≥ |β| ≥ 1. Then there exists a base B = (v1 , v2 , v3 , v4 ) of R4 for which the
representative matrices [γ]B and [q]B are as in Table 1.
Proof. We start by proving that the Jordan blocks of elements in Γ are of sizes [1, 1, 1, 1], [2, 2]
or [3, 1]. Assume that γ has Jordan blocks of sizes [2, 1, 1]. By replacing γ with γ −1
if necessary we can assume that there exists a base B = (v1 , v2 , v3 , v4 ) of R4 such that
[γ]B = diag(J2 (α), α1 , α1 ). For every i 6= 2 and every m
(2)
(v2 , vi ) = (γ m v2 , γ m vi ) = (αm v2 + mαm−1 v1 , α±m vi )
so v1 is orthogonal to v1 , v3 and v4 . Moreover, (v2 , v1 ) = 0 since for every m ≥ 1,
(3)
(v2 , v2 ) = (γ m v2 , γ m v2 ) = α2m (v2 , v2 ) + 2mα2m−1 (v2 , v1 ) + m2 α2m−2 (v1 , v1 ).
Thus, v1 is orthogonal to R4 , a contradiction.
Assume that γ has a unique Jordan block of size 4 and let B = (v1 , v2 , v3 , v4 ) be a base
of R4 such that[γ]B = J4 (1). Then for every m ≥ 1:
(4)
(v1 , v4 ) = (γ m v1 , γ m v4 ) = (v1 , v4 + mv3 +
m(m − 1)
m(m − 1)(m − 2)
v2 +
v1 )
2
6
m(m − 1)
m(m − 1)
v1 , v3 + mv2 +
v1 )
2
2
From Equation 4 it follows that v1 is orthogonal to v1 , v2 and v3 . Equation 5 then shows
that v2 is orthogonal to itself and to v3 . Thus, the 2-dimensional subspace spanned by
v1 and v2 is orthogonal to the 3-dimensional linear subspace spanned by v1 , v2 and v3 , a
contradiction.
In the rest of the proof we exhibit the existence of the required base by a case by case
verification (Lemma 3.2 and the first part of the proof imply that the cases below account
for all possible cases).
(5)
(v3 , v3 ) = (γ m v3 , γ m v3 ) = (v3 + mv2 +
Case 1: Let V := {x | (γ 2 − 2r cos θγ + r 2 )x = 0} and U := {x | (γ 2 − 2r cos θγ + r12 )x = 0}.
Then the restriction
of q to
V and to U is the zero form. There exists a base v1 , v2 of V such
c d
that [γ]v1 ,v2 =
where c and d are as in Table 1. Since (·, ·) is non-degenerate
−d c
there exists a base v3 , v4 of U such that [q]v1 ,v2 ,v3 ,v4 has the required form. Since γ belong
to the orthogonal group of q, [γ]v1 ,v2 ,v3 ,v4 must have the required form.
Cases 2 and 3 can be readily verified.
8
T. GELANDER AND C. MEIRI
Cases 4: Assume that the Jordan blocks of γ are of sizes [2, 2] and α 6= 1. Let B :=
(v1 , v2 , v3 , v4 ) be a base such that [γ]B has the required form. For every m:
(6)
(v2 , v2 ) = (γ m v2 , γ m v2 ) = (αm v2 + mαm−1 v1 , αm v2 + mαm−1 v1 )
so the restriction of q to the span of v1 and v2 is zero and similarly the restriction of q to
the span of v3 and v4 is zero. Moreover, for every m
(7)
(v2 , v3 ) = (γ m v2 , γ m v3 ) = (αm v2 + mαm−1 v1 , α−m v3 )
so (v1 , v3 ) = 0.
Moreover, α2 (v2 , v3 ) = −(v1 , v4 ) since for every m
(8)
(v2 , v4 ) = (γ m v2 , γ m v4 ) = (v2 , v4 ) + mα−1 (v1 , v4 ) + mα(v2 , v3 ).
2 ,v4 )
v we get that [q]B has the required form up to a
Finally, by replacing v4 with v4 − (v
(v2 ,v3 ) 3
scalar. In order to remedy this, we replace v3 and v4 by a scalar multiple.
Case 5: Assume that the Jordan blocks of γ are of sizes [2, 2] and γ is unipotent. Let
B := (v1 , v2 , v3 , v4 ) be a base such that [γ]B has the required form. For every m:
(9)
(v2 , v2 ) = (γ m v2 , γ m v2 ) = (v2 + mv1 , v2 + mv1 )
so v1 is orthogonal to v1 and v2 and similarly v3 is orthogonal to v3 and v4 . Moreover, for
every m
(10)
(v2 , v3 ) = (γ m v2 , γ m v3 ) = (v2 + mv1 , v3 )
so (v1 , v3 ) = 0. Since q is non-degenerate, v1 is not orthogonal to v4 and v3 is not orthogonal
(v4 ,v4 )
(v2 ,v2 )
v3 and v4 with v4 − 2(v
v1 we can assume that
to v2 . By replacing v2 with v2 − 2(v
2 ,v3 )
1 ,v4 )
(v2 , v2 ) = (v4 , v4 ) = 0. By replacing v4 with v4 −
Now, (v2 , v3 ) = −(v1 , v4 ) since for every m
(11)
(v2 ,v4 )
v
(v2 ,v3 ) 3
we can assume that (v2 , v4 ) = 0
(v2 , v4 ) = (γ m v2 , γ m v4 ) = (v2 + mv1 , v4 + mv3 ).
Thus, [q]B has the required form up to a scalar. In order to remedy the situation, we
replace v3 and v4 by a scalar multiple.
Case 6: Assume that the Jordan blocks of γ are of sizes [3, 1]. Let B := (v1 , v2 , v3 , v4 )
be a base such that [γ]B has the required form. Then for every m:
(12)
(13)
(v2 , v2 ) = (γ m v2 , γ m v2 ) = (v2 + mv1 , v2 + mv1 )
(v3 , v2 ) = (γ m v3 , γ m v3 ) = (v3 + mv2 +
m(m − 1)
v1 , v2 + mv1 )
2
m(m − 1)
m(m − 1)
v1 , v3 + mv2 +
v1 )
2
2
Equation 12 implies that (v1 , v1 ) = 0 and (v1 , v2 ) = 0. Equation 13 implies that (v2 , v2 ) =
−(v1 , v3 ). In turn, Equation 14 implies that 2(v2 , v3 ) = (v1 , v3 ). Since q is non-degenerate,
(14)
(v3 , v3 ) = (γ m v3 , γ m v3 ) = (v3 + mv2 +
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY INVARIABLE GENERATION
9
(v3 ,v3 )
v1 we can assume that (v3 , v3 ) = 0. Moreover,
(v1 , v3 ) 6= 0. By replacing v3 with v3 − 2(v
1 ,v3 )
for every m
(15)
(v4 , v3 ) = (γ m v3 , γ m v4 ) = (v4 , v3 + mv2 +
m(m − 1)
v1 )
2
so (v1 , v4 ) = 0 and (v2 , v4 ) = 0. By replacing v4 with v4 −
(v4 , v3 ) = 0.
(v3 ,v4 )
v
(v1 ,v3 ) 1
we can assume that
3.1. Completing the proof of Theorem 1.1. Let us now explain how to prove Theorem
1.1. The structure of the proof follows the exact same lines as the proof of Theorem 2.2,
only that instead of using the action on the projective space, we use the action on the
planes Grassmannian X := Gr(2, R4 ). The elements of Gr(2, R4 ) can be viewed as lines in
projective space P3 . We define two functions d, d : X × X → R by:
d(l1 , l2 ) := max{distP3 (x1 , x2 ) | x1 ∈ l1 and x2 ∈ l2 } and
d(l1 , l2 ) := min{distP3 (x1 , x2 ) | x1 ∈ l1 and x2 ∈ l2 }.
Note that d is a metric on X. For every ǫ > 0 and l ∈ X we define two open sets of X:
(l)ǫ := {t ∈ X | d(l, t) < ǫ} and [l]ǫ := {t ∈ X | d(l, t) < ǫ}.
For linearly independent u, v ∈ R4 , lu,v ∈ X denotes their projective linear span. An
element l ∈ X is called isotropic if the restriction to the corresponding plane is the zero
form. The next lemma follows from Lemma 3.3 and Table 1:
Lemma 3.4. Let γ ∈ Γ be a non-trivial element. Let B = (v1 , v2 , v3 , v4 ) be a base of R4
for which [γ]B and [q]B are as in Table 1. Let l0 , l∞ and l−∞ be the isotropic elements of
X defined in Table 1. Then
lim γ n · l0 = l∞ and
n→∞
lim γ n · l0 = l−∞
n→−∞
where the convergence is with respect to the metric d. Moreover, if ǫ > 0 is small enough
then for every 0 6= m ∈ Z, γ m · (l0 )ǫ ∩ [l0 ]ǫ = ∅.
Lemma 3.5. Let l1 , l2 ∈ X be isotropic elements and let ǫ > 0. There exists δ ∈ Γ such
that δ · (X \ [l1 ]ǫ ) ⊆ (l2 )ǫ and δ −1 · (X \ [l2 ]ǫ ) ⊆ (l1 )ǫ .
Proof. Choose an isotropic element l3 ∈ (l2 )ǫ and ǫ∗ > 0 such that the span of l1 and l3 is
R4 , [l3 ]ǫ∗ ⊂ [l2 ]ǫ and (l3 )ǫ∗ ⊂ (l2 )ǫ . It is enough to prove the lemma for l1 , l3 and ǫ∗ instead
of l1 , l2 and ǫ. For every r > 0 there exists gr ∈ SO(2, 2)(R) which preserves l1 and l3 , acts
on l1 by multiplication by r and on l3 by multiplication by 1/r. Denote ǭ := ǫ∗ /2 and fix a
large enough r such that for every m > 0, grm · (X \ [l1 ]ǭ ) ⊆ (l3 )ǭ and gr−m · (X \ [l3 ]ǭ ) ⊆ (l1 )ǭ .
Let U be a symmetric neighborhood of the identity in SO(2, 2)(R) such that for every
u ∈ U and every i ∈ {1, 3}, u · (li )ǭ ⊆ (li )ǫ∗ and u · [li ]ǭ ⊆ [li ]ǫ∗ . As in Section 2.1, by the
Poincaré recurrence theorem there exist u1 , u2 ∈ U and m > 0 such that δ := u1 grm u2 ∈ Γ.
Evidently, δ is the desired element.
10
T. GELANDER AND C. MEIRI
Corollary 3.6. Let l ∈ X be an isotropic projective line. Then for every ǫ > 0 and every
conjugacy class C 6= {Id} of Γ there exists γ ∈ C such that γ · (X \ [l]ǫ ) ⊆ (l)ǫ .
Proof. Choose γ0 ∈ C. Lemma 3.4 states that there exists an isotropic l0 ∈ X and ǫ0 > 0
such that for every 0 6= m ∈ Z, γ0m · (l0 )ǫ0 ∩ [l0 ]ǫ0 = ∅. Denote ǫ1 := min(ǫ, ǫ0 ). Lemma 3.5
implies that there exists δ ∈ γ such that δ · (X \ [l0 ]ǫ1 ) ⊆ (l)ǫ1 and δ −1 · (X \ [l]ǫ1 ) ⊆ (l0 )ǫ1 .
The element γ := δγ0 δ −1 satisfies the requirement.
Pick a sequences (li )i≥1 of isotropic elements of X and a sequence of positive numbers
(ǫi )i≥0 such that (li )ǫi ∩ [lj ]ǫj = ∅ whenever i 6= j. By Corollary 3.6, there are (γi )i≥1 ∈ Γ
representing all the non-trivial conjugacy classes such that for every i, γi ·(X \[li ]ǫi ) ⊆ (li )ǫi .
It follows that Proposition 2.7 is satisfied with Ui = (li )ǫi , i ∈ N. Hence hγi : i ∈ Ni is a
free group and in particular of infinite index in Γ. This completes the proof of Theorem
1.1.
THE CONGRUENCE SUBGROUP PROPERTY DOES NOT IMPLY INVARIABLE GENERATION
[γ]B
(1)
(2)
(3)
(4)
(5)
(6)
α = reiθ
β = re−iθ
r>1
θ 6∈ πQ
c := r cos(θ)
d := r sin(θ)
[q]B
c
−d
0
0
d
c
0
0
0
0
c
r2
−d
r2
l0 , l±∞
0
0 0 1
0 0 0 0
d
1 0 0
r2
c
0 1 0
r2
0
1
0
0
l0 := lv1 +v4 ,v2 −v3
l∞ := lv1 ,v2
l−∞ := lv3 ,v4
α>β≥1
α
0
0
0
0 0 0
β 0 0
0 β1 0
0 0 α1
0
0
0
1
0
0
1
0
0
1
0
0
1
0
0
0
l0 := lv1 +v3 ,v2 −v4
l∞ := lv1 ,v2
l−∞ := lv3 ,v4
α=β>1
α
0
0
0
0 0 0
α 0 0
0 α1 0
0 0 α1
0
0
1
0
0
1
0
0
1
0
0
0
l0 := lv1 +v3 ,v2 −v4
l∞ := lv1 ,v2
l−∞ := lv3 ,v4
α=β>1
α
0
0
0
1 0 0
α 0 0
0 α1 1
0 0 α1
α=β=1
1
0
0
0
1
1
0
0
0
0
1
0
0
0
1
1
α=β=1
1
1
0
0
0
1
1
0
0
0
0
1
1
0
0
0
0
0
0
1
11
0 0
0 α
0 0 −1 0
α
0 −1 0 0
α
α 0
0 0
l0 := lv1 +αv3 ,αv2 +v4
l∞ := lv1 ,v2
l−∞ := lv3 ,v4
0 0
0 1
0 0 −1 0
0 −1 0 0
1 0
0 0
l0 := lv1 +v2 ,v3 +v4
l∞ := lv1 ,v3
l−∞ := lv1 ,v3
0 0 2 0
0 −2 1 0
±
2 1 0 0
0 0 0 1
l0 := lv2 +√2v4 ,2v3 +v2 − 1 v1
4
l∞ := lv1 ,v2 −√2v4
l−∞ := lv1 ,v2 −√2v4
Table 1
12
T. GELANDER AND C. MEIRI
References
[D92]
J. D. Dixon, Random sets which invariably generate the symmetric group. Discrete Math. 105
(1992), 25–39.
[HNN49] G. Higman, B. H. Neumann and H. Neumann, Embedding theorems for groups, J. London Math.
Soc. 24 (1949), 247–254.
[G15]
T. Gelander, Convergence groups are not invariably generated, to appear in IMRN.
[KLS11] W. M. Kantor, A. Lubotzky and A. Shalev, Invariable generation and the Chebotarev invariant
of a finite group, J. Algebra 348 (2011), 302–314.
[KLS14] W.M. Kantor, A. Lubotzky, A. Shalev, Invariable generation on infinite groups, preprint.
[KM79] M. Kneser, Normalteiler ganzzahliger Spingruppen, J. Reine und Angew. Math. 311/312(1979),
191–214.
[R72]
M.S. Raghunathan, Discrete Subgroups of Lie Groups, Springer, New York, 1972.
[R76]
M.S. Raghunathan, On the congruence subgroup problem. Inst. Hautes Etudes Sci. Publ. Math.
No. 46 (1976), 107–161.
[R86]
M.S. Raghunathan, On the congruence subgroup problem. II. Invent. Math. 85 (1986), no. 1,
73–117.
[S14]
P. Sarnak, Notes on thin matrix groups. Thin groups and superstrong approximation, Math. Sci.
Res. Inst. Publ., 61, Cambridge Univ. Press, Cambridge, (2014), 343–362.
[W76] J. Wiegold, Transitive groups with fixed-point-free permutations, Arch. Math. (Basel) 27 (1976),
473–475.
[We77] J. Wiegold, Transitive groups with fixed-point-free permutations. II, Arch. Math. (Basel) 29
(1977), 571–573.
[WM14] D. Witte–Morris, Introduction to Arithmetic Groups.
Mathematics and Computer Science, Weizmann Institute, Rechovot 76100, Israel,
E-mail address: [email protected]
Department of Mathematics, Technion - Israel Institute of Technology, Haifa, 32000,
Israel,
E-mail address: [email protected]
| 4math.GR
|
JMLR: Workshop and Conference Proceedings 80:1–10, 2017
ACML 2017
Neural Stain-Style Transfer Learning using GAN for
Histopathological Images
Hyungjoo Cho∗
[email protected]
Seoul National University
Sungbin Lim∗
[email protected]
arXiv:1710.08543v2 [cs.CV] 25 Oct 2017
Korea University
Gunho Choi
[email protected]
Yonsei University
Hyunseok Min
[email protected]
KAIST
Abstract
Performance of data-driven network for tumor classification varies with stain-style of
histopathological images. This article proposes the stain-style transfer (SST) model based
on conditional generative adversarial networks (GANs) which is to learn not only the certain color distribution but also the corresponding histopathological pattern. Our model
considers feature-preserving loss in addition to well-known GAN loss. Consequently our
model does not only transfers initial stain-styles to the desired one but also prevent the
degradation of tumor classifier on transferred images. The model is examined using the
CAMELYON16 dataset.
Keywords: Deep Learning, Stain Normalization, Domain Adaptation, Neural Style Transfer, Generative Adversarial Network
1. Introduction
Deep learning based image recognition receives a lot of attention due to its notable application to digital histopathology including automatic tumor classification. Convolutional
neural networks(CNNs) have recently achieved state-of-the-art performance in the task
of image classification and detection, especially, replaced the traditional rule-based methods in the several contests of medical image diagnosis LeCun et al. (2015); Wang et al.
(2016). Such data-driven approach especially depends on quality of training dataset hence
it requires sensible preprocesses. In histopathology, staining e.g. haematoxylin and eosin
(H&E) is essential to examine the microscopic presence and characteristics of disease not
only for pathologists but also for neural networks. For digital histopathology, several stain
normalization preprocesses are well-known Ruifrok et al. (2001); Reinhard et al. (2001);
Ruifrok et al. (2003); Annadurai (2007); Magee et al. (2009); Macenko et al. (2009); Khan
et al. (2014); Li and Plataniotis (2015); Bejnordi et al. (2016).
Standard stain normalization algorithms are based on stain-specific color deconvolution
Ruifrok et al. (2001). Stain deconvolution requires prior knowledge of reference stain vec∗
Equal contribution
c 2017 H. Cho, S. Lim∗ , G. Choi & H. Min.
Cho Lim∗ Choi Min
Figure 1: Samples of tissue tiles from different institutes in CAMELYON16cam (2016),
17cam (2017) dataset. The first row shows normal samples, and the second row
shows tumor samples. Samples of institute 1, 2 are included cam (2016), the
others are included cam (2017) dataset.
tors for every dye present in the whole-slide images (WSI). Ruifrok et al. (2001) suggested
a manual approach to estimate the color deconvolution vectors by selecting representative
sample pixels from each stain class, and a similar approach was used in Magee et al. (2009)
for extracting stain vectors. Such manual estimation of stain vectors, however, strongly
limits their applicability in large studies. Khan et al. (2014) modified Magee et al. (2009)
by estimating stable stain matrices using an image-specific color descriptor. Combined
with a robust color classification framework based on a variety of training data from a
particular stain with nonlinear channel mappings, the method ensured smooth color transformation without introducing visual artifacts. Another approach used in Bejnordi et al.
(2016) transforms the chromatic and density distributions for each individual stain class in
the hue-saturation-density (HSD) color model. See Bejnordi et al. (2016) and references
therein.
Stain normalization methods for histopathological images have been studied extensively,
and yet these still possess challenging issues. Most of the conventional methods use various
thresholds to filter out backgrounds and other irrelevant dimensions. However, these methods cannot represent the broad feature distribution of the entire target image, thus they
require manual tuning of hyper-parameters such as thresholds. Furthermore, since nuclei
detection has a significant impact on performance of color normalization, it is unlikely to
expect good performance if there is a mistake in the nuclei detection stage. Finally, although
the major aim of most conventional approaches is to enhance the prediction performance
of classification system, these stain normalization methods and classifer work separately.
It is reported that performance of network varies with institutes even they applied same
staining methods Ciompi et al. (2017). In order to prevent such variation, it is required to
consider a domain adaption method.
In this paper, we propose a novel stain-style transfer method using deep learning, as
well as a special loss function which minimizes the difference between latent features of
input image and that of target image, thus preserves the performance of the classifier. We
2
Neural Stain-Style Transfer Network
implement fully convolutional network (FCNs) Long et al. (2015) in proposed stain-style
generator that learns the color distribution of dataset which is used to train the tumor
classifier.
Our contributions in this paper are of two areas. First, we replace the color normalization
methods with a generative model which learns certain stain-style distribution of dataset.
Second, we introduce feature-preserving loss to induce the classifier to extract better features
than different methods.
2. Stain-Style Transfer with GAN
2.1. Stain-Style of Dataset
In this section, we summarize relevant material on our model. Let I be a set of institutes
and let x ∈ X be the dataset of histological sample and the corresponding label y ∈ Y.
The class of stained images or color images with RGB channels, denoted by C = Md (R3 ),
is defined to be the set of d × d-matrix with R3 entries. Under this setting, we define the
stain-style of institute i ∈ I to be a random variable φ(i) : X → C with a probability
distribution
P(i) (c) := P(φ(i) = c),
c∈C
Since φ(i) admits a certain conditional probability distribution P(i) (·|y), the definition of
different stain-style with the same label makes sense.
2.2. Tumor Classifier Network
Suppose we trained tumor classifier network f : X → Y which infers histological pattern of
input image x ∈ X . We write f = f (i) if the classifier is especially trained on dataset which
follows stain-style φ(i) . We estimate the performance of f (i) by
h
i
(i)
Ltumor := E(x,y)∼P (x,y) `(f (i) (x), y)
(1)
where ` is a loss function for classification e.g. cross-entropy. Practically, we make classifier
to learn stained images φ(i) (x) ∈ C rather than x ∈ X . Hence one can decompose the
classifier by f (i) = fˆ(i) ◦ φ(i) where fˆ(i) : C → Y is an actual network which is trained on
dataset with stain-style φ(i) . In this case, we estimate (1) by
h
i
(i)
Ltumor ≈ Ec∼P(i) (c|y) `(fˆ(i) (c), y)
(2)
2.3. Stain-Style Transfer Network
Since the stain-styles of each institute are dissimilar, the histological pattern in image from
different institute would break up in the view of classifier network. Consequently, it would
show degraded performance Ciompi et al. (2017):
h
i
h
i
Ec∼P(j) (c|y) `(fˆ(i) (c), y) ≥ Ec∼P(i) (c|y) `(fˆ(i) (c), y) , i 6= j
3
Cho Lim∗ Choi Min
⌧ =⇣
G
G
C
(1)
(1)
C
⇣
(2)
G
(3)
Figure 2: Overview of the stain-style transfer network. The network τ is composed of two
transformations: Gray-normalization G and style-generator ζ. G standardizes
each stain-style of color images from different institutes and ζ colorizes gray images following the stain-style of certain institute.
To overcome this problem, we propose stain-style transfer (SST) network which transfers
stain-style φ(j) to the initial φ(i) . Precisely, our aim is to find a network τ : C → C which
satisfies
P(τ ◦ φ(j) = c|y) = P(j) (τ −1 (c)|y) ≈ P(i) (·|y),
∀c ∈ C
Due to the change of variable formula (Durrett, 2010, Theorem 1.6.9), (3) implies
h
i
h
i
Ec∼P(j) (τ −1 (·)|y) `(fˆ(i) (c), y) ≈ Eτ (c)∼P(i) (·|y) `(fˆ(i) (τ (c)), y)
h
i
= Ec∼P(i) (·|y) `(fˆ(i) (c), y)
(3)
(4)
hence the tumor classifier recovers its performance (2).
We emphasize that our SST network does not require the dataset of institute j to train
both fˆ(i) and τ . To make τ independent of institute j ∈ I, we employ the gray normalization
G : C → G and train stain-style generator ζ : G → C such that τ = ζ ◦ G, as illustrated in
Figure 2.
2.4. Stain-Style Generator by Conditional GAN
To train the style-generator ζ, we introduce three loss functions (a) reconstruction loss, (b)
GAN loss, and (c) feature-preserving loss.
2.4.1. Reconstruction Loss
Restricted to the initial φ(i) , SST network τ should be an reconstruction map i.e. τ ◦ φ(i) =
φ(i) . Hence we apply a reconstruction loss to minimize the L2 -distance between τ ◦ φ(i) and
4
Neural Stain-Style Transfer Network
Figure 3: Illustration of feature-preserving loss. We use the global average pooled layers of
input and generated image as the input of feature preserving loss.
its original image φ(i) using the architecture from Quan et al. (2016) which has very deep
structure with short-cut and skip connections. The reconstruction loss is denoted by
LRecon (τ, φ(i) ) := Ekτ (φ(i) ) − φ(i) k2
2.4.2. Conditional GAN Loss
As Pathak et al. showed in Pathak et al. (2016), mixing GAN loss Goodfellow et al. (2014)
with some traditional loss, such as Lrecon , improves the performance of generator. Since we
have labeled images, conditional GAN Mirza and Osindero (2014) was applied instead of
Goodfellow et al. (2014). By means of GAN, ζ is to learn a mapping from G to C and to
trick the discriminator D. Here D is to distinguish between fake and real images using the
architecture from DCGAN Radford et al. (2015). We use the following GAN loss
LGAN (τ, φ(i) ) := E[log D(ζ, φ(i) )] + E[log(1 − D(ζ, φ(i) ))]
While D learns to maximize LGAN , ζ tries to minimize it until both arrives at its optimal
state. Through the above procedure, every stained image might be transferred to have
the desired stain-style. However, this approach often tend to make frequent color images
independent of histological pattern. This phenomena is called mode collapse (of GANs)
which possibly interrupt achieving (3). Therefore we need an additional loss function.
5
Cho Lim∗ Choi Min
2.4.3. Feature-preserving Loss
As in (3), in the optimal state τ ∗ , an output of SST network τ ∗ ◦ φ(j) should approximate
target φ(i) . By the means of Kullback-Leibler divergence, (3) can be restated by
h
i
τ ∗ = arg min KL P(i) (c|y) P(j) (τ −1 (c)|y)
τ
To obtain τ ∗ , having (4) in mind, we employ the feature-presearving loss
i
h
LFP (τ, φ(i) ) := KL F fˆ(i) (c) F fˆ(i) (τ (c))
where F(fˆ(i) (·)) indicates the feature of given color image extracted from the classifier fˆ(i) .
As illustrated in Figure 3, the final layer before the activation function is used to examine
feature vector, precisely, global average pooled layer.
Consequently, the overall loss function is
L(τ, φ(i) ) := λRecon LRecon (τ, φ(i) ) + LGAN (τ, φ(i) ) + λFP LFP (τ, φ(i) )
where λRecon , λFP are the weights which are used to balance the update between different
loss functions.
3. Experiment
We perform quantitative experiment in tumor classification to evaluate the SST network. To
show the general performance of our method, we apply the extensions to vanilla models as
well as conventional method. We have 4 baseline methods: Reinhard et al. (2001), Macenko
et al. (2009), Histogram specification (HS) Annadurai (2007) and WSI color standardization
(WSICS) Bejnordi et al. (2016).
3.1. Dataset
The Camelyon16 dataset is composed of 400 slides from two different institutes, Radbound
and Utrecht. We use 180,000 patches for training, 20,000 for validation from Radbound and
140,000 patches for testing from Utrecht. The number of tumor and normal are the same.
Hypothesizing the training and validation dataset belong to a certain institue and the test
set is from another one, we can merge every stain-style into the same space by applying the
gray normalization. Both training and validation dataset are labeled, supervised learning
can be applied to train the mapping from gray image to the colored one. We used gray
normalization based on Pillow package of python which uses this formula L = 0.299 × R +
0.587 × G + 0.144 × B.
3.2. Network Architecture
In this part, we explain each network structure of classifier network and stain-style generator
which constitute SST network.
6
Neural Stain-Style Transfer Network
Figure 4: Comparison between SST and other stain normalization method: (a) Target image for transfer (b) Original input image to be transferred (c) SST (d) WSICS
(e) HS (f) Marcenko (g) Reinhard
3.2.1. Classifier Network
Classifier network carries out two tasks in experiment. Firstly, it is a discriminator which
evaluates the performance of stain-style generator ζ. Secondly, as already explained in
subsection 2.4.3, it works as a feature-extractor which is used in feature preserving loss
LF P . We use ResNet-34 from torchvision library in PyTorch as a framework.
3.2.2. Stain-Style Generator
The generator network ζ is provided an image as an input instead of a noise vector. Therefore we can use FCN type architectures and U-Net is one of the most famous network among
them. However, because of its limit of performance, we use FusionNet which has combined
the advantages of U-Net and that of ResNet. Hyperparameters of network are set as same
as Quan et al. (2016). We adapt our discriminator architectures from Radford et al. (2015)
which is based on VGG-Net without pooling layer. The hyperparameters of discriminator
are the same as those in Radford et al. (2015).
3.3. Result
Figure 4 illustrates the result of each stain normalization method on a sample image. Target image comes from Radbound which is used for training the tumor classifier. Original
image is sampled from Utrecht, used for testing the tumor classifier. Although there is no
visual difference between outputs of each method, the classification performance on these
color images varies significantly. Given the experiment results in Table 1, SST network
successfully avoids the performance degradation. SST achieves the highest performance on
original images on tumor classification with Area Under Curve(AUC) = 0.9185. This result
shows that there are difference between visual judgment and the result of classifier. In case
of WSICS’s result, which is most visually similar to SST’s, the AUC score is worse than
that of SST by about 30%. On the other hand, Macenko, which was visually the worst,
performs better than other methods except for SST. Conventional methods consider only
the physical features of input images and lose patterns which are key features for classifier’s
decision making process. In contrast, SST maintains those key features, input image’s own
patterns, and also consider the color distribution of target images as well as the contextual
information of original images.
7
Cho Lim∗ Choi Min
Table 1: Performance of tumor classifier network on different stain normalization methods.
SST network shows significant improvement compared to direct application to
original (untransferred image) and outperforms the others.
Model
AUC
Precision
Recall
Specificity
Target
0.9760
0.9114
0.9126
0.9583
Original
0.8900
0.8098
0.8111
0.8014
WSICS
0.6408
0.5989
0.5957
0.6010
SST
0.9185
0.8440
0.8460
0.8371
HS
0.4245
0.4987
0.4986
0.4162
Macenko
0.7169
0.6983
0.6956
0.6500
Reinhard
0.5611
0.6114
0.6119
0.5471
4. Conclusion
In this work, we have presented a stain style transfer approach to stain normalization for
histopathological images. To that end, we replace the stain normalization models with a
generative model which learns certain stain-style distribution of training dataset. This stain
style transfer network is considerably simpler than contemporaneous work, and produces
more realistic results without any additional labeling or annotation for training as well
as prior knowledge. Further, unlike conventional stain normalization, which acts independently of the tumor classifier, the proposed feature-preserving loss induces our coloration
in a direction that does not affect the tumor classifier. We demonstate that our model
is optimized for the performance of the tumor classifier and allows successful stain-style
transfer.
The style of chemical cell staining is mainly affected by structural information and
morphology of cells rather than factors such as cell brightness. Based on these observation
points, we converted the test image into a gray image and performed a stain style transfer
process. While this method has the advantage of making the process simpler, it has also
lost some information. To resolve the limitation, further investigation will assess direct
stain style transfer approach from color image to color image. In addition, we hope to more
closely examine parameters of our deep learning approach. Further, we will perform more
rounds of hard negative mining and consider the reliability and reproducibility of the deep
CNN models.
References
http://camelyon16.grand-challenge.org/, 2016.
http://camelyon17.grand-challenge.org/, 2017.
S Annadurai. Fundamentals of digital image processing. Pearson Education India, 2007.
Babak Ehteshami Bejnordi, Geert Litjens, Nadya Timofeeva, Irene Otte-Höller, André
Homeyer, Nico Karssemeijer, and Jeroen AWM van der Laak. Stain specific standardization of whole-slide histopathological images. IEEE transactions on medical imaging, 35
(2):404–415, 2016.
8
Neural Stain-Style Transfer Network
Francesco Ciompi, Oscar Geessink, Babak Ehteshami Bejnordi, Gabriel Silva de Souza,
Alexi Baidoshvili, Geert Litjens, Bram van Ginneken, Iris Nagtegaal, and Jeroen van der
Laak. The importance of stain normalization in colorectal tissue classification with convolutional networks. arXiv preprint arXiv:1702.05931, 2017.
Rick Durrett. Probability: theory and examples. Cambridge university press, 2010.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in
neural information processing systems, pages 2672–2680, 2014.
Adnan Mujahid Khan, Nasir Rajpoot, Darren Treanor, and Derek Magee. A nonlinear
mapping approach to stain normalization in digital histopathology images using imagespecific color deconvolution. IEEE Transactions on Biomedical Engineering, 61(6):1729–
1738, 2014.
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):
436–444, 2015.
Xingyu Li and Konstantinos N Plataniotis. A complete color normalization approach to
histopathology images using color cues computed from saturation-weighted statistics.
IEEE Transactions on Biomedical Engineering, 62(7):1862–1873, 2015.
Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for
semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 3431–3440, 2015.
Marc Macenko, Marc Niethammer, JS Marron, David Borland, John T Woosley, Xiaojun
Guan, Charles Schmitt, and Nancy E Thomas. A method for normalizing histology slides
for quantitative analysis. In Biomedical Imaging: From Nano to Macro, 2009. ISBI’09.
IEEE International Symposium on, pages 1107–1110. IEEE, 2009.
Derek Magee, Darren Treanor, Doreen Crellin, Mike Shires, Katherine Smith, Kevin Mohee, and Philip Quirke. Colour normalisation in digital histopathology images. In Proc
Optical Tissue Image analysis in Microscopy, Histopathology and Endoscopy (MICCAI
Workshop), volume 100. Citeseer, 2009.
Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint
arXiv:1411.1784, 2014.
Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros.
Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 2536–2544, 2016.
Tran Minh Quan, David GC Hilderbrand, and Won-Ki Jeong. Fusionnet: A deep fully
residual convolutional neural network for image segmentation in connectomics. arXiv
preprint arXiv:1612.05360, 2016.
9
Cho Lim∗ Choi Min
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with
deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434,
2015.
Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between
images. IEEE Computer graphics and applications, 21(5):34–41, 2001.
Arnout C Ruifrok, Dennis A Johnston, et al. Quantification of histochemical staining by
color deconvolution. Analytical and quantitative cytology and histology, 23(4):291–299,
2001.
Arnout C Ruifrok, Ruth L Katz, and Dennis A Johnston. Comparison of quantification of histochemical staining by hue-saturation-intensity (hsi) transformation and colordeconvolution. Applied Immunohistochemistry & Molecular Morphology, 11(1):85–91,
2003.
Dayong Wang, Aditya Khosla, Rishab Gargeya, Humayun Irshad, and Andrew H Beck.
Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718,
2016.
10
| 1cs.CV
|
Distributed Selection in O (log n) Time with
O (n log log n) Messages
arXiv:1511.00715v4 [cs.DS] 11 Dec 2017
Piotr Berman and Junichiro Fukuyama
Department of Computer Science and Engineering
The Pennsylvaina State University
{berman, jxf140}@cse.psu.edu
Abstract. We consider the selection problem on a completely connected
network of n processors with no shared memory. Each processor initially holds a given numeric item of b bits allowed to send a message of
max (b, lg n) bits to another processor at a time. On such a communication network G, we show that the kth smallest of the n inputs can be
detected in O (log n) time with O (n log log n) messages. The possibility
of such a parallel algorithm for this distributed k-selection problem has
been unknown despite the intensive investigation on many variations of
the selection problem carried out since 1970s. The main trick of our algorithm is to simulate the comparisons and swaps performed by the AKS
sorting network, the n-input sorting network of logarithmic depth discovered by Ajtai, Komlós and Szemerédi in 1983. We also show the universal
time lower bound lg n for many basic data aggregation problems on G,
confirming the asymptotic time optimality of our parallel algorithm.
Keywords: the selection problem, distributed selection, AKS sorting
network, communication network, comparator network
1 Introduction
The classical k-selection problem finds the kth smallest of given n numeric items.
We consider it on a completely connected network G of n processors with no
shared memory, each holding exactly one item of b bits initially. A processor
node in G may send a message of max (b, lg n)-bits at any parallel step. As the
performance metrics, we minimize the parallel running time, and/or the total
number of messages as the measure of amount of information transmitted on G.
The parallel selection problem for connected processors with no shared memory has been extensively investigated for various network topologies, cases of how
n inputs are distributed, and other constraints [1,2,3,4], as well as on the parallel comparison tree (PCT) and parallel random access machine (PRAM) models
with shared memories [5,6,7]. In this paper we focus on the above case, calling it
the k-distributed selection problem on a communication network G. As suggested
in [1], this case of G has become increasingly significant for the contemporary
distributed computing applications such as sensor networks and distributed hash
tables: It is common in their performance analysis to count messages arriving at
2
Piotr Berman and Junichiro Fukuyama
the designated destinations assuming constant time per delivery (called hops),
rather than count how many times messages are forwarded by processor nodes.
The considered network G models it well.
The distributed selection problem in this particular network case is one of the
long time research topics in parallel algorithms. The results are included in the
work such as [1,3,8,9]. The following summarizes only a few most closely related
to our interest: Let w stand for the number of processors initially holding one or
more inputs. The algorithm Frederickson
and Johnson developed in [8] finds the
k
messages on a completely connected or starkth smallest item with O w log w
shaped processor network. Santoro et al [3] discovered in 1992 an algorithm with
the expected number of messages bounded by O (w log log min(k, n − k) + w log w).
The more recent result of [1] explores a case when the processor network has any
diameter D > 1. It presents a parallel algorithm with the average time bound
O (D logD n) that is asymptotically optimal under some probabilistic
assump
tions, and one with the deterministic time bound O D log2D n .
The first contribution of this paper is a parallel algorithm for the distributed
k-selection problem that runs in time O (log n) with total O (n log log n) messages
on G. We will prove the following theorem after formulating the problem.
Theorem 1. The kth smallest of n inputs can be computed distributedly on a
communication network G in time O (log n) with O (n log log n) messages.
t
u
Such a parallel algorithm has been unknown in the long research history on
the selection problem. Satisfying the constraint oftotal O (n log log n) messages,
it improves the parallel time bound O D log2D n in [1] when log2 D log n:
On a processor network with diameter D, a message can be sent to anywhere
forwarded by O (D) processors. So the theorem means the parallel time bound
O (D log n) D log2D n.
Our algorithm simulates the comparisons and swaps performed by the AKS
sorting network, the n-input sorting network of O (log n) depth discovered by
Ajtai, Komlós and Szemerédi in 1983 [10]. Known for the difficulty of its performance analysis, the AKS sorting network itself has been a research subject in
parallel algorithm. Paterson [11] simplified its construction. Seiferas [12] further
clarified it with the estimate that the depth can be at most 7·6.07 lg n = 48.79 lg n
layers of 1/402.15-halvers. Here an ε-halver ( ∈ (0, 1)) is a comparator network of a constant depth to re-order n inputs [11,13], such that for every
z ∈ [0, n/2] ∩ Z, the left half of the output includes at most εz items among
the z largest inputs, and the right half at most εz among the z smallest inputs.
In the recent work by Goodrich [14], the constant factor of the total O (n log n)
nodes is significantly reduced although its depth bound is O (n log n). The overall research interest on the AKS sorting network has been simpler understanding
and reduction of the constant factor of the asymptotic quantities.
In order to design our parallel algorithm for the distributed selection problem,
we devise an n-input comparator
network of depth O (log log n) that reduces
strangers to at most O n log−c n for any given constant c > 0. Here a stranger
is the zth largest input (1 ≤ z ≤ n) that exists in the left half of the output
despite z > 21 n, or in the right half despite z ≤ 21 n. Formally we show:
Efficient Distributed Selection Algorithm
3
Theorem 2. For every c ∈ R+ , and n ∈ Z+ that is a sufficiently large power
of 6, there exists a comparator network H to re-order n inputs satisfying the
following two.
i) The output includes less than n lg−c n strangers. and
ii) The depth is at most 1296cd 15 lg lg n where dε stands for the minimum depth
of an ε-halver.
t
u
Such H will be constructed from given 1/5-halvers through comparator networks we devise in Section 3 as gadgets. A 1/5-halver of a depth at most 2 · 721
is built in Appendix with the classical result of Gabber and Galil on expander
graphs [15]. Consequently the depth of H is at most 2592 · 721 c lg lg n. This value
would be even far larger if ε-halvers with some smaller ε were used. Our parallel
algorithm for the distributed selection problem mimics this H also.
Our third result is the running time lower bound lg n for many data aggregation problems on G. The class of problems having the lower bound is huge
including the selection problem and many others. The following statement will
be confirmed as a corollary to the theorem we will show in Section 6.
Corollary 1. Any parallel algorithm takes at least lg n steps in the worst case
to compute each of the following three problems on a communication network
distributedly: i) the k-selection problem on n inputs each of at least dlg ne + 1
bits; ii) the problem of finding the sum of n inputs; iii) the problem of counting
numeric items among n inputs, each exceeding a given threshold.
t
u
By our formulation in Section 2, the statement assumes that each of the n
processor nodes holds one numeric item at time 0. The corollary shows the
asymptotic time optimality of our parallel algorithm as well.
The rest of the paper consists as follows. In Section 2, we define general
terminology showing related facts. After devising our gadgets from ε-halvers in
Section 3, we will prove Theorem 2 in Section 4, Theorem 1 in Section 5, and
the universal lower bound lg n in Section 6, followed by concluding remarks in
Section 7.
2 General Terminology and Related Facts
In this paper, a communication network G means a collection of n processor nodes
each two of which are connected, capable of exchanging a message of max (b, lg n)
bits at a parallel step where b ∈ Z+ is given. We consider a computational
problem P whose input is a set of n numeric items of b bits distributed over G,
i.e., every processor holds exactly one input item at time 0. A parallel algorithm
A is said to compute P on G distributedly in time t with m messages, if all
the bits of the computed result from the n inputs are stored at a designated
processor node after the tth parallel step with total m messages. Messages may
be exchanged asynchronously on G.
4
Piotr Berman and Junichiro Fukuyama
A comparator network H is a directed acyclic graph of 2-input comparators.
For notational convenience, we say that the width of H is the maximum number
of nodes of the same depth. In the standard terminology, a numeric item input
to a comparator is called wire. We may call an ordered set of wires array, for
which the set theoretical notation is used.
The AKS sorting network is a comparator network that sorts n inputs with
a depth Θ (log n) and the width at most n, performing O (n log n) comparisons.
We observe here that the kth smallest of n items distributed over G can be
detected in O (log n) time with O (n log n) messages the following way. We design
such a parallel algorithm A so that a processor node v in G simulates from time
ct to c(t + 1) (c : a constant, t ∈ N) a comparator of depth t in the AKS sorting
network H. We mean by “simulate” that v receives two items from other nodes
to send their minimum and/or maximum to anywhere on G. Due to the width
of the AKS sorting network, and since G is completely connected, n processors
in G can simulate all the swaps and copying performed by H. This way A sorts
the n inputs in O (log n) steps, then fetches the kth smallest.
Our parallel algorithm improves the above so that the number of messages
is reduced to O (n log log n).
3 Devising Gadgets from Expander Graphs for Selection
In this section, we devise some comparator networks S for the design of our
parallel algorithm. Such an S re-orders given distinct m numeric items where
m ∈ 2Z. For a constant ε ∈ (0, 1), consider a bipartite graph (V1 , V2 , E) such
that |V1 | = |V2 | = 21 m and
ε|Γ (U ) | > (1 − ε) min (ε|Vi |, |U |) ,
(1)
for every subset U of V1 or V2 , where Γ (U ) denotes the neighbor set of U . The
sets V1 and V2 represent the left and right halves of the input, respectively,
and an edge in E a comparison between the two items. It is straightforward to
check that a comparator network S derived from (1) in the obvious manner is
an ε-halver.
An expander is a graph G whose expansion ratio |Γ (U ) |/|U | is above a
constant greater than 1 for any node set U meeting some condition such as
|U | ≤ 12 |V (G)|. The bipartite graph G = (V1 , V2 , E) is an expander in this sense
whose explicit construction is given in the aforementioned Gabber-Galil result.
We demonstrate in Appendix that it leads to a 1/5-halver of depth 2 · 721 , and
also an ε-halver of a constant depth for any given ε can be explicitly constructed.
The AKS sorting network sorts numeric items with ε-halvers.
It is well-known in spectral graph theory that the minimum expansion ratio
is good if |U |/|V (G)| is bounded and the second largest eigenvalue λ (G) of G is
small. In [16], a d-regular graph called ramanujan√graph is explicitly constructed
with the asymptotically optimal bound λ (G) ≤ 2 d − 1, improving the GabberGalil result. However not much is known beyond this point regarding the explicit
construction of an ε-halver.
Efficient Distributed Selection Algorithm
5
Fig. 1. Three Comparator Networks T (C), U (C) and V (C)
Our gadgets S for the selection problem is constructed from a given ε-halver
below. For any comparator network C with 2s inputs and outputs, construct the
three comparator networks T (C), U (C) and V (C) as in Fig. 1. They are on 4s,
3s and 4s inputs with 2, 2 and 6 layers of C, respectively. Let S0 be an ε-halver.
Recursively construct
Si = V (Si−1 )
and Si0 = U (Si ),
for i ∈ Z+ ,
and define
εi = ε2i−1 + 10ε3i−1 ,
where ε0 = ε.
Consider the following property P of a comparator network.
P (q, l, s, r) q, l ∈ Z+ and s, r ∈ (0, 1) :
i) The comparator network re-orders m inputs (m ∈ qZ+ ) by l layers of εhalvers.
ii) For any given z ∈ Z+ at most m
q , there are no more than rz items each of
rank at most z in the left s of the output.
Si satisfies the property
P 2i+1 , 6i , 12 , εi , and Si0 the property
P 3 · 2i , 2 · 6i , 23 , 2εi . In Fig. 2 illustrating Si , consider each z ≤ m
q where q =
2i+1 . Let us call an item z-stranger if it has a rank at most z, existing in the
left 1 − 1q of the considered layer. The figure shows the upper bounds on the
ratios of z-strangers: Each number in black indicates the upper bound on the
number of z-strangers in the current component divided by z. Each number in
red indicates that in the current and its left components divided by z.
6
Piotr Berman and Junichiro Fukuyama
Fig. 2. Upper Bounds on the z-Stranger Ratios in Si
In particular, there are 2εi−1 z or less z-strangers
in the component L ∪
R. As it satisfies the property P 2i , 6i−1 , 12 , εi−1 , the output of L includes
2ε2i−1 z or less z-strangers.
One can check the other bounds to verify the property
P 2i+1 , 6i , 12 , εi of Si . Similarly, Si0 satisfies P 3 · 2i , 2 · 6i , 23 , 2εi .
Fig. 3. To See the Property P̂ of Ŝi
Now let
ε=
1
.
5
Construct Ŝi by recursively replacing the second layer of T (Si−1 ) by Ŝi−1 , where
Ŝ0 = S0 . Define
ε̂i = ε̂i−1
1
+ εi−1
2
for i = Z+ ,
and
ε̂0 = ε.
Efficient Distributed Selection Algorithm
7
Here we slightly generalize the definition of a stranger given in Chapter 1: The
item of rank z is a stranger if z ≤ m
2 and it is in the left half of the considered
layer, or if z > m
and
it
is
in
the
right
half.
2
i
Each Ŝi (1 ≤ i ≤ 4) satisfies P̂ 2i+1 , 6 5+4 , ε̂i−1 , ε̂i where:
Property P̂ (q, l, r, r0 ) q, l ∈ Z+ and r, r0 ∈ (0, 1) :
i) The comparator network re-orders m inputs (m ∈ qZ+ ) by l layers of εhalvers.
ii) If there are at most rm strangers in the input, there are at most r0 m
strangers in the output.
In Fig. 3 illustrating Ŝi , let l and r be the numbers of strangers output from L and
R, respectively, so l ≤ r ≤ 12 ε̂i−1 m. They are both bounded by h = 12 mε̂i−1 εi−1
since each half of the first layer1 is Si−1 . The rank of the median of m inputs is
1
4 m − r + l in the input of C that is a Ŝi−1 . So the total number of strangers in
the left half of the output is bounded by
1
1
1
mε̂i−1 + (r − l) + l ≤ mε̂i−1 + h = mε̂i .
4
4
2
i
This proves the condition ii) of P̂ 2i+1 , 6 5+4 , ε̂i−1 , ε̂i . The first condition is
straightforward to check.
In addition, construct Ŝ by vertically aligning Ŝ0 , Ŝ1 , Ŝ2 , Ŝ3 and Ŝ4 , so that
all the m outputs of Ŝi−1 are input to Ŝi . It satisfies P̂ (32, 315, 1, ε̂), where
ε̂ = ε̂4 < 0.02314.
We also denote
and ε0 = 2ε3 < 0.002644,
so S and S 0 satisfies P 16, 216, 12 , ε0 and P 24, 432, 23 , ε0 , respectively. We will
use the obtained S, S 0 and Ŝ in the next section.
S = S3 ,
S 0 = S30 ,
4 Reducing Strangers by a Comparator Network of Depth
O (log log n)
We prove Theorem 2 by explicitly building the desired comparator network H
from given ε-halvers where ε = 1/5. Assume without loss of generality that the
n inputs to H have distinct values2 .
Here are the construction rules.
1
2
Here 12 mε̂i−1 ≤ 12 m2−i for ε = 51 and i ≤ 4. So we can use the property
P 2i , 6i−1 , 12 , εi−1 of Si−1 .
For s elements of a same value g, regard that they are valued g + δ, g + 2δ, . . . , g + sδ
for an infinitesimal number δ > 0. All the statements in our proof remain true with
this change.
8
Piotr Berman and Junichiro Fukuyama
Fig. 4. Maximum Stranger Ratios in the Segment j = 7
• There are segments j = 0, 1, 2, . . . , jmax − 1 of H where jmax = bc lg lg nc. The
segment j > 0 re-orders the n outputs of j − 1 with a constant number of
layers of ε-halvers.
• The segment j = 0 consists of Ŝ on n inputs.
• Partition the left half of the input to the jth segment into subarrays
L0 , L1 , . . . , Lj−1 , Lj from the left to right in that order, satisfying
−i−1
2
n, if 0 ≤ i ≤ j − 1,
|Li | =
2−i n,
if i = j.
So |Lj | = |Lj−1 |.
• Symmetrically, define R0 , R1 , . . . , Rj−1 , Rj in the right half from the right to
left.
• The jth segment performs the following four operations.
- split: Split Lj−1 into its halves re-naming them Lj−1 and Lj . Split Rj−1
into Rj−1 and Rj also.
- compare-even-pairs: Apply S 0 or S to every pair L2i and L2i+1 (i = 0, 1, . . .)
currently existing in the left half. Apply S 0 if they have the different sizes
and S otherwise. Also perform it in the right half symmetrically.
- compare-odd-pairs: Same except for pairing L2i+1 with L2i+2 .
- swap-center: Apply Ŝ to Lj ∪ Rj .
• If j is odd, perform compare-even-pairs and swap-center simultaneously, then
compare-odd-pairs. Further perform split and compare-even-pairs, each once
in that order. If j is even, perform the same operations switching the parity.
The segment j = 7 is illustrated in Fig. 4.
Efficient Distributed Selection Algorithm
9
Each segment consists of at most 64 = 1296 layers of ε-halvers by the property
P of S and S 0 , and P̂ of Ŝ. The total depth of H does not exceed 1296cd 15 lg lg n
as claimed.
Let ε0 and ε̂ be as in the previous section, and let
∆=
1
.
10
Denote by li,j and ri,j the number of strangers in Li and Ri of the jth segment,
respectively. Our invariant is
i
X
li0 ,j ≤ ∆j−i+1 |Li |,
and
i0 =0
i
X
ri0 ,j ≤ ∆j−i+1 |Ri |,
(2)
i0 =0
for every j and i such that 0 ≤ j < jmax and 0 ≤ i ≤ j. We prove it as a lemma.
Lemma 1. (2) holds for each j and i.
Proof. Show (2) by induction on j. The 0th segment consists of Ŝ. The invariant
is true by its property P̂ (16, 315, 1, ε̂). Assume true for j − 1 and prove true for
j. We only show the induction step for the left half of the jth segment when j is
odd. The other cases are handled similarly. Fig. 4 illustrates the induction step
for j = 7.
Let li0 be the number of strangers in Li of the jth segment at the considered
moment. It meets the following bound right after compare-even-pairs and swapcenter.
0 j−i−1
, if i ≤ j − 2 and i is even,
ε ∆
0
li
∆j−i , if i ≤ j − 2 and i is odd,
≤
|Li |
ε̂ + ∆2 ,
if i = j − 1.
The third row of Fig. 4 shows the same numbers in the RHS. They mean slightly
stronger than the above: The right number in lth box (l = 1, 2, 3) indicates the
upper bound on the number of strangers in L2l−1 ∪ L2l divided by |L2l |.
The correctness of the bounds is due to the properties P and P̂ of the used
S 0 and Ŝ. The bound ε̂ + ∆2 of the case i = j − 1 is justified as follows: Its ∆2
comes from that of the case i = j − 2 as the median of Lj−1 ∪ Rj−1 differs from
that of all n by at most ∆2 |Lj−1 |. Also there are no more than |Lj−1 |ε̂ elements
in Lj−1 greater than the median of Lj−1 ∪ Rj−1 .
After compare-odd-pairs and split,
ε0 ∆j−1 ,
if i = 0,
0 0 j−i−2
ε ε∆
+ 2∆j−i , if 1 ≤ i ≤ j − 4 and i is odd,
0
li
ε0 ∆j−i−1 + 2∆j−i+1 , if 2 ≤ i ≤ j − 3 and i is even,
≤
|Li |
if i = j − 2,
ε0 ε̂ + ∆2 ,
if j − 1 ≤ i ≤ j,
2 ε̂ + ∆2 ,
as indicated in the 4th row of Fig. 4.
10
Piotr Berman and Junichiro Fukuyama
Likewise, right after the 2nd compare-even-pairs, one can check
li0
≤ ∆j−i+1 − ∆j−i+2 ,
|Li |
for each i ≤ j. These mean the jth invariant completing the induction step. The
lemma follows.
t
u
Hence, the total number of strangers in the output of H is bounded by
2∆|Ljmax −1 | < ∆2−c lg lg n+3 < n lg−c n. This proves Theorem 2.
5 Selection in O (log n) Time with O (n log log n) Messages
on a Communication Network
We show Theorem 1 in this section. Our parallel algorithm for the distributed
k-selection problem on a communication network G mimics the comparator network H given by Theorem 2 and three AKS sorting networks. Keep assuming
that n is a sufficiently large power of 6 and the inputs have distinct values. Let
q = lg n.
Choose c = 2 in Theorem 2, so the output of H includes less than n lg−2 n
strangers. Numeric expressions in this section omit obvious floor/ceiling functions if any. We first describe the algorithm focusing on the case k = n2 .
Algorithm 1 for Finding the Median of n Inputs:
1. Apply H to the given n inputs. Let L and R be its left and right halves of
the output, respectively, and C be an empty array.
n
2. Arbitrarily partition L into 2q
subarrays each of q items naming them
n .
L1 , L2 , . . . , L 2q
3. Find the maximum value li in each Li .
n by an AKS sorting network. Add the elements of L to
4. Sort the l1 , l2 , . . . , l 2q
i
−2
n
C for every i such that li is among the largest n lg n of l1 , l2 , . . . , l 2q .
5. Perform Steps 2–4 to R symmetrically to update C.
6. Sort the elements of C by another AKS sorting network. Find its median,
returning it as the median of all n.
Each AKS sorting network used above re-orders at most 2n lg−1 n inputs.
This allows us to detect the median with a sufficiently small number of comparisons, leading to O (n log log n) messages. We confirm the correctness of the
algorithm first.
Lemma 2. Algorithm 1 correctly finds the median of all n inputs.
Proof. Let i be an index such that li is not among the largest n lg−2 n of
n . If an element of L were a stranger or the median of n, the largest
l1 , l2 , . . . , l 2q
i
−2
n lg n items would be all strangers, contradicting Theorem 2. So no element
Efficient Distributed Selection Algorithm
11
of L − C is a stranger or the median. The same is true for R − C except that
the median is not included.
Steps 4 and 5 remove the same number of items from the both halves, each
neither a stranger nor the median. Hence Step 6 correctly finds the median of
all.
t
u
We now describe how to run Algorithm 1 on the communication network G.
Find our implementation below noting two remarks.
- Regard that there are 2n processor nodes in G instead of n, since any of them
at odd and even time slots can play two different roles.
- Every processor in G can help simulate any of the four comparator networks
the way mentioned in Section 2: It can receive two numeric items from other
nodes to send their minimum and/or maximum to anywhere in G.
Implementation of Algorithm 1 on G: Simulate two AKS sorting networks
in Steps 4 and 5 with extra n processors. Then select the elements of C in
O (log n) time as follows. Initially the first node of each Li is notified if li ∈ C.
Move items in L∩C by sending messages so that L∩C is held by the consecutive
smallest numbered nodes.
To find the message destinations, construct, right after li are sorted in Step
n
nodes whose leaves are the first nodes of all
4, a complete binary tree T of 2q
n . (For simplicity identify the first node with L
L1 , L2 , . . . , L 2q
i itself.) Started
from the leaves, recursively compute the number of Li ⊂ C existing in the
subtree T 0 rooted at each node of T . Then, started at the root of T , compute
the number of Li ⊂ C existing outside T 0 to the left: It is recursively sent from
the parent of the current node. Pass it to its left child. Add the total number of
Li ⊂ C under the left child and send the value to the right child.
This way every leaf of T is informed of the number of Li ⊂ C to its left. The
leaf disseminates the information to all the nodes in the same Li . Each node is
now able to compute the message destination so L ∩ C is held by the consecutive
smallest numbered nodes.
Merge L ∩ C with R ∩ C similarly. Perform Step 6 with a simulated AKS
sorting network on the obtained C. This completes the description of our implementation.
Algorithm 1 correctly runs on G in O (log n) time with O (log log n) messages.
One can check it with Lemma 2 noting that:
- The simulated H runs in O (log log n) time with O (n log log n) messages.
- Each of the three simulated AKS sorting networks runs in O (log n) time with
O (n) messages.
- It takes O (n) messages each of less than lg n bits to construct C.
This verifies Theorem 1 for k = n2 .
Selecting an item but the median is done similarly. Assume without loss of
generality that we want the kth smallest item such that k < n2 . We can detect
12
Piotr Berman and Junichiro Fukuyama
it by adding extra n − 2k elements valued −∞ to the input array. This changes
no asymptotic bounds we showed so far.
We have constructed a parallel algorithm on G that distributedly computes
the kth smallest of the n inputs in O (log n) time with O (n log log n) messages.
We now have Theorem 1.
6 The Universal Parallel Time Lower Bound lg n on a
Communication Network
In this section, we show that it takes at least lg n steps to compute many basic data aggregation problems on a communication network G including the
distributed selection problem. Consider a parallel algorithm to compute a function f (x1 , x2 , . . . , xn ) ∈ {0, 1} where x1 , x2 , . . . , xn are input numeric items of
b-bits distributed over G. Such f is said to be critical everywhere if there exist
x1 , x2 , . . . xn such that
f (x1 , x2 , . . . , xi−1 , x̂i , xi+1 , . . . , xn ) 6= f (x1 , x2 , . . . , xi−1 , xi , xi+1 , . . . , xn ) , (3)
for each index i = 1, 2, . . . , n and some b-bit numeric item x̂i .
Below we prove the time lower bound lg n to compute such f as the following
theorem.
Theorem 3. Let f be a function on n inputs critical everywhere. It takes at least
lg n parallel steps in the worst case to compute f on a communication network
distributedly.
t
u
By the theorem, we will have the statement mentioned in Section 1.
Corollary 1. Any parallel algorithm takes at least lg n steps in the worst case
to compute each of the following three problems on a communication network
distributedly: i) the k-selection problem on n inputs each of at least dlg ne + 1
bits; ii) the problem of finding the sum of n inputs; iii) the problem of counting
numeric items among n inputs, each exceeding a given threshold.
Proof. i): Given such an algorithm on n inputs x1 , x2 , . . . , xn , let f (x1 , x2 , . . . , xn )
be the least significant bit of the kth smallest input. Such a function f is critical
everywhere: Choose xi = 2(i − 1) for i = 1, 2, . . . , n. If k > 1, let x̂i = xk − 1,
otherwise x̂i = 1. (All of these numbers are dlg ne + 1-bit integers.) These satisfy
(3). By the theorem, the algorithm takes at least lg n steps to compute such f
on a communication network distributedly.
ii), iii): Shown similarly to i) by choosing f as the least significant bit of the
sum of n inputs, and that of the number of inputs exceeding a given threshold,
respectively.
t
u
Consider the class of problems to compute on a communication network
distributedly such that any particular bit of the computed result is a function
critical everywhere. It is very large containing many aggregation problems such
as the above three, each with the parallel running time lower bound lg n on G.
Efficient Distributed Selection Algorithm
13
We prove the theorem. Denote by vi the ith processor node of G, and by
xi the numeric item held by vi at time 0. Also let A be a parallel algorithm to
compute f on G distributedly. Since f is critical everywhere, there exists an input
set X = {x1 , x2 , . . . , xn } such that (3). Assume f (x1 , x2 , . . . , xn ) = 1 without
loss of generality, written as f (X) = 1.
We focus on the computation performed by A on the input set X. For time
t ≥ 0 and index i = 1, 2, . . . , n, define the set Vi,t of nodes in G recursively as
follows.
- Vi,0 = {vi } at time t = 0.
- Vi,t = Vi,t−1 ∪ Vj,t−1 if vj sends a message to vi at time t ≥ 1. Otherwise
Vi,t = Vi,t−1 .
Vi,t is the set of processors vj such that the values of xj can possibly affect the
computational result at vi at time t.
Let tmax be the time when A terminates on X, and vi be the processor that
decides f (X) = 1 at time tmax . The size of Vi,tmax must be n; otherwise A
decides f (X) = 1 at vi with no information on some input xj , contradicting
that f is critical everywhere. The following lemma formally confirms it.
Lemma 3. |Vi,tmax | = n.
Proof. By the equivalence between an algorithm and Boolean circuit [17], there
exists a Boolean circuit C to compute f (X), which is converted from A by the
reduction algorithm. Construct the subgraph of C that decides f (X) = 1 at
vi as follows: For each t = 0, 1, . . . , tmax , with the algorithm running on X at
each processor vj and the exchanged messages, inductively construct a Boolean
circuit to decide each bit of the computed result stored at vj at time t. For the
time t = tmax and processor vi , we have a circuit C 0 that decides f (X) = 1 only
from xj such that vj ∈ Vi,tmax .
The existence of C 0 means f (X) = 1 if all xj such that vj ∈ Vi,tmax have
the values specified by X. If |Vi,tmax | < n, this contradicts that f is critical
everywhere satisfying (3). Hence |Vi,tmax | = n.
t
u
We also have |Vj,t | ≤ 2t for each t and j. It is because |Vj,t | = |Vj,t−1 ∪ Vl,t−1 | ≤
2 · 2t−1 = 2t if a processor vl sends a message to vj at time t ≥ 1.
Threfore, lg n = lg |Vi,tmax | ≤ tmax . Theorem 3 follows.
7 Concluding Remarks and Open Problems
The depth bound 1296cd 51 c lg lg n we showed in Theorem 2 could be further improved by devising different gadgets using ε-halvers, or different expander graphs.
It is interesting to find its limit. It is also possible to improve the aforementioned
estimate in [12] on the constant of the O (log n) depth of a sorting network. In
addition, it is good to ask if Ω (log log n) messages are necessary to compute the
the distributed selection problem in O (log n) time on a communication network.
14
Piotr Berman and Junichiro Fukuyama
Appendix: Explicit Construction of a 1/5-Halver
An (m, q, d)-expander graph is a bipartite graph (V1 , V2 , E) such that |V1 | =
|V2 | = m, the maximum degree is at most q, and
|U |
|Γ (U ) | ≥ 1 + d 1 −
|U |,
m
for
every U ⊆ V1. Gabber and Galil demonstrated [15] how to explicitly construct
√
an m̂2 , 7, 2−2 3 -expander graph for every m̂ ∈ Z+ . Below we build a 15 -halver
from such an expander graph.
For any bipartite graph S = (V1 , V2 , E) such that |V1 | = |V2 |, denote by
G = f (S) the digraph (V1 , E 0 ) such that (vi , wj ) ∈ E iff (vi , vj ) ∈ E 0 where vi
is the ith vertex of V1 and wj the jth of V2 . Also denote S = f −1 (G). Let Gh
(h ∈ Z+ ) be the digraph (V1 , E 00 ) such that (v, w) ∈ E 00 iff there exists a path
of length h from v to w in G.
Let S be the bipartite graph from the Gabber-Galil construction, and G =
f (S). Construct
S 0 = f −1 Gh , where h = 21.
We show that the bipartite graph S 0 satisfies (1) for ε = 15 and every subset
U of V1 . It suffices to verify |Γ (U ) | ≥ c(1 − ε)m for a vertex set U of G such
that |U | ≥ cεm, and any c ∈ (0, 1]. It is further sufficient to assume c = 1 in the
following argument. Let Ui be the neighbor set of U in Gi such that |U | = εm.
With the recurrence
√
!
|Ui+1 |
2− 3
|Ui |
|Ui |
≥ 1+
1−
,
m
2
m
m
we calculate |Umh | > 0.806 > 1 − ε for h = 21 and ε = 51 . This confirms that a
maximum degree at most 7h is sufficient to achieve (1).
Switch V1 and V2 and construct S 00 = f −1 (f (S)h ). Our 15 -halver is the undirected bipartite graph (V1 , V2 , E) such that E consists all the edges in S 0 and
S 00 . It has a maximum degree at most 2 · 721 .
For building the comparator network H of Theorem 2, we must consider a
case when m is a sufficient
of 6 but not a square. If so, construct
√ large power
√
2
2− 3
an ε-halver from the b mc , 7, 2
-expander graph. It is straightforward to
check that h = 21 is still sufficient to achieve ε = 51 .
References
1. Kuhn, F., Locher, T., Wattenhofer, R.: Tight bounds for distributed selection.
In: Proc. of the 19th Annual ACM Symposium on Parallelism in Algorithms and
Architectures (SPAA), pp. 145-153. ACM Press (2007)
Efficient Distributed Selection Algorithm
15
2. Kempe, D., Dobra, A., Gehrke, J.: Gossip-based computation of aggregate information. In: Proc. of the 44th Annual IEEE Symposium on Foundations of Computer
Science (FOCS), pp. 482–491. IEEE Press (2003)
3. Santoro, N., Sidney, J.B., Sidney S.J.: A distributed selection algorithm and its
expected communication complexity. Theoretical Computer Science, vol. 100, pp.
185–204. Elsevier (1992)
4. Frederickson, G.N.: Tradeoffs for Selection in distributed networks. In: Proc. of the
2nd Annual ACM Symposium on Principles of Distributed Computing (PODC),
pp. 154–160. ACM Press (1983)
5. Pardalos, P., Rajasekaran, S., Reif, J., Rolim, J.: Handbook on randomized computing. Kluwer Academic Publishers (2001)
6. Chaudhuri, S., Hagerup, T., Raman, R.: Approximate and exact deterministic
parallel selection. LNCS, vol. 711, pp. 352–361. Springer (1993)
7. Ajtai, M., Komlós, J., Steiger, W.L., Szemerédi, E.: Optimal parallel selection has
complexity O (log log n). Journal of Computer and System Sciences, vol. 38, pp.
125–133. Academic Press (1989)
8. Frederickson, G.N., Johnson, D.B.: The complexity of selection and ranking in X
+ Y and matrices with sorted columns, J. Comput. System Sci., vol. 24 (2), pp.
197–208 Elsevier (1982).
9. Santoro, N., Suen, E.: Reduction techniques for selection in distributed files. IEEE
Transactions on Computers, vol. 38, pp. 891–896. IEEE Press (1989)
10. Ajtai, M., Komlós, J., Szemerédi, E.: Sorting in c log n parallel steps. Combinatorica, vol. 3, pp. 1–19. Springer (1983)
11. Paterson, M.: Improved sorting networks with O(log N) depth. Algorithimica, vol.
5, pp. 65–92. Springer (1990)
12. Seiferas, J. L.: Sorting networks of logarithmic depth, further simplified. Algorithmica, vol. 53, pp. 374–384. Springer (2007)
13. Baddar, S. W. A., Batcher, K. E.: Designing sorting networks. Springer (2011)
14. Goodrich, M.T.: Zig-zag sort: a simple deterministic data-oblivious sorting algorithm running in O(n log n) time. In: Proc. of Symposium on Theory of Computing
(STOC), pp. 684–693. ACM Press (2014)
15. Gabber, O., Galil, Z.: Explicit Construction of Linear-Sized Superconcentrators.
J. Computer and System Sciences, vol. 22, pp. 407–420. Academic Press (1981)
16. Lubotzky, A., Phillips, R., Sarnak, P.: Explicit Expanders and the Ramanujan
Conjectures. In: Proc. of Symposium on Theory of Computing (STOC), pp. 240–
246. ACM Press (1986)
17. Papadimitriou, C. H.: Computational Complexity. Addison-Wesley (1994)
| 8cs.DS
|
VISUAL AND TEXTUAL SENTIMENT ANALYSIS USING DEEP FUSION CONVOLUTIONAL
NEURAL NETWORKS
Xingyue Chen
Qingjie Liu∗
Yunhong Wang
arXiv:1711.07798v1 [cs.CL] 21 Nov 2017
The State Key Lab. of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China
ABSTRACT
Sentiment analysis is attracting more and more attentions and
has become a very hot research topic due to its potential applications in personalized recommendation, opinion mining,
etc. Most of the existing methods are based on either textual or visual data and can not achieve satisfactory results,
as it is very hard to extract sufficient information from only
one single modality data. Inspired by the observation that
there exists strong semantic correlation between visual and
textual data in social medias, we propose an end-to-end deep
fusion convolutional neural network to jointly learn textual
and visual sentiment representations from training examples.
The two modality information are fused together in a pooling
layer and fed into fully-connected layers to predict the sentiment polarity. We evaluate the proposed approach on two
widely used data sets. Results show that our method achieves
promising result compared with the state-of-the-art methods
which clearly demonstrate its competency.
Index Terms— sentiment analysis, visual sentiment,
deep fusion, convolutional neural network
1. INTRODUCTION
Recently, with the rapid development of the Mobile Internet and smart terminals, more and more people prefer to use
short texts and images to express their opinions and communicate with each other on social medias, such as Twitter,
Flickr, Microblog, etc. Each day, billions of messages and
posts are generated. Mining sentiment information from these
data is of great important for applications such as personalized recommendation, user modeling, crisis management,
etc [1]. It has been attracting more and more attentions and
has become a hot research topic. Plenty of works have been
published [1, 2, 3, 4, 5, 6], significantly promoting the development of sentiment analysis studies.
Most of these methods are based on text information and
consider sentiment analysis as a special case of text classification, focusing on designing specific vector representation of
sentences [2]. For example, Le and Mikolov [7] proposed an
approach to learn distributed representations for documents
∗ Corresponding
author
(a)
(b)
(c)
(d)
Fig. 1. Example image and co-related text pairs crawled from
Flickr.
and applied on sentiment analysis, achieving excellent performance. Recently, with the development of deep learning
techniques, researchers tend to use deep neural networks to
learn distributed and robust representations of text for sentiment classification. Kim [8] and Johnson [9] implemented
Convolutional Neural Networks (CNN) for text classification.
Poria et al. [10] presented a novel way of extracting features
from short texts, which is based on the activation values of an
inner layer of a deep CNN. These deep learning based methods obtain better performance than hand-crafted features in
text sentiment analysis.
Besides text data, visual content of images also convey
sentiment information [5]. The main challenge behind visual
sentiment analysis is how to learn effective representations
to bridge the ”affective gap” between low level features of
images and high level semantics [3]. Motivated by the fact
that sentiment involves high-level abstraction, such as objects
or attributes in images, Yuan et al. [11] proposed to employ
visual entities or attributes as features for visual sentiment
prediction. Borth et al. [12] employed 1,200 adjective noun
pairs (ANP) which were used as keywords to crawl images
from Flickr as mid-level features for sentiment classification.
To handle large scale weakly labeled training images, You et
al. [4] designed a progressively trained CNN architecture for
visual sentiment analysis. Compared with approaches employing low-level or mid-level features, deep learning algorithms learn hierarchical features and can achieve outstanding
performance.
In social medias, one message or post usually contains
both images and texts. They tend to have a high degree of
relevance in semantic space, providing complementary information for each other. So, it is better to take both of them
into consideration when determining sentiment. Fig. 1 shows
several examples of image and text pair crawled from Flickr1 .
It can be observed from these four examples that, both image and the corresponding text in Fig. 1(a) indicate that this
tweet carries a positive sentiment; while in Fig. 1(b) both are
negative; in Fig. 1(c), it is difficult to tell the sentiment from
the text, however, we can tell that this image expresses positive sentiment; on the contrary, in Fig. 1(d), it is hard to tell
the sentiment from the image, however the word ’amazing’ in
the text suggests an overall positive sentiment. These examples explain the motivation for our work. Since the sentiment
conveyed from visual and textual content can explain or even
strengthen each other, we learn people’s sentiment through
the available images and the informal short text.
Recent progress in sentiment classification proves the
advantage of using both text and visual modalities. Wang et
al. [13] proposed a decision level fusion method for sentiment
classification, which combines textual sentiment obtained
by N-gram features with visual sentiment obtained with
mid-level visual feature to predict final results. Later, You
et al. [14] proposed a cross-modality consistent regression
(CCR) scheme for joint textual-visual sentiment analysis.
Their approach employed deep visual and textual features to
learn a regression model. Despite effectiveness, these models ignore the intrinsic relationship between text and images
and use the visual and textual information separately. To
tackle this, You et al. [15] employed a tree-structured LSTM
for textual-visual sentiment analysis, leading to better mapping between textual words and image regions. This model
achieved the best performance over other fusion models. Although this model makes better use of the correspondence
between text and images, it lost the global information of
image in some extent, which we think is also important for
visual sentiment.
Inspired by the recent discovery that contextual information plays an import role in sentence classification [6], we
propose to utilize a structured model similar to [6] for visualtextual sentiment analysis. Different from [6], we use image
as contextual information source. Visual and textual sentiment representations are learned through multi-layer CNNs
and fused together in a pooling layer to predict sentiment
polarity. We carry out experiments on two public data sets,
including the Visual Sentiment Ontology (VSO) dataset [12]
and the Multilingual Visual Sentiment Ontology (MVSO)
dataset [16]. Our method achieves the state of the art result
on VSO dataset with 1.3% accuracy gain, demonstrating its
effectiveness.
The rest of this paper is organized as follows, Section 2
details the proposed approach. We conduct experiments on
1 https://www.flickr.com/
Section 3. And finally the paper is concluded in Secton 4.
2. DEEP FUSION SENTIMENT CLASSIFICATION
In this section, we introduce the proposed deep fusion neural
network for visual and textual sentiment analysis. The architecture of the proposed network is shown in Fig. 2, which has
two sub-networks. The upper sub-network takes text as input, and extracts vector representations of words. While the
lower sub-network takes image as input and extracts hierarchical features of image. The two representations are combined together in the following pooling layer to make full use
of two modalities information.
2.1. Deep visual sentiment feature
The recent developed CNN has been proved to be effective in
vision tasks such as image classification and object detection.
As visual sentiment analysis can be treated as a high-level
image classification problem, You et al. [4] applied CNN to
sentiment predication and achieved better performance than
using traditional low-level and mid-level features. Inspired
by the success of CNN and You’s work, in this paper, we
also implement a CNN architecture to learn high level visual
sentiment representations.
Although, more layers in CNN generally indicate a better performance, it has been reported that going deeper may
lead to worse result for visual sentiment analysis [4]. And
shallow networks may not be deep enough to learn high level
semantic representations. So, in our approach, we employ a
network that is similar to AlexNet [17], both consisting of 5
convolutional layers and taking a 244×244 pixels RGB image
as input. The nonlinearity activation function of each neuron
in this CNN is modeled by Rectified Linear Units (ReLUs),
f (x) = max(0, x).
(1)
Compared with saturating nonlinearity function, such as sigmoid, ReLUs can accelerate learning process, and prevent the
gradients vanishing during back propagation.
Each convolutional layer convolves the output of its previous layer with a set of learned kernels, followed by ReLU
non-linearity, and two optional layers: local response normalization and max pooling. The local response normalization
layer is applied across feature channels, and the max pooling
layer is applied to neighboring neurons. CNN learns hierarchical feature representations. One simple way to use these
features is making use all of them. However, since sentiment
analysis involves high level abstraction, we take the last output of max pooling layer as the visual sentiment representation.
2.2. Learning textual sentiment representation
Recently, deep learning models are also successfully applied
in many natural language processing tasks, such as sentence
Fig. 2. Framework of the proposed approach.
classification [8]. We use the same architecture as [8] to learn
textual sentiment representation.
In our work, we directly employ a pre-trained distributed
representations of words, keep the word vectors static and
learn only the other parameters of the model. Let xi ∈ Rk be
a k-dimensional word vector corresponding to the i-th word
in the sentence. A sentence of length n is represented as
have described the process by which one feature is extracted
from one filter. Our model uses multiple filters (with varying
window sizes) to obtain features. The ĥ of the different filters
is concatenated in the pooling layer. We take these features as
the textual sentiment representation.
x1:n = x1 ⊕ x2 ⊕ · · · ⊕ xn
In social medias, text and image often appear in pairs, providing complementary semantic information for each other. Ren
et al. [6] have proved that context information can be helpful
in social text sentiment analysis. Inspired by their work, we
believe the co-related image can also provide contextual information and is beneficial for improving sentiment analysis
performance. To this end, we build a fusion layer, combining
the output of image pooling layer with that of the text neural
model, before feeding them to the non-linear hidden layer. As
the output of textual pooling layer is high dimensional feature
vector, the output of visual pooling layer should be reshaped
to one-dimensional vector. The fused feature is represented
as
x = xi ⊕ xt
(6)
(2)
where ⊕ is the concatenation operator. For convenience, let
xi:i+j represents the concatenation of words xi , xi+1 , ..., xi+j .
A convolution operation involves a filter w ∈ Rhk , which is
applied to a window of h words to produce a new feature. For
example, a feature ci is generated from a window of words
xi:i+h−1 by
ci = f (w ∗ xi:i+h−1 + b)
(3)
Here * is the convolution operation, b ∈ R is a bias term
and f is a non-linear function such as the hyperbolic tangent.
This filter is applied to each possible window of words in the
sentence {x1:h , x2:h+1 , · · · , xn−h+1:n } to produce a feature
map
c = [c1 , c2 , · · · , cn−h+1 ]
(4)
2.3. Deep fusion for sentiment analysis
(5)
where ⊕ is the concatenation operator. xi and xt refer to
the output of two sub-networks, respectively. In this way, the
hidden layer automatically combines visual and textual sentimental representations.
The fusion layer is followed by three fully-connected layers. The output of the last fully-connected layer is fed to a
2-way softmax which produces a distribution over the 2 class
labels. Our training objective is to minimize the cross-entropy
loss over a set of training examples:
!
efyi
(7)
Li = − log P fj
je
For one filter , we get a three-dimensional vector and each
dimension represents the output of one pooling method. We
where we use the notation fj to mean the j-th element of the
vector of class scores f .
We exploit pooling techniques to merge the varying number
of features from the convolution layer into a vector with fixed
dimensions. A typical pooling technique is the max pooling,
which chooses the highest value on each dimension from a
set of vectors. On the other hand, min and average pooling
have also been used for sentiment classification[18], giving
significant improvements. We consider all the three pooling
methods, concatenating them together as a new hidden layer
ĥ. Formally, the values of ĥ are defined as:
ĥ = [max {c} , avg {c} , min {c}]
3. EXPERIMENTS
We evaluate the proposed deep fusion neural network on two
datasets and compare the performance of our method with 3
state of the art methods.
3.1. Datasets
There are a few public datasets available for testing. And
most of them are either texts or images. We use the following
two largest datasets which can provide texts and images for
evaluation.
VSO [12]. VSO is built on top of the visual sentiment
ontology, and contains millions of images collected by querying Flickr with thousands of adjective and noun pairs (ANPs).
Each ANP has hundreds of images and the sentiment label
of each image is decided by sentiment polarity of the corresponding ANP. The dataset provides the metadata which enables us to obtain the description of an image. We only use
image with descriptions more than 5 words and less than 150
words, that leaves us 346,139 images. The training/testing
split is 80% and 20%.
MVSO-EN [16]. The MVSO dataset consists of 15,600
concepts in 12 different languages that are strongly related to
emotions and sentiments expressed in images. Similar to the
VSO dataset, these concepts are defined in the form of ANPs,
which are crawled from Flickr. In this paper, we only use
the English dataset for our experiments. In order to avoid the
problem of excessive noise, we choose ANPs with sentiment
scores more than 0.16 and less than -0.1. We skip images
with descriptions more than 150 words and less than 5 words.
There are total 613,955 image and text pairs used in our experiment.The training/testing split is 80% and 20%.
3.2. Experimental results
To demonstrate the effectiveness of the proposed method, we
first compare it with two baseline methods, that is only using
image or text information for sentiment classification. We
call them Single Visual (SV) and Single Text (ST) models.
To show the superiority of our method, we further compare
it with 3 state of the art methods, including PCNN [4], You’s
method [14] and T-LSTM [15]. Specially, You et al. utilized
three strategies, named early fusion, which is logistic regression on concatenated visual and textual features; late fusion,
which is average of logistic regression sentiment score on
both visual and textual features; and cross-modality consistent regression (CCR).
In the textual part of our method, we employ the pretrained Word2Vec [19] model to get the distributed representations of words. The Word2Vec was pre-trained on
Wikipedia Corpus, and has a fixed size of 200 dimension.
Words do not present in pre-training dataset are initialized
randomly. The size of embedded word matrix is 150×200,
as the max-length of description is 150 words. The first layer
performs convolution over the embedded word vectors using
scalable filter size. In this work, we slide over 3, 4 or 5 words
at one time. In the visual part, the input images are first resized to 224×224 before inputing to the network. The first
five layers are same as AlexNet [17].
The model was trained on a workstation with i7-5600
CPU, 32G RAM, NVDIA 1080 GPU. We trained our model
using a mini-batch of 100. The initial learning rate is 0.0001
and exponential decays every 3000 steps with a base of 0.96.
Table 1 summarizes the performances of our approach and
comparison methods. It can be observed that the proposed
approach significantly outperforms baselines and PCNN
method, indicating fusing two modalities information does
improve sentiment prediction. This is mainly because multimodalities information enrich the semantic space. Our deep
fusion model also obtains better results than T-LSTM and
You’s method. We believe it mainly because our model utilizes the global information of image.
Table 1. Results on VSO dataset.
Methods
Baseline
ST
SV
PCNN [4]
Early Fusion
You’s [14]
Late Fusion
CCR
T-LSTM [15]
Deep fusion
Prec.
0.701
0.628
0.759
0.616
0.660
0.672
0.821
0.830
Rec.
0.680
0.640
0.826
0.646
0.629
0.678
0.833
0.857
F1
0.690
0.629
0.791
0.631
0.645
0.675
0.833
0.844
Acc.
0.695
0.631
0.781
0.621
0.650
0.672
0.833
0.847
Table 2 shows the performance of our deep fusion neural
network on MVSO-EN dataset. To our best knowledge, this
is the first performance report on MVSO data with all textual
and visual information used. We used the same experimental
setting to VSO, however obtained worse results than VSO.
This is mainly because the MVSO-EN dataset contains much
more noise than VSO.
Table 2. Results of our method on MVSO-EN dataset.
Methods
Prec. Rec.
F1
Acc.
ST 0.654 0.671 0.662 0.667
Baseline
SV 0.602 0.578 0.590 0.581
Deep fusion
0.740 0.730 0.735 0.737
4. CONCLUSIONS
In this paper, we present a new end-to-end framework for visual and textual sentiment analysis. Our method can effectively learn sentiment representations from noisy web images
with text, integrate both of them in a deep fusion layer and
predicts an overall sentiment. Extensive experimental results
demonstrate that the proposed deep fusion convolutional neural network have significantly improved the performance of
visual and textual sentiment analysis on VSO datasets, which
shows its effectiveness for this issue.
Acknowledgment
This work is surpported by the National Nature Science Foundation of China (61573045,61573045).
5. REFERENCES
[1] Fangzhao Wu and Yongfeng Huang, “Personalized microblog sentiment classification via multi-task learning,”
in AAAI, 2016, pp. 3059–3065.
[2] Bo Pang and Lillian Lee, “Opinion mining and sentiment analysis,” Foundations and trends in information
retrieval, vol. 2, no. 1-2, pp. 1–135, 2008.
[3] Jana Machajdik and Allan Hanbury, “Affective image
classification using features inspired by psychology and
art theory,” in MM. ACM, 2010, pp. 83–92.
[4] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao
Yang, “Robust image sentiment analysis using progressively trained and domain transferred deep networks,”
in AAAI, 2015, pp. 381–388.
[5] Jingwen Wang, Jianlong Fu, Yong Xu, and Tao Mei,
“Beyond object recognition: Visual sentiment analysis
with deep coupled adjective and noun neural networks,”
in IJCAI, 2016, pp. 3484–3490.
[6] Yafeng Ren, Yue Zhang, Meishan Zhang, and
Donghong Ji, “Context-sensitive twitter sentiment classification using neural network,” in AAAI, 2016, pp.
215–221.
[7] Quoc V Le and Tomas Mikolov, “Distributed representations of sentences and documents.,” in ICML, 2014,
vol. 14, pp. 1188–1196.
[8] Yoon Kim, “Convolutional neural networks for sentence
classification,” arXiv preprint arXiv:1408.5882, 2014.
[9] Rie Johnson and Tong Zhang, “Effective use of word
order for text categorization with convolutional neural
networks,” arXiv preprint arXiv:1412.1058, 2014.
[10] Soujanya Poria, Erik Cambria, and Alexander Gelbukh,
“Deep convolutional neural network textual features and
multiple kernel learning for utterance-level multimodal
sentiment analysis,” in EMNLP, 2015, pp. 2539–2544.
[11] Jianbo Yuan, Sean Mcdonough, Quanzeng You, and
Jiebo Luo, “Sentribute: image sentiment analysis from
a mid-level perspective,” in WISDOM. ACM, 2013, pp.
10–17.
[12] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel,
and Shih-Fu Chang, “Large-scale visual sentiment ontology and detectors using adjective noun pairs,” in MM.
ACM, 2013, pp. 223–232.
[13] Min Wang, Donglin Cao, Lingxiao Li, Shaozi Li, and
Rongrong Ji, “Microblog sentiment analysis based on
cross-media bag-of-words model,” in ICIMCS. ACM,
2014, pp. 76–80.
[14] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao
Yang, “Cross-modality consistent regression for joint
visual-textual sentiment analysis of social multimedia,”
in WSDM. ACM, 2016, pp. 13–22.
[15] Quanzeng You, Liangliang Cao, Hailin Jin, and Jiebo
Luo, “Robust visual-textual sentiment analysis: When
attention meets tree-structured recursive neural networks,” in MM. ACM, 2016, pp. 1008–1017.
[16] Brendan Jou, Tao Chen, Nikolaos Pappas, Miriam Redi,
Mercan Topkara, and Shih-Fu Chang, “Visual affect
around the world: A large-scale multilingual visual sentiment ontology,” in MM. ACM, 2015, pp. 159–168.
[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, “Imagenet classification with deep convolutional
neural networks,” in NIPS, 2012, pp. 1097–1105.
[18] Duy-Tin Vo and Yue Zhang, “Target-dependent twitter
sentiment classification with rich automatic features,” in
IJCAI, 2015, pp. 1347–1353.
[19] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean, “Distributed representations of
words and phrases and their compositionality,” in NIPS,
2013, pp. 3111–3119.
| 1cs.CV
|
Non-existence of boundary maps for some
hierarchically hyperbolic spaces
arXiv:1610.07691v1 [math.GT] 25 Oct 2016
Sarah C. Mousley
Abstract
We provide negative answers to questions posed by Durham, Hagen, and Sisto on the
existence of boundary maps for some hierarchically hyperbolic spaces, namely maps from rightangled Artin groups to mapping class groups. We also prove results on existence of boundary
maps for free subgroups of mapping class groups.
1
Introduction
Let Γ be a finite graph with vertex set V (Γ) = {s1 , . . . , sk }. The right-angled Artin group
determined by Γ, denoted by A(Γ), is the group with the following presentation:
A(Γ) = hs1 , . . . , sk : [si , sj ] = 1 ⇔ si sj is an edge in Γi.
Let S = Sg,n be a connected, oriented surface of genus g with n punctures, and let Mod(S)
denote the mapping class group of S. Clay, Leininger, and Mangahas [5] and Koberda [7]
construct “nice” embeddings of right-angled Artin groups to mapping class groups. In [1]
and [2], a geometric structure called a hierarchically hyperbolic space (HHS) was introduced.
Important examples of spaces that are HHS’s include mapping class groups of surfaces and rightangled Artin groups. In [6] Durham, Hagen, and Sisto construct a boundary for hierarchically
hyperbolic spaces (see Section 2). In that paper, the authors ask the following question,
motivated by a desire to develop a notion of geometrically finite subgroups of mapping class
groups.
Question 1.1. Let A(Γ) be a right-angled Artin group embedded in Mod(S) in the sense of
either Clay, Leininger, and Mangahas [5] or Koberda [7]. Does the embedding A(Γ) → Mod(S)
extend continuously to an injective map ∂A(Γ) → ∂M od(S)?
We prove that in general the answer to this question is no by providing, for each type of
embedding, an explicit example where the embedding does not extend continuously.
Theorem 1.2. There exists a surface S, a right-angled Artin group Γ, a Clay, Leininger, and
Mangahas embedding φ : A(Γ) → Mod(S), and a Koberda embedding φ0 : A(Γ) → Mod(S) such
that, regardless of the HHS structure on A(Γ), neither φ nor φ0 extends continuously to a map
∂A(Γ) → ∂Mod(S).
We also prove the following result which gives a complete characterization of Koberda
embeddings of free groups, which send all generators to powers of Dehn twists, that have
continuous extensions.
Theorem 1.3. Let {α1 , . . . , αk } be a collection of pairwise intersecting curves in S and Γ the
graph with V (Γ) = {s1 , . . . , sk } and no edges. For sufficiently large N , the homomorphism
φ : A(Γ) → Mod(S)
defined by
φ(si ) = TαNi for all i
is injective by the work of Koberda [7]. Moreover, φ extends continuously to a map ∂A(Γ) →
∂Mod(S) if and only if {α1 , . . . , αk } pairwise fill S, where A(Γ) is equipped with any HHS
structure.
1
In fact, we prove something stronger than Theorem 1.3. We prove a non-existence result
(Theorem 5.3) for a class of Koberda embeddings of right-angled Artin groups that are not
necessary free groups. We also prove an existence result (Theorem 6.1) for a class of embeddings
of free groups that includes the Koberda embeddings described in Theorem 1.3 as well as a
class of Clay, Leininger, Mangahas embeddings.
In Section 2 we will recall relevant definitions and theorems and introduce notation. Section
3 will establish a handful of lemmas that will be used for proving Theorem 1.2. Section 4 is
devoted to proving Theorem 1.2 for a Clay, Leininger, Mangahas embedding, and in Section
5 we prove Theorem 1.2 for a Koberda embedding. Using similar techniques, we then prove
that a more general class of Koberda embeddings of right-angled Artin groups do not extend
continuously (Theorem 5.3), which will imply one direction of Theorem 1.3. In Section 6 we
will prove Theorem 6.1, which will imply the other direction of Theorem 1.3.
Remark: Koberda [7] proved that both types of embeddings we discuss are injective. We
call the embeddings that send generators of our right-angled Artin group to mapping classes
that are pseudo-Anosov on subsurfaces Clay, Leininger, Mangahas embeddings primarily to
distinguish the two types, but also to emphasize that these types of embeddings have nice
geometric properties (see Theorem 2.5).
Acknowledgments: The author was supported by the Department of Defense (DoD)
through the National Defense Science and Engineering Graduate Fellowship (NDSEG) Program
and by a Research Assistantship through NSF Grant number DMS-1510034. The author would
like to thank her PhD advisor Chris Leininger for his numerous insights which inspired a great
deal of this paper and also for his patience and consistent support. The author would also like
to thank Mark Hagen and Matt Durham for helpful conversations.
2
Background
In this section, we recall some needed definitions and theorems.
A,B
Notation: Let f, g : X → R be functions. Given constants A ≥ 1 and B ≥ 0, we write f g
to mean f (x) ≥ A1 g(x) − B for all x ∈ X, and will just write f g when the constants are
understood.
2.1
Curves and subsurfaces
Throughout this paper, we let S = Sg,n denote a connected, oriented surface of genus g with n
punctures. Define the complexity of S to be ξ(S) = 3g −3+n. We will always assume ξ(S) ≥ 1.
Additionally, we fix a complete hyperbolic metric on S. That is, we assume that S is of the
form S = H2 /Λ, where Λ ⊆ Isom+ (H2 ) and Λ acts properly discontinuously and freely on H2 .
For i = 1, 2 let γei be a bi-infinite path in H2 with ends limiting to distinct points xi and yi
on ∂H2 . We say that γe1 and γe2 link if the geodesic connecting x1 to y1 intersects the geodesic
connecting x2 to y2 in the interior of H2 .
By a curve in S, we will always mean the geodesic representative in the homotopy class of
an essential, simple, closed curve in S. By a multicurve in S, we will always mean a collection
of pairwise disjoint curves in S. We write i(α, β) to denote the geometric intersection number
of curves α and β. We say that a pair of curves α and β fill S if for every curve γ in S we have
i(γ, α) > 0 or i(γ, β) > 0.
A non-annular subsurface Y of S is a component of S after removing a (possibly empty)
collection of pairwise disjoint curves on S. Additionally, we require that Y satisfies ξ(Y ) ≥ 1;
in particular, we do not consider a pair of pants to be a subsurface. We define ∂Y to be the
collection of curves in S that are disjoint from Y and also are contained in the closure of Y ,
treating Y as a subset of S. When Y 6= S, the path metric completion of Y is a surface with
2
boundary, and the image of this boundary under the map induced by the inclusion Y ⊆ S is
∂Y .
An annular subsurface of S is define as follows. Let α be a curve in S. Choose a component
α
e of the preimage of α in H2 , and let h ∈ Λ be a primitive isometry with axis α
e. Define
Y = (H2 − {x, y})/hhi,
where x and y are the fixed points of h on ∂H2 . Observe that Y is a compact annulus and
int(Y ) → S is a covering. We say that Y is the annular subsurface of S with core curve α. We
define ∂Y to be α.
For any subsurface Y of S, we will write Y ⊆ S, even though when Y is an annulus, Y is
not a subset of S.
Given f ∈ Mod(S) and a curve or simple bi-infinite geodesic γ in S, we define f (γ) to
be the curve or simple bi-infinite geodesic obtained as follows. Consider a component γ
e of
the preimage of γ in H2 . Choose a representative ψ in the isotopy class of f and lift it to a
map ψe : H2 → H2 . We define f (γ) to be the image in S of the geodesic in H2 that connects
e γ ) on ∂H2 . Given Y ⊆ S, if Y is non-annular we let f (Y ) denote the
the endpoints of ψ(e
non-annular subsurface in its isotopy class. If Y is an annulus with core curve α, we let f (Y )
denote the annular subsurface of S with core curve f (α).
2.2
Curve complex
Let Y be a subsurface of S. If Y satisfies ξ(Y ) ≥ 1, the curve complex of Y , denoted C(Y ),
is the simplicial complex whose vertices are curves contained in Y , and if ξ(Y ) > 1, a set of
vertices form a simplex if and only if they are pairwise disjoint. If ξ(Y ) = 1, then we define
the simplices of C(Y ) differently. In the case that Y is a once punctured torus, a set of vertices
form a simplex if and only if they pairwise intersect exactly once. If Y is a four times punctured
sphere, a set of vertices form a simplex if and only if they pairwise intersect exactly twice.
Now let Y be a compact annulus. Consider all embedded arcs in Y that connect one
boundary component to the other. We define two arcs to be equivalent if one can be homotoped
to the other, fixing the endpoints of the arcs throughout the homotopy. In this case, the curve
complex of Y is the simplicial complex whose vertices are equivalence classes of arcs, and a set
of vertices form a simplex if and only if for each pair of vertices there exist representative arcs
of each whose restrictions to int(Y ) are disjoint. The following simple formula will be useful
to us: given inequivalent arcs α, β in C(Y ),
dC(Y ) (α, β) = |α · β| + 1,
(1)
where α · β denotes the algebraic intersection number of α and β.
2.3
Markings and subsurface projection
A marking µ on S is a maximal collection of pairwise disjoint curves in S, denoted base(µ),
together with another collection of associated curves called transversals: for each β ∈ base(µ)
its associated transversal γβ is a curve that intersects β minimally (i.e. once or twice) and is
disjoint from all other curves in base(µ).
Let Y be a subsurface of S and β a multicurve in S. We will now define the projection of β
to Y , which we will denote by πY (β). Suppose Y is not an annulus and β is a single curve. If
β is disjoint from Y , define πY (β) = ∅. If β is contained in Y , define πY (β) = β. Otherwise,
β ∩ Y is a collection of essential arcs in Y with endpoints on ∂Y . For each such arc γ, take
the geodesic representatives of the boundary components of a small regular neighborhood of
γ ∪ ∂Y that are contained in Y . Define πY (β) to be the collection of all such curves over all
arcs γ in β ∩ Y . If β is a multicurve, define πY (β) to be the union of the projections to Y of
each curve in β.
3
Now let Y be an annular subsurface with core curve α and int(Y ) → S the associated
covering. Let β be a multicurve or a bi-infinite, simple geodesic in S. Consider the full preimage
of β in int(Y ). Each component is an arc in int(Y ) which we view as having endpoints on the
boundary of Y . In this case, we define πY (β) to be the (equivalence classes of) arcs in this
collection that have an endpoint on each boundary component of Y . When convenient, we will
write πα (β) instead of πY (β).
We now describe how to project a marking µ to Y ⊆ S. If Y is non-annular or Y is
an annulus whose core curve is not contained in base(µ), we define πY (µ) = πY (base(µ)).
Otherwise, Y is an annulus with core curve α ∈ base(µ), and we define πY (µ) to be πY (γα ),
where γα is the transversal associated to α.
Given any subsurface Y ⊆ S, we define
dY (µ, µ0 ) = diamC(Y ) (πY (µ) ∪ πY (µ0 )),
where µ and µ0 are markings, collections of curves, or (when Y is an annulus) bi-infinite simple
geodesics in S. A useful fact about subsurface projection is the following. For all f ∈ Mod(S)
dY (f (µ), f (µ0 )) = df −1 (Y ) (µ, µ0 ).
In this paper, we utilize the following theorem, which involves subsurface projections.
Theorem 2.1 (Lemma 2.3 in [10]). For all subsurfaces W of S, given any marking or multicurve µ such that πW (µ) 6= ∅, we have that diamC(W ) (πW (µ)) ≤ 2. If W is an annulus, then
diamC(W ) (πW (µ)) ≤ 1.
f
Masur and Minsky [10] define the marking graph of S, denoted M(S),
to be the graph
whose vertices are markings and vertices are adjacent if one can be obtained from the other by
f
an elementary move; see [10] for a complete definition. Giving M(S)
the path metric dM(S)
f
f
and Mod(S) a word metric dMod(S) , there is an action of Mod(S) on M(S) by isometries for
which every orbit map is a quasi-isometry. The following theorem gives a relationship between
f
distances in M(S)
and subsurface projections.
Theorem 2.2 (Lemma 3.5 in [10]). For any subsurface W of S and any markings µ and µ0
(µ, µ0 ).
on S, we have that dW (µ, µ0 ) ≤ 4dM(S)
f
We say that distinct subsurfaces X and Y are disjoint if πX (∂Y ) = ∅ and πY (∂X) = ∅. We
say that X is a proper subsurface of Y , denoted X ( Y , if πY (∂X) 6= ∅ and πX (∂Y ) = ∅. We
say that X and Y are overlapping, denoted X t Y , if πY (∂X) 6= ∅ and πX (∂Y ) 6= ∅. In the
case where X and Y are not annuli, these relationships, respectively, are disjointness, proper
containment, and intersection without containment as subsets of S. We say X and Y fill S if
for every curve γ in S we have πX (γ) 6= ∅ or πY (γ) 6= ∅.
The following theorems will be used to prove our results. The first theorem was proved
in [4] and later a simpler proof with constructive constants appeared in [8].
Theorem 2.3 (Behrstock inequality: Theorem 4.3 in [4], Lemma 2.13 in [8]). Let X and Y
be overlapping subsurfaces of S and µ a marking on S. Then
dX (µ, ∂Y ) ≥ 10
implies that
dY (µ, ∂X) ≤ 4.
Theorem 2.4 (Bounded Geodesic Image Theorem: Theorem 3.1 in [10]). There exists a
constant K0 depending only on S such that the following is true. Let Y and Z be subsurfaces
of S with Y a proper subsurface of Z. Let v1 , . . . , vn be any geodesic segment in C(Z) satisfying
πY (vi ) 6= ∅ for all 1 ≤ i ≤ n. Then
diamC(Y ) (πY (v1 ) ∪ . . . ∪ πY (vn )) ≤ K0 .
4
2.4
Partial order on subsurfaces
Let µ, µ0 be markings on S and K ≥ 20. Let Ω(K, µ, µ0 ) denote the collection of subsurfaces Y
of S such that dY (µ, µ0 ) ≥ K. Behrstock, Kleiner, Minsky, and Mosher [3] define the following
partial order on Ω(K, µ, µ0 ). Given X, Y ∈ Ω(K, µ, µ0 ) such that X t Y , define X ≺ Y if and
only if one of the following equivalent conditions is satisfied:
dX (µ, ∂Y ) ≥ 10,
dX (∂Y, µ0 ) ≤ 4,
dY (µ, ∂X) ≤ 4,
or dY (∂X, µ0 ) ≥ 10.
That these conditions are equivalent is a consequence of Theorem 2.3; see Corollary 3.7 in [5].
2.5
Embedding RAAGs in Mod(S)
If f ∈ Mod(S) is such that there exists a representative in the isotopy class of f that pointwise
fixes the complement of a non-annular subsurface Y , we say that f is supported on Y . Given
such an f , we define the translation length of f on C(Y ) to be
dY (µ, f n (µ))
,
n→∞
n
τY (f ) = lim
where µ is any marking on S. If f ∈ Mod(S) is a power of a Dehn twist about a curve α, we
say that f is supported on the annular subsurface Y with core curve α, and define τY (f ) to be
the absolute value of the power. In either case, we say that Y fully supports f if τY (f ) > 0.
By the work of Masur and Minsky [9], when Y is non-annular, Y fully supports f if and only
if f is pseudo-Anosov on Y .
Clay, Leininger, and Mangahas [5] proved the following result, which allows us to find
quasi-isometrically embedded right-angled Artin subgroups inside Mod(S).
Theorem 2.5 (Theorem 2.2 in [5]). Let Γ be a finite graph with V (Γ) = {s1 , . . . , sk }, and let
{X1 , . . . , Xk } be a collection of non-annular subsurfaces of S. Suppose si sj is an edge in Γ if
and only if Xi and Xj are disjoint, and si sj is not an edge in Γ if and only if Xi t Xj or
i = j. Then there exists a constant C > 0 such that the following holds. Let {f1 , . . . , fk } be a
set of mapping classes of S such that fi is pseudo-Anosov on Xi and satisfies τXi (fi ) ≥ C for
all i. Then the homomorphism
φ : A(Γ) → Mod(S)
defined by
φ(si ) = fi
for all i
is a quasi-isometric embedding, and in particular is injective.
Koberda [7] also has a result which produces right-angled Artin subgroups of Mod(S).
Below we give a special case of Koberda’s result that we will use.
Theorem 2.6 (Theorem 1.1 in [7]). Let {α1 , . . . , αk } be a collection of distinct curves in S. Let
Γ be the graph with V (Γ) = {s1 , . . . , sk } and with si sj an edge in Γ if and only if i(αi , αj ) = 0.
Then for sufficiently large N , the homomorphism
φ : A(Γ) → Mod(S)
defined by
φ(si ) = TαNi
for all i,
is injective, where Tαi denotes a Dehn twist about αi .
2.6
Gromov boundary of hyperbolic spaces
A geodesic metric space X is Gromov hyperbolic (or just hyperbolic) if there exists a δ ≥ 0
such that given any geodesic triangle in X, each side is contained in the δ-neighborhood of the
union of the other two sides. Given a Gromov hyperbolic space (X, dX ) and points x, y, z ∈ X,
the Gromov product of x and y with respect to z is defined as
(x, y)z =
1
(dX (x, z) + dX (y, z) − dX (x, y)) .
2
5
We say that a sequence (xn ) in X converges at infinity if lim inf (xi , xj )z = ∞ for some (any)
i,j→∞
z ∈ X. We define two such sequences (xn ) and (yn ) to be equivalent if lim inf (xi , yj )z = ∞
i,j→∞
for some (any) z ∈ X. The Gromov boundary of X is the collection of all such sequences up
to this equivalence, and is denoted ∂G X or just ∂X when it is clear from context that we are
using the Gromov boundary.
One Gromov hyperbolic space that this paper is concerned with is the curve complex of S,
which was proved to be Gromov hyperbolic by Masur and Minsky [10]. We can now state a
corollary of Theorem 2.4 that will be useful later.
Corollary 2.7. Let X and Y be subsurfaces of S with X a proper subsurface of Y . Suppose
(µn )n∈N is a sequence of markings on S such that πY (µn ) → λ for some λ ∈ ∂C(Y ). Then
diamC(X) (πX (µ1 ) ∪ πX (µ2 ) ∪ . . .) < ∞.
Proof. For each n, choose αn ∈ πY (µn ). Because πY (µn ) → λ ∈ ∂C(Y ), we can choose L large
so that for all n ≥ L we have
(αn , αL )α1 ≥ 2 + dY (∂X, α1 ),
(2)
where the Gromov product is computed in C(Y ). Consider n ≥ L. Let γn be a geodesic in
C(Y ) with endpoints αn and αL . If there exists a vertex v on γn with πX (v) = ∅, then v and
∂X form a multicurve in Y, which implies that
1
dY (αn , α1 ) + dY (αL , α1 ) − dY (αn , αL )
(αn , αL )α1 =
2
1
dY (αn , v) + dY (v, α1 ) + dY (αL , v) + dY (v, α1 ) − (dY (αn , v) + dY (v, αL ))
≤
2
= dY (v, α1 ) ≤ dY (v, ∂X) + dY (∂X, α1 ) ≤ 1 + dY (∂X, α1 ).
But this contradicts Inequality (2), so we conclude that πX (v) 6= ∅ for all v on γn . We can now
apply Theorems 2.1 and 2.4 to see that for all n ≥ L
dX (µn , µL ) ≤ diamC(X) (πX (µn )) + dX (αn , αL ) + diamC(X) (πX (µL )) ≤ 2 + K0 + 2,
where K0 is an in Theorem 2.4. Therefore,
diamC(X) (πX (µ1 ) ∪ πX (µ2 ) ∪ . . .) ≤ diamC(X) (πX (µ1 ) ∪ . . . ∪ πX (µL )) + 2(K0 + 4) < ∞.
2.7
Hierarchically hyperbolic spaces
In [1] Behrstock, Hagen, and Sisto define the notion of a hierarchically hyperbolic space.
Roughly, a hierarchically hyperbolic space is a quasi-geodesic metric space X , equipped with
additional structure which we will call a hierarchically hyperbolic space (HHS) structure. An
b
HHS structure consists of an index set G and for each W ∈ G a Gromov hyperbolic space CW
b
CW
and a projection map πW : X → 2
. The elements of G and the projection maps must satisfy
a long list of properties. See [1] and [2].
The first example of a hierarchically hyperbolic space is Mod(S), where here G is the
b is the curve graph of W for W ∈ G, and projection πW
collection of all subsurfaces of S, CW
f
is given by composing an orbit map for the action of Mod(S) on M(S)
with the subsurface
projection map defined in Section 2.3. The work of Masur and Minsky [9], [10] and Behrstock [4]
imply that Mod(S) is a hierarchically hyperbolic space. See [2] Section 11 for details. In fact,
the notion of hierarchical hyperbolicity was motivated by a desire to generalize some of the
machinery surrounding mapping class groups.
6
In [1] it is shown that a large class of CAT(0) cube complexes can be equipped with a hierarchically hyperbolic structure, including the universal covers of Salvetti complexes associated
to right-angled Artin groups. The CAT(0) cube complex we are primarily concerned with is
the Cayley graph X of A(Γ) when Γ has no edges (that is, A(Γ) is a free group). We equip
A(Γ) with a hierarchically hyperbolic structure by equipping X with such a structure and then
associating A(Γ) with X.
2.8
Boundary of hierarchically hyperbolic spaces
In [6] the authors construct a boundary for hierarchically hyperbolic spaces. Here we will
describe convergence in this boundary for Mod(S) and for free groups. With the exception of
Theorem 5.3, these will be the only examples we will need.
As a set, the HHS boundary of Mod(S) is defined as follows:
(
∂Mod(S) =
X
cY λY : cY ≥ 0 and λY ∈ ∂C(Y ) for all Y,
Y ⊆S
X
cY = 1,
Y ⊆S
)
0
and if cY 0 , cY > 0, then Y and Y are disjoint or equal .
In [6], the authors define a topology on Mod(S) ∪ ∂Mod(S). In this topology, Definition 2.10
of [6] tells us that a sequence of mapping classes (gn )n∈N in Mod(S) converges to a point
k
k
X
X
ci λi in ∂Mod(S), where ci > 0 for all i,
ci = 1, and λi ∈ ∂C(Yi ) for pairwise disjoint
i=1
i=1
subsurfaces Y1 , . . . , Yk , if and only if the following statements hold: For a fixed marking µ on
S,
1. lim πYi (gn µ) = λi for each i = 1, . . . , k,
n→∞
2. lim
n→∞
ci
dYi (µ, gn µ)
=
for each i, j = 1, . . . , k, and
dYj (µ, gn µ)
cj
dW (µ, gn µ)
= 0 for every (any) i = 1, . . . , k and every subsurface W ⊆ S that is
dYi (µ, gn µ)
disjoint from Yj for all j = 1, . . . , k.
3. lim
n→∞
Let Γ be a graph with no edges, and let A(Γ) be the corresponding free group, equipped
with an HHS structure. The HHS boundary of A(Γ) will be denoted by ∂A(Γ). We do not
define ∂A(Γ) here because Theorem 4.3 in [6] implies that the identity map A(Γ) → A(Γ)
extends to a homeomorphism A(Γ) ∪ ∂G A(Γ) → A(Γ) ∪ ∂A(Γ). Thus, two sequences in A(Γ)
converge to the same point in ∂G A(Γ) if and only if they converge to the same point in ∂A(Γ).
(See Section 2 of [6] for the definition of ∂A(Γ).)
Another useful fact on convergence is that Mod(S) ∪ ∂Mod(S) and A(Γ) ∪ ∂A(Γ) are sequentially compact (see Theorem 3.4 of [6]).
To understand Question 1.1 and the statements of our theorems, one last definition is
needed.
Definition 2.8. Let φ : A(Γ) → Mod(S) be an injective homomorphism and let A(Γ) and
Mod(S) be equipped with any fixed HHS structures. We say that φ extends continuously to a
map ∂A(Γ) → ∂Mod(S) if there exists a function φ : A(Γ) ∪ ∂A(Γ) → Mod(S) ∪ ∂Mod(S) such
that (1) φ|A(Γ) = φ, (2) φ(∂A(Γ)) ⊆ ∂Mod(S), and (3) φ is continuous at each point in ∂A(Γ).
Remark 2.9. To establish that φ : A(Γ) → Mod(S) extends continuously, it is enough to show
that for all x ∈ ∂A(Γ), given any two sequences (xn ) and (yn ) in A(Γ) that converge to x, we
have that (φ(xn )) and (φ(yn )) converge to the same point in ∂Mod(S). This follows from a
diagonal sequence argument (see the end of the proof of Theorem 5.6 in [6] for details).
7
3
Lemmas on subsurface projections
The following lemmas are the heart of our proof of Theorem 1.2.
Lemma 3.1. Suppose X and Y are disjoint subsurfaces of S, and if Y is an annulus, then the
core of Y is not contained in ∂X. If µ and µ0 are markings and f ∈ M od(S) a mapping class
supported on X, then |dY (µ, f (µ0 )) − dY (µ, µ0 )| ≤ 4.
Proof. If Y is not an annulus, then πY (f (µ0 )) = πY (µ0 ) so the claim clearly holds. Assume
then that Y is an annular subsurface of S with core α, and let int(Y ) → S be the associated
covering. Because πX (α) = ∅ and α is not in ∂X, we can find a curve γ in S, distinct from α,
that intersects α and satisfies πX (γ) = ∅. If X is not an annulus, define Z to be the component
of S − X that contains α. If X is an annulus with core β, let Z be the component of S
containing α after removing a small regular neighborhood of β. The neighborhood should be
taken small enough so that γ is contained in Z. Let α
e be the component of the preimage of α
e be the component of the preimage of Z in int(Y ) that
in int(Y ) that is a closed curve. Let Z
contains α
e.
Abusing notation, we let f denote a representative in the isotopy class of f that fixes Z
pointwise. Let fe : int(Y ) → int(Y ) denote the lift of f that fixes a point on α
e, and thus fixes
e pointwise. Let γ
Z
e be a component of the preimage of γ in int(Y ) that intersects α
e. Then γ
e
e
e
is contained in Z, implying that f fixes γ
e pointwise.
It then follows that for β 0 ∈ πY (µ0 ), we have (after replacing fe(β 0 ) with a some representative in its homotopy class) that fe(β 0 ) and β 0 intersect at most once. Thus, by Equation (1)
we have dC(Y ) (fe(β 0 ), β) = 1 + |fe(β 0 ) · β 0 | ≤ 2. Now apply the triangle inequality and Theorem
2.1 to see that
|dY (µ, f (µ0 )) − dY (µ, µ0 )| ≤ dY (µ0 , f (µ0 ))
≤ diamC(Y ) (πY (µ0 )) + dC(Y ) (β 0 , fe(β 0 )) + diamC(Y ) (πY (f (µ0 )))
≤ 1 + 2 + 1 = 4.
Lemma 3.2. Given a homomorphism φ : A(Γ) → Mod(S) and a marking µ on S, there exists
a constant M ≥ 1 such that the following holds. Let y1 . . . yn ∈ A(Γ), where each yi ∈ V (Γ).
Then dW (µ, φ(y1 . . . yn )µ) ≤ M n for all subsurfaces W ⊆ S.
Proof. Define M = 4 max{dM(S)
(µ, φ(x)µ) : x ∈ V (Γ)}. By the triangle inequality and Theof
rem 2.2,
dW (µ, φ(y1 . . . yn )µ) ≤
≤
=
n
X
i=1
n
X
i=1
n
X
dW (φ(y1 . . . yi−1 )µ, φ(y1 . . . yi )µ)
4dM(S)
(φ(y1 . . . yi−1 )µ, φ(y1 . . . yi )µ)
f
4dM(S)
(µ, φ(yi )µ) ≤ M n.
f
i=1
Lemma 3.3. Let φ : A(Γ) → Mod(S) be a homomorphism. Let (gn )n∈N be a sequence of
elements in A(Γ) and µ a marking on S. Suppose for some subsurface W ⊆ S there exist
A,B
constants A ≥ 1 and B ≥ 0, that do not depend on n, such that dW (µ, φ(gn )µ) ||gn ||, where
||gn || denotes the word length of gn with respect to the standard generating set V (Γ) for A(Γ).
8
Further suppose that lim ||gn || = ∞ and that (πW (φ(gn )µ))n∈N converges to some point λW
n→∞
in ∂C(W ). Then all accumulation
points of (φ(gn ))n∈N in Mod(S) ∪ ∂Mod(S) are in ∂Mod(S)
X
and are of the form
cY λY , where cW > 0.
Y ⊆S
Proof. After passing to a subsequence, we may assume that (φ(gn ))n∈N converges. By
assumption, lim dW (µ, φ(gn )µ) = ∞.
Combine this with Theorem 2.2 to see that
n→∞
f
lim dM(S)
(µ, φ(gn )µ) = ∞. Because M(S)
is quasi-isometric to Mod(S) via orbit maps, it
f
n→∞
follows that lim dMod(S) (1, φ(gn )) = ∞. Thus, it must be that lim φ(gn ) ∈ ∂Mod(S).
n→∞
n→∞
X
Suppose lim φ(gn ) =
cY λY for constants cY ≥ 0 and λY ∈ ∂C(Y ). We will now argue
n→∞
Y ⊆S
that cW > 0. Let Z ⊆ S be such that cZ > 0. If W = Z, we are done. So we assume W 6= Z.
By definition of the topology on Mod(S) ∪ ∂Mod(S), we have that lim πZ (φ(gn )µ) = λZ . If
n→∞
W ( Z, then Corollary 2.7 implies that diamC(W ) (πW (φ(g1 )µ) ∪ πW (φ(g2 )µ) ∪ . . .) < ∞. But
this cannot be since πW (φ(gn )µ) → λW ∈ ∂C(W ). Similarly, we cannot have Z ( W for then
Corollary 2.7 implies that diamC(Z) (πZ (φ(g1 )µ) ∪ πZ (φ(g2 )µ) ∪ . . .) < ∞, contradicting that
πZ (φ(gn )µ) → λZ ∈ ∂C(Z). Now suppose that Z t W . Then by Theorem 2.3, after passing to
a subsequence, we have that
dW (∂Z, φ(gn )µ) ≤ 10 for all n,
or
dZ (∂W, φ(gn )µ) ≤ 10 for all n.
If dW (∂Z, φ(gn )µ) ≤ 10 for all n, then for all n
dW (µ, φ(gn )µ) ≤ dW (µ, ∂Z) + dW (∂Z, φ(gn )µ) ≤ dW (µ, ∂Z) + 10,
contradicting that πW (φ(gn )µ) → λW ∈ ∂C(W ). Similarly, if dZ (∂W, φ(gn )µ) ≤ 10 for all n,
then dZ (µ, φ(gn )µ) is bounded independent of n contradicting that πZ (φ(gn )µ) → λZ ∈ ∂C(Z).
So it is not the case that Z t W . Therefore it must be that W and Z are disjoint for all Z ⊆ S
with cZ > 0.
A,B
Fix Z ⊆ S with cZ > 0. Lemma 3.2 together with the fact that dW (µ, φ(gn )µ) ||gn ||
implies that
1
||gn || − B
dW (µ, φ(gn )µ)
≥ A
,
(3)
dZ (µ, φ(gn )µ)
M ||gn ||
where M ≥ 1 is as in Lemma 3.2. Since ||gn || → ∞, Equation (3) implies
dW (µ, φ(gn )µ)
≥ lim
n→∞ dZ (µ, φ(gn )µ)
n→∞
lim
1
A ||gn ||
−B
> 0.
M ||gn ||
Therefore by definition of the topology of Mod(S) ∪ ∂Mod(S), we have cW > 0 as desired.
4
Clay, Leininger, Mangahas RAAGs
In this section, we prove the first part of Theorem 1.2. We begin with a description of a Clay,
Leininger, Mangahas embedding φ : A(Γ) → Mod(S).
Embedding construction: Let Γ be the graph with vertex set V (Γ) = {a, b} and no
edges. Let S = H2 /Λ, Xa , and Xb be the surfaces indicated in Figure 1. For short, let Xab
]
denote Xa ∪ Xb . Let S^
− Xab be a component of the preimage of S − Xab in H2 , and let ∂X
ab
2
^
be a geodesic in H that is in the boundary of S − Xab .
]
Let γ
e be a geodesic in H2 that links with ∂X
ab and maps to a simple bi-infinite geodesic γ
^
in S. Further suppose that γ
e ∩ (S − Xab ) is an infinite ray and let p be its endpoint on ∂H2 .
9
S:
Xb
γ
∂Xab
Xa
Figure 1: Overlapping subsurfaces Xa and Xb of surface S, curve ∂Xab , and
bi-infinite simple geodesic γ.
For example, take γ to be the simple bi-infinite geodesic in S with one end spiraling around
a curve essential in S − Xab and the other end spiraling around a curve in Xa as in Figure 1,
and take γ
e to be an appropriate lift of γ. Choose fb ∈ Mod(S) so that fb is pseudo-Anosov on
Xb . To simplify arguments, we abuse notation and let fb denote a representative in the isotopy
class of fb that fixes all points outside Xb . This ensures that f˜b fixes S^
− Xab pointwise, where
2
2
˜
]
fb : H → H is the lift of fb fixing some point on ∂Xab . Thus, the extension of feb to ∂H2
]
fixes pointwise p and the endpoints x and y of ∂X
ab . Additionally, we choose fb to have the
following properties:
]
1. f˜b (e
γ ) links with h(e
γ ), where h ∈ Λ is a primitive isometry with axis ∂X
ab , and
2. τXb (fb ) ≥ C, where C is as in Theorem 2.5.
We note that a pseudo-Anosov on Xb satisfying (1) can be obtained from any mapping
class that is pseudo-Anosov on Xb by post-composing with some number of Dehn twists (or
inverse Dehn twists) about ∂Xab . Finally, a pseudo-Anosov on Xb satisfying (1) and (2) can
be obtained from one satisfying (1) by passing to a sufficiently high power.
Let fa ∈ Mod(S) be any mapping class that is pseudo-Anosov on Xa and satisfies τXa (fa ) ≥
C. Theorem 2.5 says that the homomorphism
φ : A(Γ) → Mod(S)
defined by
φ(a) = fa , φ(b) = fb
is a quasi-isometric embedding.
Equip A(Γ) with any HHS structure. In the remainder of this section we will prove the
following theorem, which proves the first part of Theorem 1.2.
Theorem 4.1. The sequences (an )n∈N and (an bn )n∈N converge to the same point in ∂A(Γ),
but (φ(an ))n∈N and (φ(an bn ))n∈N do not converge to the same point in ∂Mod(S).
We will divide the proof of Theorem 4.1 into two propositions.
Proposition 4.2. The sequences (an )n∈N and (an bn )n∈N converge to the same point in ∂A(Γ).
Proof. Let X be the Cayley graph of A(Γ). By the discussion in Section 2.8, to show that
(an )n∈N and (an bn )n∈N converge to the same point in ∂A(Γ), it is enough to show that they
converge to the same point in ∂G X. Now the Gromov product
(ai , aj bj )1 = min(i, j) → ∞ as i, j → ∞.
Thus, lim an = lim an bn in ∂G X, as desired.
n→∞
n→∞
Throughout the rest of this section µ will denote a fixed marking on S. To continue, we
require the following lemma.
10
Lemma 4.3. There exist constants A ≥ 1 and B ≥ 0 such that for all n ≥ 1 we have
A,B
d∂Xab (µ, φ(an bn )µ) n. Consequently, after passing to a subsequence, (π∂Xab (φ(an bn )µ))n∈N
converges to a point in ∂C(∂Xab ).
Proof. We begin by establishing the following claim.
n
Claim 1: Let n ≥ 1. Then feb (e
γ ) has endpoint p and links with hi (e
γ ) for all 1 ≤ i ≤ n.
Proof of Claim 1: By our choice of feb and γ
e, we know the claim holds for n = 1. Let n ≥ 2.
n−1
e
Inductively, suppose that fb
(e
γ ) has endpoint p and links with hi (e
γ ) for all 1 ≤ i ≤ n − 1.
2
]
Let I be the interval in ∂H that connects the endpoints of ∂Xab and does not contain p,
oriented from the repelling fixed point of h to the attracting fixed point. We will use interval
notation when speaking about connected subsets of I. Now feb extends continuously to a
homeomorphism of ∂H2 , which we will also denote by feb , and because feb fixes the endpoints
]
of ∂X
e in I,
ab , this extension restricts to a homeomorphism of I. Let z be the endpoint of γ
n−1
i
e
and let x ∈ ∂I be the attracting fixed point of h. Because fb
(e
γ ) links with h (e
γ ) for all
1 ≤ i ≤ n − 1 and has endpoint p, we have
(feb
n−1
(z), x] ⊆ (hi (z), x]
for all
0 ≤ i ≤ n − 1.
(4)
Since feb (e
γ ) has endpoint p and links with h(e
γ ), it must be that feb (z) ∈ (hz, x]. It follows from
e
e
this, the fact that fb and h fix x, that fb and h commute by uniqueness of map lifting, and (4),
that for all 0 ≤ i ≤ n − 1
n
n−1
n−1
n−1
feb (z) = feb
(feb (z)) ∈ feb
(h(z), x] = h(feb
(z), x] ⊆ h(hi (z), x] = (hi+1 (z), x].
(5)
n
n
Because feb fixes p, we have feb (p) = p. This combined with (5) implies that feb (e
γ ) links with
hi+1 (e
γ ) for all 0 ≤ i ≤ n − 1, proving Claim 1.
n
By Claim 1, after replacing feb (e
γ ) with the geodesic connecting its endpoints, the images of
n
2
e
fb (e
γ ) and γ
e in (H −{x, y})/hhi intersect each other at least n times, and all these intersections
have the same sign. Now apply Equation (1) to see that
d∂Xab (γ, φ(bn )γ) ≥ n + 1.
It follows that
d∂Xab (µ, φ(bn )µ) ≥ d∂Xab (γ, φ(bn )γ) − d∂Xab (µ, γ) − d∂Xab (φ(bn )µ, φ(bn )γ)
≥ n + 1 − 2d∂Xab (µ, γ)
(6)
Lemma 3.1 says that |d∂Xab (µ, φ(an bn )µ)−d∂Xab (µ, φ(bn )µ)| ≤ 4. This together with Equation
(6) implies that
d∂Xab (µ, φ(an bn )µ)) n.
From this and that fact that C(∂Xab ) is quasi-isometric to R it is immediate that
(π∂Xab (φ(an bn )µ))n∈N has a subsequence converging to a point in ∂C(∂Xab ).
Proposition 4.4. The sequences (φ(an ))n∈N and (φ(an bn ))n∈N do not converge to the same
point in Mod(S) ∪ ∂Mod(S).
Proof. After passing to a subsequence, we may assume that (φ(an ))n∈N and (φ(an bn ))n∈N
converge to points p and q respectively in Mod(S) ∪ ∂Mod(S) and, by Lemma 4.3, that
(π∂Xab (φ(an bn )µ))n∈N converges to a point in ∂C(∂Xab ). Lemmas 3.3 and 4.3 imply that
11
q is in ∂Mod(S). Say q =
X
cqY λqY , where cqY ≥ 0 and λqY ∈ ∂C(Y ) for all Y ⊆ S. Then
Y ⊆S
Lemmas 3.3 and 4.3 also imply that cq∂Xab > 0.
Now if p were in Mod(S), then
we would be done since clearly then p 6= q. So we will assume
X
that p ∈ ∂Mod(S), and let p =
cpY λpY . Now observe that by Lemma 3.1 and Theorem 2.1
Y ⊆S
d∂Xab (µ, φ(an )µ) ≤ d∂Xab (µ, µ) + 4 ≤ 5.
Thus, (π∂Xab (φ(an )µ))n∈N does not limit to a point on ∂C(∂Xab ). So by definition of the
topology of Mod(S) ∪ ∂Mod(S), it must be that cp∂Xab = 0. Since cq∂Xab > 0, we see that p 6= q,
which completes the proof.
5
Koberda RAAGs
In this section we complete the proof of Theorem 1.2. Following this, we will discuss how to
use similar techniques to prove a large class of Koberda embeddings do not extend.
Let α and β be the pair of intersecting curves on S = H2 /Λ depicted in Figure 2. Let Γ
be the graph with V (Γ) = {a, b} and no edges. For sufficiently large N , Theorem 2.6 says that
the homomorphism
φ : A(Γ) → Mod(S)
defined by φ(a) = Tα N and φ(b) = Tβ N
is injective, where Tα and Tβ denote Dehn twists about α and β respectively. Throughout this
section, we let µ be a fixed marking on S. Equip A(Γ) with an HHS structure.
In this section we prove the following theorem, which will complete the proof of Theorem
1.2.
Theorem 5.1. There exists g ∈ A(Γ) such that the sequences (an )n∈N and (an g n )n∈N converge
to the same point in ∂A(Γ), but (φ(an ))n∈N and (φ(an g n ))n∈N do not converge to the same point
in ∂Mod(S).
As a step towards proving Theorem 5.1, we prove the following lemma in which we construct
g ∈ A(Γ).
Lemma 5.2. There exist constants A ≥ 1 and B ≥ 0 and a word g ∈ A(Γ) such that for all
A,B
n ≥ 1 we have dη (µ, φ(an g n )µ) n, where η is the curve shown in Figure 2. Consequently,
after passing to a subsequence, (πη (φ(an g n )µ))n∈N converges to a point in ∂C(η).
Proof. We will prove that there exist constants c1 , c2 , c3 such that g = bc1 ac2 bc3 has the desired
properties.
e be a component of the preimage of A in H2 . Let βe
Let A be the annulus in Figure 2. Let A
e and let
be a component of the preimage of β such that a segment of βe is in the boundary of A,
e Let h ∈ Λ be a primitive
ηe denote the component of the preimage of η in the boundary of A.
e
isometry with axis ηe. Let α
e be the component of the preimage of α that links with βe and h(β)
e
and contains a segment that is in the boundary of A.
Let Yα be the component of S −α that contain η. To simplify arguments, we let φ(a) denote
g : H2 → H2 be the lift of
a representative in its isotopy class that fixes Yα pointwise. Let φ(a)
φ(a) that fixes some point on α
e. Similarly define Yβ to be the component of S − β containing
g be
η, choose a representative in the isotopy class of φ(b) that fixes Yβ pointwise, and let φ(b)
e It then follows that
the lift of φ(b) that fixes some point on β.
g = 1 on Y
fα
φ(a)
and
12
g = 1 on Y
fβ ,
φ(b)
c3
βe
hβe
c
c
β
g 2 φ(b)
g 3 (e
φ(a)
γ)
g (e
φ(b)
γ)
α
α
e
γ
e
c
c
c
g 3 (e
g 2 φ(b)
g 1 φ(a)
γ)
φ(b)
e
A
ηe
A
η
h(e
γ)
γ
p
spacespacespacespacespaceH2
−−−−−−−−−−−→
S
Figure 2: Curves α, β, and η, bounding an annulus A, and simple bi-infinite
geodesic γ on surface S, and the universal cover H2 of S as in Lemma 5.2.
where for i ∈ {α, β} we let Yei denote the component of the preimage of Yi in H2 whose boundary
g fixes the endpoints of ηe.
contains ei. Observe that for i ∈ {a, b} we have that φ(i)
Choose a geodesic γ
e in H2 that links with both βe and ηe and maps to a simple bi-infinite
fα ∩ Y
fβ is an infinite ray, and let p denote its endpoint
geodesic in S. Further, suppose that γ
e∩Y
on ∂H2 . For example, take γ to be the simple bi-infinite geodesic in S with one end spiraling
around a curve essential in Yα ∩ Yβ and the other end spiraling around a curve essential in
S −Yβ as in Figure 2, and take γ
e to be an appropriate component of the preimage of γ. Observe
g
g
that φ(a) and φ(b) must fix p.
c
c
c
g 3 (e
g 2 φ(b)
g 3 (e
Now choose c3 ∈ Z so that φ(b)
γ ) links with α
e. Then choose c2 ∈ Z so that φ(a)
γ)
c1
c2
c3
g φ(a)
g φ(b)
g (e
e Finally, choose c1 ∈ Z so that φ(b)
links with h(β).
γ ) links with h(e
γ ). See Figure
2.
To simplify notation, define
g = bc1 ac2 bc3 ∈ A(Γ)
and
c
c
c
g = φ(b)
g 1 φ(a)
g 2 φ(b)
g 3.
φ(g)
n
g (e
As in Lemma 4.3, we have that φ(g)
γ ) has endpoint p and links with hi (e
γ ) for all 1 ≤ i ≤ n,
n
implying that dη (γ, φ(g )γ) ≥ n + 1. It follows that
dη (µ, φ(g n )µ) ≥ dη (γ, φ(g n )γ) − dη (µ, γ) − dη (φ(g n )µ, φ(g n )γ) ≥ n + 1 − 2dη (µ, γ).
(7)
Now Lemma 3.1 says that |dη (µ, φ(an g n )µ) − dη (µ, φ(g n )µ)| ≤ 4. This together with Equation
(7) implies that dη (µ, φ(an g n )µ) n. From this and that fact that C(η) is quasi-isometric to R,
it is immediate that (πη (φ(an g n )µ))n∈N has a subsequence converging to a point in ∂C(η).
We can now prove Theorem 5.1.
Proof of Theorem 5.1. Let g ∈ A(Γ) be as in Lemma 5.2. By the discussion in Section 2.8,
to show that (an )n∈N and (an g n )n∈N converge to the same point ∂A(Γ) it is enough to show
that they converge to the same point in ∂G X, where X is the Cayley graph of A(Γ). Now the
Gromov product
(ai , aj g j )1 = (ai , aj (bc1 ac2 bc3 )j )1 = min(i, j) → ∞ as i, j → ∞.
13
Therefore lim an = lim an g n in ∂G X, as desired.
n→∞
n→∞
To finish this proof, we mimic the proof of Proposition 4.4. Replacing b with g, and ∂Xab
with η, and Lemma 4.3 with Lemma 5.2, we find that (φ(an ))n∈N and (φ(an g n ))n∈N do not
converge to the same point in ∂Mod(S).
Our techniques used to prove Theorem 5.1 can be used to prove a more general statement
on non-existence of boundary maps for right-angled Artin groups that are not necessarily free
groups. To prove this more general statement, one needs to understand HHS structures for all
right-angled Artin groups. In the following theorem, by a standard HHS structure on A(Γ),
we mean one induced by a factor system generated by a rich family of subgraphs of Γ. We
refer the reader to [1], specifically Proposition 8.3 and Remark 13.2, for details and to [6] for a
general description of the corresponding HHS boundary. In the proof of the following theorem,
we freely use definitions and notations used in [1] and [6].
Theorem 5.3. Let {α1 , . . . , αk } be any collection pairwise distinct of curves in S. Let Γ be the
graph with V (Γ) = {s1 , . . . , sk } and si sj an edge in Γ if and only and i(αi , αj ) = 0. Give A(Γ)
a standard HHS structure, or if A(Γ) is a free group, any HHS structure. If there exists distinct
intersecting curves αi and αj that do not fill S, then any corresponding Koberda embedding
φ : A(Γ) → Mod(S) does not extend continuously to a map ∂A(Γ) → Mod(S).
Proof. Consider the subgraph Λ of Γ with V (Λ) = {si , sj }. Contained in the Salvetti complex
SΓ associated to Γ there is a subcomplex that is the Salvetti complex associated to A(Λ). We
fΛ denote the lift of this subcomplex to the universal cover S
fΓ of SΓ that contains 1. Let
let S
R be a rich family of induced subgraphs of Γ, and let F be the corresponding factor system in
fΓ . Lemma 8.4 of [1] tells us that
S
fΛ : F ∈ F}
F 0 = {F ∩ S
fΛ . Associating A(Γ) and A(Λ) with S
fΓ and S
fΛ respectively, we equip
is a factor system in S
each with the HHS structures corresponding to their respective factor systems. We first argue
that the inclusion map A(Λ) → A(Γ) extends continuously to a map ∂A(Λ) → ∂A(Γ). If φ
extends continuously to a map ∂A(Γ) → ∂Mod(S), it will follow that A(Λ) → Mod(S) extends
continuously to a map ∂A(Λ) → ∂Mod(S); we will show that this is impossible.
First, consider A(Λ) → A(Γ). Given U ∈ F 0 such that U is not a 0-cube, define π(U )
fΛ . Observe that U
to be the parallelism class of the ⊆-minimal F ∈ F such that U = F ∩ S
and V are nested (respectively orthogonal) if and only if π(U ) and π(V ) are nested (respectively orthogonal). This together with Lemma 10.11 of [6] implies that (A(Λ) → A(Γ), π) is
a hieromorphism. Theorem 5.6 of [6] gives a condition guaranteeing that a hieromorphism
extends continuously. In our case, if the following claims are true, we can apply Theorem 5.6
to conclude that A(Λ) → A(Γ) extends continuously.
Claim 1: π is injective.
Proof of Claim 1: Suppose U, V ∈ F 0 and π(U ) = π(V ). Then π(U ) v π(V ) and π(V ) v π(U ).
Thus, U ⊆ V and V ⊆ U , implying U = V , as desired.
Claim 2: If [F ] ∈ F is not a class of 0-cubes and there exists no U ∈ F 0 satisfying π(U ) = [F ],
f
then diamCF
b (πF (SΛ )) is bounded above uniformly for some (any) F ∈ [F ].
Proof of Claim 2: Let [F ] ∈ F be as in Claim 1. First, suppose there exists F ∈ [F ] such
fΛ 6= ∅. By Lemma 8.5 in [1], we have gF (S
fΛ ) ⊆ F ∩ S
fΛ . If F ∩ S
fΛ is a 0-cube,
that F ∩ S
f
then diamCF
b (πF (SΛ )) ≤ 1, so the claim holds. Otherwise, there must exists F ∈ F such that
f
fΛ . It follows that CF is coned off in CF
b and that gF (S
fΛ ) ⊆ F .
F ( F and F ∩ SΛ = F ∩ S
f
This implies that diamCF
b (πF (SΛ )) ≤ 4.
14
fΛ = ∅ for all F ∈ [F ]. Choose g ∈ A(Γ) and Γ0 ∈ R so that g Sf
Now assume F ∩ S
Γ0 ∈ [F ].
An argument like that in the proof of Proposition 8.3 of [1] shows that
fΛ ) = g(SeΓ0 ∩Λ∩Lkg ) ⊆ g(SeΓ0 ∩Lkg ),
ggSg0 (S
(8)
Γ
fΛ ) = {g}, implying that
where Lkg denotes the link of g. Now if Γ0 ∩ Λ ∩ Lkg = ∅, then ggSg0 (S
Γ
fΛ )) ≤ 1. Assume then that Γ0 ∩ Λ ∩ Lkg 6= ∅. Then by definition of R and
diam b g (π g (S
C(g SΓ0 )
g SΓ0
F, we have that Γ0 ∩ Lkg ∈ R and g(SeΓ0 ∩Lkg ) ∈ F − {0-cubes}. If g(SeΓ0 ∩Lkg ) is not a proper
0
f
f
subcomplex of g Sf
Γ0 , then Γ ⊆ Lkg, implying that SΓ0 is parallel to g SΓ0 (see Lemma 2.4 in [1]).
f
f
e
f
0
0
But this cannot be because SΓ ∩ SΛ = SΓ ∩Λ 6= ∅ and no factor parallel to g Sf
Γ0 intersects SΛ
e
f
e
0
0
0
non-trivially. Therefore, g(SΓ ∩Lkg ) must be a proper subcomplex of g SΓ . Thus, Cg(SΓ ∩Lkg )
b Sf
f
is coned off in C(g
Γ0 ). This together with (8) implies that diamC(g
b S
g0 ) (πg S
g0 (SΛ )) ≤ 4,
Γ
Γ
completing the proof of Claim 2.
We now argue that A(Λ) → Mod(S) does not extend continuously to a map ∂A(Λ) →
∂Mod(S). Let η denote a geodesic representative of an essential boundary component of a
small regular neighborhood of αi ∪ αj . Using the proof techniques of Lemma 5.2, we can
construct g ∈ A(Λ) so that dη (µ, φ(sni g n )µ) grows linearly in n. For later convenience, we
construct g so that when written in reduced form, the first letter of g is s±1
j . As in Proposition
4.4, we see that the sequences (φ(sni )) and (φ(sni g n )) do not converge to the same point in
Mod(S) ∪ ∂Mod(S). Now observe that (sni ) and (sni g n ) converge to the same point in ∂G A(Λ).
Therefore, by the discussion in Section 2.8, (sni ) and (sni g n ) converge to the same point in
∂A(Λ). We have now established that A(Λ) → Mod(S) does not extend continuously to a map
∂A(Λ) → ∂Mod(S). Therefore, A(Γ) → Mod(S) does not extend continuously when A(Γ) is
equipped with a standard HHS structure.
Now suppose A(Γ) is a free group equipped with any HHS structure. Then by the discussion
in Section 2.8, because (sni ) and (sni g n ) converge to the same point in ∂G A(Γ), we have that
(sni ) and (sni g n ) converge to the same point in ∂A(Γ). Because (φ(sni )) and (φ(sni g n )) do not
converge to the same point in ∂Mod(S), it follows that A(Γ) → Mod(S) does not extend
continuously.
6
Existence of boundary maps for some free groups
In this section, we show that a class of embeddings of free groups in Mod(S), which include a
class of Koberda embeddings and a class of Clay, Leininger, and Mangahas embeddings, extend
continuously.
Throughout this section, let Γ be the graph with V (Γ) = {s1 , . . . , sk } and no edges, and let
A(Γ) denote the corresponding right-angled Artin group (a rank k free group). Equip A(Γ) with
an HHS structure. Let {X1 , . . . , Xk } be a collection of pairwise distinct, pairwise overlapping,
and pairwise filling subsurfaces of S and {f1 , . . . , fk } a collection of mapping class such that
fi is fully supported on Xi . Let µ be a fixed marking on S. The main theorem of this section
is the following, which implies the remaining direction of Theorem 1.3 in the introduction.
Theorem 6.1. Let A(Γ) be the rank k free group equipped with any HHS structure. Let
{X1 , . . . , Xk } a collection of pairwise distinct, pairwise overlapping, and pairwise filling subsurfaces of S, and {f1 , . . . , fk } a collection of mapping class such that fi is fully supported on
Xi . There exists a C > 0 such that if τXi (fi ) ≥ C for all i, then the homomorphism
φ : A(Γ) → Mod(S)
defined by
φ(si ) = fi for all i
is a quasi-isometric embedding and extends continuously to a map ∂A(Γ) → ∂Mod(S).
15
We emphasize the arguments we will use to establish that φ is a quasi-isometric embedding
are essentially the same as those used by Clay, Leininger, and Mangahas to prove Theorem
2.5. In particular, when the the Xi are all non-annular, that φ is a quasi-isometric embedding
is Theorem 2.5. To prove Theorem 6.1, we require the following proposition.
Proposition 6.2. There exists K > 0 such that the following holds. For each 1 ≤ i ≤ k,
assume τXi (fi ) ≥ 2K. Let φ : A(Γ) → Mod(S) be the homomorphism defined by φ(si ) = fi for
±1
all i. Consider g1 . . . gk ∈ A(Γ), where for each i we have gi = xei i for some xi ∈ {s±1
1 , . . . , sk }
ek
e1
and ei > 0, and xi 6= xi+1 , and x1 . . . xk is a reduced word. Let Yi be the subsurface of S that
fully supports φ(xi ). Then
1. For each 1 ≤ i ≤ k, we have dφ(g1 ...gi−1 )Yi (µ, φ(g1 . . . gk )µ) ≥ Kei ,
2. For all 1 ≤ i < j ≤ k, we have φ(g1 . . . gi−1 )Yi ≺ φ(g1 . . . gj−1 )Yj , where ≺ denotes the
partial order on Ω(K, µ, φ(g1 . . . gk )µ), and
3. The homomorphism φ : A(Γ) → Mod(S) is a quasi-isometric embedding.
Proof. Define K = K0 + 20 + 2 max{dXi (µ, ∂Xj ) : 1 ≤ i, j ≤ k and i 6= j}, where K0 is
maximum of the constants in Theorem 6.12 of [10] and Theorem 2.4. Statements (1) and (2)
of this proposition are essentially Theorem 5.2 in [5]. The difference is that Theorem 5.2 does
not allow for the homomorphism to send a generator to a power of a Dehn twist. The only
obstruction to Theorem 5.2 holding for homomorphisms φ of this type is the following. Suppose
Xi is the subsurface that fully supports φ(si ) and let σ ∈ A(Γ) be a non-empty word in letters
commuting with si , not including si . If Xi is non-annular, then dXi (φ(σ)µ0 , µ00 ) = dXi (µ0 , µ00 )
for any markings µ0 , µ00 . This not necessarily true if Xi is an annulus. However, this issue does
not arise for us because A(Γ) a free group implies no such σ exists. Thus, the arguments used
to prove Theorem 5.2 in [5] also prove our Statements (1) and (2). The proof of our Statement
(3) is the same as the proof in [5] of Theorem 2.5, using our Statements (1) instead of their
Theorem 5.2.
The proof of the next lemma is essentially contained in the proof of Theorem 6.1 in [5]. We
include a proof here for completeness.
Lemma 6.3. Let φ : A(Γ) → Mod(S), g1 . . . gk ∈ A(Γ), and Yi be as in Proposition 6.2. Let G
be a geodesic in C(S) with one end in πS (µ) and one end in πS (φ(g1 . . . gk )µ). Then for each
1 ≤ i ≤ k, there exists a curve γi on G such that πφ(g1 ...gi−1 )Yi (γi ) = ∅. If |i − j| ≥ 3 and γi
and γj are two such curves, then γi 6= γj .
Proof. Fix 1 ≤ i ≤ k. By way of contradiction, suppose for all curves v on G, we have
πφ(g1 ...gi−1 )Yi (v) 6= ∅. Then Theorem 2.4 and Theorem 2.1 together imply that
dφ(g1 ...gi−1 )Yi (µ, φ(g1 . . . gk )µ) ≤ 4 + K0 .
But Proposition 6.2 says dφ(g1 ...gi−1 )Yi (µ, φ(g1 . . . gk )µ) ≥ K > K0 + 4, a contradiction. Thus,
there must exists a curve γi on G such that πφ(g1 ...gi−1 )Yi (γi ) = ∅, as desired. Note that this
implies that γi and ∂φ(g1 . . . gi−1 )Yi form a multicurve.
Now consider γi and γj , where 1 ≤ i < j ≤ k and |i − j| ≥ 3. We will show that γi and
γj are distinct curves. To the contrary, suppose γi = γj . Because of the filling assumption
on {X1 , . . . , Xk }, the pair of subsurfaces Yi+1 and Yi+2 fill S. Thus, φ(g1 . . . gi+1 )Yi+1 =
φ(g1 . . . gi )Yi+1 and φ(g1 . . . gi+1 )Yi+2 are also a pair of subsurfaces that fill S. Thus, it must
be that πφ(g1 ...gn−1 )Yn (γ) 6= ∅ for some n ∈ {i + 1, i + 2}. In any case, i < n < j.
In the remainder of this proof, to simplify notation, for each ` we define Y` = φ(g1 . . . g`−1 )Y` .
By Proposition 6.2, we have
Yi ≺ Yn ≺ Yj ,
16
where ≺ is the partial order on Ω(K, µ, φ(g1 . . . gk )µ). In particular, these three subsurfaces
are pairwise overlapping. This together with the assumption that γi = γj and Theorem 2.1
implies that
dYn (∂Yi , ∂Yj ) ≤ dYn (∂Yi , γi ) + dYn (γj , ∂Yj ) ≤ 2 + 2 = 4.
It follows from this and the definition of ≺ that
dYn (µ, φ(g1 . . . gk )µ) ≤ dYn (µ, ∂Yi ) + dYn (∂Yi , ∂Yj ) + dYn (∂Yj , φ(g1 . . . gk )µ) ≤ 4 + 4 + 4 = 12.
But this cannot be, because dYn (µ, φ(g1 . . . gk )µ) ≥ K ≥ 20 by Proposition 6.2. Therefore, γi
and γj are distinct curves.
We have now developed the tools we will need to prove Theorem 6.1.
Proof of Theorem 6.1. Define C = 2K, where K is as in Proposition 6.2 and for each 1 ≤ i ≤ k,
assume that τXi (fi ) ≥ C. By Proposition 6.2, φ is a quasi-isometric embedding.
Let X denote the Cayley graph of A(Γ). Choose x ∈ ∂G X. Let γ be the infinite geodesic
ray in X based at 1 limiting to a point x in ∂G X. We think of γ as an infinite word of the
±1
form y1 y2 y3 . . ., where each yi ∈ {s±1
1 , . . . , sk } and the word y1 y2 . . . yi is a reduced word for
all i. By construction, the sequence (y1 . . . yn ) in converges to x in X ∪ ∂G X. Let (hn ) be
another sequence in A(Γ) that converges to x in X ∪ ∂G X. We will show that (φ(hn )) and
(φ(y1 . . . yn )) converge to the same point in ∂Mod(S). By the discussion in Section 2.8, this
will prove the theorem. We will consider two case: (1) There does not exist N ≥ 1 such that
yi = yN for all i ≥ N , and (2) such an N exists. In both cases, we will assume each hn is
en,i
written in the form hn = gn,1 . . . gn,N (n) , where for all i we have gn,i = xn,i
for some en,i > 0
en,N (n)
e,1
±1
±1
and xn,i ∈ {s1 , . . . , sk } satisfying xn,i 6= xn,i+1 , and xn,1 . . . xn,N (n) is a reduced word.
Case 1: Suppose there does not exist N ≥ 1 such that yi = yN for all i ≥ N . Then we
can think of γ as an infinite word of the form g1 g2 g3 . . ., where gi = xei i for some ei > 0 and
e1
ei
±1
xi ∈ {s±1
1 , . . . , sk } satisfying xi 6= xi+1 , and x1 . . . xi is a reduced word for all i. Define Yi
to be the subsurface that fully supports φ(xi ). For short, we let Yi denote φ(g1 . . . gi−1 )Yi .
Because (hn ) and (y1 . . . yn ) converge to the same point in ∂G X and X is a tree, hn and
y1 . . . yn must agree on longer and longer initial segments as n → ∞. In particular, given L ≥ 1,
there exists M such that for all n ≥ M , we have gn,1 . . . gn,L = g1 . . . gL . Consider n ≥ M and
k ≥ e1 + · · · + eL . Choose a curve β ∈ base(µ). Given σ ∈ A(Γ), let G(σ) denote some choice
of geodesic in C(S) with endpoints β and φ(σ)β. By Lemma 6.3, for all 1 ≤ i ≤ L there exist
curves γi and γi0 on G(y1 . . . yk ) and G(hn ) respectively such that πYi (γi ) = ∅ and πYi (γi0 ) = ∅.
Observe that
dS (γi , ∂Yi ) ≤ 1 and dS (γi0 , ∂Yi ) ≤ 1.
Choose γr to be the curve in {γi : 1 ≤ i ≤ L} closest to φ(y1 . . . yk )β. Lemma 6.3 tells us
that if |i − j| ≥ 3, then γi 6= γj . So necessarily dS (β, γr ) ≥ L/3. Thus, the Gromov product,
computed in C(S),
1
dS (β, φ(y1 . . . yk )β) + dS (β, φ(hn )β) − dS (φ(y1 . . . yk )β, φ(hn )β)
(φ(y1 . . . yk )β, φ(hn )β)β =
2
1
≥
dS (β, γr ) + dS (γr , φ(y1 . . . yk )β) + dS (β, γr0 ) + dS (γr0 , φ(hn )β)−
2
0
0
dS (φ(y1 . . . yk )β, γr ) + dS (γr , ∂Yr ) + dS (∂Yr , γr ) + dS (γr , φ(hn )β)
1
0
≥
dS (β, γr ) + dS (β, γr ) − 2
2
1
≥ (L/3 − 2).
2
17
It follows that
lim inf (φ(y1 . . . yk )β, φ(hn )β))β = ∞.
(9)
k,n→∞
Because (hn ) is an arbitrary sequence converging to x, we could have taken it to be (y1 . . . yn ).
Thus, Equation (9) tells us two things: (1) (φ(y1 . . . yn )µ) converges to a point in ∂C(S), and
(2) (φ(y1 . . . yn )µ) and (φ(hn )µ) converge to the same point in ∂C(S). By definition of the
topology on Mod(S) ∪ ∂Mod(S), this tells us that (φ(y1 . . . yn )) and (φ(hn )) converge to the
same point in ∂Mod(S).
Case 2: Assume there exists N ≥ 1 such that yi = yN for all i ≥ N . Corollary
6.2 in [6] tells us that the action of Mod(S) by left multiplication extends to an action of
Mod(S) on Mod(S) ∪ ∂Mod(S) by homeomorphisms. Consequently, if we can show that
(φ((y1 . . . yN −1 )−1 hn ))n∈N and (φ(yN . . . yn ))n∈N converge to the same point in ∂Mod(S), then
(φ(hn ))n∈N and (φ(y1 . . . yn ))n∈N must converge to the same point in ∂Mod(S). Furthermore,
((y1 . . . yN −1 )−1 hn )n∈N and (yN . . . yn )n∈N converge to the same point in ∂G X. Thus, without
loss of generality we assume N = 1. By our assumption, y1 . . . yn = y1n for all n.
Let Y be the subsurface that fully supports φ(y1 ) and let ∂Y = {β1 , . . . , β` }. Then
dY (µ, φ(y1n )µ)
>0
n→∞
n
lim
and
πY (φ(y1n )µ) → λY for some λY ∈ ∂C(Y ).
(10)
Further observe that for all i
dβi (µ, φ(y1n )µ)
≥ 0.
(11)
n→∞
n
If (11) is an equality, let λi be any point in ∂C(βi ). Otherwise, define λi ∈ ∂C(βi ) to be
lim πβi (φ(y1n )µ). For all subsurfaces W disjoint from Y and not an annulus with core curve in
lim
n→∞
∂Y , Lemma 3.1 and Theorem 2.1 imply that dW (µ, φ(y1n )µ) ≤ dW (µ, µ) + 4 ≤ 6. Consequently,
lim φ(y1n ) = cY λY +
n→∞
`
X
i=1
ci λi , where cY +
`
X
i=1
ci = 1 and
dβi (µ, φ(y1n )µ)
ci
= lim
.
n→∞ dY (µ, φ(y n )µ)
cY
1
(y1n )
Because (hn ) and
converge to the same point in ∂G X, given any L ≥ 1, for all
sufficiently large n we have xn,1 = y1 and en,1 ≥ L. So by removing finitely many initial terms
e
from hn , for convenience we may assume that gn,1 = y1n,1 for all n. Observe that en,1 → ∞ as
n → ∞. It is immediate from this and the definition of the topology of Mod(S)∪∂Mod(S) that
lim φ(gn,1 ) = lim φ(y1n ). Thus, to finish the proof, we must show lim φ(gn,1 ) = lim φ(hn ).
n→∞
n→∞
n→∞
n→∞
By passing to subsequences, we may assume that either N (n) = 1 for all n or N (n) ≥ 2 for all
n. If the former holds, then hn = gn,1 , and we are done. Assume then that N (n) ≥ 2 for all n.
To proceed, we require the following claims.
Claim 1: dY (φ(gn,1 )µ, φ(hn )µ) is bounded above, independent of n.
Claim 2: Let W be a subsurface that is disjoint from Y . Then dW (φ(gn,1 )µ, φ(hn )µ) is
bounded above, independent of n.
We postpone the proofs of these claims and for now assume they are true. First, observe
that Claim 1 and (10) imply that πY (φ(hn )µ) → λY . If Inequality (11) is strict, then Claim
2 implies that πβi (φ(hn )µ) → λi . Further observe that Claims 1 and 2 imply that for all W
disjoint from Y
dW (µ, φ(gn,1 )µ)
=
lim
n→∞ dY (µ, φ(gn,1 )µ)
dW (µ, φ(gn,1 )µ)
dW (µ, φ(hn )µ)
lim
n→∞
dW (µ, φ(hn )µ)
en,1
en,1
=
= lim
.
n→∞ dY (µ, φ(hn )µ)
dY (µ, φ(gn,1 )µ)
dY (µ, φ(hn )µ)
lim
lim
n→∞
n→∞
en,1
en,1
lim
n→∞
It follows that lim φ(gn,1 ) = lim φ(hn ) as desired.
n→∞
n→∞
To finish the proof, we will now prove Claims 1 and 2. For each n, let Zn denote the
subsurface that fully supports φ(xn,2 ).
18
Proof of Claim 1: Fix n ≥ 1. Because Y fully supports φ(xn,1 ), by Proposition 6.2, we know
Y ≺ φ(gn,1 )Zn , where ≺ denotes the partial order on Ω(K, µ, φ(hn )µ).
Thus,
dY (∂φ(gn,1 )Zn , φ(hn )µ) ≤ 4. Therefore,
dY (φ(gn,1 )µ, φ(hn )µ) ≤ dY (φ(gn,1 )µ, ∂φ(gn,1 )Zn ) + dY (∂φ(gn,1 )Zn , φ(hn )µ) ≤ dY (µ, ∂Zn ) + 4.
There are finitely many possibilities for Zn , so this completes the proof of Claim 1.
Proof of Claim 2: Fix n ≥ 1. Because Y and Zn fill S and Y and W are disjoint, it must be
that πZn (∂W ) 6= ∅. There are two cases to consider: (1) W t Zn and (2) W ( Zn . First,
suppose that W t Zn . It then follows from Proposition 6.2, Theorem 2.1, and the definition
of K that
dZn (∂W, φ(gn,2 . . . gn,N (n) )µ) ≥ dZn (µ, φ(gn,2 . . . gn,N (n) )µ) − dZn (∂Y, ∂W ) − dZn (µ, ∂Y )
≥ K − 2 − K/2 ≥ 10.
Thus Theorem 2.3 implies that dW (∂Zn , φ(gn,2 . . . gn,N (n) )µ) ≤ 4. From this and Theorem 2.2
we find that
dW (φ(gn,1 )µ, φ(hn )µ) = dW (µ, φ(gn,2 . . . gn,N (n) )µ)
≤ dW (µ, ∂Zn ) + dW (∂Zn , φ(gn,2 . . . gn,N (n) )µ)
(µ, µi ) : 1 ≤ i ≤ k} + 4,
≤ 4 max{dM(S)
f
(12)
where µi is a fixed choice of marking with ∂Xi ⊆ base(µi ) for each 1 ≤ i ≤ k. This provides a
uniform bound in the case that W t Zn .
Now suppose that W ( Zn . First, observe that because Zn fully supports φ(xn,2 ), the
sequence (πZn (φ(xn,2 )m µ))m∈N converges to a point in ∂C(Zn ). Thus, by Corollary 2.7 there
exists a constant M , that depends on W and xn,2 , such that dW (µ, φ(gn,2 )µ) ≤ M for all
n. Note that there are only finitely many possibilities for xn,2 , so M can be chosen to be
independent of n. This implies that
dW (φ(gn,1 )µ, φ(hn )µ) ≤ dW (µ, φ(gn,2 . . . gn,N (n) )µ)
= dW (µ, φ(gn,2 )µ) + dW (φ(gn,2 )µ, φ(gn,2 . . . gn,N (n) )µ)
≤ M + dφ(gn,2 )−1 W (µ, φ(gn,3 . . . gn,N (n) )µ).
Now if N (n) = 2, then we can apply Theorem 2.1 to see that
dφ(gn,2 )−1 W (µ, φ(gn,3 . . . gn,N (n) )µ) = dφ(gn,2 )−1 W (µ, µ) ≤ 2,
and Claim 2 is established. Suppose then that N (n) ≥ 3. Let Vn denote the subsurface that
fully supports φ(xn,3 ). Observe that because τZn (φ(xn,2 )) ≥ 2K and ∂Y and ∂W form a
multicurve, we have
dZn (∂φ(gn,2 )−1 W, ∂Vn ) ≥ dZn (∂W, ∂φ(gn,2 )−1 W ) − dZn (∂W, ∂Y ) − dZn (µ, ∂Y ) − dZn (µ, ∂Vn )
≥ 2K − 2 − K/2 − K/2 > 2.
This together with Theorem 2.2 establishes that ∂φ(gn,2 )−1 W and ∂Vn do not form a multicurve. Thus, φ(gn,2 )−1 W t Vn . So to bound dφ(gn,2 )−1 W (µ, φ(gn,3 . . . gn,N (n) )µ) from above independent of n, we can use the same techniques used above to bound dW (µ, φ(gn,2 . . . gn,N (n) )µ)
when W t Zn . This completes the proof of Claim 2.
References
[1] Jason Behrstock, Mark F. Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces
I: curve complexes for cubical groups. arXiv:1412.2171v3, 2014.
19
[2] Jason Behrstock, Mark F. Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces
II: Combination theorems and the distance formula. arXiv:1509.00632v2, 2014.
[3] Jason Behrstock, Bruce Kleiner, Yair Minsky, and Lee Mosher. Geometry and rigidity of
mapping class groups. Geom. Topol., 16(2):781–888, 2012.
[4] Jason A. Behrstock. Asymptotic geometry of the mapping class group and Teichmüller
space. Geom. Topol., 10:1523–1578, 2006.
[5] Matt T. Clay, Christopher J. Leininger, and Johanna Mangahas. The geometry of rightangled Artin subgroups of mapping class groups. Groups Geom. Dyn., 6(2):249–278, 2012.
[6] Matthew G. Durham, Mark F. Hagen, and Alessandro Sisto. Boundaries and automorphisms of hierarchically hyperbolic spaces. arXiv:1604.01061v2, 2016.
[7] Thomas Koberda. Right-angled Artin groups and a generalized isomorphism problem for
finitely generated subgroups of mapping class groups. Geom. Funct. Anal., 22(6):1541–
1590, 2012.
[8] Johanna Mangahas. A recipe for short-word pseudo-Anosovs.
135(4):1087–1116, 2013.
Amer. J. Math.,
[9] Howard A. Masur and Yair N. Minsky. Geometry of the complex of curves. I. Hyperbolicity.
Invent. Math., 138(1):103–149, 1999.
[10] Howard A. Masur and Yair N. Minsky. Geometry of the complex of curves. II. Hierarchical
structure. Geom. Funct. Anal., 10(4):902–974, 2000.
20
| 4math.GR
|
Structure and Interpretation of Computer Programs∗
arXiv:0803.4025v1 [cs.SE] 27 Mar 2008
Ganesh Narayan, Gopinath K
Computer Science and Automation
Indian Institute of Science
{nganesh, gopi}@csa.iisc.ernet.in
Abstract
Call graphs depict the static, caller-callee relation between “functions” in a program. With most source/target
languages supporting functions as the primitive unit of composition, call graphs naturally form the fundamental control
flow representation available to understand/develop software. They are also the substrate on which various interprocedural analyses are performed and are integral part
of program comprehension/testing. Given their universality
and usefulness, it is imperative to ask if call graphs exhibit
any intrinsic graph theoretic features – across versions, program domains and source languages. This work is an attempt to answer these questions: we present and investigate
a set of meaningful graph measures that help us understand
call graphs better; we establish how these measures correlate, if any, across different languages and program domains; we also assess the overall, language independent
software quality by suitably interpreting these measures.
1
Introduction
Complexity is one of the most pertinent characteristics
of computer programs and, thanks to Moore’s law, computer programs are becoming ever larger and complex; it’s
not atypical for a software product to contain hundreds of
thousands, even millions of lines of code where individual
components interact in myriad of ways. In order to tackle
such complexity, variety of code organizing motifs were
proposed. Of these motifs, functions form the most fundamental unit of source code: software is organized as set
of functions – of varying granularity and utility, with functions computing various results on their arguments. Critical
feature of this organizing principle is that functions themselves can call other functions. This naturally leads to the
notion of function call graph where individual functions are
nodes, with edges representing caller-callee relations; in∗ In
reverence to the Wizard Book.
Sridhar V
Applied Research Group
Satyam Computers
[email protected]
degree depicts the number of functions that could call the
function and outdegree depicts the number of functions that
this function can call. Since no further restrictions are employed, the caller-callee relation induces a generic graph
structure, possibly with loops and cycles.
In this work we study the topology of such (static) call
graphs. Our present understanding of call graphs is limited;
we know: that call graphs are directed and sparse; can have
cycles and often do; are not strongly connected; evolve over
time and could exhibit preferential attachment of nodes and
edges. Apart from these basic understanding, we do not
know much about the topology of call graphs.
2
Contributions
In this paper we answer questions pertaining to topological properties of call graphs by studying a representative set
of open-source programs. In particular, we ask following
questions: What is the structure of call graphs? Are there
any consistent properties? Are some properties inherent to
certain programming language/problem class? In order to
answer these questions, we investigate set of meaningful
metrics from plethora of graph properties [9]. Our specific
contributions are:
1) We motivate and provide insights as to why certain
call graph properties are useful and how they could help us
develop better and robust software. 2) We compare graph
structure induced by different language paradigms under an
eventual but structurally immediate structure – call graphs.
The authors are unaware of any study that systematically
compare the call graphs of different languages; in particular,
the “call graph” structure of functional languages. 3) Our
corpus, being varied and large, is far more statistically representative compared to the similar studies ([24], [4],[18]).
4) We, apart from confirming previous results in a rigorous
manner, also compute new metrics to capture finer aspects
of graph structure. 5) As a side effect, we provide a potential means to assess software quality, independent of the
source language.
Rest of the paper is organized as follows. We begin by
justifying the utility of our study and proceed to introduce
relevant structural measures in section 4. Section 5 discusses the corpus and methodology. We then present our
the measurements and interpretations (Section 6). We conclude with section 7 and 8.
3
Motivation
Call graphs define the set of permissible interactions and
information flows and could influence software processes in
non trivial ways. In order to give the reader an intuitive understanding as to how graph topology could influence software processes, we present following four scenarios where
it does.
Bug Propagation Dynamics Consider how a bug in
some function affects the rest of the software. Let foo call
bar and bar could return an incorrect value because of a
bug in bar. if foo is to incorporate this return value in
its part of computation, it is likely to compute wrong answer as well; that is, bar has infected foo. Note that such
an infection is contagious and, in principle, bar can infect
any arbitrary function fn as long as fn is connected to bar.
Thus connectedness as graph property trivially translates to
infectability. Indeed, with appropriate notions of infection
propagation and immunization, one could understand bug
expression as an epidemic process. It is well known that
graph topology could influence the stationary distribution
of this process. In particular, the critical infection rate – the
infection rate beyond which an infection is not containable
– is highly network specific; in fact, certain networks are
known to have zero critical thresholds [5]. It pays to know
if call graphs are instances of such graphs.
Software Testing: Different functions contribute differently to software stability. Certain functions that, when
buggy, are likely to render the system unusable. Such functions, functions whose correctness is central to statistical
correctness of the software, are traditionally characterized
by per-function attributes like indegree and size. Such simple measure(s), though useful, fail to capture the transitive dependencies that could render even a not-so-well connected function an Achilles heel. Having unambiguous metrics that measure a node’s importance helps making software testing more efficient. Centrality is such a measure
that gives a node’s importance in a graph. Once relevant
centrality measures were assigned, one could expend relatively more time testing central functions. Or, equally,
test central functions and their called contexts for prevalent error modes like interface nonconformity, context disparity and the likes ([21], [7]). By considering node centralities, one could bias the testing effort to achieve similar
confidence levels without a costlier uniform/random testing
schedule; though most developers intuitively know the im-
portance of individual functions and devise elaborate test
cases to stress these functions accordingly, we believe such
an idiosyncratic methodology could be safely replaced by
an informed and statistically tenable biasing based on centralities. Centrality is also readily helpful in software impact
analysis.
Software Comprehension: Understanding call graph
structure helps us to construct tools that assist the developers in comprehending software better. For instance, consider a tool that magically extracts higher-level structures
from program call graph by grouping related, lower-level
functions. Such a tool, for example, when run on a kernel
code base, would automatically decipher different logical
subsystems, say, networking, filesystem, memory management or scheduling. Devising such a tool amounts to finding
appropriate similarity metric(s) that partitions the graph so
that nodes within a partition are “more” similar compared
to nodes outside. Understandably, different notions of similarities entail different groupings. Recent studies show how
network structure controls such grouping [2] and how per
node graph metrics can be used to improve the developerperceived clustering validity ([26], [17]).
Inter Procedural Analysis Call graph topology could
influences both precision and convergence of Inter Procedural Analysis (IPA). When specializing individual procedures
in a program, procedures that have large indegree could end
up being less optimal: dataflow facts for these functions
tend to be too conservative as they are required to be consistent across a large number of call sites. By specifically
cloning nodes with large indegree and by distributing the
indegrees “appropriately” between these clones, one could
specialize individual clones better. Also, number of iterations an iterative IPA takes compute a fixed-point depends
on the max(longest path length, largest cycle).
4
Statistical Properties of Interest
As with most nascent sciences, graph topology literature is strewn with notions that are overlapping, correlated
and misused gratuitously; for clarity, we restrict ourselves
to following structural notions. A note on usage: we employ graphs and networks interchangeably; G = (V, E),
| V |= n and | E |= m; (i, j) implies i calls j; di denotes
the degree of vertex i and dij denotes the geodesic distance
between i and j; N (i) denotes the immediate neighbours
of i; graphs are directed and simple: for every (i1 , j1 ) and
(i2 , j2 ) present, either (i1 6= i2 ) or (j1 6= j2 ) is true.
Graphs, in general, could be modeled as random, small
world, power-law, or scale rich, each permitting different
dynamics.
Random graphs: random graph model [11], is perhaps
the simplest network model: undirected edges are added
at random between a fixed number n of vertices to create
a network in which each of the 12 n(n − 1) possible edges
is independently present with some probability p, and the
vertex degree distribution follows Poisson in the limit of
large n.
Small world graphs: exhibit high degree of clustering
−1
and have
=
Pmean−1geodesic distance ` – defined as, `
1
d
–
in
the
range
of
log
n;
that
is,
number
i6=j ij
n(n+1)
of vertices within a distance r of a typical central vertex
grows exponentially with r [19].
It should be noted that a large number of networks, including random networks, have ` in the range of log n or,
even, log log n. In this work, we deem a network to be
small world if ` grows sub logarithmically and the network
exhibits high clustering.
Power law networks: These are networks whose degree
distribution follow the discrete CDF: P [X > x] ∝ cx−γ ,
where c is a fixed constant, and γ is the scaling exponent.
When plotted as a double logarithmic plot, this CDF appears as a straight line of slope −γ. The sole response of
power-law distributions to conditioning is a change in scale:
for large values of x, P [X > x|X > Xi ] is identical to
the (unconditional) distribution P [X > x]. This “scale invariance” of power-law distributions is attributed as scalefreeness. Note that this notion of scale-freeness does not
depict the fractal-like self similarity in every scale.
Graphs with similar degree distributions differ widely
in other structural aspects; rest of the definitions introduce
metrics that permit finer classifications.
degree correlations: In many real-world graphs, the
probability of attachment to the target vertex depends
also on the degree of the source vertex: many networks
show assortative mixing on their degrees, that is, a preference for high-degree nodes to attach to other highdegree node; others show disassortative mixing where
high-degree nodes consistently attach to low-degree ones.
Following measure, a variant of Pearson correlation coefficientP[20], givesP the degree correlation.
ρ =
1
(ji +ki )]2
m−1
ji ki −[m−1
i
i 2
P
P
, where ji , ki are the
1
1
2
2
−1
−1
2
m
(ji
i 2
+ki )−[m
(ji +ki )]
i 2
degrees of the vertices at the ends of ith edge, with i =
1 · · · m. ρ takes values in the range −1 ≤ ρ ≤ 1, with
ρ > 0 signifying assortativity and ρ < 0 signifying dissortativity. ρ = 0 when there is no discernible correlation
between degrees of nodes that share an edge.
scale free metric: a useful measure capturing the fractal
nature ofP
graphs is scale-free metric s(g) [16], defined as:
s(g) =
(i,j)∈E di dj , along with its normalized variant
S(g) = ss(g)
; smax is the maximal s(g) and is dictated by
max
the type of network understudy1 . Rest of the paper will use
the normalized variant.
s(g) is maximal when nodes with similar degree conPn
1
2
For unrestricted graphs, smax =
i=1
(di /2).di .
nect to each other [13]; thus, S(g) is close to one for networks that are fractal like, where the connectivity, at all degrees, stays similar. On the other hand, in networks where
nodes repeatedly connect to dissimilar nodes, S(g) is close
to zero. Networks that exhibit power-law, but have have a
scale free metric S(g) close to zero are called scale rich;
power-law networks whose S(g) value is close to one are
called scale-free. Measures S(g) and ρ are similar and are
correlated; but they employ different normalizations and are
useful in discerning different features [16].
clustering coefficient: is a measure of how clustered,
or locally structured, a graph is: it depicts how, on an average, interconnected each node’s neighbors are. Specifically,
if node v has kv immediate neighbors, then the clustering
coefficient for that node, Cv , is the ratio of number of edges
present between its neighbours Ev to the total possible connections between v’s neighbours, that is, kv (kv − 1)/2. The
whole graph clusteringDcoefficient,
E C, is the average of Cv s:
2Ev
that is, C = hCv iv = kv (kv −1) .
v
clustering profile: C has limited use when immediate connectivity is sparse. In order to understand interconnection profile of transitively connected
neighbours, we
P
use clustering profile [1]: Ckd =
d
C (i) =
{i|di =k}
C d (i)
|{i|di =k}|
|{(j,k);j,k∈N (i)|djk ∈G(V \i)=d}|
, where
. Note that, by this
(|N2(i)|)
definition, clustering coefficient C is simply Ck1 ,
centrality: of a node is a measure of relative importance
of the node within the graph; central nodes are both points
of opportunities – that they can reach/influence most nodes
in the graph, and of constraints – that any perturbation in
them is likely to have greater impact in a graph. Many centrality measures exist and have been successfully used in
many contexts ([6], [10]). Here we focus on betweenness
centrality Bu (of node u), defined as the ratio of number of
geodesic paths that pass through the node (u) to that of the
P
total number of geodesic paths: that is, Bu = ij σ(i,u,j)
σ(i,j) ;
nodes that occur on many shortest paths between other vertices have higher betweenness than those that do not.
connected components: size and number of connected
components gives us the macroscopic connectivity of the
graph. In particular, number and size of strongly connected
components gives us the extent of mutual recursion present
in the software. Number of weakly connected component
gives us the upper bound on amount of runtime indirection
resolutions possible.
edge reciprocity: measures if the edges are reciprocal,
that is, if (i, j) ∈ E, is (j, i) also ∈ E? A robust measure for reciprocity is defined as [12]: ρ = %−ā
1−ā where
P
aij aji
% = ijm
and ā is mean of values in adjacency matrix.
This measure is absolute: ρ greater than zero imply larger
reciprocity than random networks and ρ less than zero im-
InDegree Distribution
OutDegree Distribution
10000
100000
C.Linux
C++.Coin
OCaml.Coq
Haskell.Yarrow
C.Linux
C++.Coin
OCaml.Coq
Haskell.Yarrow
10000
1000
Frequency
Frequency
1000
100
100
10
10
1
1
1
10
100
1000
1
10
InDegree
Figure 1. Indegree Distribution
ply smaller reciprocity than random networks.
5
100
OutDegree
Figure 2. Outdegree Distribution
connected components for our measurements. Component
statistics were computed for the whole data set.
Corpora & Methodology
6
We studied 35 open source projects. The projects are
written in four languages: C, C++, OCaml and Haskel.
Appendix 8 enlists these software, their source language,
versiom, domain and size: number of nodes N and the
number of edges M . Most programs used are large, used
by tens of thousands of users, written by hundreds of developers and were developed over years. These programs
are actively developed and supported. Most of these programs – from proof assistant to media player, provide varied functionalities and have no apparent similarity or overlap in usage/philosophy/developers; if any, they exhibit
greater orthogonality: Emacs Vs Vim, OCaml Vs GCC,
Postgres Vs Framerd, to name a few. Many are stand-alone
programs while few, like glibc and ffmpeg, are provided
as libraries. Some programs, like Linux and glibc, have
machine-dependent components while others like yarrow
and psilab are entirely architecture independent.
In essence, our sample is unbiased towards applications,
source languages, operating systems, program size, program features and developmental philosophy. The corpus
versions and age vary widely: some are few years old while
others, like gcc, Linux kernel and OCamlc, are more than a
decade old. We believe that any invariant we find in such a
varied collection is likely universal.
We used a modified version of CodeViz [27] to extract
call graphs from C/C++ sources. For OCaml and Haskell,
we compiled the sources to binary and used this modified
CodeViz to extract call graph from binaries. OCaml programs were compiled using ocamlopt while for Haskell we
used GHC. A note of caution: to handle Haskell’s laziness,
GHC uses indirect jumps. Our tool, presently, could handle
such calls only marginally; we urge the reader to be mindful
of measures that are easily perturbed by edge additions.
We used custom developed graph analysis tools to measure most of the properties; where possible we also used
the graph-tool software [14]. We used the largest weakly
Interpretation
In the following section we walk through the results, discuss what these results mean and why they are of interest to
language and software communities. Note that most plots
have estimated sample variance as the confidence indicator.
Also, most graphs run a horizontal line that separates data
from different languages.
Degree Distribution: Fitting samples to a distribution
is impossibly thorny: any sample is finite, but of the distributions there are infinitely many. Despite the hardness of
this problem, many of the previous results were based either
on visual inspection of data or on linear regression, and are
likely to be inaccurate [8].
We use cumulative distribution to fit the data and we
compute the likelihood measures for other distributions in
order to improve the confidence using [8]. Figures 1 and
2 depict how four programs written in four different language paradigms compare; the indegree distribution permits
power-law (2.3 ≤ γ '≤ 2.9) while the outdegree distribution permits exponential distribution (Haskell results are
coarse, but are valid). This observation, that in and out degree distributions differ consistently across languages, is expected as indegree and outdegree are conditioned very differently during the developmental process.
Outdegree has a strict budget; large, monolithic functions are difficult to read and reuse. Thus outdegree is minimized on a local, immediate scale. On the other hand,
large indegree is implicitly encouraged, up to a point; indegree selection, however, happens in a non-local scale, over
a much larger time period; usually backward compatibility permits lazy pruning/modifying of such nodes. Consequently one would expect the variability of outdegree – as
depicted by the length of the errorbar, to be far less compared to that of the indegree. This is consistent with the
observation (Fig. 3). Note that the tail of the outdegree is
prominent in OCaml and C++: languages that allow highly
1000
Average Degree
Epidemic Threshold
HaXml
DrIFT
Frown
yarrow
ott
psilab
glsurf
fftw
coqtop
ocamlc
knotes
doxygen
coin
cgal
kcachegrind
cccc
gnuplot
openssl
sim
vim70
emacs
framerd
postgresql
gimp
fvwm
gcc
MPlayer
glibc
ffmpeg
sendmail
bind
httpd
linux
scheme
110
idegree
odegree
Largest Eigen Value of the Adjacency Matric
100
90
80
70
60
50
40
30
C
C++
OCaml
Haskell
20
10
0
1
2
3
Average Degree
4
5
6
0
2000
4000
6000
8000
10000
12000
14000
16000
Graph/Matrix Size - N
Figure 3. Average Degree
Figure 4. Epidemic Threshold Vs N
stylized call composition.
Such observations are critical as distributions portend the
accuracy of sample estimates. In particular, such distributions as power-law that permits non-finite mean and variance – consequently eluding central limit theorem, are very
poor candidates for simple sampling based analyses; understanding the degree distribution is of both empirical and theoretical importance.
Consider the bug propagation process delineated in Section 3. Assuming that the inter-node bug propagation is
Markovian, we could construct an irreducible, aperiodic,
finite state space Markov chain (not unlike [6]) with bug
introduction rate β and debugging (immunization) rate δ as
parameters. Note that this Markov chain has two absorbing states: all-infected or all-cured. Equipped with these
notions, we could ask what is the minimal critical infection rate βc beyond which no amount of immunization will
help to save the software; below βc the system exponentially converges to the good, all-cured absorbing state. It is
known that for a sufficiently large power-law network with
exponent in the range 2 < γ ≤ 3, βc is zero [5]. Thus
one is tempted to conclude that, provided Markovian assumption holds, it is statistically impossible to construct an
all-reliable program. However that would be inaccurate as
the sum of indegree and outdegree distribution2 indegree
and outdegree need not follow power-law. However a recent study [25] establishes that, for finite networks, βc is
bounded by the spectral diameter of the graph; in partic1
ular, βc = λ1,A
, where λ1,A is the largest eigenvalue of
the adjacency matrix. Figure 4 depicts the relation between
λ1,A and the graph size, n. For a “robust” software, we require βc to be large, or equally, λ1,A to be small. However,
it is evident from the plot that larger the graph, higher the
λ1,A . This trend is observed uniformly across languages.
Thus, we are to conclude that large programs tend to be
more fragile, confirming the established wisdom. Another
equally important inference one can make from the indegree distribution is that uniform fault testing is bound to fail:
should one is to build a statistically robust software, testing
efforts ought to be heavily biased. These two inferences
align closely with the common wisdom, except that these
inferences are rigorously established (and party explained)
using the statistical nature of call graphs.
Scale Free Metric: Fig. 5 shows how scale-free metric
for symmetrized call graphs vary with different programs.
Two observations are critical: First, S(g) is close to zero.
This implies call graphs are scale-rich and not scale-free.
This is of importance because in a truly scale-free networks,
epidemics are even harder to handle; hubs are connected to
hubs and the Markov chain rapidly converged to the allinfected absorption state. In scale-rich networks, as hubs
tend to connect to lesser nodes, the rate of convergence is
less rapid. Second, S(g) appears to be language independent3 . Both near zero and higher S(g)s appear in all languages. Thus call graphs, though follow power-law for indegree, are not fractal like in the self-similarity sense.
Degree Correlation: Fig 7 show how input-input (i-i)
and output-output (o-o) degrees correlate with each other.
These sets are weakly assortative, signifying hierarchical
organization.
But finer picture evolves as far as languages are concerned. C programs appears to have very similar i-i and
o-o profiles with o-o correlation being smaller and comparable to i-i correlation. In addition, C’s correlation measure
is consistently less than that of other languages and is close
to zero; thus, C programs exhibit as much i-i/o-o correlation
as that of a random graph of similar size. In other words,
if foo calls bar, the number of calls bar makes is independent of the number of calls foo made; this implies less
hierarchical program structure as one would like the level
n functions to receive fewer calls compared to level n − 1
functions. For instance, variance(list) is likely to receive fewer calls compared to sum(list); we would also
like level n functions to have higher outdegree compared
to level n − 1 functions. Thus, in a highly hierarchical design, i-i and o-o correlations would be mildly assortative,
with i-i being more assortative. For C++, i-i and o-o dif-
2 Bug propagation is symmetric: foo and bar can pass/return bugs to
one another.
3 Except
Haskell; but this could be an artifact of edge limited sample.
18000
0.25
HaXml
HaXml
DrIFT
Frown
yarrow
ott
psilab
glsurf
fftw
coqtop
ocamlc
knotes
doxygen
coin
cgal
kcachegrind
cccc
gnuplot
openssl
sim
vim70
emacs
framerd
postgresql
gimp
fvwm
gcc
MPlayer
glibc
ffmpeg
sendmail
bind
httpd
linux
scheme
-0.05
S(g)
Scale Free Metric
0.2
0.15
0.1
0.05
0
0
5
10
15
20
25
30
35
Programs
in
out
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Assortativity
Figure 5. Scale Free Metric
Figure 7. Assortativity Coefficient
1
10
mean.geodesic
log(n)
9
0.1
Mean Geodesic Distance
Clustering Coefficient
8
0.01
0.001
1e-04
1e-05
7
6
5
4
3
CC
RandCC
2
Program
Figure 6. Clustering Coefficient
fer and are not ordered consistently. OCaml and Haskell
exhibit marked difference in correlations: as with C, the oo correlation is close to zero; but, i-i correlation is orders
of magnitude higher than o-o correlation. That is, OCaml
forces nodes with “proportional” indegree to pair up. If foo
is has an indegree X, bar is likely to receive, say, 2X indegree. One could interpret this result as a sign of stricter
hierarchical organization in functional languages.
Clustering Coefficient: Fig. 6 depicts how the call
graph clustering coefficients compare to clustering coefficients of random networks of same size. Computed clustering coefficients are orders of magnitude higher than their
random counterpart signifying higher degree of clustering.
Also, observe that `, as depicted is Fig. 8, is in the order of
log n. Together these observations make call graphs decidedly small world, irrespective of the source language.
We also have observed that average clustering coefficient
for nodes of particular degree, C(di ) follows power-law.
That is, the plot of di to C(di ) follows the power-law with
C(di ) ∝ d−1
i : high degree nodes exhibit lesser clustering and lower degree notes exhibit higher clustering. It is
also observed that OCaml’s fit for this power-law is the one
that had least misfit. Though we need further samples to
confirm it, we believe functional languages exhibit cleaner,
non-interacting hierarchy compared to both procedural and
OO languages.
Component Statistics: Fig. 9 gives us the components
statistics for the data set. It depicts the number of weakly
1
scheme
linux
httpd
bind
sendmail
ffmpeg
glibc
MPlayer
gcc
fvwm
gimp
postgresql
framerd
emacs
vim70
sim
openssl
gnuplot
cccc
kcachegrind
cgal
coin
doxygen
knotes
ocamlc
coqtop
fftw
glsurf
psilab
ott
yarrow
Frown
DrIFT
HaXml
HaXml
scheme
linux
httpd
bind
sendmail
ffmpeg
glibc
MPlayer
gcc
fvwm
gimp
postgresql
framerd
emacs
vim70
sim
openssl
gnuplot
cccc
kcachegrind
cgal
coin
doxygen
knotes
ocamlc
coqtop
fftw
glsurf
psilab
ott
yarrow
Frown
DrIFT
HaXml
HaXml
1e-06
Figure 8. Harmonic Geodesic Mean
connected components (#WCC), number of strongly connected components (#SCC), and fraction of nodes in the
largest strongly connected component (%SCC).
#WCC is lower in C and OCaml. For C++ and Haskell,
#WCC is higher compared to rest of the sample. This is
an indication of lazy call resolution, coinciding with the delayed/lazy bindings encouraged by both the languages. The
#SCC values are highest for OCaml. This observation, combined with reciprocity of OCaml programs, makes OCaml
a language that encourages recursion at varying granularity.
On the other end, C++ rates least against #SCC values.
Another important aspect in Fig. 9 is the observed values for %SCC; this fraction varies, surprisingly, from 1% to
30% of total number of nodes. C leads the way with some
applications, notably vim and Emacs, measuring as much as
20 to 30% for %SCC. OCaml follows C with a moderate 2
to 6% while C++ measures 1% to 3%. We do not yet know
why one third of an application cluster to form a SCC. Also,
%SCC values say that certain languages, notably OCaml,
and programs domains (Editors: Vim and Emacs) exhibit
significant mutual connectivity.
Edge Reciprocity: Fig. 11 shows the plot of edge reciprocity for various programs. Edge reciprocity is a measure of direct mutual recursion in the software. High reciprocity in a layered system implies layering inversion and
we would, ideally, like a program to have negative reciprocity.
100000
0.045
"o" u 1:4
#WCC
#SCC
%SCC
10000
0.04
0.035
0.03
Reciprocity
100
10
1
0.1
0.025
0.02
0.015
0.01
0.005
1e-04
0
scheme
linux
httpd
bind
sendmail
ffmpeg
glibc
MPlayer
gcc
fvwm
gimp
postgresql
framerd
emacs
vim70
sim
openssl
gnuplot
cccc
kcachegrind
cgal
coin
doxygen
knotes
ocamlc
coqtop
fftw
glsurf
psilab
ott
yarrow
Frown
DrIFT
HaXml
HaXml
0.01
0.001
scheme
linux
httpd
bind
sendmail
ffmpeg
glibc
MPlayer
gcc
fvwm
gimp
postgresql
framerd
emacs
vim70
sim
openssl
gnuplot
cccc
kcachegrind
cgal
coin
doxygen
knotes
ocamlc
coqtop
fftw
glsurf
psilab
ott
yarrow
Frown
DrIFT
HaXml
HaXml
%SCC : #SCC : #CC
1000
Figure 9. Component Statistics
Figure 11. Edge Reciprocity
0.003
6e-08
C++.Coin
C.Linux
0.0025
Betweenness Centrality
Betweenness Centrality
5e-08
4e-08
3e-08
2e-08
0.002
0.0015
0.001
0.0005
1e-08
0
0
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
0
20000
200
300
400
500
Figure 12. Betweenness - C++.Coin
Figure 10. Betweenness - C.Linux
Most programs exhibit close to zero reciprocity: most
call graphs exhibit as much reciprocity as that of random
graphs of comparable size. None exhibit negative reciprocity, implying no statistically significant preferential selection to not to violate strict layering.
The software that had least reciprocity is the Linux kernel. Recursion of any kind is abhorred inside kernel as
kernel-stack is a limited resource; besides, in a environment
where multiple contexts/threads communicate using shared
memory, mutual recursion could happen through continuation flow, not just as explicit control flow. Functional languages like OCaml naturally show higher reciprocity. Another curious observation is that compilers, both OCamlc
and gcc, appear to have relatively higher reciprocity. This
is the second instance where applications (Compilers: GCC
and OCamlc) determining the graph property; this could be
seen as a reflection of how compilers work: a great deal of
the lexing, parsing and semantic algorithms that compilers
are based on follow rich mutually recursive mathematical
definitions.
Clustering Profile: As we see in Fig 13, Clustering Profile indeed gives us a better insight. Y axis depicts the average clustering coefficient for nodes, say i and j, that are
connected by geodesic distance dij . In all the graphs observed, this average clustering increases up to dij =3 and
falls rapidly as dij increases further. We measured clustering profile for degrees one to ten and the clustering profile
appears to be unimodal, reaching the maximum at dij =3,
100
irrespective of language/program domain. It suggests that
maximal clustering occurs between nodes that are separated
exactly by five hops: clustering profile for a node u is measured with u removed; so dij =3 is 5 hops in the original
graph. However exciting we find this result to be, we currently have no explanation for this phenomenon.
Betweenness: Fig. 10 to 15 depict how betweenness
centrality is distributed – in different programs, written in
different different languages. Note that betweenness is not
distributed uniformly: it follows a rapidly decaying exponential distribution. This confirms our observation that importance of functions is distributed non-uniformly. Thus,
by concentrating test efforts in functions that have higher
betweenness – functions that are central to most paths – we
could test the software better, possibly with less effort. An
interesting line of investigation is to measure the correlation
between various centrality measures and actual per function
bug density in a real-world software.
7
Related Work
Understanding graph structures originating from various
fields is an active field of research with vast literature; there
is a renewed enthusiasm in studying graph structure of software and many studies, alongside ours, report that software
graphs exhibit small-world and power-law properties.
[18] studies the call graphs and reports that both indegree
and outdegree distributions follow power-law distributions
600
0.5
1
2
3
4
5
6
7
8
9
10
0.45
0.4
Clustering Profile
0.35
0.3
0.25
0.2
0.15
0.1
0.05
scheme
linux
httpd
bind
sendmail
ffmpeg
glibc
MPlayer
gcc
fvwm
gimp
postgresql
framerd
emacs
vim70
sim
openssl
gnuplot
cccc
kcachegrind
cgal
coin
doxygen
knotes
ocamlc
coqtop
fftw
glsurf
psilab
ott
yarrow
Frown
DrIFT
HaXml
0
Figure 13. Clustering Profile for Neighbours reachable in k+2 hops
0.008
4e-05
OCaml.Coq
Haskell.Yarrow
0.007
3.5e-05
3e-05
Betweenness Centrality
Betweenness Centrality
0.006
0.005
0.004
0.003
2.5e-05
2e-05
1.5e-05
0.002
1e-05
0.001
5e-06
0
0
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
Figure 14. Betweenness - OCaml.Coq
Figure 15. Betweenness - Haskell.Yarrow
and the graph exhibits hierarchical clustering. But [23] suggests that indegree alone follows power-law while the outdegree admits exponential distribution. [23] also suggests a
growing network model with copying, as proposed in [15],
would consistently explain the observations.
More recently, [4] studies the degree distributions of various meaningful relationships in a Java software. Many relationships admit power-law distributions in their indegree
and exponential distribution in their out-degree. [22] studies the dynamic, points-to graph of objects in Java programs
and found them to follow power-law.
Note that most work, excepting [4], do not rigorously
compare the likelihood of other distributions to explain the
same data. Power-law is notoriously difficult to fit and even
if power-law is a genuine fit, it might not be the best fit [8].
evolving graph system with distinct characteristics, a viewpoint we think is of importance in developing and maintaining large software systems.
There is lot that needs to be done. First, we need to measure the correlation between these precise quantities and
the qualitative, rule of thumb understanding that developers usually possess. This helps us making such qualitative,
albeit useful, observations rigorous. Second, we need to
verify our finding over a much larger set to improve the
inference confidence. Finally, graphs are extremely useful
objects that are analysed in a variety of ways, each exposing relevant features; of these variants, the authors find two
fields very promising: topological and algebraic graph theories. In particular, studying call graphs using a variant of
Atkin’s A-Homotopy theory is likely to yield interesting results [3]. Also, spectral methods applied to call graphs is an
area that we think is worth investigating.
8
Conclusion & Future Work
We have studied the structural properties of large software systems written in different languages, serving different purposes. We measured various finer aspects of these
large systems in sufficient detail and have argued why such
measures could be useful; we also depicted situations where
such measurements are practically beneficial. We believe
our study is a step towards understanding software as an
References
[1] A. Abdo and A. de Moura. Clustering as a measure of the local topology of networks. In arXiv:physics/0605235, 2006.
[2] S. Asur, S. Parthasarathy, and D. Ucar. An ensemble approach for clustering scalefree graphs. In LinkKDD workshop, 2006.
10000
[3] H. Barcelo, X. Kramer, R. Laubenbacher, and C. Weaver.
Foundations of a connectivity theory for simplicial complexes. In Advances in Applied Mathematics, 2006.
[4] G. Baxter and et al. Understanding the shape of java software. In OOPSLA, 2006.
[5] M. Boguna, R. Pastor-Satorras, and A. Vespignani. Absence
of epidemic threshold in scale-free networks with connectivity correlations. In arXiv:cond-mat/0208163, 2002.
[6] S. Brin and L. Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN
Systems, 30(1–7):107–117, 1998.
[7] J. A. Brretzen and R. Conradi. Results and experiences from
an empirical study of fault reports in industrial projects. In
PROFES, 2006.
[8] A. Clauset and et al. Power-law distributions in empirical
data. In arXiv.org:physics/0706.1062, 2007.
[9] L. da F. Costa, F. A. Rodrigues, G. Travieso, and P. R. V.
Boas. Characterization of complex networks: A survey of
measurements. Advances In Physics, 56:167, 2007.
[10] F. S. Diego Puppin. The social network of java classes. In
ACM Symposium on Applied Computing, 2006.
[11] P. Erdös and A. Rnyi. On random graphs. In Publicationes
Mathematicae 6, 1959.
[12] D. Garlaschelli and M. I. Loffredo. Patterns of link reciprocity in directed networks. Physical Review Letters, 2004.
[13] D. Hrimiuc. The rearrangement inequality - a tutorial:
pims.math.ca/pi/issue2/page21-23.pdf, 2007.
[14] http://projects.forked.de/graph tool/.
[15] P. Krapivsky and S. Redner. Network growth by copying.
Phys Rev E Stat Nonlin Soft Matter Phys, 2005.
[16] L. Li, D. Alderson, R. Tanaka, J. Doyle, and W. Willinger. Towards a theory of scale-free graphs. In arXiv:condmat/0501169, 2005.
[17] Y. Matsuo. Clustering using small world structure. In
Knowledge-Based Intelligent Information and Engineering
Systems, 2002.
[18] C. R. Myers. Software systems as complex networks. In
Physical Review, E68, 2003.
[19] M. Newman. The structure and function of complex networks. In SIAM Review, 2003.
[20] M. E. J. Newman. Assortative mixing in networks. Physical
Review Letters, 89:208701, 2002.
[21] D. E. Perry and W. M. Evangelist. An empirical study of
software interface faults. In Symposium on New Directions
in Computing, 1985.
[22] A. Potanin, J. Noble, M. Frean, and R. Biddle. Scale-free
geometry in oo programs. Comm. ACM, May 2005.
[23] S. Valverde and R. V. Sole. Logarithmic growth dynamics
in software networks. In arXiv:physics/0511064, Nov 2005.
[24] S. Valverde and R. V. Sole. Hierarchical small worlds in
software architecture. In arXiv:cond-mat/0307278, 2006.
[25] Y. Wang, D. Chakrabarti, C. Wang, and C. Faloutsos. Epidemic spreading in real networks: An eigenvalue viewpoint.
In Symposium on Reliable Distributed Computing, 2003.
[26] A. Y. Wu, M. Garland, and J. Han. Mining scale-free networks using geodesic clustering. In ACM SIGKDD, 2004.
[27] www.csn.ul.ie/ mel/projects/codeviz/.
Appendix I
*
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Lang
C
C
C
C
C
C
C
C
C
C
C
C
C
C
C
C
C
C
C++
C++
C++
C++
C++
C++
OCaml
OCaml
OCaml
OCaml
OCaml
OCaml
Haskell
Haskell
Haskell
Haskell
Haskell
Name.Version
scheme.mit.7.7.1
linux.2.6.12.rc2
httpd.2.2.4
bind.9.4.1
sendmail.8.12.8
ffmpeg.2007.05
glibc.2.3.6
MPlayer.1.0rc1
gcc.4.0.0
fvwm.2.5.18
gimp.2.3.9
postgresql.8.2.3
framerd.2.6.1
emacs.21.4
vim70
sim.outorder.3.0
openssl.0.9.8e
gnuplot.4.2.2
cccc.3.1.4
kcachegrind.0.4
cgal.3.3
coin.2.4.6
doxygen.1.5.3
knotes.3.3
ocamlc.opt.3.09
coqtop.opt.8
fftw.3.2alpha2
glsurf.2.0
psilab.2.0
ott.0.10.11
yarrow.1.2
Frown.0.6.1
DrIFT.2.2.1
HaXml.1.Validate
HaXml.1.Xtract
Appln. Domain
Interpreter
Kernel
Web Sever
Name Server
Mail Server
Media Codecs
C Lib
Media Player
C Compiler
Win Manager
Image Editor
R-DBMS
OO-DMBS
“Editor”
Editor
µarch/ISA Sim
Crypto Lib
Graph Plotting
Code Metrics
Cache Analyser
CompGeom Lib
OpenGL 3D Lib
Doc Generator
PIM
OCaml Compiler
Theorem Prover
FFT Computing
OpenGL Surface
Numeric Envirn
PL/Calculi Tool
Theorem Prover
Parser Gen
Typed Preproc
XML Validate
XML grep
N
2512
20165
1396
4534
783
4207
4401
7985
10848
3312
16021
8517
3490
3872
4489
442
7078
2191
1654
2593
3151
12963
11723
2174
4397
16126
585
4003
1888
4300
9397
6796
1428
4117
3909
M
5610
70010
4014
18874
4064
11692
13972
21744
48847
12052
88473
41189
17048
13154
18368
1089
21827
7045
5627
8054
7690
51877
31889
4942
12732
51092
1011
9173
4341
11193
15199
10218
3292
7624
5242
| 6cs.PL
|
1
Dynamic Parallel and Distributed Graph Cuts
Miao Yu1,2 , Shuhan Shen1,4 , and Zhanyi Hu? 1,3,4
arXiv:1512.00101v2 [cs.DS] 12 Sep 2016
1
National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing
100190, China
2
Zhongyuan University of Technology, Zhengzhou 450007, China
3
CAS Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences,
Beijing 100190, China
4
University of Chinese Academy of Sciences, Beijing 100049, China
{myu, shshen, huzy}@nlpr.ia.ac.cn
Abstract—Graph cuts are widely used in computer vision. To
speed up the optimization process and improve the scalability for
large graphs, Strandmark and Kahl [1], [2] introduced a splitting
method to split a graph into multiple subgraphs for parallel
computation in both shared and distributed memory models.
However, this parallel algorithm (the parallel BK-algorithm) does
not have a polynomial bound on the number of iterations and is
found to be non-convergent in some cases [3] due to the possible
multiple optimal solutions of its sub-problems.
To remedy this non-convergence problem, in this work, we first
introduce a merging method capable of merging any number of
those adjacent sub-graphs that can hardly reach agreement on
their overlapping regions in the parallel BK-algorithm. Based on
the pseudo-boolean representations of graph cuts, our merging
method is shown to be effectively reuse all the computed flows
in these sub-graphs. Through both splitting and merging, we
further propose a dynamic parallel and distributed graph cuts
algorithm with guaranteed convergence to the globally optimal
solutions within a predefined number of iterations. In essence, this
work provides a general framework to allow more sophisticated
splitting and merging strategies to be employed to further boost
performance. Our dynamic parallel algorithm is validated with
extensive experimental results.
Index Terms—graph cuts, parallel computation, convergence,
Markov random field
I. I NTRODUCTION
Graph cuts optimization plays an important role in solving
the following energy minimization problem, which is usually
derived from a Maximum A Posteriori (MAP) estimation of
a Markov Random Field (MRF) [4]:
X
E(X) =
ψc (xc ),
(1)
c∈C
where ψc (·) is the potential function, xc is the set of variables
defined on clique c, and C is the set of all the cliques. The
most recent computer vision problems often require the energy
models to contain a massive number of variables, have large
label spaces and/or include higher-order interactions, which
invariably enlarge the scale of the graph and increase the
number of calls of graph cuts algorithm for solving the energy
functions in the form (??). Consequently, determining how to
increase the scalability and speed up the optimization process
of graph cuts becomes an urgent task. However, processor
makers favor multi-core chip designs to manage CPU power
dissipation. Therefore, fully exploiting the modern multicore/multi-processor computer architecture to further boost the
efficiency and scalability of graph cuts has attracted much
attention recently.
A. Related work
Thanks to the duality relationship between maxflow and
minimal s−t cut, the maxflow algorithms, usually categorized
as augmenting-path-based [5], [6], [7], [8] and push-relabelbased [9], [10], [11], [12], have been a focus for parallelizing
in the literature.
Push-relabel-based methods [9], [10], [11], [12] are relatively easy to parallelize due to their memory locality feature.
Therefore, a number of parallelized maxflow algorithms have
been developed [13], [14], [15], [16]. These algorithms generally exhibit superior performance on huge 3D grids with high
connectivity. Nevertheless, they are usually unable to achieve
the same efficiency on 2D grids or moderately sized sparse 3D
grids, which is common in many computer vision problems,
as the state-of-the-art serial augmenting path method [6], [8]
could do on a commodity multi-core platform. The GPU
implementation [17] often fails to produce the correct results
unless the amount of regularization is low [1]. The data
transfer between the main memory and the graphics card is
also a bottleneck for the GPU-based implementation.
Perhaps the most widely used graph cuts solver in the computer vision community is the algorithm proposed by Boykov
and Kolmogorov [6] (called the BK-algorithm). This is a serial
augmenting path maxflow algorithm that effectively reuses
the two search trees originated from s and t, respectively.
To parallelize the BK-algorithm for further efficiency, the
graph is usually split into multiple parts, either disjoint or
overlapping. Liu and Sun [18] uniformly partitioned the graph
into a number of disjoint subgraphs, concurrently ran the BKalgorithm in each subgraph to obtain short-range search trees
within each subgraph, and then adaptively merged adjacent
subgraphs to enlarge the search range until only one subgraph
remained and all augmenting paths were found. For some 2D
image segmentation cases, this algorithm can achieve a nearlinear speedup with up to 4 computational threads. However,
2
this method requires a shared-memory model, which makes it
difficult to use on distributed platforms. To make the algorithm
applicable to both shared and distributed models, Strandmark
and Kahl [1], [2] proposed a new parallel and distributed BKalgorithm (called the parallel BK-algorithm), which splits the
graph into overlapped subgraphs based on dual decomposition [19], [20]. The BK-algorithm is then run in a parallel and
iterative fashion. Unfortunately, due to the possible multiple
optimal solutions of the BK-algorithm, which is an inherent
characteristic of all graph cuts methods, the parallel BKalgorithm may fail to converge in some cases [3].
By combining push-relabel and path augmentation methods,
Shekhovtsov and Hlavac [3] proposed an algorithm to further
reduce the number of message exchanges in the region pushrelabel algorithm [16]. Bhusnurmath and Taylor [21] reformulated graph cuts as an `1 minimization problem and provided
a highly parallelized implementation. However, even its GPUbased implementation is not significantly faster than the BKalgorithm for any type of graph.
II. C ONVERGENCE PROBLEM IN THE PARALLEL
BK- ALGORITHM
Because our work intends to remedy the non-convergence
problem of the parallel BK-algorithm [1], [2], a brief review
of it is first provided, followed by a convergence analysis of
the parallel BK-algorithm.
A. A brief review of the parallel BK-algorithm
By formulating the graph cuts problem defined on graph
G(V, C), with vertex set V = {s, t} ∪ V and the capacity set
C, as a linear program EV (x), Strandmark and Kahl [1], [2]
introduced a splitting method to split graph G(V, C) into a set
of overlapping subgraphs. Without loss of generality, only the
two-subgraph case is discussed here, denoted as G1 (V1 , C1 )
and G2 (V2 , C2 ), where V1 = {s, t} ∪ V1 , V2 = {s, t} ∪ V2 and
V1 ∩V2 6= ∅. If the separable condition holds, stated as ∀cij >
0, ∃Vk =⇒ i ∈ Vk ∧j ∈ Vk , finding the s−t cut with minimal
cost in graph G(V, C) can be reformulated as maximizing
a concave non-differential dual function g(λ), and it can be
solved in the following iterative way: 1) calculate the optimal
∗
∗
solutions xV1 and xV2 from the two sub-problems EV1 (x1 |λ)
2
and EV2 (x | − λ) defined on G1 (V1 , C1 ) and G2 (V2 , C2 ),
respectively, which are in the same form as the linear program
representation of graph cuts such that the BK-algorithm can
be applied on these two subgraphs simultaneously to obtain
an ascent direction; 2) update the dual variables λ along this
ascent direction to modify EV1 (x1 |λ) and EV2 (x2 | − λ) to
initiate the next iteration. The part of Fig. 3 before “Merge
Subgraphs” illustrates this process, where the linear program
representation E for each graph (subgraph) is depicted above
its corresponding box.
The above process is repeated until the following stopping
rule is satisfied:
î ∗ (k) ∗ (k) ó
xV1 − xV2
= 0,
(2)
V1 ∩V2
where [ · ]V1 ∩V2 is the projection on V1 ∩ V2 . Once the above
condition holds, the supergradient of g(λ) equals 0, which
means the maximum value is reached.
(a) One thread, 0.060 s
(b) Two threads, 0.047 s
(c) Three threads, 0.138 s
(d) Four threads, ∞
Fig. 1: Solving an identical graph cuts problem using the parallel
BK-algorithm with different numbers of threads.
B. Convergence problem
In theory, the parallel BK-algorithm is intended to maximize
a concave non-differential function g(λ) whose global optimal
solution can be found by the supergradient ascent method.
However, this algorithm may have problems converging to the
optimal solution for the following two reasons: 1) Its subproblems, EV1 (x1 |λ) and EV2 (x2 | − λ), may have multiple
optimal solutions. Fig. 3 of Strandmark and Kahl’s original
paper [1] gives such an example. 2) The heuristic step size
rule used by the parallel BK-algorithm, has a much faster
convergence rate in practice, but it has no theoretical guarantee
of convergence.
Fig. 1 shows an example of failure of the parallel BKalgorithm. Note that unless otherwise stated, all the experiments in this paper are conducted on a Sun Grid Engine
cluster, with each compute node having 2×Intel R Xeon R E52670 v2 @ 2.5GHz (up to 20 Cores and 40 Threads) and
128GB memory. The graph associated with the image of Fig.
1 is constructed as follows: The edge costs are determined by
the image gradient and the pixels in the leftmost and rightmost
columns are connected to the source and the sink respectively.
The red curves in Fig. 1 depict the final “cut”, and the dotted
blue lines are the boundaries of the neighboring parts. Fig.
1(a) is the result of the serial BK-algorithm. Its final “cut”
serves as the ground-truth and its running time serves as the
baseline in this example. Vertical splitting would lead to the
“worst-case scenario”, since all the possible s − t paths are
severed and all flows have to be communicated between the
threads. Part of this example comes from Fig. 8 of Strandmark
and Kahl’s original paper [1], where it is used to demonstrate
the robustness of the parallel BK-algorithm to a poor splitting
choice. They showed that if the graph is vertically split into
only two parts, the parallel BK-algorithm could still obtain the
same “cut”, but could do so approximately 22% faster than
the serial BK-algorithm, as shown in Fig. 1(b). However, we
find that if the number of split parts is further increased, the
3
TABLE I: Number of images that failed to converge in the parallel
BK-algorithm, where PRB represents the segmentation problems and
TRDS represents the number of computational threads
TRDS
PRB
seg1
seg2
2
3
4
5
6
7
8
1
16
5
500
14
500
30
500
67
500
92
500
138
3
2
6
performance of the parallel BK-algorithm begins to degrade.
With three threads, as shown in Fig. 1(c), the parallel BKalgorithm takes more than twice the computational time of
the serial BK-algorithm to obtain the same “cut”. If the graph
is vertically split into four parts, as shown in Fig. 1(d), the
parallel BK-algorithm is unable to converge.
To further assess the impact of the methods of splitting
and the number of computational threads on the convergence
of the parallel BK-algorithm, the following two fore-/background segmentation problems are carried out on the enlarged
Berkeley segmentation dataset [22], which consists of 500
images. The graph construction for the first segmentation
problem, denoted as seg1, is the same as the above experiment.
The graph construction for the second segmentation problem,
denoted as seg2, is a lightly different from seg1. In seg2, each
pixel is connected to both the source and the sink, and the
edge capacity depends on the value of the pixel. Therefore,
vertically splitting is a reasonable splitting choice. We tried
2 to 8 computational threads of the parallel BK-algorithm to
solve these two segmentation problems on all 500 images.
The number of images that failed to converge in each case
are presented in Table I. The maximum allowed number of
iterations is set to 1000, which is the same as that used in its
original implementation.
It can be seen from Table I that improper splitting and the
thread number both have a great impact on the convergence
of the parallel BK-algorithm, where all the images failed to
converge in seg1 once the thread number exceeded 3. Even if
the graph is split properly, the failure of convergence seems
to be unavoidable with a moderate number of computation
threads. In seg2, where the vertical splitting was reasonable,
for approximately 3% of the images, the method failed to
converge with 4 computational threads, and the failure rate
increased to nearly 30% with 8 computational threads. The
lack of convergence guarantee is a great deficiency of the parallel BK-algorithm, which severely hampers its applicability.
III. A MERGING METHOD
To remedy the non-convergence problem in the parallel
BK-algorithm, we first introduce a merging method that can
deal with the suspected neighboring subgraphs causing nonconvergence and then give a correctness and efficiency analysis
of the merging method based on pseudo-boolean representations of graph cuts.
As indicated by the experiments in the previous section, splitting a graph into more subgraphs often increases
the speedup, but it also increases the possibility of nonconvergence. However, it is difficult to know the optimal
3
1
3
2
5
0
7
5
4
4
2
7
s
6
4
1
5
0
0
s
s
0
1
2
3
2
6
1
5
6
1
1
6
5
4
5
2
2
7
3
7
10
5
4
4
6
5
4
6
1
6
5
3
6
6
t
t
t
(a) G1 (V1 , C1 )
(b) G2 (V2 , C2 )
(c) G12 (V12 , C12 )
Fig. 2: Merging two neighboring subgraphs G1 (V1 , C1 ) and
G2 (V2 , C2 ) into a single graph G12 (V12 , C12 ).
number of subgraphs at the beginning of the parallel BKalgorithm. Is it possible to dynamically adjust the current
splitting of a graph when over-splitting is detected during
the running time, so as to further speed up the computation
while simultaneously guaranteeing convergence? To this end,
a merging method is proposed in this paper.
Merging two subgraphs into a single graph is performed
as follows: given two neighboring subgraphs G1 (V1 , C1 ) and
G2 (V2 , C2 ), denote the merged graph as G12 (V12 , C12 ), where
V1 = {s, t} ∪ V1 , V2 = {s, t} ∪ V2 , V1 ∩ V2 6= ∅ and V12 =
{s, t} ∪ V12 . The merged vertex set V12 = V1 ∪ V2 . Only if
the two endpoints are all within V1 ∩ V2 is the edge capacity
of the merged graph G12 the summation of the capacities of
the same edge in G1 and G2 . Otherwise, the edge capacity
of the merged graph G12 is simply that from only one of the
two subgraphs. Fig. 2 gives an example. Merging more than
two subgraphs into a single graph can be performed similarly,
where the vertices in the overlapped region are absorbed and
the edge capacities in the overlapped region are summed. It is
worth noting that the merging method can also be applied to
N-D graphs (subgraphs), just like the splitting method of the
parallel BK-algorithm.
Fig. 3 shows the application of the merging method in the
parallel BK-algorithm with only two subgraphs. When there is
evidence that the two subgraphs may have difficulties reaching
agreement on their optimal values, the merging operation is
invoked to obtain a merged graph G0 from which the optimal
∗
values xV are obtained. This naturally raises the following key
question: will the merged graph G0 obtain the same optimal
solutions as the original graph G, despite that the original was
first split into two subgraphs, both of which then experienced
maxflow computation and updating of many rounds?
There are two related questions: How should one decide
which subgraphs to merge? What is the best time for merging?
Although there are many cues for over-splitting, such as the
existence of a large number of vertices within the overlapped
region that admit multiple optimal solutions or the scarcity
of pushed flows in the subgraphs, there is currently no clearcut answer. However, we use a simple strategy based on the
following assumption: the number of vertices in the overlapped
region that disagree on their optimal values is expected to
decrease with iterations if the current graph splitting is proper.
The strategy is detailed in Algorithm 1, where numDif f is
used to record the minimal number of disagreement vertices
4
EV1 (x1 j¸(1) )
(1)
G1
EV1 (x1 j¸(2) )
BKalgorithm
(1)
EV (x)
f G(xV)
0
(1 )
¤ (1)
G1 ; xV1
(1)
f G2 (xV2 )
BKalgorithm
(k0 )
¤ (k)
G1 ; xV1
BKalgorithm
(10 )
¤ (1)
G2 ; xV2
(10 )
f G2 (xV2 )
update
graph
(2)
G2
(2)
f G2 (xV2 )
(k+1)
G1
(k+1)
f G1
f G1 (xV1 )
Does the stopping
rule hold?
NO
update
graph
(k0 )
(2)
(xV1 )
NO
G0
EV2 (x2 j ¡ ¸(k+1) )
EV2 (x2 j ¡ ¸(2) )
EV2 (x2 j¸(1) )
(1)
(2)
G1
f G1 (xV1 )
f G1 (xV1 )
Does the stopping
rule hold?
G2
update
graph
(10 )
f G1 (xV1 )
G
Merge
Subgraphs
s
¤
1 (k+1)
EV1 (x j¸
)
xV
A Single Iteration
BKalgorithm
(k0 )
BKalgorithm
Initial
Splitting
¤ (k)
G2 ; xV2
update
graph
(k0 )
0
f G (xV)
(k+1)
G2
(k+1)
f G2
f G2 (xV2 )
(xV2 )
Fig. 3: Application of the merging method in the parallel BK-algorithm with only two subgraphs.
of all the iterations. If numDif f remains non-decreasing in
some successive iterations, which is controlled by ITER, then
merging should be invoked to ensure better graph splitting.
In the merging operation, every two adjacent subgraphs are
merged into one graph in order to maintain work load balance.
Clearly, the number of disagreement vertices will decrease at
least by 1 for every ITER−1 iterations, or every two adjacent
subgraphs are merged. Therefore, Algorithm 1 will converge
within no more than M × (ITER − 1) + logN
2 iterations,
where M is the number of disagreement vertices in the first
iteration and N is the number of subgraphs into which the
original graph is split in the initial splitting stage. Since a naive
merging strategy is used and the convergence guarantee holds,
Algorithm 1 is named the “Naive Converged Parallel BKAlgorithm”. Obviously, more-suitable merging strategies can
further boost the performance, this issue will be investigated
in our further work.
Algorithm 1 Naive Converged Parallel BK-Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
Set numDif f := +∞; iter := 0;
Split graph G into N overlapped subgraphs G1 , . . . , GN ;
repeat
Run the BK-algorithm concurrently on all the subgraphs
∗ (k)
(k)
Gi to obtain xVi
(k0 )
Update all the residual subgraphs Gi
to obtain
(k+1)
Gi
;
Count the number of nodes that disagree on their
optimal values, denoted as nDif f ;
if nDif f ≥ numDif f then
+ + iter;
if iter == ITER then
Every two neighboring subgraphs are merged into
one graph;
end if
else
numDif f := nDif f ;
iter := 0;
end if
until nDif f == 0;
IV. P SEUDO - BOOLEAN REPRESENTATION - BASED
INVARIANCE ANALYSIS FOR GRAPH CUTS ALGORITHMS
As noted in the previous section, the correctness of our
merging methods depends on whether the merged graph obtains the same optimal solutions as that of the original graph.
To answer this question, we propose a new pseudo-boolean
representation, named restricted homogeneous posiforms, to
track the changes for all the graphs (subgraphs) under the
operations of the parallel graph cuts algorithms. We then
develop an invariance analysis method.
A. Restricted homogeneous posiforms for graph cuts
Since each s − t cut in a graph is a partition of all
its vertices to either S or T , it can be represented by a
realization of a set of boolean variables, xV = {xi |i ∈ V}.
Therefore, a function with its arguments being boolean to
represent an s − t cut and its value being real representing
the cost of this s − t cut can be used to express a graph cuts
problem. Such a function is usually called a pseudo-boolean
function in the combinatorial optimization community [23],
[24], [25], whose representations are often categorized into
two types: multi-linear polynomials whose form is uniquely
determined by a pseudo-boolean function and posiforms that
can uniquely determine a pseudo-boolean function, although
a pseudo-boolean function can have many different posiforms
representing it. Among these many possible posiforms, we
define the restricted homogeneous posiforms, which can have a
one-to-one correspondence with the graph on which the graph
cuts problem is defined.
Theorem 1 A graph cuts problem defined on graph G(V, C),
with V = {s, t} ∪ V and non-negative edge capacities,
can be uniquely represented using the following restricted
homogeneous posiforms:
X
X
X
φG
ai xi +
aj x̄j +
aij x̄i xj ,
(3)
h (xV ) =
i∈V
j∈V
i,j∈V
where xi ∈ {0, 1}, x̄i = 1 − xi , ai ≥ 0, aj ≥ 0, aij ≥ 0.
Moreover, the components of φG
h (xV ) have the following oneto-one correspondence with edges of graph G(V, C), as shown
in Table II.
5
TABLE II: The correspondence of components of φG
h and edges of
G, where xi = 0 implies vertex i belongs to S and xi = 1 implies
vertex i belongs to T in an s − t cut.
Comp. φG
h (xV )
Edge G(V, C)
Relations
ai xi
aj x̄j
aij x̄i xj
(s, i)
(j, t)
(i, j)
ai = csi , csi ∈ C
aj = cjt , cjt ∈ C
aij = cij , cij ∈ C
Proof: The cost of an s − t cut is
X
CS,T =
cij
(4)
Theorem 1 states that each unary term in the restricted
homogeneous posiforms, as in (??), corresponds to an t-link
edge in the graph, whereas each pairwise term corresponds
to an n-link edge, and their coefficients are the capacities of
those edges in the graph. Therefore, any changes in the graph
will be reflected in its restricted homogeneous posiforms, and
vice versa.
In addition to the above proposed restricted homogeneous
posiforms, a pseudo-boolean function can also be uniquely
represented as a multi-linear polynomial [23], [24], [25]:
i∈S,j∈T
and the relations between the values of xi and the parts to
which the node i belongs are:
®
0 ⇐⇒ i ∈ S
xi =
.
(5)
1 ⇐⇒ i ∈ T
Only if φG
h (xV ) equals the cost, up to a constant, of the s−t cut
defined by the realization of xV , for every possible realizations
of xV , can φG
h (xV ) be the pseudo-boolean function of the
graph cuts problem defined on graph G(V, C). Since Table
II defines a one-to-one relationship between the components
of the restricted homogeneous posiforms and the edges of the
graph G(V, C), it suffices to verify the equivalence between
the value of each component of φG
h (xV ) and the cost of its
corresponding edge in the s − t cut under all the possible
realizations of xV . This can be done for the three components
listed in Table II.
1). ai xi
®
0 ⇐⇒ xi = 0
The value of ai xi is
, its correspondai ⇐⇒ xi = 1
ing
® edge (s, i) contributes to the cost of an s − t cut by:
0 ⇐⇒ i ∈ S
, which can be seen from the definition of
csi ⇐⇒ i ∈ T
the cost of an s − t cut in (??). From the relations of (??) and
the equivalence of ai and csi , it is clear that the value of ai xi
equals the cost of its corresponding edge (s, i) in the s − t cut
under all the possible realizations of xV .
2). aj x̄j
It can be verified in a similar way as in the previous case
that the value of aj x̄j equals the cost of its corresponding
edge (j, t) in the s − t cut under all the possible realizations
of xV .
3). aij x̄i xj
®
aij ⇐⇒ xi = 0, xj = 1
The value of aij x̄i xj is
. Its
0
⇐⇒ otherwise
corresponding
edge (i, j) contributes to the cost of an s − t
®
cij ⇐⇒ i ∈ S, j ∈ T
cut by:
. From the relations of (??)
0 ⇐⇒ otherwise
and the equivalence of aij and cij , it is clear that the value of
aij x̄i xj equals the cost of its corresponding edge (i, j) in the
s − t cut under all the possible realizations of xV .
For each of the three components in the restricted homogeneous posiforms (??), its value under all the possible
realizations of xV equals the contribution of its corresponding
edge to the cost of the s − t cut. Therefore, Theorem 1 is
proved.
G
f (x) = l0 +
n
X
li xi +
j=1
X
lij xi xj , xi ∈ {0, 1},
(6)
ij
with lij ≤ 0. In fact, lij ≤ 0 is the necessary and sufficient
condition for f G (x) to represent a graph cuts problem defined
on a graph G with non-negative edge capacities, as proved by
Freeman and Drineas [26]
V2
V1
s
0
1
V1\V2
5
t
V2nV1
(b) G(V, C)
s
s
0
5
2
3
3
6
4
4
2
7
2
7
5
5
1
5
6
6
(a) Partition of the set V
1
1
3
6
V1nV2
6
5
4
4
6
3
2
5
4
2
7
5
7
10
1
3
2
2
3
6
5
4
1
2
3
6
5
6
2
t
t
(c) G1 (V1 , C1 )
(d) G2 (V2 , C2 )
Fig. 4: Splitting graph G(V, C) into two subgraphs G1 (V1 , C1 ) and
G2 (V2 , C2 ) .
B. Invariance analysis for parallel graph cuts algorithm
By the above multi-linear polynomials and proposed restricted homogeneous posiforms, here we give an invariance
analysis method for parallel graph cuts algorithm.
There are four types of operations in the naive converged
parallel BK-algorithm: splitting, maxflow computation, graph
updating and merging subgraphs. The first three types of
operations are also used in the parallel BK-algorithm. Fig.
3 is the pipeline of the naive converged parallel BK-algorithm
in the two subgraphs case. Without loss of generality, only the
6
two subgraphs case is analyzed here, but the conclusions can
be easily generalized to more subgraphs cases.
Suppose the original graph is G, the two subgraphs in
(k)
(k)
the kth iteration are G1 , G2 , and their corresponding re(k)
G
1
stricted homogeneous posiforms are φG
(xV1 ) and
h (xV ), φh
It is easy to verify that the following relations hold for the
coefficients of f G (xV ) and of φG
h (xV ):
X
X
X
l=
b
ai +
a
¯i +
Ê
ai ,
(13)
¨
¨
¨
i∈V ∩V
i∈V \V
i∈V \V
1
(k)
G
X
X
i,j∈V1 ∩V2
i or j∈V1\V2
(k)
G
φh 1
(xV1 ) =
(k)
b
a1i xi
˙
i∈V1\V2
X
(k)
¯lij = −āij , ∀i, j ∈ V1 ∩V2 ,
(17)
b
lij = −b
aij , ∀i or j ∈ V1 \V2 ,
(18)
Ê
lij = −Ê
aij , ∀i or j ∈ V2 \V1 .
(19)
(8)
G
(k)
of f G1 (xV1 ) and of φh 1 (xV1 ):
X 1(k)
X
(k)
(k)
l1 =
b
ai +
a
¯1i ,
¨
¨
i∈V ∩V
i∈V \V
1
(k)
(k)
b
a1i − b
a1i
li1 = b
˙
¨
X (k)
2(k)
Ê
ai xi +
a
¯2i xi +
˙
˙
i∈V1 ∩V2
i∈V2\V1
X 2(k)
X (k)
Ê
ai x̄i +
a
¯2i x̄i +
¨
¨
i∈V1 ∩V2
i∈V2\V1
X
X
(k)
(k)
Ê
a2ij x̄i xj .
ā2ij x̄i xj +
¯lij xi xj +
(k)
1
2
X
+
(20)
2
(k)
b
a1ki , ∀i ∈ V1 \V2 ,
(21)
k∈V1
i∈V1 ∩V2
X
(k)
b
li1 xi +
¯l1 xi xj +
ij
(9)
(k)
G2 ,
X
X
X
(k)
Ê
li2 xi +
+
X
2(k)
Ê
lij
xi xj .
i or j∈V2\V1
(k)
1(k)
b
lij
= −b
a1ij , ∀i or j ∈ V1 \V2 .
(24)
(k)
G2
(k)
(k)
(k)
b
li2 = b
a2i − b
a2i
˙
¨
1
1
+
X
(k)
b
a2ki , ∀i ∈ V2 \V1 ,
(12)
(26)
k∈V2
1
(11)
(25)
2
X (k)
(k)
(k)
¯l2(k) = a
¯2i − a
¯2i +
ā2ki , ∀i ∈ V1 ∩V2 ,
i
˙
¨
k∈V ∩V
¯l2(k) xi +
i
i∈V1 ∩V2
i∈V2\V1
¯l2(k) xi xj
ij
i,j∈V1 ∩V2
X
(23)
2
¯l1(k) xi +
i
1(k)
b
lij
xi xj ,
¯l1(k) = −ā1(k) , ∀i, j ∈ V1 ∩V2 ,
ij
ij
(k)
(10)
(22)
2
And those between f G2 (xV2 ) and φh (xV2 ):
X 2(k)
X
(k)
(k)
l2 =
b
ai +
a
¯2i ,
¨
¨
i∈V ∩V
i∈V \V
i or j∈V1\V2
f G2 (xV2 ) = l2 +
X
Ê
lij xi xj ,
i∈V1 ∩V2
(k)
i,j∈V1 ∩V2
(k)
1
i or j∈V2\V1
i∈V1\V2
X
i∈V2\V1
X
b
lij xi xj +
i or j∈V1\V2
(k)
X (k)
(k)
(k)
¯l1(k) = a
¯1i − a
¯1i +
ā1ki , ∀i ∈ V1 ∩V2 ,
i
˙
¨
k∈V ∩V
i or j∈V2\V1
X
(xV1 ) = l1 +
(k)
(16)
X
i∈V1\V2
f
(15)
X
Ê
li = Ê
ai − Ê
ai +
Ê
aki , ∀i ∈ V2 \V1 ,
˙
¨
k∈V
(k)
The multi-linear polynomials of graphs G,
and
(k)
(k)
denoted as f G (xV ), f G1 (xV1 ) and f G2 (xV2 ) respectively,
can be expressed as:
X
X
X
b
¯li xi +
Ê
f G (xV ) = l +
li x i +
li xi +
(k)
G1
āki , ∀i ∈ V1 ∩V2 ,
1 ∩V2
Similarly, the following relations hold for the coefficients
i or j∈V1\V2
(k)
G1
i,j∈V1 ∩V2
¯li = a
¯i − a
¯i +
˙
¨
k∈V
(7)
(k)
X
i,j∈V1 ∩V2
X
1
X
2
(k)
+
a
¯1i xi +
˙
i∈V1 ∩V2
i,j∈V1 ∩V2
G
(14)
X
a
¯1i x̄i +
b
a1i x̄i +
¨
¨
i∈V1 ∩V2
i∈V1\V2
X
X
(k)
1(k)
b
a1ij x̄i xj ,
āij x̄i xj +
φh 2 (xV2 ) =
1
i or j∈V2\V1
(k)
X
2
b
li = b
ai − b
ai +
b
aki , ∀i ∈ V1 \V2 ,
˙
¨
k∈V
X
b
ai xi +
a
¯ i xi +
Ê
ai xi +
˙
˙
˙
i∈V1 ∩V2
i∈V1\V2
i∈V2\V1
X
X
X
b
ai x̄i +
a
¯i x̄i +
Ê
ai x̄i +
¨
¨
¨
i∈V1 ∩V2
i∈V1\V2
i∈V2\V1
X
X
X
āij x̄i xj +
b
aij x̄i xj +
Ê
aij x̄i xj ,
=
2
X
φh 2 (xV2 ) respectively. Since V = {s, t} ∪ V and V1 \ V2 ,
V1 ∩ V2 and V2 \ V1 are partitions of V, as shown in Fig. 4(a).
The three restricted homogeneous posiforms can be equally
written as:
φG
h (xV )
1
2
(27)
2
¯l2(k) = −ā2(k) , ∀i, j ∈ V1 ∩V2 ,
ij
ij
(28)
(k)
2(k)
b
lij
= −b
a2ij , ∀i or j ∈ V2 \V1 .
(29)
Thus far, the relations between the coefficients of
(k)
G1
the restricted homogeneous posiforms φG
(xV1 ),
h (xV ), φh
G
(k)
φh 2 (xV2 ) and those of the multi-linear polynomials f G (xV ),
(k)
(k)
f G1 (xV1 ), f G2 (xV2 ) have been established. The following
7
four propositions state the changes of the multi-linear polynomials for each of the four operations in Fig. 3.
(1)
(1)
Proposition 1 The coefficients of f G1 (xV1 ), f G2 (xV2 ) and
f G (xV ), which are the multi-linear polynomials of the split
(1)
(1)
subgraph G1 , G2 and the original graph G, respectively,
satisfy the following relations:
(1)
(1)
1) l1 + l2 = l.
(1)
1(1)
2) b
li1 = b
li , ∀i ∈ V1 \V2 ; b
lij
=b
lij , ∀i or j ∈ V1 \V2 .
(1)
(1)
2
2
Ê
Ê
Ê
Ê
3) li = li , ∀i ∈ V2 \V1 ; lij = lij , ∀i or j ∈ V2 \V1 .
(1)
(1)
1(1)
2(1)
= ¯lij
= 12 ¯lij , ∀i, j ∈
4) ¯li1 = ¯li2 = 21 ¯li , ∀i ∈ V1∩V2 ; ¯lij
V1 ∩V2 .
Proof: The original graph G was first split into two
subgraphs. Then, after K iterations of maxflow computation
(K+1)
(K+1)
and updating, the two subgraphs G1
and G2
were
obtained. The following relations among the coefficients of
(K+1)
(K+1)
f G1
(xV ) , f G2
(xV ) and f G (xV ) can be drawn from
Proposition.1 – 3:
(K+1)
1(K+1)
b
li1
=b
li , ∀i ∈ V1\V2 ; b
lij
=b
lij , ∀i or j ∈ V1\V2 , (31)
(K+1)
2(K+1)
Ê
li2
=Ê
li , ∀i ∈ V2\V1 ; Ê
lij
=Ê
lij , ∀i or j ∈ V2\V1 , (32)
Proof: See Appendix A.
Proposition 2 For all augmenting path maxflow algorithms,
pushing a flow F from s to t in graph G results in a residual
graph G0 , where the multi-linear polynomial of the residual
graph G0 differs from that of the original graph G by a
constant term of F.
¯l1(K+1) = ¯l2(K+1) = 1 ¯lij , ∀i, j ∈ V1 ∩V2 ,
ij
ij
2
(33)
K
X
(m+1)
¯l1(K+1) = 1 ¯li +
4λi
, ∀i ∈ V1 ∩V2 ,
i
2
m=1
(34)
K
X
(m+1)
¯l2(K+1) = 1 ¯li −
4λi
, ∀i ∈ V1 ∩V2 ,
i
2
m=1
(35)
l1
Proof: See Appendix B.
(K+1)
(1)
= l1
K
X
−
F1 ,
(m)
(36)
(m)
(37)
m=1
(k+1)
(k+1)
Proposition
3 The
coefficients of f G1
, f G2
and those
(k0 )
(k0 )
G1
G2
of f
, f
, which are updated by the dual variables
λ(k) , satisfy the following relations:
(k0 )
(k+1)
(k+1)
, ∀i ∈ V1 ∩V2 .
= ¯li1 + 4λi
1) ¯li1
0
(k+1)
(k )
(k+1)
2) ¯li2
= ¯li2 − 4λi
, ∀i ∈ V1 ∩V2 .
3) all the other coefficients remain unchanged.
(k+1)
(k+1)
(k)
where 4λi
= λi
− λi , ∀i ∈ V1 ∩V2 .
l2
(K+1)
(1)
= l2
K
X
−
F2 .
m=1
(K+1)
(K+1)
When the two subgraphs G1
and G2
are merged into
0
G0 , according to Proposition 4, the coefficients of f G (xV ) are:
(K+1)
l0 = l1
(1)
= l1
Proof: See Appendix C.
(K+1)
+ l2
(1)
+ l2
−
0
Proposition 4 The coefficients of f G (xV ), which is the multilinear polynomial of the merged graph G0 , have the following
(k+1)
(k+1)
relations with those of f G1
(xV1 ) and f G2
(xV2 ):
(k+1)
(k+1)
1) l0 = l1
+ l2
.
(k+1)
0
1(k+1)
2) b
li0 = b
li1
, ∀i ∈ V1 \V2 ; b
lij
=b
lij
, ∀i or j ∈ V1 \V2 .
(k+1)
0
2
0
2(k+1)
Ê
Ê
Ê
Ê
3) li = li
, ∀i ∈ V2 \V1 ; lij = lij , ∀i or j ∈ V2 \V1 .
(k+1)
(k+1)
¯l0 = ¯l1(k+1) +
4) ¯li0 = ¯li1
+ ¯li2
, ∀i ∈ V1 ∩ V2 ;
ij
ij
(k+1)
¯l2
,
∀i,
j
∈
V
∩V
.
1
2
ij
=l−
K
X
(m)
F1
K
X
m=1
K
X
−
m=1
b
li0
=
0
1
b
lij
=b
lij
−
K
X
(m)
F2
m=1
(m)
F2 ,
∈ V1 \V2 ;
(39)
=b
lij , ∀i or j ∈ V1 \V2 .
(K+1)
Ê
li0 = Ê
li2
=Ê
li , ∀i ∈ V2 \V1 ;
(K+1)
0
2
Ê
lij = Ê
lij
=Ê
lij , ∀i or j ∈ V2 \V1 .
Proof: See Appendix D.
(38)
m=1
(K+1)
b
li1
=b
li , ∀i
(K+1)
(m)
F1
(40)
¯l0 = ¯l1(K+1)+ ¯l2(K+1)
i
i
i
C. Correctness and efficiency of the merging
The following theorem states the relations between the
multi-linear polynomials of the original graph G and that of
the merged graph G0 shown in Fig. 3.
0
Theorem 2 The multi-linear polynomials f G (xV ) of the
merged graph G0 satisfies the following equation:
0
f G (xV ) = f G (xV ) +
K
X
m=1
(m)
F1
+
K
X
(m)
F2 ,
(30)
m=1
where f G (xV ) is the multi-linear polynomials of the original
(m)
(m)
graph G, F1 and F2 are the flows computed in the mth
(m)
(m)
iteration on subgraphs G1 and G2 , respectively, and K is
the number of iterations of the two subgraphs before merging.
=
= ¯li ,
K
X
1¯
(m+1)
li +
4λi
2
m=1
!
+
K
1¯ X
(m+1)
li −
4λi
2
m=1
∀i ∈ V1 ∩V2 ,
!
(41)
¯l0 = ¯l1(K+1)+ ¯l2(K+1) = 1 ¯lij + 1 ¯lij = ¯lij , ∀i, j ∈ V1 ∩V2 . (42)
ij
ij
ij
2
2
0
Therefore, only the constant term of f G (xV ) differs from that
P
(m) PK
(m)
of f G (xV ) by K
+ m=1 F2 . All the other terms
m=1 F1
have the same coefficients, and Theorem 2 is proved.
PK
P
(m)
(m)
Note that
+ K
is the accumulated
m=1 F1
m=1 F2
flows in the subgraphs, which is independent of the variables
∗
xV . Since only the optimal arguments xV at which f G (xV )
reaches its global minimum, not the global minimum itself,
matters for a graph cuts problem defined on graph G, if the
8
pseudo-boolean functions of two graph cuts problems only
differ by flows which do not depend on their arguments
xV , the two graph cuts problems are equivalent. And by the
convention of the combinatorial optimization community, this
difference is often referred to as a constant. An illustrative
example is given in Appendix E. Therefore, Theorem 2 reveals
the following two important things: 1) The two graph cuts
problems, which are defined on the original graph G and the
merged graph G0 , are equivalent because their corresponding
pseudo-boolean functions only differ by a constant. Hence
the same global optimal solution can be obtained from the
merged graph G0 . 2) Moreover, this constant is the summation
of all the flows computed in the two subgraphs in all the K
iterations, indicating that the flows are reused when computing
the maxflow for the merged graph G0 . As a result, computing
the maxflow for the merged graph G0 will be much faster than
computing the maxflow for the original graph G from scratch.
Theorem 2 also gives a good explanation of the parallel BKalgorithm from the perspective of flows. The algorithm can
be regarded as decomposing the maximum flow computation
of the original graph into subgraphs. Once the summation
of all the accumulated flows in the subgraphs reaches the
maximum flow of the original graph, the parallel BK-algorithm
converges. Therefore, whenever the subgraphs can easily accumulate flows, the parallel BK-algorithm can easily converge.
To evaluate the effectiveness of reusing flows in the proposed merging method, we once again apply the two fore/back-ground image segmentation methods (seg1 and seg2
used in section II-B) on all 500 images of the Berkeley
segmentation dataset [22]. Here, the naive converged parallel
BK-algorithm (Algorithm 1), whose N is fixed at 2 and whose
ITER is set to 1, 5, 15 and 20, respectively, is used. Fig. 5
shows the distribution of the relative times, which are defined
as the ratio of the times for computing the maxflow of the
merged graph G0 to that of the original graph G. The figure
also shows the relative used flows, which are defined as the
ratio of all the accumulated flows in the subgraphs before
PK
P
(m)
(m)
merging, that is
+ K
in (??), to the
m=1 F1
m=1 F2
maxflows of the original graph G.
The relative times for seg1 when ITER = 1, 5, 10, as shown
in Fig. 5(a), are roughly equal to 1, which means there is no
speedup in computing the merged graph G0 . By examining
Fig. 5(b), we found that the reused flows in these three cases
are very small (their histograms are all concentrated in the
first bin). It is precisely this no-flow to reuse that accounts for
the ineffectiveness in reusing flows. Because for the original
graph G of seg1 only the pixels in the leftmost column are
connected to the source and the pixels in the rightmost column
are connected to the sink, the maxflows of the two subgraphs
(1)
(1)
in the first iteration are all equal to 0, i.e., F1 = F2 = 0,
(1)
because none of the pixels of G1 are connected to the sink
(1)
and none of the pixels of G2 are connected to the source.
Once the subgraphs accumulated a certain amount of flows,
a sustained acceleration in the computations of the merged
graph occurred, as shown in Fig. 5(a),(b) for the ITER = 20
case.
The flows in the subgraphs for seg2 are fairly easy to
accumulate, as can be seen in Fig. 5(d), where even in the
ITER = 1 case, over 90% of the images have accumulated
over 95% of the flows of the original graph. This leads to a
significant speedup in the merged graph, where the maxflow
computations for approximately 40% of the merged graphs are
over 10 times faster than those of the original graphs, and over
90% of the images could achieve a speedup factor of at least
3. Although further increase of the flows in the subgraphs is
relatively slow, once this increase becomes apparent, it will be
reflected in the improvement of the relative times as well, as
shown in Fig. 5(c),(d) for the ITER = 20 case.
The experiments here show that whenever there are accumulated flows in the subgraphs, these flows will be reused
effectively in the merged graph. It is worth noting that the
correctness and efficitiveness of the merging method can be
easily generalized to cases of more than two subgraphs.
V. DYNAMIC PARALLEL GRAPH CUTS ALGORITHM
With the two key building blocks – splitting and merging at
hand, our dynamic parallel graph cuts algorithm is described
in Algorithm 2.
Algorithm 2 Dynamic Parallel Graph Cuts Algorithm
9:
10:
Set nDif f := +∞;
Initial splitting;
while nDif f > 0 do
A subgraph can be further split into a number of
subgraphs;
repeat
Compute maxflow for all the subgraphs;
Update all the subgraphs accordingly;
Count the number of nodes that disagree on their
optimal values, denoted as nDif f ;
until nDif f == 0 or some other conditions are met
Any number of neighboring subgraphs can be merged;
11:
end while
1:
2:
3:
4:
5:
6:
7:
8:
In essence, Algorithm 2 is a general parallel framework,
where lines 4, 6, 9 and 10 do not specify their concrete
realization but give the user complete freedom to choose their
own implementations for specific problem and specific parallel
and distributed platforms. Line 6 specifies the used maxflow
solver. Lines 4, 9 and 10 jointly determine when and how the
subgraphs are merged or further split: line 9 specifies the time
for merging and its realization is given in line 10, and line 4
determines the dynamic splitting strategy.
It can be seen from lines 6 – 8 of Algorithm 2 that a series
of similar maxflow problems (more specifically, graph cuts
problems in adjacent iterations only differing in a tiny fraction
of t-link values) need to be solved in order to obtain the global
optimal solutions of the original graph. Therefore, any graph
cuts algorithm that is efficient in this dynamic setting can be
used as the maxflow solver in line 6. In other words, any such
serial graph cuts algorithm can be parallelized using Algorithm
2 to further boost its efficiency and scalability.
9
(a) Relative times for seg1
(b) Relative reused flows for seg1
(c) Relative times for seg2
(d) Relative reused flows for seg2
Fig. 5: Relative times and relative reused flows for the two segmentations. In all the figures, ITER is the iteration control parameter used
in Algorithm 1.
Since the BK-algorithm [6] has some efficient dynamic
variants [27], [28], [29], it is parallelized (used as the maxflow
solver) by the parallel BK-algorithm [1], [2]. Pseudoflowbased maxflow algorithms [11], which can often be efficiently
warm started after solving a similar problem, can also be
parallelized using Algorithm 2. For example, the most recently
proposed Excesses Incremental Breadth-First Search (Excesses
IBFS) algorithm [30], which is a generalization of the incremental breadth-first search (IBFS) algorithm [8] to maintain
a pseudoflow, could act as a better candidate maxflow solver
in line 6 of Algorithm 2. The reason is that the performance
of the Excesses IBFS is competitive (usually faster) compared
with other dynamic graph cuts algorithms. More importantly,
it has a polynomial running time guarantee whereas there is
no known polynomial time limit for the BK-algorithm.
Dynamic merging aims to dynamically remedy the oversplitting during the running time, and it can make the dynamic
parallel graph cuts algorithm converge to the globally optimal
solutions within a predefined number of iterations. Here is
a general procedure: 1) the graph is initially split into N
subgraphs, and no further splitting is permitted; 2) invoke
the merging operations every K iterations; 3) for a merging
operation, every neighboring ` subgraphs are merged into a
single graph. In the worst case, the merging operations can
be invoked logN
` times, leaving only one graph, from which
the global optimal solutions can be obtained. Therefore, the
dynamic parallel graph cuts algorithm can converge within no
more than K logN
` iterations under these settings. By setting
K = 1 and ` = N , only one iteration is needed in this case.
Dynamic splitting is also important in some cases. Since
the connection strengths may vary dramatically even in a
uniformly connected graph, it is difficult to design an optimal
scheme to divide the graph into subgraphs at the beginning.
In addition, since the commonly used equal division strategy
in most cases can only balance the memory storage, not the
computational load, and because the situation could be further
worsened by the dynamic merging process, dynamic splitting
is needed for dynamic workload balancing.
Incorporating dynamic merging and dynamic splitting enables the parallel granularity, load balance and configuration
of the subgraph splitting to be dynamically adjusted during
the parallel computation. Hence Algorithm 2 is here called
dynamic. Our dynamic parallel graph cuts algorithm is a
general framework that could be tailored to fit different types
of graphs and platforms for better efficiency and convergence
rate.
To summarize, our proposed dynamic parallel graph cuts
algorithm has the following three main advantages: 1) any
serial graph cuts solver that is efficient in the dynamic setting
can be parallelized using the dynamic parallel graph cuts
algorithm; 2) the dynamic parallel graph cuts algorithm guarantees convergence, moreover, it can converge to the global
optimum within a predefined number of iterations; 3) the
subgraph splitting configuration can be adjusted freely during
the parallelization process.
VI. E XPERIMENTS
Although advanced splitting and merging strategies could
be cleverly used to adjust the subgraph splitting configuration
during the whole parallelization process, finding such a general
strategy is not an easy task. For a fair comparison with the
parallel BK-algorithm, only the BK-algorithm is used as the
graph cuts solver in the iterations of all the parallel graph
cuts algorithms. Therefore, in all the following experiments,
only the naive converged parallel BK-algorithm (Algorithm
1) is used, in comparison with the parallel BK-algorithm, to
10
TABLE III: Number of images that have triggered merging operations
for seg2.
ITER
TRDS
4
8
15
20
25
30
successImages
290
197
89
91
49
66
40
54
486
362
evaluate the effectiveness of our introduced merging method.
There are only two parameters for all these experiments: ITER
controls the starting time of the merging operations in the naive
converged parallel BK-algorithm, and TRDS is the number of
the used computational threads for the two parallel algorithms.
A. Image segmentation
The performance of the naive converged parallel BKalgorithm is first evaluated on the two image segmentation
problems, seg1 and seg2, on all 500 images of the Berkeley
segmentation dataset [22], which are used in section II-B
to evaluate the convergence problem in the parallel BKalgorithm. TRDS is set to 4 and 8, respectively. And for each
settings of TRDS, ITER is set to 15, 20, 25, 30, respectively.
We further define three parameters, tBK , tPBK and tCPBK , to
represent the time elapsed for maxflow calculations using
the BK-algorithm, the parallel BK-algorithm and the naive
converged parallel BK-algorithm, respectively. This image
dataset can be divided into two parts according to whether
the images can converge or not in the parallel BK-algorithm
– referred to as successImages and failedImages, respectively.
The relative times (tCPBK /tBK ) for seg1 on the failedImages, which contains all the images of the dataset in
these cases under 4 and 8 computational threads, are shown
in Fig. 6(a) and Fig. 6(b), respectively. Although all these
images that failed to converge in the parallel BK-algorithm
can certainly converge in the naive converged parallel BKalgorithm, they usually take a little more time than in the serial
BK-algorithm, as shown in Fig. 6(a),(b). The improper settings
of ITER, e.g., ITER = 15, will degrade the performance of
the naive converged parallel BK-algorithm. This is possibly
because the subgraphs could not accumulate enough flows
for reuse, and the merging operations would be invoked too
many times, more than the necessary, if ITER is too small.
However, the naive converged parallel BK-algorithm can come
up with similar results with a wide range of ITER, e.g.,
ITER = 20, 25, 30.
For seg2, successImages accounts for the majority of the
500 images. Table III shows the number of images that belong
to successImages but that also triggered the merging operations
in the naive converged parallel BK-algorithm under different
ITER and TRDS settings. It is evident from Table III that
the small ITER values will lead to a large number of images
to invoke the merging operations. Experiments are carried out
on successImages and failedImages, as shown in Fig. 6(c)-(d),
and Fig. 6(e)-(f), respectively. The relative times (tPBK /tBK )
of the parallel BK-algorithm on successImages are also shown
in Fig. 6(c)-(d), shown as ORG in black, for comparison. It
can be seen from Fig. 6(c)-(d) that except for the ITER = 15
case, which is slightly inferior to the parallel BK-algorithm,
the naive converged parallel BK-algorithm performs equally
as well as the parallel BK-algorithm under all the other ITER
and TRDS settings. Moreover, the naive converged parallel
BK-algorithm is approximately 30% faster, on average, than
the serial BK-algorithm on failedImages, as shown in Fig.
6(e)-(f). This is a promising result. It shows that the naive
converged parallel BK-algorithm not only has convergence
guarantee but may also have a considerable speedup even
for these images for which the parallel BK-algorithm failed
to converge. Concerning the parameter selection of the naive
converged parallel BK-algorithm, because the algorithm is not
sensitive to the settings of ITER, it can be set to a number
in the range 15 − 30, and TRDS can be set to a number
in the range 2 − 8. Usually, the larger TRDS is, the faster
the naive converged parallel BK-algorithm is. However, the
speedup gain grows very slowly if TRDS is further increased.
B. Semantic segmentation
Among many other multi-label problems in computer vision [?], where maximum flow solvers are invoked many
times as a subroutine for each problem instance, such as
stereo, image denoising, semantic segmentation might be one
of the most challenging applications for the two parallelized
BK-algorithms because of the following two difficulties: 1)
typically, only a few labels (object classes) can appear in an
image, but the label set is relatively large, so as to contain
all the possible labels that can occur in the whole image set;
2) each label that appears is usually concentrated on a small
part of the image except for the background. Due to the first
difficulty, very few flows can be pushed from source to sink
for the graphs built for the move-making energies that moved
to the non-existent labels of the image, which makes the overhead much more significant for creating, synchronizing and
terminating multiple threads. The latter difficulty makes the
workload very unbalanced among the computational threads.
We used the semantic segmentation algorithm proposed by
Shotton [31] on the MSRC-21 dataset [31], where the unary
potentials for pixels are the negative log-likelihood of the
corresponding output of the classifiers derived from TextonBoost [31]. The most discriminative weak classifiers are found
using multi-class Gentle Ada-Boost [32], and the pairwise
potentials are the contrast sensitive potentials [33] defined
on the eight-neighborhood system. The energy functions are
solved via α-expansion [34].
Fig. 7(a) shows the relative times for the parallel BKalgorithm using 2, 4, 8 and 12 computational threads on all
the test images of the MSRC-21 dataset. It can be seen that
the parallel BK-algorithm is always slower than the serial
BK-algorithm, except for the 8 computational threads case, in
which the parallel BK-algorithm is a bit faster than the BKalgorithm. Although the acceleration is poor, only one out of
the 295 test images is non-convergent for the 4, 8 and 12 computational threads cases. Moreover, only when α = 16, which
corresponds to solving the ‘road’-expansion energy function,
does the parallel BK-algorithm failed to converge. This nonconvergent example is shown in Fig. 8, where Fig. 8(a) is the
11
(a) Relative times for seg1 on failedImages of 4
computational threads. The red, blue, green and
cyan lines represent tCPBK /tBK with ITER
set to 15, 20, 25 and 30, respectively, and their
medians are 1.58, 1.23, 1.34 and 1.17.
(b) Relative times for seg1 on failedImages of 8
computational threads. The red, blue, green and
cyan lines represent tCPBK /tBK with ITER
set to 15, 20, 25 and 30, respectively, and their
medians are 1.89, 1.74, 1.91 and 2.03.
(c) Relative times for seg2 on successImages of
4 computational threads. The ORG in black
represents tPBK /tBK , and its median is 0.56.
The red, blue, green and cyan lines represent
tCPBK /tBK with ITER set to 15, 20, 25 and 30,
respectively, and their medians are 0.68, 0.58,
0.57 and 0.55.
(d) Relative times for seg2 on successImages of
8 computational threads. The ORG in black
represents tPBK /tBK , and its median is 0.52.
The red, blue, green and cyan lines represent
tCPBK /tBK with ITER set to 15, 20, 25 and 30,
respectively, and their medians are 0.56, 0.50,
0.50 and 0.49.
(e) Relative times for seg2 on all 14 failedImages of 4 computational threads. The red, blue,
green and cyan lines represent tCPBK /tBK with
ITER set to 15, 20, 25 and 30, respectively, and
their medians are 0.75, 0.71, 0.73 and 0.70.
(f) Relative times for seg2 on all 138 failedImages of 8 computational threads. The red, blue,
green and cyan lines represent tCPBK /tBK with
ITER set to 15, 20, 25 and 30, respectively, and
their medians are 0.66, 0.69, 0.72 and 0.73.
Fig. 6: Relative times for the two segmentation methods
(a) Relative times (tPBK /tBK ) for semantic segmentation on all the images
of the MSRC-21 dataset that can converge in the parallel BK-algorithm
with TRDS set to 2, 4, 8 and 12. The medians are 1.32, 1.14, 0.97 and
1.29.
(b) Relative times (tCPBK /tPBK ) on the unnecessary merged images in
the naive converged parallel BK-algorithm with TRDS set to 8 and 12
and ITER set to 15 and 25. The medians are 1.01, 1.00, 0.96 and
0.96.
Fig. 7: Relative times for semantic segmentation on the MSRC-21 dataset
original image and Fig. 8(b) is the hand-labeled ground truth.
The different pixel colors represent different object classes,
whose correspondence is shown in Fig. 8(h). The black pixels
are the ‘void’ pixels, which are ignored when computing
semantic segmentation accuracies and are treated as unlabeled
pixels during training. The semantic segmentation result when
the serial BK-algorithm is used as the graph cuts solver is
shown in Fig. 8(c). It is the benchmark of the correctness for
all the other parallel graph cuts algorithms, and its running
time acts as the baseline. The semantic segmentations by the
parallel BK-algorithm with 4 and 8 computational threads are
shown in Fig. 8(d) and Fig. 8(e) respectively. Fig. 8(f) shows
the result of the naive converged parallel BK-algorithm with
8 computational threads. The dotted blue lines in Fig. 8(d)–(f)
are the boundaries of the neighboring subgraphs. In contrast
to Fig. 1, the graph is horizontally split for parallel graph cuts
computation.
Since the parallel BK-algorithm is not convergent by only
expanding the road class, the differences between the semantic
segmentations of the parallel BK-algorithm and those of the
BK-algorithm and the naive converged parallel BK-algorithm
lie only on the left border between ‘building’ and ‘road’, as
shown in the white dotted boxes in Fig. 8(b)–(f). For better
clarity, these white boxes are enlarged in Fig. 8(g), where the
result of the naive converged parallel BK-algorithm, which
is the same as that of the BK-algorithm, is shown in the
upper left. Since the parallel BK-algorithm obtained the same
incorrect result under 4 and 8 threads, it is in the bottom
left. The image and the ground truth are in the upper right
and bottom right, respectively. It is clear that the accuracies,
defined as the intersection vs union, of both the building and
road classes of the parallel BK-algorithm are worse than those
of the naive converged parallel BK-algorithm and the BKalgorithm. Because the road pixels in the white box of the
12
(a) Original image
(b) Ground truth
(c) The BK-algorithm
(d) The parallel BK-algorithm, 4 threads
(e) The parallel BK-algorithm, 8 threads
(f) The naive converged parallel BK-algorithm, 8
threads
(g) Zoom in
(h) Color-coding legend
FigureFig.
1: Example
resultsexample
of ourofnew
simultaneous
object class recognition and segmentation
8: Non-converged
the parallel
BK-algorithm
algorithm. Up to 21 object classes (color-coded in the key) are recognized, and the corresponding object
instances segmented in the images. For clarity, textual labels have been superimposed on the resulting
segmentations.
Note, foraccuracy
instance, how
the airplane
has been correctly
recognized
and separated
ground truth are all labeled ‘void’,
only the building
the parallel
BK-algorithm
on average
but are slightly
shorterfrom the
sky, and
the grass
theseofexperiments
one singlewith
learned
multi-class model has
of the parallel BK-algorithm,building,
which isthe
85.7%,
is worse
thanlawn.
thanInthose
the parallel only
BK-algorithm
12 computational
been used to segment all the test images. Further results from this system are given in Figure 18.
those of the naive converged parallel BK-algorithm and the threads. This experiment shows that even in the most chalBK-algorithm, which are all 86.4%. The running time of the lenging case, the naive converged parallel BK-algorithm could
BK-algorithm in this example
is 0.564
whiletothe
have the
convergence
guarantee
with almost
no additional
time recognimination,
andseconds,
to be robust
occlusion.
Our
focus problems
typically
associated
with object
is notlonger:
only the
accuracy
of segmentation
and the
recogtionBK-algorithm.
techniques that rely on sparse features (such as
parallel BK-algorithm takes much
0.998
seconds
with cost over
parallel
nition,
but alsoThe
thetime
efficiency
of the algorithm, which [33, 36]). These problems are mainly related to tex4 threads and 1.150 seconds with
8 threads.
the naive
becomes
important
converged parallel BK-algorithm
takenparticularly
is comparable
to that when dealing with tureless or very highly textured image regions. FigC. Deployment onure
distributed
large image
collections
or 15,
video
2 shows platforms
some examples of images with which
of the BK-algorithm: with 4 threads
if ITER
is set to
20,sequences.
At a local level, the appearance of an image patch those techniques would very likely struggle. In con25, it takes 0.560, 0.556 and 0.571 seconds, respectively; with
The dynamic parallel graph cuts algorithm (Algorithm 2)
leads to ambiguities in its class label. For example, trast, our technique based on dense features is ca8 threads if ITER is set to 15, 20, 25, it takes 0.613, 0.610 can be deployed on distributed platforms, where both splitting
a window could be part of a car, a building or an pable of coping with both textured and untextured
and 0.618 seconds, respectively.
and merging
can objects,
take place
within
the same
or or selfand with
multiple
objectsmachine
which interairplane. To overcome these ambiguities,
it is necacross
different
machines.
Fig.
9
gives
such
an
example
on
an
The naive converged parallel
BK-algorithm
does
not
have
essary to incorporate longer range information such occlude, while retaining high efficiency.
MPICH2
cluster, where
yellow
stripes areinthethis
overlapping
the convergence problem andasisthe
much
faster
thanof the
paral- and
spatial
layout
an object
also contextual
Thethe
main
contributions
paper are threeregions
sameThe
machine
and the red
are the
lel BK-algorithm on the failure
example.
However,
for the image.
information
from
the surrounding
Towithin
achievethefold.
most significant
is astripes
novel type
of feature,
overlapping
regionswhich
between
two
this, we
construct a discriminative
for labeling
we call
themachines.
texture-layout filter. These features
other images, on which the parallel
BK-algorithm
successfully model
images
whichBK-algorithm
exploits all three
types ofThere
information:
record patterns
of iftextons,
and exploit
the textural
converged, the naive converged
parallel
may
is no substantial
difference
the merging
operations
appearance,
layout,
Our place
tech- within
appearance
of themachine.
object, itsThere
layout,
unnecessarily invoke mergingtextural
operations,
depending
on and
the context.
only take
the same
areand
twoits textunique
can extra
modelcost.
veryTo
long-range
ral subgraphs
context. Our
is a new dissetting of ITER, and must pay
some
evaluatecontextual
ways torelationmerge the
thatsecond
spreadcontribution
across different
extending
half the
size of machines,
the image.depending
criminative
model
combines
filthis “cost”, Fig.7(b) shows theships
relative
times over
of tCPBK
/tPBK
on whether
the that
merged
graph texture-layout
can fit in
Additionally, our technique overcomes several ters with lower-level image features, in order to proon all these images that get converged by the parallel BK- the memory of one of the machines. If it can, merging acts
algorithm but unnecessarily invoked merging operations by in the same way as on the shared memory platforms, except
the naive converged parallel BK-algorithm with TRDS set to that MPI [35] is2used for the inter-machine communication.
8 and 12 and ITER set to 15 and 25.
Fig. 9 illustrates such a case, where the rightmost subgraph
It can be seen from Fig. 7(b) that for both ITER settings, the in the left machine is merged into the machine on the right
running times of the naive converged parallel BK-algorithm side. The communication overheads in this case are significant
with 8 computational threads are roughly equal to those of and largely depend on the size of the subgraph being merged.
13
V3
V6
V4
V1
MPI
V2 \V3
V3 \V4
V3 \V4
V6
(V3nV2) [ V4
MPI
V1 \V2
V5
V2
V5
V2
V1
V4 \V5
V5 \V6
(a) Before merging
V1 \V2
V2 \V3
V2 \V3
V4 \V5
V5 \V6
(b) After merging
Fig. 9: Dynamic parallel graph cuts algorithm on an MPICH2 cluster
If the merged graph cannot fit in the memory of a single
machine, the current splitting can only be “adjusted”, not
simply merged. “Adjusting” is achieved via first merging a
small portion with the overlapped region and then re-splitting
on the new boundary. Therefore only a small part rather than
the whole subgraph is merged in this case. It is worth noting
that Theorem 2 holds in all cases, such that the merged graph
is a reparameterization of the original graph with less flows.
Also, note that besides the parallel BK-algorithm, any other
distributed graph cuts algorithms could also be used here.
We use some of the big instances in [36] to evaluate our
naive converged parallel BK-algorithm on a 2 × 2 grid of an
MPICH2 cluster. The whole graph is split across the four
machines and further split into two subgraphs within each
machine. In contrast to all the above experiments, ITER has
a great impact on the performance of the naive converged
parallel BK-algorithm in this distributed setting. The LB07bunny-lrg dataset, which was solved in 6.3 seconds by the
parallel BK-algorithm, was solved in 6.4, 6.4 and 6.6 seconds
by the naive converged parallel BK-algorithm when ITER was
set to 25, 20 and 15, respectively. However, further reducing
ITER would dramatically increase the running time. The naive
converged parallel BK-algorithm took 25.5 seconds when
ITER = 10, whereas it took 70 seconds when ITER = 5.
The underlying reason is that network speed is much slower
than RAM speed, such that merging subgraphs across different
machines is a time-consuming operation. When ITER is
relatively small, subgraphs across different machines will be
merged. For instance, in the case of ITER = 5, merging
operation was invoked three times such that all eight subgraphs across four machines were merged into one graph, and
communication took more than 40 out of a total 70 seconds.
Thus the naive converged parallel BK-algorithm may pay some
higher running time cost for the guaranteed convergence under
the above simple merging and splitting strategies.
VII. C ONCLUSION AND FURTHER WORK
To remedy the non-convergence problem in the parallel
BK-algorithm, we first introduced a merging method along
with the naive converged parallel BK-algorithm that simply
merges every two neighboring subgraphs once the algorithm is
found to have difficulty converging under the current subgraph
splitting configuration. We then introduced a new pseudoboolean representation for graph cuts, namely the restricted
homogeneous posiforms. We further developed an invariance
analysis method for graph cuts algorithms and proved the
correctness and effectiveness of our proposed merging method.
Based on the merging method, we proposed a general parallelization framework called the dynamic parallel graph cuts
algorithm, which allows the subgraph splitting configuration
to be adjusted freely during the parallelization process and has
guaranteed convergence at the same time. Extensive experiments for image segmentation and semantic segmentation with
different number of computational threads demonstrated that
even under the simplest merging strategy, our proposed naive
converged parallel BK-algorithm not only has guaranteed
convergence but also has competitive computational efficiency
compared with the parallel BK-algorithm.
How to split a graph for parallel computation and how
to dynamically adjust the subgraph splitting configuration
during the parallelization process in order to further boost the
efficiency and accelerate the convergence rate are our future
research directions.
ACKNOWLEDGMENT
This work was supported by National Natural Science Foundation of China under grants 61333015, 61421004, 61473292.
We also thank Dr. Nilanjan Ray and the anonymous reviewers
for their valuable comments.
R EFERENCES
[1] P. Strandmark and F. Kahl, “Parallel and distributed graph cuts by dual
decomposition,” in Computer Vision and Pattern Recognition (CVPR),
2010 IEEE Conference on. IEEE, 2010, pp. 2085–2092. 1, 2, 9
[2] P. Strandmark, F. Kahl, and T. Schoenemann, “Parallel and distributed
vision algorithms using dual decomposition,” Computer Vision and
Image Understanding, vol. 115, no. 12, pp. 1721–1732, 2011. 1, 2,
9
[3] A. Shekhovtsov and V. Hlavac, “A distributed mincut/maxflow algorithm
combining path augmentation and push-relabel,” International Journal
of Computer Vision, vol. 104, no. 3, pp. 315–342, 2013. 1, 2
[4] A. Blake, P. Kohli, and C. Rother, Markov random fields for vision and
image processing. MIT Press, 2011. 1
[5] D. D. Sleator and R. E. Tarjan, “A data structure for dynamic trees,”
in Proceedings of the thirteenth annual ACM symposium on Theory of
computing. ACM, 1981, pp. 114–122. 1
[6] Y. Boykov and V. Kolmogorov, “An experimental comparison of mincut/max-flow algorithms for energy minimization in vision,” Pattern
Analysis and Machine Intelligence, IEEE Transactions on, vol. 26, no. 9,
pp. 1124–1137, 2004. 1, 9
[7] L. R. Ford and D. R. Fulkerson, “Maximal flow through a network,”
Canadian journal of Mathematics, vol. 8, no. 3, pp. 399–404, 1956. 1
[8] A. V. Goldberg, S. Hed, H. Kaplan, R. E. Tarjan, and R. F. Werneck,
“Maximum flows by incremental breadth-first search,” in Algorithms–
ESA 2011. Springer, 2011, pp. 457–468. 1, 9
[9] A. V. Goldberg and R. E. Tarjan, “A new approach to the maximumflow problem,” Journal of the ACM (JACM), vol. 35, no. 4, pp. 921–940,
1988. 1
14
[10] B. V. Cherkassy and A. V. Goldberg, “On implementing push-relabel
method for the maximum flow problem,” in Proceedings of the
4th International IPCO Conference on Integer Programming and
Combinatorial Optimization. London, UK, UK: Springer-Verlag,
1995, pp. 157–171. [Online]. Available: http://dl.acm.org/citation.cfm?
id=645586.659457 1
[11] D. S. Hochbaum, “The pseudoflow algorithm: A new algorithm for the
maximum-flow problem,” Operations research, vol. 56, no. 4, pp. 992–
1009, 2008. 1, 9
[12] A. V. Goldberg, “The partial augment–relabel algorithm for the maximum flow problem,” in Algorithms-ESA 2008. Springer, 2008, pp.
466–477. 1
[13] ——, “Processor-efficient implementation of a maximum flow algorithm,” Information processing letters, vol. 38, no. 4, pp. 179–185, 1991.
1
[14] D. A. Bader and V. Sachdeva, “A cache-aware parallel implementation
of the push-relabel network flow algorithm and experimental evaluation
of the gap relabeling heuristic.” in ISCA PDCS, 2005, pp. 41–48. 1
[15] R. Anderson and J. C. Setubal, “A parallel implementation of the pushrelabel algorithm for the maximum flow problem,” Journal of parallel
and distributed computing, vol. 29, no. 1, pp. 17–26, 1995. 1
[16] A. Delong and Y. Boykov, “A scalable graph-cut algorithm for nd grids,”
in Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE
Conference on. IEEE, 2008, pp. 1–8. 1, 2
[17] V. Vineet and P. Narayanan, “Cuda cuts: Fast graph cuts on the gpu,” in
Computer Vision and Pattern Recognition Workshops, 2008. CVPRW’08.
IEEE Computer Society Conference on. IEEE, 2008, pp. 1–8. 1
[18] J. Liu and J. Sun, “Parallel graph-cuts by adaptive bottom-up merging,”
in Computer Vision and Pattern Recognition (CVPR), 2010 IEEE
Conference on. IEEE, 2010, pp. 2181–2188. 1
[19] H. Everett III, “Generalized lagrange multiplier method for solving
problems of optimum allocation of resources,” Operations research,
vol. 11, no. 3, pp. 399–417, 1963. 2
[20] N. Komodakis, N. Paragios, and G. Tziritas, “Mrf optimization via dual
decomposition: Message-passing revisited,” in Computer Vision, 2007.
ICCV 2007. IEEE 11th International Conference on. IEEE, 2007, pp.
1–8. 2
[21] A. Bhusnurmath and C. J. Taylor, “Graph cuts via ell 1 norm minimization,” Pattern Analysis and Machine Intelligence, IEEE Transactions on,
vol. 30, no. 10, pp. 1866–1871, 2008. 2
[22] P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik, “Contour detection
and hierarchical image segmentation,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 33, no. 5, pp. 898–916, May 2011. [Online]. Available:
http://dx.doi.org/10.1109/TPAMI.2010.161 3, 8, 10
[23] E. Boros and P. L. Hammer, “Pseudo-boolean optimization,” Discrete
applied mathematics, vol. 123, no. 1, pp. 155–225, 2002. 4, 5
[24] P. L. Hammer, “Some network flow problems solved with pseudoboolean programming,” Operations Research, vol. 13, no. 3, pp. 388–
399, 1965. 4, 5
[25] E. Boros, P. L. Hammer, and G. Tavares, “Preprocessing of unconstrained quadratic binary optimization,” Tech. Rep., 2006. 4, 5
[26] D. Freedman and P. Drineas, “Energy minimization via graph cuts:
settling what is possible,” in Computer Vision and Pattern Recognition,
2005. CVPR 2005. IEEE Computer Society Conference on, vol. 2, June
2005, pp. 939–946 vol. 2. 5
[27] P. Kohli and P. H. Torr, “Dynamic graph cuts for efficient inference
in markov random fields,” Pattern Analysis and Machine Intelligence,
IEEE Transactions on, vol. 29, no. 12, pp. 2079–2088, 2007. 9
[28] ——, “Efficiently solving dynamic markov random fields using graph
cuts,” in Computer Vision, 2005. ICCV 2005. Tenth IEEE International
Conference on, vol. 2. IEEE, 2005, pp. 922–929. 9
[29] O. Juan and Y. Boykov, “Active graph cuts,” in Computer Vision and
Pattern Recognition, 2006 IEEE Computer Society Conference on, vol. 1.
IEEE, 2006, pp. 1023–1029. 9
[30] A. V. Goldberg, S. Hed, H. Kaplan, P. Kohli, R. E. Tarjan, and R. F.
Werneck, “Faster and more dynamic maximum flow by incremental
breadth-first search,” in Algorithms–ESA 2015. Springer, 2015, pp.
619–630. 9
[31] J. Shotton, J. Winn, C. Rother, and A. Criminisi, “Textonboost: Joint
appearance, shape and context modeling for multi-class object recognition and segmentation,” in Computer Vision–ECCV 2006. Springer,
2006, pp. 1–15. 10
[32] A. Torralba, K. P. Murphy, and W. T. Freeman, “Sharing features: efficient boosting procedures for multiclass object detection,” in Computer
Vision and Pattern Recognition, 2004. CVPR 2004. Proceedings of the
2004 IEEE Computer Society Conference on, vol. 2. IEEE, 2004, pp.
II–762. 10
[33] Y. Boykov and M.-P. Jolly, “Interactive graph cuts for optimal boundary
amp; region segmentation of objects in n-d images,” in Computer Vision,
2001. ICCV 2001. Proceedings. Eighth IEEE International Conference
on, vol. 1, 2001, pp. 105–112 vol.1. 10
[34] Y. Boykov, O. Veksler, and R. Zabih, “Fast approximate energy minimization via graph cuts,” Pattern Analysis and Machine Intelligence,
IEEE Transactions on, vol. 23, no. 11, pp. 1222–1239, 2001. 10
[35] M. Snir, S. Otto, S. Huss-Lederman, D. Walker, and J. Dongarra, MPIThe Complete Reference, Volume 1: The MPI Core, 2nd ed. Cambridge,
MA, USA: MIT Press, 1998. 12
[36] U. of Western Ontario, “Max-flow problem instances in vision,” in
http://vision.csd.uwo.ca/data/maxflow. 13
A PPENDIX A
P ROOF OF P ROPOSITION 1
Proof: Fig.4(b)–(d) give an example of splitting, where
the capacities of the edges in the two split subgraphs, whose
two terminals are common for the two subgraphs, are weighted
by 1/2 and all the other capacities simply come from those of
their corresponding edges in the original graph. Therefore, the
following relations hold for the coefficients of the restricted
(1)
(1)
homogeneous posiforms of the two split subgraphs G1 , G2
and the original graph G:
(1)
(1)
ai , ∀i ∈ V1\V2 ,
(43)
ai , b
a1i = b
b
a1i = b
¨
˙
¨
˙
(1)
(1)
2
2
Ê
ai = Ê
ai , Ê
ai = Ê
ai , ∀i ∈ V2\V1 ,
(44)
˙
˙
¨
¨
(1)
(1)
(1)
(1)
1
1
a
¯1i = a
¯2i = a
¯i , a
¯1i = a
¯i , ∀i ∈ V1 ∩V2 , (45)
¯2i = a
2
2
˙
¨
¨
¨
˙
˙
(1)
b
a1ij = b
aij , ∀i or j ∈ V1 \ V2 ,
(1)
Ê
a2ij = Ê
aij , ∀i or j ∈ V2 \ V1 ,
(46)
(47)
1
āij , ∀i, j ∈ V1 ∩V2 .
(48)
2
And from the relations of the coefficients of the restricted
homogeneous posiforms and the multi-linear polynomials,
stated in (??)–(??), it is easy to verify that all the relations
listed in this theorem hold.
(1)
(1)
ā1ij = ā2ij =
A PPENDIX B
P ROOF OF P ROPOSITION 2
Proof: All the flows are pushed through the augmenting
paths in the augmenting path maxflow algorithms. Without loss
of generality, suppose s → i1 → i2 → · · · → ik → t is such
an augmenting path in graph G with the minimal capacity of
forward arc (edge) greater than or equal to F. Therefore a flow
of F can be pushed along this augmenting path, resulting in a
residual graph G0 . It is clear that graph G and residual graph
G0 only differ in these forward and backward arcs, such that
the restricted homogeneous posiforms of the two graphs can
be written as:
a
r
G
G
φG
h (x) = φh (x) + φh (x),
a
0
0
r
0
G
G
φG
h (x) = φh (x) + φh (x),
a
a
0
(49)
(50)
G
where φG
h (x) and φh (x) contain all the components corresponding to all the forward and backward arcs ralong the
augmenting
path of graph G and G0 respectively. φG
h (x) and
r 0
G
φh (x) contain all the rest components of graph G and G0 ,
15
r
r
a
0
G
G
such that φG
h (x) = φh (x). Suppose that φh (x) is in the
following form:
a
φG
h (x) = ai1 xi1 +ai1 i2 x̄i1 xi2 +ai2 i1 x̄i2 xi1 +· · ·+aik x̄k , (51)
where all the coefficients of the forward arcs (e.g. ai1 i2 ) are
greater than or equal to F. Pushing a flow F in this augmenting
path will cause all the capacities of the forward arcs reduced
by F and all the
capacities of the backward arcs increased by
a 0
F. Therefore φG
h (x) is in the following form
a
A PPENDIX E
A N I LLUSTRATIVE E XAMPLE OF E QUIVALENT G RAPH
C UTS P ROBLEMS
Here we give an illustrative example on the equivalence
of graph cuts problems. For the three graph cuts problems
defined on the graphs shown in Fig. E.1(a)-(c) respectively,
their unique restricted homogeneous posiforms, according to
Theorem 1, are
0
φG
h (x) = (ai1 − F)xi1 + (ai1 i2 − F)x̄i1 xi2 +
(ai2 i1 + F)x̄i2 xi1 + · · · + (aik − F)x̄k .
φG
h (x)
−
a
0
φG
h (x)
(53)
it is easy to verify xi1 + x̄i1 xi2 − x̄i2 xi1 + · · · + x̄k ≡ 1, hence
a
0
G
φG
h (x) − φh (x) ≡ F.
(54)
2
2
1
2
1
5
2
t
0
3
3
1
2
0
2
4
t
(a) Original graph G0
0
(b) Residual graph G1
0
t
(c) Residual graph G2
Fig. E.1: Three equivalent graph cuts problems.
0
G
Or φG
h (x) = φh (x) + F, and Proposition 2 is proved.
A PPENDIX C
P ROOF OF P ROPOSITION 3
Proof: Updating operations only modify the T-link capacities of all the nodes within V1 ∩V2 in the following way:
(k)
6
s
0
2
=
F × (xi1 + x̄i1 xi2 − x̄i2 xi1 + · · · + x̄k ) ,
(k+1)
4
1
1
a
s
0
Therefore,
a
s
(52)
(k+1)
a
¯1i
=a
¯1i + 4λi
, ∀i ∈ V1 ∩V2 ,
(55)
˙
˙
(k+1)
(k)
(k+1)
a
¯2i
=a
¯2i + 4λi
, ∀i ∈ V1 ∩V2 .
(56)
¨
¨
And from (??) and (??), it is easy to verify that all the
conclusions in this proposition hold.
0
φG
h (xV ) = 6x̄1 + 4x2 + 2x̄2 + x̄1 x2 + 2x̄2 x1 ,
(63)
1
φG
h (xV ) = 5x̄1 + 3x2 + 2x̄2 + 2x̄1 x2 + x̄2 x1 ,
(64)
2
φG
h (xV )
(65)
= 4x̄1 + 3x̄1 x2 .
And their corresponding multi-linear polynomials are
f G0 (xV ) = 8 − 4x1 + 3x2 − 3x1 x2 ,
G1
(66)
(xV ) = 7 − 4x1 + 3x2 − 3x1 x2 ,
(67)
f G2 (xV ) = 4 − 4x1 + 3x2 − 3x1 x2 .
(68)
f G0 (xV ) = f G1 (xV ) + 1 = f G2 (xV ) + 4.
(69)
f
Obviously,
A PPENDIX D
P ROOF OF P ROPOSITION 4
Proof: From the merging operation described in section
III and Theorem 1, the following relations hold for the
coefficients of the restricted homogeneous posiforms of the
(k+1)
merged graph G0 and those of the two subgraphs G1
,
(k+1)
G2
(k+1)
(k+1)
b
a0i = b
a1i
, b
a0i = b
a1i
, ∀i ∈ V1\V2
˙
˙
¨
¨
0
2(k+1)
0
2(k+1)
Ê
ai = Ê
ai
, Ê
ai = Ê
ai
, ∀i ∈ V2\V1
˙
˙
¨
¨
(k+1)
(k+1)
a
¯0i = a
¯1i
+a
¯2i
˙
˙ (k+1) ˙ (k+1) , ∀i ∈ V1 ∩V2
a
¯0i = a
¯2i
+a
¯1i
¨
¨
¨
(k+1)
b
a0ij = b
a1ij
, ∀i or j ∈ V1 \ V2
(k+1)
Ê
a0ij = Ê
a2ij
(k+1)
ā0ij = ā1ij
, ∀i or j ∈ V2 \ V1
(k+1)
+ ā2ij
, ∀i, j ∈ V1 ∩V2
(57)
(58)
(59)
(60)
(61)
(62)
And all the four equalities in Proposition 4 can be verified by
combining the relations of the coefficients of the restricted
homogeneous posiforms and the multi-linear polynomials,
stated in (??)–(??).
Therefore, f G0 (xV ) and f G1 (xV ) only differ by a constant
of 1, and f G0 (xV ) and f G2 (xV ) only differ by a constant of
4. Note that “1” equals the total flow pushed through the path
s → 2 → 1 → t in G0 , by which the residual graph G1 is
produced, and “4” equals the sum of the total flows pushed
through the path s → 2 → 1 → t, and those through the path
s → 2 → t in G0 , by which the residual graph G2 is produced.
Proposition 2 showed that pushing a flow F in graph G will
only reduce the multi-linear polynomial of G by a constant of
F. Here, the flow F varies by each augmenting path operation,
but it is independent of the variables of the pseudo-boolean
function representing the graph cuts problem. Hence the three
graph cuts problems are equivalent.
16
Miao Yu received his B.B.A. and M.S. both from
Southwest Jiaotong University in 2004 and 2007
respectively, and the Ph.D. from the Institute of
Automation, Chinese Academy of Sciences in 2016.
Since 2007, he has been with the Zhongyuan University of Technology, where he is now a lecturer.
His research interests include parallel computing,
probabilistic graphical models, higher order energy
minimization and scene understanding.
Shuhan Shen received his B.S. and M.S. both
from Southwest Jiao Tong University in 2003 and
2006 respectively, and the Ph.D. from Shanghai Jiao
Tong University in 2010. Since 2010, he has been
with the National Laboratory of Pattern Recognition
at Institute of Automation, Chinese Academy of
Sciences, where he is now an associate professor.
His research interests are in 3D computer vision,
which include image based 3D modeling of large
scale scenes, 3D perception for intelligent robot, and
3D semantic reconstruction.
Zhanyi Hu received his B.S. Degree in Automation
from the North China University of Technology in
1985, the Ph.D. Degree (Docteur d’État) in Computer Science from the University of Liege, Belgium,
in Jan. 1993. Since 1993, he has been with the Institute of Automation, Chinese Academy of Sciences.
From 1997 to 1998, he was a Visiting Scientist in
the Chinese University of Hong Kong. From 2001 to
2005, he was an executive panel member of National
High-Tech R& D Program (863 Program). From
2005 to 2010, he was a member of the Advisory
Committee of the National Natural Science Foundation of China. Dr. Hu is
now a research professor of computer vision, a deputy editor-in-chief for
Chinese Journal of CAD& CG, and an associate editor for Science in China,
and Journal of Computer Science and Technology. He was the organization
committee co-chair of ICCV2005, and a program co-chair of ACCV2012. Dr.
Hu has published more than 150 peer-reviewed journal papers, including IEEE
PAMI, IEEE IP, IJCV. His current research interests include biology-inspired
vision and large scale 3D reconstruction from images.
| 8cs.DS
|
2012 5th International Conference on BioMedical Engineering and Informatics (BMEI 2012)
Detect adverse drug reactions for the drug Pravastatin
Yihui Liu l' 2
'Institute of Intelligent Information Processing,Shandong
Polytechnic University, China
[email protected]
drug reaction (ADR) is widely concerned for
public health issue. ADRs are one of most common causes to
withdraw some drugs from market. Prescription event
monitoring (PEM) is an important approach to detect the
adverse drug reactions. The main problem to deal with this
method is how to automatically extract the medical events or
side effects from high-throughput medical data, which are
collected from day to day clinical practice. In this study we
propose an original approach to detect the ADRs using feature
matrix and feature selection. The experiments are carried out on
the drug Pravastatin. Major side effects for the drug are
detected. The detected ADRs are based on computerized
method, further investigation is needed.
Abstract—Adverse
Keywords-adverse drug reaction; feature matrix; feature
selection; Pravastatin
Uwe Aickelin 2
Department of Computer Science,
University of Nottingham, UK
2
(QLDS) [20]. But their methods achieve low accuracies for
detecting ADRs.
In this paper we propose feature selection approach to
detect ADRs from The Health Improvement Network
(THIN) database. First feature matrix, which represents the
medical events for the patients before and after taking drugs,
is created by linking patients' prescriptions and corresponding
medical events together. Then significant features are selected
based on feature selection methods, comparing the feature
matrix before patients take drugs with one after patients take
drugs. Finally the significant ADRs can be detected from
thousands of medical events based on corresponding features.
Experiments are carried out on the drug Pravastatin. Good
performance is achieved.
II. FEATUREMATRIXAND FEATURESELECTION
I. INTRODUCTION
Adverse drug reaction (ADR) is widely concerned for
public health issue. ADRs are one of most common causes to
withdraw some drugs from market [1]. Now two major
methods for detecting ADRs are spontaneous reporting
system (SRS) [2,3,4,5], and prescription event monitoring
(PEM) [6,7,8,9,10]. The World Health Organization (WHO)
defines a signal in pharmacovigilance as "any reported
information on a possible causal relationship between an
adverse event and a drug, the relationship being unknown or
incompletely documented previously" [11]. For spontaneous
reporting system, many machine learning methods are used to
detect ADRs, such as Bayesian confidence propagation neural
network (BCPNN) [12,13], decision support method [14],
genetic algorithm [15], knowledge based approach [16,17], etc.
One limitation is the reporting mechanism to submit ADR
reports [14], which has serious underreporting and is not able
to accurately quantify the corresponding risk. Another
limitation is hard to detect ADRs with small number of
occurrences of each drug-event association in the database.
Prescription event monitoring has been developed by the
Drug Safety Research Unit (DSRU) and is the first large-scale
systematic post-marketing surveillance method to use event
monitoring in the UK [8]. The health data are widely and
routinely collected. It is a key step for PEM methods to
automatically detect the ADRs from thousands of medical
events. In paper [18,19], MUTARA and HUNT, which are
based on a domain-driven knowledge representation
Unexpected Temporal Association Rule, are proposed to
signal unexpected and infrequent patterns characteristic of
ADRs, using Queensland Hospital morbidity data, more
commonly referred to as the Queensland Linked Data Set
A.
The THIN Database
The Health Improvement Network (THIN) is a
collaboration product between two companies of EPIC and
InPS. EPIC is a research organization, which provides the
electronic database of patient care records from UK and other
countries. There are 'Therapy' and 'Medical' databases in THIN
data. The "Therapy" database contains the details of
prescriptions issued to patients. Information of patients and the
prescription date for the drug can be obtained. The "Medical"
database contains a record of symptoms, diagnoses, and
interventions recorded by the general practice (GP) and/or
primary care team. Each symptom for patients forms a record.
By linking patient identifier, their prescriptions, and their
corresponding medical events (symptoms) together, feature
matrix to characterize the symptoms during the period before or
after patients take drugs is built.
B.
The Extraction of Feature Matrix
To detect the ADRs of drugs, first feature matrix is
extracted from THIN database, which describes the medical
events that patients occur before or after taking drugs. Then
feature selection method of Student's t-test is performed to
select the significant features from feature matrix containing
thousands of medical events. Figure 1 shows the process to
detect the ADRs using feature matrix. Feature matrix A
(Figure 2) describes the medical events for each patient
during 60 days before they take drugs. Feature matrix B
reflects the medical events during 60 days after patients take
drugs. In order to reduce the effect of the small events, and
save the computation time and space, we set 100 patients as a
group. Matrix X and Y are feature matrix after patients are
divided into groups.
D. Feature Selection Based on Student's t-test
Matrix X
Matrix A
GroupA
Significant
DrugID
Feature
THIN
database
selection
Group B
Matrix B
D r u g ID
Therapy
Features
Matrtg
4
Prescription
date
Feature extraction and feature selection are widely used in
biomedical data processing [21-27]. In our research we use
Student's t-test [28] feature selection method to detect the
significant ADRs from thousands of medical events. Student's
t-test is a kind of statistical hypothesis test based on a normal
distribution, and is used to measure the difference between
two kinds of samples.
E. Other parameters
•
Medical
A or B
Patients
The variable of ratio R, is defined to evaluate significant
changes of the medical events, using ratio of the patient
number after taking the drug to one before taking the drug.
The variable R represents the ratio of patient number after
taking the drug to the number of whole population having
one particular medical symptom.
2
Figure 1. The process to detect ADRs. Matrix A and B are
feature matrix before patients take drugs or after patients take
drugs. The time period of observation is set to 60 days.
Matrix X and Y are feature matrix after patients are divided
into groups. We set 100 patients as one group.
rv'ed ca
events
The ratio variables R, and R are defined as follows:
2
Ri= [N ,/sTB if NB *0;
NA
R2 =
Patients
0
0
0
Figure 2. Feature matrix. The row of feature matrix represents
the patients, and the column of feature matrix represents the
medical events. The element of events is 1 or 0, which
represents the patient having or not having the corresponding
medical symptom or event.
C. Medical Events and Readcodes
Medical events or symptoms are represented by medical
codes or Readcodes. There are 103387 types of medical
events in "Readcodes" database. The Read Codes used in
general practice (GP), were invented and developed by Dr
James Read in 1982. The NHS (National Health Service) has
expanded the codes to cover all areas of clinical practice. The
code is hierarchical from left to right or from level 1 to level
5. It means that it gives more detailed information from level
1 to level 5. Table 1 shows the medical symptoms based on
Readcodes at level 3 and at level 5. 'Other soft tissue
disorders' is general term using Readcodes at level 3. 'Foot
pain', 'Heel pain', etc., give more details using Readcodes at
level 5.
TABLE I.
TABLE I. MEDICAL EVENTS BASED ON READCODES AT
LEVEL 3 AND LEVEL 5.
Level
Level 3
Muscle
pain
Level 5
Readcodes
N24..00
N245.16
N245111
N245.13
N245700
N245.15
Medical events
Other soft tissue disorders
Leg pain
Toe pain
Foot pain
Shoulder pain
Heel pain
if NB = 0;
NAI N
where NB and NA represent the numbers of patients before
or after they take drugs for having one particular medical
event respectively. The variable N represents the number of
whole population who take drugs.
III . EX PERI MENT S A ND RESULT S
The drug Pravastatin is used to test our proposed method,
using 475 GPs data in THIN database. Student's t-test is
performed to select the significant features from feature
matrix, which represent the medical events having the
significant changes after patients take the drugs.
Drugs.com provides access to healthcare information
tailored for a professional audience, sourced solely from the
most trusted, well-respected and independant agents such as
the Food and Drug Administration (FDA), American Society
of Health-System Pharmacists, Wolters Kluwer Health,
Thomson Micromedex, Cerner Multum and Stedman's. In the
Drugs.com [29], side effects for Pravastatin are listed, such as
severe allergic reactions (rash; hives; itching; difficulty
breathing; tightness in the chest; swelling of the mouth, face,
lips, or tongue); change in the amount of urine produced;
chest pain; dark urine; fever, chills, or persistent sore throat;
flu-like symptoms; muscle pain, tenderness, or weakness (with
or without fever or fatigue); pale stools; red, swollen,
blistered, or peeling skin; severe stomach pain; vision changes;
yellowing of the eyes or skin, etc.
There are seven currently prescribed forms of statin drugs.
They are Rosuvastatin, Atorvastatin, Simvastatin, Pravastatin,
Fluvastatin, Lovastatin, Pitavastatin, muscle pain and
musculoskeletal events [29, 30] are two of the main side
effects of statin drugs.
1189
In experiments, we use two kinds of feature matrix. One
feature matrix is based on all medical events using Readcodes at
level 1-5 in order to observe the detailed symptoms. Another
one is based on the medical events using Readcodes at level 13,
combining the detailed information of Readcodes at level 4 and
5 into one general term using Readcodes at level 3. 10875
patients from half data of 475GPs are taking Pravastatin. 13438
medical events or symptoms are obtained using Readcodes at
level 1-5, and the feature matrix of 10875x13438 is formed to
describe the detailed information of patients. After grouping
them, 108x13435 feature matrix is obtained. For Readcodes at
level 1-3, 108x2764 feature matrix is obtained.
Table 2 shows the top 20 detected results in ascending
order of p value of Student's t-test, using Readcodes at level 15
and at level 1-3. The detected results are using p values less than
0.05, which represent the significant change after patients take
the drug. Table 3 shows the results in descending order of the
ratio of the number of patients after taking the drug to one
before taking the drug. Our detected results are consistent with
the listed results in [29,30]. Table 4 shows cancer information
related to the patients who take the drug Pravastatin.
IV. CONCLUSIONS
In this study we propose a novel method to successfully
detect the ADRs using feature matrix and feature selection. A
feature matrix, which characterizes the medical events before
patients take drugs or after patients take drugs, is created
from THIN database. The feature selection method of
Student's t-test is used to detect the significant features from
thousands of medical events. The significant ADRs, which
are corresponding to significant features, are detected. The
detected ADRs are based on our proposed method, further
investigation is needed.
REFERENCES
[1] G. Severino, and M.D. Zompo, "Adverse drug reactions: role of
pharmacogenomics," Pharmacological Research, vol. 49, pp. 363-373,
2004.
[2] Y. Qian, X. Ye, W. Du, J. Ren, Y. Sun, H. Wang, B. Luo, Q. Gao, M.
Wu, and J. He, "A computerized system for signal detection in
spontaneous
reporting
system
of
Shanghai
China,"
Pharmacoepidemiology and Drug Safety, vol. 18, pp. 154-158, 2009.
[3] K S. Park, and 0. Kwon, "The state of adverse event reporting and
signal generation of dietary supplements in Korea," Regulatory
Toxicology and Pharmacology, vol. 57, pp. 74-77, 2010.
[4] A. Farcas, A. Sinpetrean, C. Mogosan, M. Palage, 0. Vostinaru, M.
Bojita, and D. Dumitrascu, "Adverse drug reactions detected by
stimulated spontaneous reporting in an internal medicine department
in Romania," European Journal of Internal Medicine, vol. 21, pp. 453457, 2010.
[5] A. Szarfman , S.G. Machado, and R.T. O'Neill, "Use of screening
algorithms and computer systems to efficiently signal higher-thanexpected combinations of drugs and events in the US FDA's
spontaneous reports database," Drug Saf, vol. 25, pp. 381-392, 2002.
[6] R. Kasliwal, L.V. Wilton, V. Cornelius, B. Aurich-Barrera, S.A.W.
Shakir, "Safety profile of Rosuvastatin-results of a prescription-event
monitoring study of 11680 patients," Drug Safety, vol. 30, pp. 157-170,
2007.
[7] R.D. Mann, K. Kubota, G. Pearce, and L Wilton, "Sahneterol: a study
by prescription-event monitoring in a UK cohort of 15,407 patients," J
Clin Epidemiol, vol. 49, pp. 247-250, 1996.
[8] N.S. Rawson, G.L. Pearce, and W.H. Inman , "Prescription-event
monitoring: Methodology and recent progress," Journal of Clinical
Epidemiology, vol. 43, pp. 509-522, 1990.
[9] B. Aurich-Barrera, L Wilton, D. Brown, and S. Shakir, "Paediatric
postmarketing phannacovigilance using prescription-event monitoring:
comparison of the adverse event profile of lamotrigine prescribed to
children and adults in England," Drug Saf. , vol. 33, pp. 751-763, 2010.
[10] T,J. Hannan, "Detecting adverse drug reactions to improve patient
outcomes," International Journal of Medical Informatics, vol. 55, pp.
61-64, 1999.
[11] R.H. Meyboom, M. Lindquist, A.C. Egberts, and I.R. Edwards, "Signal
detection and follow-up in phannacovigilance," Drug Saf. , vol. 25, pp.
459-465, 2002.
[12] A. Bate, M. Lindquist, I.R. Edwards, S. Olsson, R. Orre, A. Lansner,
and R.M. De Freitas, "A Bayesian neural network method for adverse
drug reaction signal generation," Eur J Clin Phannacol, vol. 54, pp.315321, 1998.
[13] R. Orre, A. Lansner, A. Bate, and M. Lindquist, "Bayesian neural
networks with confidence estimations applied to data mining,"
Computational Statistics & Data Analysis, vol. 34, pp. 473-493, 2000.
[14] M. Hauben and A. Bate, "Decision support methods for the detection
of adverse events in post-marketing data," Drug Discovery Today, vol.
14, pp. 343-357, 2009.
[15] Y. Koh, C.W. Yap, and S.C. Li, "A quantitative approach of using
genetic algorithm in designing a probability scoring system of an
adverse drug reaction assessment system," International Journal of
Medical Informatics, vol. 77, pp. 421-430, 2008.
[16] C. Henegar, C. Bousquet, A.L. Lillo-Le, P. Degoulet, and M.C. Jaulent,
"A knowledge based approach for automated signal generation in
phannacovigilance," Stud Health Technol Inform, vol. 107, pp. 626630, 2004.
[17] C. Bousquet , C. Henegar , A.L. Lou& , P. Degoulet, and M.C. Jaulent,
Implementation of automated signal generation in phannacovigilance
using a knowledge-based approach. International Journal of Medical
Informatics vol.74, pp. 563-571, 2007.
[18] H. Jin, J. Chen, H. He, G.J. Williams, C. Kelman, C.M. O'Keefe,
"Mining unexpected temporal associations: applications in detecting
adverse drug reactions," IEEE Trans Inf Technol Biomed, vol. 12, pp.
488-500, 2008.
[19] H. Jin, J. Chen, H. He, C. Kelman, D. McAullay, C.M. O'Keefe,
"Signaling Potential Adverse Drug Reactions from Administrative
Health Databases," IEEE Trans. Knowl. Data Eng., vol. 22, pp. 839853, 2010.
[20] G. Williams , D. Vickers , C. Rainsford , L Gu , H. He , R. Baxter, and
S. Hawkins, Bias in the Queensland linked data set, CSIRO Math. Inf.
Sci., 2002
[21] Y. Liu, "Feature extraction and dimensionality reduction for mass
spectrometry data," Computers in Biology and Medicine, vol. 39, pp.
818-823, 2009.
[22] Y. Liu, "Detect key genes information in classification of microarray
data," EURASIP Journal on Advances in Signal Processing, 2008.
Available: doi:10.1155/2008/612397.
[23] Y. Liu and L Bai, "Find significant gene information based on changing
points of microarray data," IEEE Transactions on Biomedical
Engineering, vol. 56, pp. 1108-1116, 2009.
[24] Y. Liu, M. Muftah, T. Das, L Bai, K Robson, and D. Auer,
"Classification of MR Tumor Images Based on Gabor Wavelet
Analysis," Journal of Medical and Biological Engineering, vol. 32, pp.
22-28, 2012.
[25] Y. Liu, "Dimensionality reduction and main component extraction of
mass spectrometry cancer data," Knowledge-Based Systems, vol. 26,
pp. 207-215, 2012.
[26] Y. Liu, "Wavelet feature extraction for high-dimensional microarray
data," Neurocomputing, vol. 72, pp. 985-990, 2009.
[27] Y. Liu, W. Aubrey, K Martin, A. Sparkes, C. Lu, and R. King, "The
analysis of yeast cell morphology features in exponential and stationary
Phase," Journal of Biological Systems, vol. 19, pp. 561-575, 2011.
1190
[28] E. Kreyszig, Introductory Mathematical Statistics, John Wiley, New
York, 1970.
TABLE II.
[29] http://www.drugs.com/sfx/pravastatin-side-effects.html.
[30] http://www.statinanswers.com/effects.htm
THE TOP POTENTIAL ADRS FOR PRAVASTATIN BASED ON P VALUE OF STUDENT'S T-TEST.
Rank
NA
R1
R2
Readcodes
Medical events
NB
1
1Z12.00
Chronic kidney disease stage 3
84
532
6.33
5.14
2
C34..00
Gout
85
280
329
2.71
3
1M10.00
Knee pain
102
421
4.13
4.07
4
1969.00
Abdominal pain
81
389
4.80
3.76
5
A53..11
Shingles
16
153
9.56
1.48
6
1D14.00
C/O: a rash
135
497
3.68
4.80
7
M03z000
Cellulitis NOS
37
252
6.81
2.43
8
N245.17
Shoulder pain
130
488
3.75
4.71
F4C0.00
Acute conjunctivitis
69
288
2.78
9
4.17
10
N131.00
Cervicalgia - pain in neck
129
419
325
4.05
11
1C9..00
Sore throat symptom
61
280
4.59
2.71
12
C10F.00
Type 2 diabetes mellitus
183
507
2.77
4.90
13
F46..00
Cataract
33
219
6.64
2.12
14
H06z000
Chest infection NOS
132
688
5.21
6.65
Level
15
N143.00
Sciatica
78
291
3.73
2.81
1-5
16
H01..00
Acute sinusitis
53
218
4.11
2.11
17
H060.00
Acute bronchitis
58
329
5.67
3.18
18
16C6.00
Back pain without radiation NOS
107
407
3.80
3.93
19
J16y400
Dyspepsia
122
384
3.15
3.71
20
N245.13
Foot pain
36
205
5.69
1.98
1
171..00
Cough
473
1799
3.80
16.54
2
H06..00
Acute bronchitis and bronchiolitis
383
1574
4.11
14.47
3
N21..00
Peripheral enthesopathies and allied syndromes
184
685
3.72
6.30
1Z1..00
139
758
5.45
6.97
4
Chronic renal impairment
5
M22..00
Other dermatoses
97
476
4.91
4.38
Level
6
1B8..00
Eye symptoms
125
542
4.34
4.98
1-3
7
N24..00
Other soft tissue disorders
750
1933
2.58
17.77
8
H05..00
Other acute upper respiratory infections
179
862
4.82
7.93
9
F4C..00
Disorders of conjunctiva
128
537
420
4.94
10
19F..00
Diarrhoea symptoms
147
633
4.31
5.82
11
1B1..00
General nervous symptoms
307
1052
3.43
9.67
12
N14..00
Other and unspecified back disorders
237
876
3.70
8.06
13
M03..00
Other cellulitis and abscess
79
404
5.11
3.71
14
16C..00
Backache symptom
233
851
3.65
7.83
15
1M1..00
Pain in lower limb
125
536
4.29
4.93
16
N09..00
Other and unspecified joint disorders
320
1058
3.31
9.73
17
K19..00
Other urethral and urinary tract disorders
179
725
4.05
6.67
18
196..00
Type of GIT pain
149
588
3.95
5.41
19
1D1..00
C/O: a general symptom
352
1126
320
10.35
4.52
ABO..00
Dermatophytosis including tinea or ringworm
87
393
3.61
20
Variable NB and NA represent the numbers of patients before or after they take drugs for having one particular medical event. Variable R1represents the ratio of the numbers
of patients after taking drugs to the numbers of patients before taking drugs. Variable R2 represents the ratio of the numbers of patients after taking drugs to the number of the
whole population.
TABLE III.
Level
1-5
Rank
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Readcodes
F591.00
J574700
16J4.00
D010.00
M22z.11
A080300
B49..00
1NO3.00
F420600
F42y900
K190311
173C.12
SKz..00
1CA2.11
J578.00
SE3..11
16J7.00
1B16.11
THE POTENTIAL ADRs FOR PRAVASTATIN BASED ON DESCENDING ORDER OF R1 VALUE.
Medical events
Sensorineural hearing loss
Anal pain
Swollen knee
Pernicious anaemia
Actinic keratosis
Infectious gastroenteritis
Malignant neoplasm of urinary bladder
C/O: dry skin
Non proliferative diabetic retinopathy
Macular oedema
Recurrent UTI
SOBOE
Injury NOS
Voice hoarseness
Colonic polyp
Arm bruise
Swollen foot
Agitated - symptom
1191
NB
1
1
0
1
1
1
1
1
0
1
1
0
1
3
2
1
0
1
NA
36
26
25
25
25
24
23
22
21
20
19
19
18
53
35
17
17
17
RI
36.00
26.00
25.00
25.00
25.00
24.00
23.00
22.00
21.00
20.00
19.00
19.00
18.00
17.67
17.50
17.00
17.00
17.00
R2
0.33
0.24
0.23
0.23
0.23
0.22
0.21
0.20
0.19
0.18
0.17
0.17
0.17
0.49
0.32
0.16
0.16
0.16
Level
1-3
19
20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
S64..13
G86..11
161..00
SE3..00
B49..00
F4B..00
1C7..00
1BD..00
SKz..00
D31..00
M25..00
HOz..00
BB1..00
F41..00
G41..00
K08..00
B7J..00
1NO..00
SA1..00
S93..00
S24..00
SD7..00
1A3..00
B83..00
F17..00
K11..00
F33..00
E26..00
A38..00
K13..00
F40..00
F12..00
Head injury
Lymphoedema
Appetite symptom
Contusion, upper limb
Malignant neoplasm of urinary bladder
Comeal opacity and other disorders of cornea
Snoring symptoms
Harmful thoughts
Injury NOS
Purpura and other haemorrhagic conditions
Disorders of sweat glands
Acute respiratory infection NOS
[M]Epithelial neoplasms NOS
Retinal detachments and defects
Chronic puhnonary heart disease
Impaired renal function disorder
Haemangiomas and lymphangiomas of any site
Skin symptoms
Open wound of knee, leg and ankle
Open wound of fmger(s) or thumb
Fracture of carpal bone
Superficial injury of foot and toe(s)
Micturition stream
Carcinoma in situ of breast and genitourinary system
Autonomic nervous system disorders
Hydronephrosis
Nerve root and plexus disorders
Physiological malfunction arising from mental factors
Septicaemia
Other kidney and ureter disorders
Disorders of the globe
Parkinson's disease
1
1
1
1
1
1
1
1
1
2
2
1
1
1
1
1
1
2
0
0
1
1
3
0
0
0
1
1
2
0
0
3
17
16
34
28
27
21
18
18
18
33
33
16
15
15
15
15
15
29
14
14
14
14
39
13
13
13
13
13
24
12
11
32
17.00
16.00
34.00
28.00
27.00
21.00
18.00
18.00
18.00
16.50
16.50
16.00
15.00
15.00
15.00
15.00
15.00
14.50
14.00
14.00
14.00
14.00
13.00
13.00
13.00
13.00
13.00
13.00
12.00
12.00
11.00
10.67
0.16
0.15
0.31
0.26
0.25
0.19
0.17
0.17
0.17
0.30
0.30
0.15
0.14
0.14
0.14
0.14
0.14
0.27
0.13
0.13
0.13
0.13
0.36
0.12
0.12
0.12
0.12
0.12
0.22
0.11
0.10
0.29
TABLEIV. THEPOTENTIALADRs RELATEDTO CANCERFORPRAVASTATIN BASEDONP VALUEOFSTUDENT ST-TEST.
Cancer
( level 15)
Rank
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Readcodes
B33..11
B46..00
B49..00
BB2A.00
B34..00
BB2..12
B32..00
B13..00
B22z.00
B10..00
BB5..11
B22..00
B585.00
B141.00
B834.00
BB12.00
BB31.00
B49z.00
B133.00
BB5M.00
B590.11
BB13.00
BBEF.11
B577.00
H51y700
B222.00
BB5y000
1J0..00
B134.00
BB2B.00
Medical events
Basal cell carcinoma
Malignant neoplasm of prostate
Malignant neoplasm of urinary bladder
[M]Squamous cell carcinoma NOS
Malignant neoplasm of female breast
[M]Squamous cell neoplasms
Malignant melanoma of skin
Malignant neoplasm of colon
Malignant neoplasm of bronchus or lung NOS
Malignant neoplasm of oesophagus
[M]Adenocarcinomas
Malignant neoplasm of trachea, bronchus and lung
Secondary malignant neoplasm of bone and bone marrow
Malignant neoplasm of rectum
Carcinoma in situ of prostate
[M]Carcinoma NOS
[M]Basal cell carcinoma NOS
Malignant neoplasm of urinary bladder NOS
Malignant neoplasm of sigmoid colon
[M]Tubular adenomas and adenocarcinomas
Carcinomatosis
[M]Carcinoma, metastatic, NOS
[M]Lentigo maligna
Secondary malignant neoplasm of liver
Malignant pleural effusion
Malignant neoplasm of upper lobe, bronchus or lung
[M]Basal cell adenocarcinoma
Suspected malignancy
Malignant neoplasm of caecum
[M]Squamous cell carcinoma, metastatic NOS
1192
NB
29
8
1
2
7
2
0
2
4
1
4
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
NA
126
47
23
29
30
17
9
15
16
9
15
6
6
6
5
5
7
6
4
4
4
6
3
3
3
3
3
3
3
3
R1
4.34
5.88
23.00
14.50
4.29
8.50
9.00
7.50
4.00
9.00
3.75
6.00
6.00
6.00
5.00
5.00
7.00
6.00
4.00
4.00
4.00
6.00
3.00
3.00
3.00
3.00
3.00
3.00
3.00
3.00
R2
1.16
0.43
0.21
0.27
0.28
0.16
0.08
0.14
0.15
0.08
0.14
0.06
0.06
0.06
0.05
0.05
0.06
0.06
0.04
0.04
0.04
0.06
0.03
0.03
0.03
0.03
0.03
0.03
0.03
0.03
| 5cs.CE
|
Published as a conference paper at ICLR 2017
M AKING N EURAL P ROGRAMMING A RCHITECTURES
G ENERALIZE VIA R ECURSION
arXiv:1704.06611v1 [cs.LG] 21 Apr 2017
Jonathon Cai, Richard Shin, Dawn Song
Department of Computer Science
University of California, Berkeley
Berkeley, CA 94720, USA
{jonathon,ricshin,dawnsong}@cs.berkeley.edu
A BSTRACT
Empirically, neural networks that attempt to learn programs from data have exhibited poor generalizability. Moreover, it has traditionally been difficult to reason
about the behavior of these models beyond a certain level of input complexity. In
order to address these issues, we propose augmenting neural architectures with
a key abstraction: recursion. As an application, we implement recursion in the
Neural Programmer-Interpreter framework on four tasks: grade-school addition,
bubble sort, topological sort, and quicksort. We demonstrate superior generalizability and interpretability with small amounts of training data. Recursion divides
the problem into smaller pieces and drastically reduces the domain of each neural network component, making it tractable to prove guarantees about the overall
system’s behavior. Our experience suggests that in order for neural architectures
to robustly learn program semantics, it is necessary to incorporate a concept like
recursion.
1
I NTRODUCTION
Training neural networks to synthesize robust programs from a small number of examples is a challenging task. The space of possible programs is extremely large, and composing a program that performs robustly on the infinite space of possible inputs is difficult—in part because it is impractical
to obtain enough training examples to easily disambiguate amongst all possible programs. Nevertheless, we would like the model to quickly learn to represent the right semantics of the underlying
program from a small number of training examples, not an exhaustive number of them.
Thus far, to evaluate the efficacy of neural models on programming tasks, the only metric that has
been used is generalization of expected behavior to inputs of greater complexity (Vinyals et al.
(2015), Kaiser & Sutskever (2015), Reed & de Freitas (2016), Graves et al. (2016), Zaremba et al.
(2016)). For example, for the addition task, the model is trained on short inputs and then tested on
its ability to sum inputs with much longer numbers of digits. Empirically, existing models suffer
from a common limitation—generalization becomes poor beyond a threshold level of complexity.
Errors arise due to undesirable and uninterpretable dependencies and associations the architecture
learns to store in some high-dimensional hidden state. This makes it difficult to reason about what
the model will do when given complex inputs.
One common strategy to improve generalization is to use curriculum learning, where the model is
trained on inputs of gradually increasing complexity. However, models that make use of this strategy
eventually fail after a certain level of complexity (e.g. the single-digit multiplication task in Zaremba
et al. (2016), the bubble sort task in Reed & de Freitas (2016), and the graph tasks in Graves et al.
(2016)). In this version of curriculum learning, even though the inputs are gradually becoming more
complex, the semantics of the program is succinct and does not change. Although the model is
exposed to more and more data, it might learn spurious and overly complex representations of the
program, as suggested in Zaremba et al. (2016). That is to say, the network does not learn the true
program semantics.
In this paper, we propose to resolve these issues by explicitly incorporating recursion into neural
architectures. Recursion is an important concept in programming languages and a critical tool to
1
Published as a conference paper at ICLR 2017
reduce the complexity of programs. We find that recursion makes it easier for the network to learn
the right program and generalize to unknown situations. Recursion enables provable guarantees on
neural programs’ behavior without needing to exhaustively enumerate all possible inputs to the programs. This paper is the first (to our knowledge) to investigate the important problem of provable
generalization properties of neural programs. As an application, we incorporate recursion into the
Neural Programmer-Interpreter architecture and consider four sample tasks: grade-school addition,
bubble sort, topological sort, and quicksort. Empirically, we observe that the learned recursive programs solve all valid inputs with 100% accuracy after training on a very small number of examples,
out-performing previous generalization results. Given verification sets that cover all the base cases
and reduction rules, we can provide proofs that these learned programs generalize perfectly. This is
the first time one can provide provable guarantees of perfect generalization for neural programs.
2
2.1
T HE P ROBLEM AND O UR A PPROACH
T HE P ROBLEM OF G ENERALIZATION
When constructing a neural network for the purpose of learning a program, there are two orthogonal
aspects to consider. The first is the actual model architecture. Numerous models have been proposed
for learning programs; to name a few, this includes the Differentiable Neural Computer (Graves
et al., 2016), Neural Turing Machine (Graves et al., 2014), Neural GPU (Kaiser & Sutskever, 2015),
Neural Programmer (Neelakantan et al., 2015), Pointer Network (Vinyals et al., 2015), Hierarchical
Attentive Memory (Andrychowicz & Kurach, 2016), and Neural Random Access Machine (Kurach
et al., 2016). The architecture usually possesses some form of memory, which could be internal
(such as the hidden state of a recurrent neural network) or external (such as a discrete “scratch pad”
or a memory block with differentiable access). The second is the training procedure, which consists
of the form of the training data and the optimization process. Almost all architectures train on
program input/output pairs. The only model, to our knowledge, that does not train on input-output
pairs is the Neural Programmer-Interpreter (Reed & de Freitas, 2016), which trains on synthetic
execution traces.
To evaluate a neural network that learns a neural program to accomplish a certain task, one common
evaluation metric is how well the learned model M generalizes. More specifically, when M is
trained on simpler inputs, such as inputs of a small length, the generalization metric evaluates how
well M will do on more complex inputs, such as inputs of much longer length. M is considered to
have perfect generalization if M can give the right answer for any input, such as inputs of arbitrary
length.
As mentioned in Section 1, all approaches to neural programming today fare poorly on this generalization issue. We hypothesize that the reason for this is that the neural network learns to spuriously
depend on specific characteristics of the training examples that are irrelevant to the true program
semantics, such as length of the training inputs, and thus fails to generalize to more complex inputs.
In addition, none of the current approaches to neural programming provide a method or even aim to
enable provable guarantees about generalization. The memory updates of these neural programs are
so complex and interdependent that it is difficult to reason about the behaviors of the learned neural
program under previously unseen situations (such as problems with longer inputs). This is highly
undesirable, since being able to provide the correct answer in all possible settings is one of the most
important aspects of any learned neural program.
2.2
O UR A PPROACH U SING R ECURSION
In this paper, we propose that the key abstraction of recursion is necessary for neural programs to
generalize. The general notion of recursion has been an important concept in many domains, including mathematics and computer science. In computer science, recursion (as opposed to iteration)
involves solving a larger problem by combining solutions to smaller instances of the same problem.
Formally, a function exhibits recursive behavior when it possesses two properties: (1) Base cases—
terminating scenarios that do not use recursion to produce answers; (2) A set of rules that reduces all
other problems toward the base cases. Some functional programming languages go so far as not to
define any looping constructs but rely solely on recursion to enable repeated execution of the same
code.
2
Published as a conference paper at ICLR 2017
In this paper, we propose that recursion is an important concept for neural programs as well. In
fact, we argue that recursion is an essential element for neural programs to generalize, and makes it
tractable to prove the generalization of neural programs. Recursion can be implemented differently
for different neural programming models. Here as a concrete and general example, we consider a
general Neural Programming Architecture (NPA), similar to Neural Programmer-Interpreter (NPI)
in Reed & de Freitas (2016). In this architecture, we consider a core controller, e.g., an LSTM
in NPI’s case, but possibly other networks in different cases. There is a (changing) list of neural
programs used to accomplish a given task. The core controller acts as a dispatcher for the programs.
At each time step, the core controller can decide to select one of the programs to call with certain
arguments. When the program is called, the current context including the caller’s memory state is
stored on a stack; when the program returns, the stored context is popped off the stack to resume
execution in the previous caller’s context.
In this general Neural Programming Architecture, we show it is easy to support recursion. In particular, recursion can be implemented as a program calling itself. Because the context of the caller
is stored on a stack when it calls another program and the callee starts in a fresh context, this enables recursion simply by allowing a program to call itself. In practice, we can additionally use tail
recursion optimization to avoid problems with the call stack growing too deep. Thus, any general
Neural Programming Architecture supporting such a call structure can be made to support recursion.
In particular, this condition is satisfied by NPI, and thus the NPI model naturally supports recursion
(even though the authors of NPI did not consider this aspect explicitly).
By nature, recursion reduces the complexity of a problem to simpler instances. Thus, recursion
helps decompose a problem and makes it easier to reason about a program’s behavior for previously
unseen situations such as longer inputs. In particular, given that a recursion is defined by two
properties as mentioned before, the base cases and the set of reduction rules, we can prove a recursive
neural program generalizes perfectly if we can prove that (1) it performs correctly on the base cases;
(2) it learns the reduction rules correctly. For many problems, the base cases and reduction rules
usually consist of a finite (often small) number of cases. For problems where the base cases may be
extremely large or infinite, such as certain forms of motor control, recursion can still help reduce the
problem of generalization to these two aspects and make the generalization problem significantly
simpler to handle and reason about.
As a concrete instantiation, we show in this paper that we can enable recursive neural programs in the
NPI model, and thus enable perfectly generalizable neural programs for tasks such as sorting where
the original, non-recursive NPI program fails. As aforementioned, the NPI model naturally supports
recursion. However, the authors of NPI did not consider explicitly the notion of recursion and as a
consequence, did not learn recursive programs. We show that by modifying the training procedure,
we enable the NPI model to learn recursive neural programs. As a consequence, our learned neural
programs empirically achieve perfect generalization from a very small number of training examples.
Furthermore, given a verification input set that covers all base cases and reduction rules, we can
formally prove that the learned neural programs achieve perfect generalization after verifying its
behavior on the verification input set. This is the first time one can provide provable guarantees of
perfect generalization for neural programs.
We would also like to point out that in this paper, we provide as an example one way to train a
recursive neural program, by providing a certain training execution trace to the NPI model. However,
our concept of recursion for neural programs is general. In fact, it is one of our future directions to
explore new ways to train a recursive neural program without providing explicit training execution
traces or with only partial or non-recursive traces.
3
3.1
A PPLICATION TO L EARNING R ECURSIVE N EURAL P ROGRAMS WITH NPI
BACKGROUND : NPI A RCHITECTURE
As discussed in Section 2, the Neural Programmer-Interpreter (NPI) is an instance of a Neural
Programmer Architecture and hence it naturally supports recursion. In this section, we give a brief
review of the NPI architecture from Reed & de Freitas (2016) as background.
3
Published as a conference paper at ICLR 2017
We describe the details of the NPI model relevant to our contributions. We adapt machinery from
the original paper slightly to fit our needs. The NPI model has three learnable components: a
task-agnostic core, a program-key embedding, and domain-specific encoders that allow the NPI to
operate in diverse environments.
The NPI accesses an external environment, Q, which varies according to the task. The core module
of the NPI is an LSTM controller that takes as input a slice of the current external environment, via
a set of pointers, and a program and arguments to execute. NPI then outputs the return probability
and next program and arguments to execute. Formally, the NPI is represented by the following set
of equations:
st = fenc (et , at )
ht = flstm (st , pt , ht−1 )
rt = fend (ht ), pt+1 = fprog (ht ), at+1 = farg (ht )
t is a subscript denoting the time-step; fenc is a domain-specific encoder (to be described later) that
takes in the environment slice et and arguments at ; flstm represents the core module, which takes
in the state st generated by fenc , a program embedding pt ∈ RP , and hidden LSTM state ht ; fend
decodes the return probability rt ; fprog decodes a program key embedding pt+1 ;1 and farg decodes
arguments at+1 . The outputs rt , pt+1 , at+1 are used to determine the next action, as described in
Algorithm 1. If the program is primitive, the next environmental state et+1 will be affected by pt
and at , i.e. et+1 ∼ fenv (et , pt , at ). As with the original NPI architecture, the experiments for this
paper always used a 3-tuple of integers at = (at (1), at (2), at (3)).
Algorithm 1 Neural programming inference
1: Inputs: Environment observation e, program p, arguments a, stop threshold α
2: function RUN(e, p, a)
3:
h ← 0, r ← 0
4:
while r < α do
5:
s ← fenc (e, a), h ← flstm (s, p, h)
6:
r ← fend (h), p2 ← fprog (h), a2 ← farg (h)
7:
if p is a primitive function then
8:
e ← fenv (e, p, a).
9:
else
10:
function RUN(e, p2 , a2 )
A description of the inference procedure is given in Algorithm 1. Each step during an execution
of the program does one of three things: (1) another subprogram along with associated arguments
is called, as in Line 10, (2) the program writes to the environment if it is primitive, as in Line 8,
or (3) the loop is terminated if the return probability exceeds a threshold α, after which the stack
frame is popped and control is returned to the caller. In all experiments, α is set to 0.5. Each time a
subprogram is called, the stack depth increases.
The training data for the Neural Programmer-Interpreter consists of full execution traces for the
program of interest. A single element of an execution trace consists of a step input-step output pair,
which can be synthesized from Algorithm 1: this corresponds to, for a given time-step, the step
input tuple (e, p, a) and step output tuple (r, p2 , a2 ). An example of part of an addition task trace,
written in shorthand, is given in Figure 1. For example, a step input-step output pair in Lines 2 and
3 of the left-hand side of Figure 1 is (ADD1, WRITE OUT 1). In this pair, the step input runs
a subprogram ADD1 that has no arguments, and the step output contains a program WRITE that
has arguments of OUT and 1. The environment and return probability are omitted for readability.
Indentation indicates the stack is one level deeper than before.
It is important to emphasize that at inference time in the NPI, the hidden state of the LSTM controller
is reset (to zero) at each subprogram call, as in Line 3 of Algorithm 1 (h ← 0). This functionality
1
The original NPI paper decodes to a program key embedding kt ∈ RK and then computes a program
embedding pt+1 , which we also did in our implementation, but we omit this for brevity.
4
Published as a conference paper at ICLR 2017
Non-Recursive
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Recursive
1
2
3
4
5
6
7
8
9
10
11
12
13
14
ADD
ADD1
WRITE OUT 1
CARRY
PTR CARRY LEFT
WRITE CARRY 1
PTR CARRY RIGHT
LSHIFT
PTR INP1 LEFT
PTR INP2 LEFT
PTR CARRY LEFT
PTR OUT LEFT
ADD1
...
ADD
ADD1
WRITE OUT 1
CARRY
PTR CARRY LEFT
WRITE CARRY 1
PTR CARRY RIGHT
LSHIFT
PTR INP1 LEFT
PTR INP2 LEFT
PTR CARRY LEFT
PTR OUT LEFT
ADD
...
Figure 1: Addition Task. The non-recursive trace loops on cycles of ADD1 and LSHIFT, whereas
in the recursive version, the ADD function calls itself (bolded).
is critical for implementing recursion, since it permits us to restrict our attention to the currently
relevant recursive call, ignoring irrelevant details about other contexts.
3.2
R ECURSIVE F ORMULATIONS FOR NPI PROGRAMS
We emphasize the overall goal of this work is to enable the learning of a recursive program. The
learned recursive program is different from neural programs learned in all previous work in an important aspect: previous approaches do not explicitly incorporate this abstraction, and hence generalize
poorly, whereas our learned neural programs incorporate recursion and achieve perfect generalization.
Since NPI naturally supports the notion of recursion, a key question is how to enable NPI to learn
recursive programs. We found that changing the NPI training traces is a simple way to enable this.
In particular, we construct new training traces which explicitly contain recursive elements and show
that with this type of trace, NPI easily learns recursive programs. In future work, we would like to
decrease supervision and construct models that are capable of coming up with recursive abstractions
themselves.
In what follows, we describe the way in which we constructed NPI training traces so as to make
them contain recursive elements and thus enable NPI to learn recursive programs. We describe the
recursive re-formulation of traces for two tasks from the original NPI paper—grade-school addition
and bubble sort. For these programs, we re-use the appropriate program sets (the associated subprograms), and we refer the reader to the appendix of Reed & de Freitas (2016) for further details on
the subprograms used in addition and bubble sort. Finally, we implement recursive traces for our
own topological sort and quicksort tasks.
Grade School Addition. For grade-school addition, the domain-specific encoder is
fenc (Q, i1 , i2 , i3 , i4 , at ) = M LP ([Q(1, i1 ), Q(2, i2 ), Q(3, i3 ), Q(4, i4 ), at (1), at (2), at (3)]),
where the environment Q ∈ R4×N ×K is a scratch-pad that contains four rows (the first input number, the second input number, the carry bits, and the output) and N columns. K is set to 11, to
represent the range of 10 possible digits, along with a token representing the end of input.2 At
any given time, the NPI has access to values pointed to by four pointers in each of the four rows,
represented by Q(1, i1 ), Q(2, i2 ), Q(3, i3 ), and Q(4, i4 ).
The non-recursive trace loops on cycles of ADD1 and LSHIFT. ADD1 is a subprogram that adds
the current column (writing the appropriate digit to the output row and carrying a bit to the next
column if needed). LSHIFT moves the four pointers to the left, to move to the next column. The
program terminates when seeing no numbers in the current column.
Figure 1 shows examples of non-recursive and recursive addition traces. We make the trace recursive
by adding a tail recursive call into the trace for the ADD program after calling ADD1 and LSHIFT,
2
The original paper uses K = 10, but we found it necessary to augment the range with an end token, in
order to terminate properly.
5
Published as a conference paper at ICLR 2017
Partial Recursive
Non-Recursive
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
BUBBLESORT
BUBBLE
PTR 2 RIGHT
BSTEP
COMPSWAP
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
BSTEP
COMPSWAP
SWAP 1 2
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
RESET
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
PTR 3 RIGHT
BUBBLE
...
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
BUBBLESORT
BUBBLE
PTR 2 RIGHT
BSTEP
COMPSWAP
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
BSTEP
COMPSWAP
SWAP 1 2
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
RESET
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
PTR 3 RIGHT
BUBBLESORT
BUBBLE
...
Full Recursive
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
BUBBLESORT
BUBBLE
PTR 2 RIGHT
BSTEP
COMPSWAP
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
BSTEP
COMPSWAP
SWAP 1 2
RSHIFT
PTR 1 RIGHT
PTR 2 RIGHT
BSTEP
RESET
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
LSHIFT
PTR 1 LEFT
PTR 2 LEFT
LSHIFT
PTR 3 RIGHT
BUBBLESORT
BUBBLE
...
Figure 2: Bubble Sort Task. The non-recursive trace loops on cycles of BUBBLE and RESET.
The difference between the partial recursive and full recursive versions is in the indentation of Lines
10-15 and 20-22 (bolded), since in the full recursive version, BSTEP and LSHIFT are made tail
recursive; the final calls to BSTEP and LSHIFT return immediately as they occur after the pointer
reaches the end of the array. Also note that COMPSWAP conditionally swaps numbers under the
bubble pointers.
as in Line 13 of the right-hand side of Figure 1. Via the recursive call, we effectively forget that the
column just added exists, since the recursive call to ADD starts with a new hidden state for the
LSTM controller. Consequently, there is no concept of length relevant to the problem, which has
traditionally been an important focus of length-based curriculum learning.
Bubble Sort. For bubble sort, the domain-specific encoder is
fenc (Q, i1 , i2 , i3 , at ) = M LP ([Q(1, i1 ), Q(1, i2 ), i3 == length, at (1), at (2), at (3)]),
where the environment Q ∈ R1×N ×K is a scratch-pad that contains 1 row, to represent the state of
the array as sorting proceeds in-place, and N columns. K is set to 11, to denote the range of possible
numbers (0 through 9), along with the start/end token (represented with the same encoding) which
is observed when a pointer reaches beyond the bounds of the input. At any given time, the NPI has
access to the values referred to by two pointers, represented by Q(1, i1 ) and Q(1, i2 ),. The pointers
at index i1 and i2 are used to compare the pair of numbers considered during the bubble sweep,
swapping them if the number at i1 is greater than that in i2 . These pointers are referred to as bubble
pointers. The pointer at index i3 represents a counter internal to the environment that is incremented
once after each pass of the algorithm (one cycle of BUBBLE and RESET); when incremented a
number of times equal to the length of the array, the flag i3 == length becomes true and terminates
the entire algorithm.
The non-recursive trace loops on cycles of BUBBLE and RESET, which logically represents one
bubble sweep through the array and reset of the two bubble pointers to the very beginning of the
array, respectively. In this version, there is a dependence on length: BSTEP and LSHIFT are called
a number of times equivalent to one less than the length of the input array, in BUBBLE and RESET
respectively.
Inside BUBBLE and RESET, there are two operations that can be made recursive. BSTEP, used
in BUBBLE, compares pairs of numbers, continuously moving the bubble pointers once to the right
each time until reaching the end of the array. LSHIFT, used in RESET, shifts the pointers left until
reaching the start token.
6
Published as a conference paper at ICLR 2017
We experiment with two levels of recursion—partial and full. Partial recursion only adds a tail
recursive call to BUBBLESORT after BUBBLE and RESET, similar to the tail recursive call
described previously for addition. The partial recursion is not enough for perfect generalization, as
will be presented later in Section 4. Full recursion, in addition to making the aforementioned tail
recursive call, adds two additional recursive calls; BSTEP and LSHIFT are made tail recursive.
Figure 2 shows examples of traces for the different versions of bubble sort. Training on the full
recursive trace leads to perfect generalization, as shown in Section 4. We performed experiments
on the partially recursive version in order to examine what happens when only one recursive call is
implemented, when in reality three are required for perfect generalization.
Algorithm 2 Depth First Search Topological Sort
1: Color all vertices white.
2: Initialize an empty stack S and a directed acyclic graph DAG to traverse.
3: Begin traversing from Vertex 1 in the DAG.
4: function T OPO S ORT(DAG)
5:
while there is still a white vertex u: do
6:
color[u] = grey
7:
vactive = u
8:
do
9:
if vactive has a white child v then
10:
color[v] = grey
11:
push vactive onto S
12:
vactive = v
13:
else
14:
color[vactive ] = black
15:
Write vactive to result
16:
if S is empty then pass
17:
else pop the top vertex off S and set it to vactive
18:
while S is not empty
Topological Sort. We choose to implement a topological sort task for graphs. A topological sort
is a linear ordering of vertices such that for every directed edge (u, v) from u to v, u comes before
v in the ordering. This is possible if and only if the graph has no directed cycles; that is to say, it
must be a directed acyclic graph (DAG). In our experiments, we only present DAG’s as inputs and
represent the vertices as values ranging from 1, . . . , n , where the DAG contains n vertices.
Directed acyclic graphs are structurally more diverse than inputs in the two tasks of grade-school
addition and bubble sort. The degree for any vertex in the DAG is variable. Also the DAG can have
potentially more than one connected component, meaning it is necessary to transition between these
components appropriately.
Algorithm 2 shows the topological sort task of interest. This algorithm is a variant of depth first
search. We created a program set that reflects the semantics of Algorithm 2. For brevity, we refer
the reader to the appendix for further details on the program set and non-recursive and recursive
trace-generating functions used for topological sort.
For topological sort, the domain-specific encoder is
fenc (DAG, Qcolor , pstack , pstart , vactive , childList, at )
= M LP ([Qcolor (pstart ), Qcolor (DAG[vactive ][childList[vactive ]]), pstack == 1, at (1), at (2), at (3)]),
where Qcolor ∈ RU ×4 is a scratch-pad that contains U rows, each containing one of four colors
(white, gray, black, invalid) with one-hot encoding. U varies with the number of vertices in the
graph. We further have Qresult ∈ NU , a scratch-pad which contains the sorted list of vertices at
the end of execution, and Qstack ∈ NU , which serves the role of the stack S in Algorithm 2. The
contents of Qresult and Qstack are not exposed directly through the domain-specific encoder; rather,
we define primitive functions which manipulate these scratch-pads.
The DAG is represented as an adjacency list where DAG[i][j] refers to the j-th child of vertex i.
There are 3 pointers (presult , pstack , pstart ), presult points to the next empty location in Qresult ,
7
Published as a conference paper at ICLR 2017
pstack points to the top of the stack in Qstack , and pstart points to the candidate starting node for
a connected component. There are 2 variables (vactive and vsave ); vactive holds the active vertex
(as in Algorithm 2) and vsave holds the value of vactive before executing Line 12 of Algorithm 2.
childList ∈ NU is a vector of pointers, where childList[i] points to the next child under consideration for vertex i.
The three environment observations aid with control flow in Algorithm 2. Qcolor (pstart ) contains
the color of the current start vertex, used in the evaluation of the condition in the WHILE loop in Line
5 of Algorithm 2. Qcolor (DAG[vactive ][childList[vactive ]]) refers to the color of the next child of
vactive , used in the evaluation of the condition in the IF branch in Line 9 of Algorithm 2. Finally,
the boolean pstack == 1 is used to check whether the stack is empty in Line 18 of Algorithm 2.
An alternative way of representing the environment slice is to expose the values of the absolute
vertices to the model; however, this makes it difficult to scale the model to larger graphs, since large
vertex values are not seen during training time.
We refer the reader to the appendix for the non-recursive trace generating functions. In
the non-recursive trace, there are four functions that can be made recursive—TOPOSORT,
CHECK CHILD, EXPLORE, and NEXT START, and we add a tail recursive call to each of
these functions in order to make the recursive trace. In particular, in the EXPLORE function,
adding a tail recursive call resets and stores the hidden states associated with vertices in a stack-like
fashion. This makes it so that we only need to consider the vertices in the subgraph that are currently relevant for computing the sort, allowing simpler reasoning about behavior for large graphs.
The sequence of primitive operations (MOVE and WRITE operations) for the non-recursive and
recursive versions are exactly the same.
Quicksort. We implement a quicksort task, in order to demonstrate that recursion helps with learning divide-and-conquer algorithms. We use the Lomuto partition scheme; the logic for the recursive
trace is shown in Algorithm 3. For brevity, we refer the reader to the appendix for information about
the program set and non-recursive and recursive trace-generating functions for quicksort. The logic
for the non-recursive trace is shown in Algorithm 4 in the appendix.
Algorithm 3 Recursive Quicksort
1: Initialize an array A to sort.
2: Initialize lo and hi to be 1 and n, where n is the length of A.
3:
4: function Q UICK S ORT(A, lo, hi)
5:
if lo < hi: then
6:
p = PARTITION(A, lo, hi)
7:
QUICKSORT (A, lo, p − 1)
8:
QUICKSORT (A, p + 1, hi)
9:
10: function PARTITION(A, lo, hi)
11:
pivot = lo
12:
for j ∈ [lo, hi − 1] : do
13:
if A[j] ≤ A[hi] then
14:
swap A[pivot] with A[j]
15:
pivot = pivot + 1
16:
swap A[pivot] with A[hi]
17:
return pivot
For quicksort, the domain-specific encoder is
fenc (Qarray , QstackLo , QstackHi , plo , phi , pstackLo , pstackHi , ppivot , pj , at ) =
M LP ([Qarray (pj ) ≤ Qarray (phi ), pj == phi ,
QstackLo (pstackLo − 1) < QstackHi (pstackHi − 1), pstackLo == 1, at (1), at (2), at (3)]),
where Qarray ∈ RU ×11 is a scratch-pad that contains U rows, each containing one of 11 values (one
of the numbers 0 through 9 or an invalid state). Our implementation uses two stacks QstackLo and
8
Published as a conference paper at ICLR 2017
QstackHi , each in RU , that store the arguments to the recursive QUICKSORT calls in Algorithm 3;
before each recursive call, the appropriate arguments are popped off the stack and written to plo and
phi .
There are 6 pointers (plo , phi , pstackLo , pstackHi , ppivot , pj ). plo and phi point to the lo and hi
indices of the array, as in Algorithm 3. pstackLo and pstackHi point to the top (empty) positions
in QstackLo and QstackHi . ppivot and pj point to the pivot and j indices of the array, used in
the PARTITION function in Algorithm 3. The 4 environment observations aid with control flow;
QstackLo (pstackLo − 1) < QstackHi (pstackHi − 1) implements the lo < hi comparison in Line 5 of
Algorithm 3, pstackLo == 1 checks if the stacks are empty in Line 18 of Algorithm 4, and the other
observations (all involving ppivot or pj ) deal with logic in the PARTITION function.
Note that the recursion for quicksort is not purely tail recursive and therefore represents a more
complex kind of recursion that is harder to learn than in the previous tasks. Also, compared to the
bubble pointers in bubble sort, the pointers that perform the comparison for quicksort (the COMPSWAP function) are usually not adjacent to each other, making quicksort less local than bubble
sort. In order to compensate for this, ppivot and pj require special functions (MOVE PIVOT LO
and MOVE J LO) to properly set them to lo in Lines 11 and 12 of the PARTITION function in
Algorithm 3.
3.3
P ROVABLY P ERFECT G ENERALIZATION
We show that if we incorporate recursion, the learned NPI programs can achieve provably perfect
generalization for different tasks. Provably perfect generalization implies the model will behave
correctly, given any valid input. In order to claim a proof, we must verify the model produces
correct behavior over all base cases and reductions, as described in Section 2.
We propose and describe our verification procedure. This procedure verifies that all base cases and
reductions are handled properly by the model via explicit tests. Note that recursion helps make this
process tractable, because we only need to test a finite number of inputs to show that the model
will work correctly on inputs of unbounded complexity. This verification phase only needs to be
performed once after training.
Formally, verification consists of proving the following theorem:
∀i ∈ V, M (i) ⇓ P (i)
where i denotes a sequence of step inputs (within one function call), V denotes the set of valid
sequences of step inputs, M denotes the neural network model, P denotes the correct program, and
P (i) denotes the next step output from the correct program. The arrow in the theorem refers to
evaluation, as in big-step semantics. The theorem states that for the same sequence of step inputs,
the model produces the exact same step output as the target program it aims to learn. M , as described
in Algorithm 1, processes the sequence of step inputs by using an LSTM.
Recursion drastically reduces the number of configurations we need to consider during the verification phase and makes the proof tractable, because it introduces structure that eliminates infinitely
long sequences of step inputs that would otherwise need to be considered. For instance, for recursive
addition, consider the family F of addition problems an an−1 . . . a1 a0 + bn bn−1 . . . b1 b0 where no
CARRY operations occur. We prove every member of F is added properly, given that subproblems
S = {an an−1 + bn bn−1 , an−1 an−2 + bn−1 bn−2 , . . . , a1 a0 + b1 b0 } are added properly.
Without using a recursive program, such a proof is not possible, because the non-recursive program
runs on an arbitrarily long addition problem that creates correspondingly long sequences of step
inputs; in the non-recursive formulation of addition, ADD calls ADD1 a number of times that is
dependent on the length of the input. The core LSTM module’s hidden state is preserved over all
these ADD1 calls, and it is difficult to interpret with certainty what happens over longer timesteps
without concretely evaluating the LSTM with an input of that length. In contrast, each call to
the recursive ADD always runs for a fixed number of steps, even on arbitrarily long problems
in F , so we can test that it performs correctly on a small, fixed number of step input sequences.
This guarantees that the step input sequences considered during verification contain all step input
sequences which arise during execution of an unseen problem in F , leading to generalization to any
problem in F . Hence, if all subproblems in S are added correctly, we have proven that any member
of F will be added correctly, thus eliminating an infinite family of inputs that need to be tested.
9
Published as a conference paper at ICLR 2017
To perform the verification as described here, it is critical to construct V correctly. If it is too small,
then execution of the program on some input might require evaluation of M (i) on some i ∈
/ V , and
so the behavior of M (i) might deviate from P (i). If it is too large, then the semantics of P might not
be well-defined on some elements in V , or the spurious step input sequences may not be reachable
from any valid problem input (e.g., an array for quicksort or a DAG for topological sort).
To construct this set, by using the reference implementation of each subprogram, we construct a
mapping between two sets of environment observations: the first set consists of all observations
that can occur at the beginning of a particular subprogram’s invocation, and the second set contains
the observations at the end of that subprogram. We can obtain this mapping by first considering
the possible observations that can arise at the beginning of the entry function (ADD, BUBBLESORT, TOPOSORT, and QUICKSORT) for some valid program input, and iteratively applying
the observation-to-observation mapping implied by the reference implementation’s step output at
that point in the execution. If the step output specifies a primitive function call, we need to reason
about how it can affect the environment so as to change the observation in the next step input. For
non-primitive subprograms, we can update the observation-to-observation mapping currently associated with the subprogram and then apply that mapping to the current set. By iterating with this
procedure, and then running P on the input observation set that we obtain for the entry point function, we can obtain V precisely. To make an analogy to MDPs, this procedure is analogous to how
value iteration obtains the correct value for each state starting from any initialization.
An alternative method is to run P on many different program inputs and then observe step input
sequences which occur, to create V . However, to be sure that the generated V is complete (covers
all the cases needed), we need to check all pairs of observations seen in adjacent step inputs (in particular, those before and after a primitive function call), in a similar way as if we were constructing
V from scratch. Given a precise definition of P , it may be possible to automate the generation of V
from P in future work.
Note that V should also contain the necessary reductions, which corresponds to making the recursive
calls at the correct time, as indicated by P .
After finding V , we construct a set of problem inputs which, when executed on P , create exactly the
step input sequences which make up V . We call this set of inputs the verification set, SV .
Given a verification set, we can then run the model on the verification set to check if the produced
traces and results are correct. If yes, then this indicates that the learned neural program achieves
provably perfect generalization.
We note that for tasks with very large input domains, such as ones involving MNIST digits or speech
samples, the state space of base cases and reduction rules could be prohibitively large, possibly
infinite. Consequently, it is infeasible to construct a verification set that covers all cases, and the
verification procedure we have described is inadequate. We leave this as future work to devise a
verification procedure more appropriate to this setting.
4
E XPERIMENTS
As there is no public implementation of NPI, we implemented a version of it in Keras that is as
faithful to the paper as possible. Our experiments use a small number of training examples.
Training Setup. The training set for addition contains 200 traces. The maximum problem length
in this training set is 3 (e.g., the trace corresponding to the problem “109 + 101”).
The training set for bubble sort contains 100 traces, with maximum problem length of 2 (e.g., the
trace corresponding to the array [3,2]).
The training set for topological sort contains 6 traces, with one synthesized from a graph of size 5
and the rest synthesized from graphs of size 7.
The training set for quicksort contains 4 traces, synthesized from arrays of length 5.
The same set of problems was used to generate the training traces for all formulations of the task,
for non-recursive and recursive versions.
10
Published as a conference paper at ICLR 2017
Table 1: Accuracy on Randomly Generated Problems for Bubble Sort
Length of Array
Non-Recursive
Partially Recursive
Full Recursive
2
3
4
8
20
90
100%
6.7%
10%
0%
0%
0%
100%
23%
10%
0%
0%
0%
100%
100%
100%
100%
100%
100%
We train using the Adam optimizer and use a 2-layer LSTM and task-specific state encoders for the
external environments, as described in Reed & de Freitas (2016).
4.1
R ESULTS ON G ENERALIZATION OF R ECURSIVE N EURAL P ROGRAMS
We now report on generalization for the varying tasks.
Grade-School Addition. Both the non-recursive and recursive learned programs generalize on
all input lengths we tried, up to 5000 digits. This agrees with the generalization of non-recursive
addition in Reed & de Freitas (2016), where they reported generalization up to 3000 digits. However,
note that there is no provable guarantee that the non-recursive learned program will generalize to all
inputs, whereas we show later that the recursive learned program has a provable guarantee of perfect
generalization.
In order to demonstrate that recursion can help learn and generalize better, for addition, we trained
only on traces for 5 arbitrarily chosen 1-digit addition sum examples. The recursive version can generalize perfectly to long problems constructed from these components (such as the sum “822+233”,
where “8+2” and “2+3” are in the training set), but the non-recursive version fails to sum these long
problems properly.
Bubble Sort. Table 1 presents results on randomly generated arrays of varying length for the
learned non-recursive, partially recursive, and full recursive programs. For each length, we test each
program on 30 randomly generated problems. Observe that partially recursive does slightly better
than non-recursive for the setting in which the length of the array is 3, and that the fully recursive
version is able to sort every array given to it. The non-recursive and partially recursive versions are
unable to sort long arrays, beyond length 8.
Topological Sort. Both the non-recursive and recursive learned programs generalize on all graphs
we tried, up to 120 vertices. As before, the non-recursive learned program lacks a provable guarantee
of generalization, whereas we show later that the recursive learned program has one.
In order to demonstrate that recursion can help learn and generalize better, we trained a non-recursive
and recursive model on just a single execution trace generated from a graph containing 5 nodes3 for
the topological sort task. For these models, Table 2 presents results on randomly generated DAGs
of varying graph sizes (varying in the number of vertices). For each graph size, we test the learned
programs on 30 randomly generated DAGs. The recursive version of topological sort solves all graph
instances we tried, from graphs of size 5 through 70. On the other hand, the non-recursive version
has low accuracy, beginning from size 5, and fails completely for graphs of size 8 and beyond.
Quicksort. Table 3 presents results on randomly generated arrays of varying length for the learned
non-recursive and recursive programs. For each length, we test each program on 30 randomly generated problems. Observe that the non-recursive program’s correctness degrades for length 11 and
beyond, while the recursive program can sort any given array.
3
The corresponding edge list is [(1, 2), (1, 5), (2, 4), (2, 5), (3, 5)].
11
Published as a conference paper at ICLR 2017
Table 2: Accuracy on Randomly Generated Problems for Topological Sort
Number of Vertices
Non-Recursive
Recursive
5
6
7
8
70
6.7%
6.7%
3.3%
0%
0%
100%
100%
100%
100%
100%
Table 3: Accuracy on Randomly Generated Problems for Quicksort
Length of Array
Non-Recursive
Recursive
3
5
7
11
15
20
22
25
30
70
100%
100%
100%
73.3%
60%
30%
20%
3.33%
3.33%
0%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
As mentioned in Section 2.1, we hypothesize the non-recursive programs do not generalize well
because they have learned spurious dependencies specific to the training set, such as length of the
input problems. On the other hand, the recursive programs have learned the true program semantics.
4.2
V ERIFICATION OF P ROVABLY P ERFECT G ENERALIZATION
We describe how models trained with recursive traces can be proven to generalize, by using the
verification procedure described in Section 3.3. As described in the verification procedure, it is
possible to prove our learned recursive program generalizes perfectly by testing on an appropriate
set of problem inputs, i.e., the verification set. Recall that this verification procedure cannot be
performed for the non-recursive versions, since the propagation of the hidden state in the core LSTM
module makes reasoning difficult and so we would need to check an unbounded number of examples.
We describe the base cases, reduction rules, and the verification set for each task in Appendix A.6.
For each task, given the verification set, we check the traces and results of the learned, to-be-verified
neural program (described in Section 4.1; and for bubble sort, Appendix A.6) on the verification
set, and ensure they match the traces produced by the true program P . Our results show that for all
learned, to-be-verified neural programs, they all produced the same traces as those produced by P
on the verification set. Thus, we demonstrate that recursion enables provably perfect generalization
for different tasks, including addition, topological sort, quicksort, and a variant of bubble sort.
Note that the training set can often be considerably smaller than the verification set, and despite
this, the learned model can still pass the entire verification set. Our result shows that the training
procedure and the NPI architecture is capable of generalizing from the step input-output pairs seen
in the training data to the unseen ones present in the verification set.
5
C ONCLUSION
We emphasize that the notion of a neural recursive program has not been presented in the literature
before: this is our main contribution. Recursion enables provably perfect generalization. To the best
of our knowledge, this is the first time verification has been applied to a neural program, provid12
Published as a conference paper at ICLR 2017
ing provable guarantees about its behavior. We instantiated recursion for the Neural ProgrammerInterpreter by changing the training traces. In future work, we seek to enable more tasks with
recursive structure. We also hope to decrease supervision, for example by training with only partial
or non-recursive traces, and to develop novel Neural Programming Architectures integrated directly
with a notion of recursion.
ACKNOWLEDGMENTS
This material is in part based upon work supported by the National Science Foundation under Grant
No. TWC-1409915, DARPA under Grant No. FA8750-15-2-0104, and Berkeley Deep Drive. Any
opinions, findings, and conclusions or recommendations expressed in this material are those of the
author(s) and do not necessarily reflect the views of National Science Foundation and DARPA.
R EFERENCES
Marcin Andrychowicz and Karol Kurach. Learning efficient algorithms with hierarchical attentive
memory. CoRR, abs/1602.03218, 2016. URL http://arxiv.org/abs/1602.03218.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. CoRR, abs/1410.5401,
2014. URL http://arxiv.org/abs/1410.5401.
Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwiska, Sergio Gmez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
Adri Puigdomnech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain,
Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis. Hybrid computing using a neural network with dynamic external memory. Nature, 538
(7626):471–476, October 2016. ISSN 0028-0836, 1476-4687. doi: 10.1038/nature20101. URL
http://www.nature.com/doifinder/10.1038/nature20101.
Lukasz Kaiser and Ilya Sutskever. Neural gpus learn algorithms. CoRR, abs/1511.08228, 2015.
URL http://arxiv.org/abs/1511.08228.
Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random access machines. ERCIM
News, 2016(107), 2016. URL http://ercim-news.ercim.eu/en107/special/
neural-random-access-machines.
Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs
with gradient descent, 2015.
Scott Reed and Nando de Freitas. Neural programmer-interpreters. ICLR, 2016.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 2692–2700, 2015. URL
http://papers.nips.cc/paper/5866-pointer-networks.
Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms
from examples. In Proceedings of the 33nd International Conference on Machine Learning, ICML
2016, New York City, NY, USA, June 19-24, 2016, pp. 421–429, 2016. URL http://jmlr.
org/proceedings/papers/v48/zaremba16.html.
13
Published as a conference paper at ICLR 2017
A
A.1
APPENDIX
P ROGRAM S ET FOR N ON -R ECURSIVE T OPOLOGICAL S ORT
Program
TOPOSORT
Descriptions
Perform topological
sort on graph
TRAVERSE
Traverse graph until
stack is empty
Check if a white
child exists; if so, set
childList[vactive ] to
point to it
Repeatedly traverse
subgraphs until stack
is empty
Interact with stack,
either pushing or
popping
Move pstart until
reaching a white
vertex. If a white
vertex is found, set
pstart to point to it;
this signifies the start
of a traversal of a
new connected component. If no white
vertex is found, the
entire execution is
terminated
Write a value either
to environment (e.g.,
to color a vertex)
or variable (e.g., to
change the value of
vactive )
Move a pointer
(e.g.,
pstart
or
childList[vactive ])
up or down
CHECK CHILD
EXPLORE
STACK
NEXT START
WRITE
MOVE
Calls
TRAVERSE,
NEXT START,
WRITE, MOVE
CHECK CHILD, EXPLORE
MOVE
Arguments
NONE
STACK,
CHECK CHILD,
WRITE, MOVE
WRITE, MOVE
NONE
MOVE
NONE
NONE
Described below
NONE
Described below
NONE
NONE
PUSH, POP
Argument Sets for WRITE and MOVE.
WRITE. The WRITE operation has the following arguments:
ARG 1 (Main Action): COLOR CURR, COLOR NEXT, ACTIVE START, ACTIVE NEIGHB,
ACTIVE STACK, SAVE, STACK PUSH, STACK POP, RESULT
COLOR CURR colors vactive , COLOR NEXT colors Vertex DAG[vactive ][childList[vactive ]],
ACTIVE START
writes
pstart
to
vactive ,
ACTIVE NEIGHB
writes
DAG[vactive ][childList[vactive ]] to vactive , ACTIVE STACK writes Qstack (pstack ) to
vactive , SAVE writes vactive to vsave , ST ACK P U SH pushes vactive to the top of the
stack, ST ACK P OP writes a null value to the top of the stack, and RESU LT writes vactive to
Qresult (presult ).
ARG 2 (Auxiliary Variable): COLOR GREY, COLOR BLACK
COLOR GREY and COLOR BLACK color the given vertex grey and black, respectively.
14
Published as a conference paper at ICLR 2017
MOVE.
The MOVE operation has the following arguments:
ARG 1 (Pointer): presult , pstack , pstart , childList[vactive ], childList[vsave ]
Note that the argument is the identity of the pointer, not what the pointer points to; in other words,
ARG 1 can only take one of 5 values.
ARG 2 (Increment or Decrement): UP, DOWN
A.2
A.2.1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
T RACE -G ENERATING F UNCTIONS FOR T OPOLOGICAL S ORT
N ON -R ECURSIVE T RACE -G ENERATING F UNCTIONS
// Top level topological sort call
TOPOSORT() {
while (Qcolor (pstart ) is a valid color): // color invalid when all vertices explored
WRITE(ACTIVE_START)
WRITE(COLOR_CURR, COLOR_GREY)
TRAVERSE()
MOVE(pstart , UP)
NEXT_START()
}
TRAVERSE() {
CHECK_CHILD()
EXPLORE()
}
CHECK_CHILD() {
while (Qcolor (DAG[vactive ][childList[vactive ]]) is not white and is not invalid): // color invalid when all children explored
MOVE(childList[vactive ], UP)
}
EXPLORE() {
do
if (Qcolor (DAG[vactive ][childList[vactive ]]) is white):
WRITE(COLOR_NEXT, COLOR_GREY)
STACK(PUSH)
WRITE(SAVE)
WRITE(ACTIVE_NEIGHB)
MOVE(childList[vsave ], UP)
else:
WRITE(COLOR_CURR, COLOR_BLACK)
WRITE(RESULT)
MOVE(presult , UP)
if(pstack == 1):
break
else:
STACK(POP)
CHECK_CHILD()
while (true)
}
STACK(op) {
if (op == PUSH):
WRITE(STACK_PUSH)
MOVE(pstack , UP)
}
if (op == POP):
WRITE(ACTIVE_STACK)
WRITE(STACK_POP)
MOVE(pstack , DOWN)
NEXT_START() {
while(Qcolor (pstart ) is not white and is not invalid): // color invalid when all vertices explored
MOVE(pstart , UP)
}
15
Published as a conference paper at ICLR 2017
A.2.2
R ECURSIVE T RACE -G ENERATING F UNCTIONS
Altered Recursive Functions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
// Top level topological sort call
TOPOSORT() {
if (Qcolor (pstart ) is a valid color): // color invalid when all vertices explored
WRITE(ACTIVE_START)
WRITE(COLOR_CURR, COLOR_GREY)
TRAVERSE()
MOVE(pstart , UP)
NEXT_START()
TOPOSORT() // Recursive Call
}
CHECK_CHILD() {
if (Qcolor (DAG[vactive ][childList[vactive ]]) is not white and is not invalid): // color invalid when all children explored
MOVE(childList[vactive ], UP)
CHECK_CHILD() // Recursive Call
}
EXPLORE() {
if (Qcolor (DAG[vactive ][childList[vactive ]]) is white):
WRITE(COLOR_NEXT, COLOR_GREY)
STACK(PUSH)
WRITE(SAVE)
WRITE(ACTIVE_NEIGHB)
MOVE(childList[vsave ], UP)
else:
WRITE(COLOR_CURR, COLOR_BLACK)
WRITE(RESULT)
MOVE(presult , UP)
if(pstack == 1):
return
else:
STACK(POP)
CHECK_CHILD()
EXPLORE() // Recursive Call
}
NEXT_START() {
if (Qcolor (pstart ) is not white and is not invalid): // color invalid when all vertices explored
MOVE(pstart , UP)
NEXT_START() // Recursive Call
}
A.3
N ON -R ECURSIVE Q UICKSORT
Algorithm 4 Iterative Quicksort
1: Initialize an array A to sort and two empty stacks Slo and Shi .
2: Initialize lo and hi to be 1 and n, where n is the length of A.
3:
4: function PARTITION(A, lo, hi)
5:
pivot = lo
6:
for j ∈ [lo, hi − 1] : do
7:
if A[j] ≤ A[hi] then
8:
swap A[pivot] with A[j]
9:
pivot = pivot + 1
10:
swap A[pivot] with A[hi]
11:
return pivot
12:
13: function Q UICK S ORT(A, lo, hi)
14:
while Slo and Shi are not empty: do
15:
Pop states off Slo and Shi , writing them to lo and hi.
16:
p = PARTITION(A, lo, hi)
17:
Push p + 1 and hi to Slo and Shi .
18:
Push lo and p − 1 to Slo and Shi .
16
Published as a conference paper at ICLR 2017
A.4
P ROGRAM S ET FOR Q UICKSORT
Program
QUICKSORT
Descriptions
Run the quicksort
routine in place for
the array A, for
indices from lo to hi
PARTITION
Runs the partition
function.
At end,
pointer ppivot is
moved to the pivot
Runs the FOR loop
inside the partition
function
Compares
A[pivot] ≤ A[j]; if
so, perform a swap
and increment ppivot
Sets ppivot to lo index
Sets pj to lo index
Sets pj to −∞
Pushes lo/hi states
onto stacks Slo and
Shi according to
argument (described
below)
Moves pointer one
unit up or down
Swaps elements at
given array indices
Write
a
value
either
to
stack
(e.g., QstackLo or
QstackHi ) or to
pointer (e.g.,
to
change the value of
phi )
COMPSWAP LOOP
COMPSWAP
SET PIVOT LO
SET J LO
SET J NULL
STACK
MOVE
SWAP
WRITE
Calls
Non-Recursive: PARTITION,
STACK,
WRITE
Recursive: same as
non-recursive version,
along with QUICKSORT
COMPSWAP LOOP,
MOVE PIVOT LO,
MOVE J LO, SWAP
Arguments
Implicitly: array
A to sort, lo, hi
COMPSWAP, MOVE
NONE
SWAP, MOVE
NONE
NONE
NONE
NONE
NONE
WRITE, MOVE
NONE
NONE
Described below
NONE
Described below
NONE
Described below
NONE
Described below
NONE
Argument Sets for STACK, MOVE, SWAP, WRITE.
STACK.
The STACK operation has the following arguments:
ARG 1 (Operation): STACK PUSH CALL1, STACK PUSH CALL2, STACK POP
STACK PUSH CALL1 pushes lo and pivot − 1 to QstackLo and QstackHi . STACK PUSH CALL2
pushes pivot + 1 and hi to QstackLo and QstackHi . STACK POP pushes −∞ values to QstackLo
and QstackHi .
MOVE.
The MOVE operation has the following arguments:
ARG 1 (Pointer): pstackLo , pstackHi , pj , ppivot
Note that the argument is the identity of the pointer, not what the pointer points to; in other words,
ARG 1 can only take one of 4 values.
ARG 2 (Increment or Decrement): UP, DOWN
17
Published as a conference paper at ICLR 2017
SWAP.
The SWAP operation has the following arguments:
ARG 1 (Swap Object 1): ppivot
ARG 2 (Swap Object 2): phi , pj
WRITE. The WRITE operation has the following arguments:
ARG 1 (Object to Write): ENV STACK LO, ENV STACK HI, phi , plo
ENV STACK LO and ENV STACK HI represent QstackLo (pstackLo ) and QstackHi (pstackHi ), respectively.
ARG 2 (Object to Copy): ENV STACK LO PEEK, ENV STACK HI PEEK, phi , plo , ppivot − 1,
ppivot + 1, RESET
ENV STACK LO PEEK and ENV STACK HI PEEK represent QstackLo (pstackLo − 1) and
QstackHi (pstackHi − 1), respectively. RESET represents a −∞ value.
Note that the argument is the identity of the pointer, not what the pointer points to; in other words,
ARG 1 can only take one of 4 values, and ARG 2 can only take one of 7 values.
A.5
A.5.1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
T RACE -G ENERATING F UNCTIONS FOR Q UICKSORT
N ON -R ECURSIVE T RACE -G ENERATING F UNCTIONS
Initialize plo to 1 and phi to n (length of array)
Initialize pj to −∞
QUICKSORT() {
while (pstackLo 6= 1):
if (QstackLo (pstackLo − 1) < QstackHi (pstackHi − 1)):
STACK(STACK_POP)
else:
WRITE(phi , ENV_STACK_HI_PEEK)
WRITE(plo , ENV_STACK_LO_PEEK)
STACK(STACK_POP)
PARTITION()
STACK(STACK_PUSH_CALL2)
STACK(STACK_PUSH_CALL1)
}
PARTITION() {
SET_PIVOT_LO()
SET_J_LO()
COMPSWAP_LOOP()
SWAP(ppivot , phi )
SET_J_NULL()
}
COMPSWAP_LOOP() {
while (pj 6= phi ):
COMPSWAP()
MOVE(pj , UP)
}
COMPSWAP() {
if (A[pj ] ≤ A[phi ]):
SWAP(ppivot , pj )
MOVE(ppivot , UP)
}
STACK(op) {
if (op == STACK_PUSH_CALL1):
WRITE(ENV_STACK_LO, plo )
WRITE(ENV_STACK_HI, ppivot − 1)
MOVE(pstackLo , UP)
MOVE(pstackHi , UP)
if (op == STACK_PUSH_CALL2):
WRITE(ENV_STACK_LO, ppivot + 1)
WRITE(ENV_STACK_HI, phi )
MOVE(pstackLo , UP)
MOVE(pstackHi , UP)
}
if (op == STACK_POP):
WRITE(ENV_STACK_LO, RESET)
WRITE(ENV_STACK_HI, RESET)
MOVE(pstackLo , DOWN)
MOVE(pstackHi , DOWN)
18
Published as a conference paper at ICLR 2017
A.5.2
R ECURSIVE T RACE -G ENERATING F UNCTIONS
Altered Recursive Functions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Initialize plo to 1 and phi to n (length of array)
Initialize pj to −∞
QUICKSORT() {
if (QstackLo (pstackLo − 1) < QstackHi (pstackHi − 1)):
PARTITION()
STACK(STACK_PUSH_CALL2)
STACK(STACK_PUSH_CALL1)
WRITE(phi , ENV_STACK_HI_PEEK)
WRITE(plo , ENV_STACK_LO_PEEK)
QUICKSORT() // Recursive Call
STACK(STACK_POP)
WRITE(phi , ENV_STACK_HI_PEEK)
WRITE(plo , ENV_STACK_LO_PEEK)
QUICKSORT() // Recursive Call
STACK(STACK_POP)
}
COMPSWAP_LOOP() {
if (pj 6= phi ):
COMPSWAP()
MOVE(pj , UP)
COMPSWAP_LOOP() // Recursive Call
}
A.6
BASE CASES , REDUCTION RULES , AND VERIFICATION SETS
In this section, we describe the space of base cases and reduction rules that must be covered for each
of the four sample tasks, in order to create the verification set.
For addition, we analytically determine the verification set. For tasks other than addition, it is difficult to analytically determine the verification set, so instead, we randomly generate input candidates
until they completely cover the base cases and reduction rules.
Base Cases and Reduction Rules for Addition. For the recursive formulation of addition, we
analytically construct the set of input problems that cover all base cases and reduction rules. We
outline how to construct this set.
It is sufficient to construct problems where every transition between two adjacent columns is covered. The ADD reduction rule ensures that each call to ADD only covers two adjacent columns,
and so the LSTM only ever runs for a fixed number of steps necessary to process these two columns.
We construct input problems by splitting into two cases: one case in which the left column contains
a null value and another in which the left column does not contain any null values. We then construct
problem configurations that span all possible valid environment states (for instance, in order to force
the carry bit in a column to be 1, one can add the sum “1+9” in the column to the right).
The operations we need to be concerned most about are CARRY and LSHIFT, which induce partial
environment states spanning two columns. It is straightforward to deal with all other operations,
which do not induce partial environment states.
Under the assumption that there are no leading 0’s (except in the case of single digits) and the
two numbers to be added have the same number of digits, the verification set for addition contains
20,181 input problems. The assumption of leading 0’s can be easily removed, at the cost of slightly
increasing the size of the verification set. We made the assumption of equivalent lengths in order to
parametrize the input format with respect to length, but this assumption can be removed as well.
Base Cases and Reduction Rules for Bubble Sort. The original version of the bubblesort implementation exposes the values within the array. While this matches the description from Reed &
de Freitas (2016), we found that this causes an unnecessary blowup in the size of V and makes it
much more difficult to construct the verification set. For purposes of verification, we replace the
domain-specific encoder with the following:
fenc (Q, i1 , i2 , i3 , at ) = M LP ([Q(1, i1 ) ≤ Q(1, i2 ), 1 ≤ i1 ≤ length, 1 ≤ i2 ≤ length,
i3 == length, at (1), at (2), at (3)]),
19
Published as a conference paper at ICLR 2017
Table 4: Accuracy on Randomly Generated Problems for Variant of Bubble Sort
Length of Array
Non-Recursive
Recursive
2
3
4
5
6
7
8
9
10
12
15
70
100%
100%
100%
100%
90%
86.7%
6.7%
0%
0%
0%
0%
0%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
which directly exposes which of the two values pointed to is larger. This modification also enables
us to sort arrays containing arbitrary comparable elements.
By reasoning about the possible set of environment observations created by all valid inputs, we
construct V using the procedure described in Section 3.3. Using this modification, we constructed a
verification set consisting of one array of size 10.
We also report on generalization results for the non-recursive and recursive versions of this variant of
bubble sort. Table 4 demonstrates that the accuracy of the non-recursive program degrades sharply
when moving from arrays of length 7 to arrays of length 8. This is due to the properties of the training
set—we trained on 2 traces synthesized from arrays of length 7 and 1 trace synthesized from an array
of length 6. Table 4 also demonstrates that the (verified) recursive program generalizes perfectly.
Base Cases and Reduction Rules for Topological Sort. For each function we use to implement
the recursive version of topological sort, we need to consider the set of possible environment observation sequences we can create from all valid inputs and test that the learned program produces the
correct behavior on each of these inputs. We have three observations: the color of the start node, the
color of the active node’s next child to be considered, and whether the stack is empty. Naı̈vely, we
might expect to synthesize and test an input for any sequence created by combining the four possible
colors in two variables and another boolean variable for whether the stack is empty (so 32 possible
observations at any point), but for various reasons, most of these combinations are impossible to
occur at any given point in the execution trace.
Through careful reasoning about the possible set of environment observations created by all valid
inputs, and how each of the operations in the execution trace affects the environment, we can construct V using the procedure described in Section 3.3. We then construct a verification set of size 73
by ensuring that randomly generated graphs cover the analytically derived V . The model described
in the training setup of Section 4 (trained on 6 traces) was verified to be correct via the matching
procedure described in Section 4.2.
Base Cases and Reduction Rules for Quicksort. As with the others, we apply the procedure
described in Section 3.3 to construct V and then empirically create a verification set which covers
V . The verification set can be very small, as we found a 10-element array ([8,2,1,2,0,8,5,8,3,7]) is
sufficient to cover all of V . We note that an earlier version of quicksort we tried lacked primitive
operations to directly move a pointer to another, and therefore needed more functions and observations. As this complexity interfered with determining the base cases and reductions, we changed the
algorithm to its current form. Even though the earlier version also generalized just as well in practice, relatively small differences in the formulation of the traces and the environment observations
can drastically change the difficulty of verification.
20
| 9cs.NE
|
1
Generalized Common Informations:
Measuring Commonness by the Conditional
Maximal Correlation
arXiv:1610.09289v3 [cs.IT] 24 Jul 2017
Lei Yu, Houqiang Li, Senior Member, IEEE, and Chang Wen Chen, Fellow, IEEE
Abstract
In literature, different common informations were defined by Gács and Körner, by Wyner, and by Kumar, Li, and
Gamal, respectively. In this paper, we define two generalized versions of common informations, named approximate
and exact information-correlation functions, by exploiting the conditional maximal correlation as a commonness or
privacy measure. These two generalized common informations encompass the notions of Gács-Körner’s, Wyner’s,
and Kumar-Li-Gamal’s common informations as special cases. Furthermore, to give operational characterizations of
these two generalized common informations, we also study the problems of private sources synthesis and common
information extraction, and show that the information-correlation functions are equal to the minimum rates of
commonness needed to ensure that some conditional maximal correlation constraints are satisfied for the centralized
setting versions of these problems. As a byproduct, the conditional maximal correlation has been studied as well.
Index Terms
Common information, conditional maximal correlation, information-correlation function, sources synthesis, information extraction
I. I NTRODUCTION
Common information, as an information measure on the common part between two random variables, was first
investigated by Gács and Körner [1] in content of distributed common information extraction problem: extracting
a same random variable from each of two sources individually. The common information of the sources is defined
by the maximum information of the random variable that can be extracted from them. For correlated memoryless
sources X, Y (taken from finite alphabets), [1] shows that the Gács-Körner common information between them is
CGK (X; Y ) =
sup
H(f (X)).
(1)
f,g:f (X)=g(Y )
Lei Yu is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore (e-mail: [email protected]).
This work was done when he was at University of Science and Technology of China. Houqiang Li is with the Department of Electronic
Engineering and Information Science, University of Science and Technology of China, Hefei, China (e-mail: [email protected]). Chang Wen
Chen is with Department of Computer Science and Engineering, State University of New York at Buffalo, Buffalo, NY, USA (e-mail:
[email protected]).
July 25, 2017
DRAFT
It also can be expressed as
CGK (X; Y ) =
I(XY ; U ),
(2)
H(f (X, U ) |U )
(3)
inf
PU |XY :CGK (X;Y |U )=0
(the proof of (2) is given in Appendix A), where
CGK (X; Y |U ) :=
sup
f,g:f (X,U )=g(Y,U )
denotes the conditional common information between X, Y given U . The constraint CGK (X; Y |U ) = 0 in (2)
implies all the common information between X, Y is contained in U .
Wyner [3] studied distributed source synthesis (or distributed source simulation) problem, and defined common
information in a different way. Specifically, he defined common information as the minimum information rate needed
to generate sources in a distributed manner with asymptotically vanishing normalized relative entropy between the
induced distribution and some target joint distribution. Given a target distribution PXY , this common information
is proven to be
CW (X; Y ) =
inf
PU |XY :X→U →Y
I(XY ; U ).
(4)
Furthermore, as a related problem, the problem of exactly generating target sources was studied by Kumar, Li, and
Gamal recently [12]. The notion of exact common information (rate) (denoted as KKLG (X; Y )) is introduced, which
is defined to be the minimum code rate to ensure the induced distribution is exactly (instead approximately) same
to some target joint distribution. By comparing these common informations, it is easy to show that CGK (X; Y ) ≤
I(X; Y ) ≤ CW (X; Y ) ≤ KKLG (X; Y ) ≤ H(XY ).
Observe that in the definitions of Gács-Körner and Wyner common informations, different dependency constraints
are used. Gács-Körner common information requires the common variable U to be some function of each of the
sources (or equivalently, there is no conditional common information given U ); while Wyner common information
requires the sources conditionally independent given the common variable U . These two constraints are closely
related to an important dependency measure, Hirscbfeld-Gebelein-Renyi maximal correlation (or simply maximal
correlation). This correlation measures the maximum (Pearson) correlation between square integrable real-valued
random variables generated by the individual random variables. According to the definition, maximal correlation is
invariant on bijective mappings (or robust to bijective transform), hence it reveals some kind of intrinsic dependency
between two sources. This measure was first introduced by Hirschfeld [5] and Gebelein [4], then studied by Rényi
[6], and recently it has been exploited to some interesting problems of information theory, such as measure of
non-local correlations [9], maximal correlation secrecy [10], converse result of distributed communication [14],
etc. Furthermore, maximal correlation also indicates the existence of Gács-Körner or Wyner common information:
There exists Gács-Körner common information between two sources if and only if the maximal correlation between
them equals one; and there exists Wyner common information between two sources if and only if the maximal
correlation between them is positive.
The common informations proposed by Gács and Körner and by Wyner (or by Kumar, Li, and Gamal) are
defined in two different problems: distributed common information extraction and distributed source synthesis. In
these problems, the common informations are defined from different points of view. One attempt to unify them can
2
be found in [11], where Kamath and Anantharam converted common information extraction problem into a special
case of distributed source synthesis problem by specifying the synthesized distribution to be that of the common
randomness. In this paper, we attempt to give another unification of the existing common informations. Specifically,
we unify and generalize the Gács-Körner and Wyner common informations by defining a generalized common
information, (approximate) information-correlation function. In this generalized definition, the conditional maximal
correlation (the conditional dependency of the sources given the common randomness) is exploited to measure the
privacy (or commonness), and the mutual information is used to measure the information amount of such common
randomness. The Gács-Körner common information and Wyner common information are two special and extreme
cases of our generalized definition with correlation respectively being 0 and 1−1 , and hence both of them can be
seen as hard-measures of common information. However, in our definition, correlation could be any number between
0 and 1, hence our definition gives a soft-measure of common information. Our results give a more comprehensive
answer to the classic problem: What is the common information between two correlated sources? Furthermore,
similarly we also unify and generalize the Gács-Körner and Kumar-Li-Gamal common informations into another
generalized common information, (exact) information-correlation function. To give an operational interpretation
of the approximate and exact generalized common informations, we also study common information extraction
problem and private sources synthesis problem, and show that the information-correlation functions correspond to
the minimum achievable rates under privacy constraints for the centralized case of each problem.
The rest of this paper is organized as follows. Section II summarizes definitions and properties of maximal
correlation. Section III defines information-correlation function and provides the basic properties. Sections IV and
V investigate the private sources synthesis problem and common information extraction problem respectively. Finally,
Section VI gives the concluding remarks.
A. Notation and Preliminaries
We use PX (x) to denote the probability distribution of random variable X, which is also shortly denoted as PX
U
or P (x). We also use PX and QX to denote different probability distribution with common alphabet X . We use PX
to denote the uniform distribution over the set X , unless otherwise stated. We use fP or fQ to denote a quantity or
operation f that is defined on pmf P or Q. The total variation distance between two probability measures P and
Q with common alphabet is defined by
kP − QkT V := sup |P (A) − Q(A)|
(5)
A∈F
where F is the σ-algebra of the probability space.
In this paper, some achievability schemes involves a random codebook C (or a random binning B). For simplicity,
we also denote the induced conditional distribution PX|C=c (given C = c) as PX (suppressing the condition C = c),
which can be seen as a random pmf.
For any pmfs PX and QX on X , we write PX ≈ QX if kPX − QX kT V < for non-random pmfs, or
EC kPX − QX kT V < for random pmfs. For any two sequences of pmfs PX (n) and QX (n) on X (n) (where X (n) is
1 1−
implies the correlation approaching 1 from the left.
3
arbitrary and it differs from X n which is a Cartesian product), we write PX (n) ≈ QX (n) if limn→∞ kPX (n) − QX (n) kT V =
0 for non-random pmfs, or limn→∞ EC kPX (n) − QX (n) kT V = 0 for random pmfs.
The following properties of total variation distance hold.
Property 1. [19], [22] Total variation distance satisfies:
1) If the support of P and Q is a countable set X , then
1X
|P (x) − Q(x)|.
2
kP − QkT V =
(6)
x∈X
2) Let > 0 and let f (x) be a function with bounded range of width b > 0. Then
PX ≈ QX ⇒ |EP f (X) − EQ f (X)| < b,
(7)
where EP indicates that the expectation is taken with respect to the distribution P.
3) PX (n) ≈ QX (n) ⇒ PX (n) PY (n) |X (n) ≈ QX (n) PY (n) |X (n) ,
PX (n) PY (n) |X (n) ≈ QX (n) QY (n) |X (n) ⇒ PX (n) ≈ QX (n) .
4) For any two sequences of non-random pmfs PX (n) Y (n) and QX (n) Y (n) , if PX (n) PY (n) |X (n) ≈ QX (n) QY (n) |X (n) ,
then there exists a sequence x(n) ∈ X (n) such that PY (n) |X (n) =x(n) ≈ QY (n) |X (n) =x(n) .
5) If PX (n) ≈ QX (n) and PX (n) PY (n) |X (n) ≈ PX (n) QY (n) |X (n) , then PX (n) PY (n) |X (n) ≈ QX (n) QY (n) |X (n) .
II. (C ONDITIONAL ) M AXIMAL C ORRELATION
In this section, we first define several correlations, including (Pearson) correlation, correlation ratio, and maximal
correlation, and then study their properties. These concepts and properties will be used to define and investigate
information-correlation functions in subsequent sections.
In this section, we assume all alphabets are general (not limited to finite or countable) unless otherwise stated.
A. Definition
Definition 1. For any random variables X and Y with alphabets X ⊆ R and Y ⊆ R, the (Pearson) correlation of
X and Y is defined by
ρ(X; Y ) =
√
cov(X,Y )
var(X)
√
var(Y )
, if var(X)var(Y ) > 0,
if var(X)var(Y ) = 0.
0,
Moreover, the conditional correlation of X and Y given another random variable U is defined by
√ E[cov(X,Y
√ |U )]
, if E[var(X|U )]E[var(Y |U )] > 0,
E[var(X|U )] E[var(Y |U )]
ρ(X; Y |U ) =
0,
if E[var(X|U )]E[var(Y |U )] = 0.
(8)
(9)
Definition 2. For any random variables X and Y with alphabets X ⊆ R and Y, the correlation ratio of X on Y
is defined by
θ(X; Y ) = sup ρ(X; g(Y )),
g
4
(10)
where the supremum is taken over all the functions g : Y 7→ R. Moreover, the conditional correlation ratio of X
on Y given another random variable U with alphabet U is defined by
(11)
θ(X; Y |U ) = sup ρ(X; g(Y, U )|U ),
g
where the supremum is taken over all the functions g : Y × U 7→ R.
Remark 1. Note that in general θ(X; Y ) 6= θ(Y ; X) and θ(X; Y |U ) 6= θ(Y ; X|U ).
Definition 3. For any random variables X and Y with alphabets X and Y, the maximal correlation of X and Y
is defined by
(12)
ρm (X; Y ) = sup ρ(f (X); g(Y )),
f,g
where the supremum is taken over all the functions f : X 7→ R, g : Y 7→ R. Moreover, the conditional maximal
correlation of X and Y given another random variable U with alphabet U is defined by
(13)
ρm (X; Y |U ) = sup ρ(f (X, U ); g(Y, U )|U ),
f,g
where the supremum is taken over all the functions f : X × U 7→ R, g : Y × U 7→ R.
It is easy to verify that
(14)
ρm (X; Y |U ) = sup θ(f (X, U ); Y |U ).
f
Note that the unconditional versions of correlation coefficient, correlation ratio, and maximal correlation have
been well studied in literature. The conditional versions are first introduced by Beigi and Gohari recently [9],
where it is named as maximal correlation of a box and used to study the problem of non-local correlations. In this
paper, we will well study conditional maximal correlation (and conditional correlation ratio), and give some useful
properties.
B. Properties
According to the definition, maximal correlation remains the same after applying bijective transform (one-to-one
correspondence) on each of the variables. Hence it is robust to bijective transform. Furthermore, for finite valued
random variables maximal correlation ρm (X; Y |U ) can be characterized by the second largest singular value λ2 (u)
of the matrix Qu with entries
Qu (x, y) := p
p(x, y|u)
p(x|u)p(y|u)
=p
p(x, y, u)
p(x, u)p(y, u)
.
(15)
Lemma 1. (Singular value characterization). For any random variables X, Y, U,
ρm (X; Y |U ) =
sup
λ2 (u).
(16)
u:P (u)>0
Remark 2. This shows the conditional maximal correlation is consistent with the unconditional version (U = ∅)
[2]
ρm (X; Y ) = λ2 .
5
(17)
Furthermore, for any random variables X, Y, U with finite alphabets, the supremum in (12), (13) and (16) is actually
a maximum.
The proof of this lemma is given in Appendix B. This lemma gives a simple approach to compute (conditional)
maximal correlation. Observe that λ2 (u) is equal to the maximal correlation ρm (X; Y |U = u) between X and Y
under condition U = u, and under distribution PXY |U =u . Hence Lemma 1 leads to the following result.
Lemma 2. (Alternative characterization). For any random variables X, Y, U,
ρm (X; Y |U ) =
sup
u:P (u)>0
(18)
ρm (X; Y |U = u).
Note that the right-hand side of (18) was first defined by Beigi and Gohari [9]. This lemma implies the equivalence
between the conditional maximal correlation defined by us and that defined by Beigi and Gohari.
Furthermore, Lemmas 1 and 2 also hold for continuous random variables, if the constraint of P (u) > 0 is
replaced with p(u) > 0. Here p(u) denotes the probability density function (pdf) of U . Notice that Lemmas 1 and
2 imply that ρm (X; Y |U ) can be different for different distributions of X, Y , even if the distributions are only
different up to a zero measure set. In measure theory, people usually do not care the difference with zero measure.
Therefore, we refine the definition of conditional maximal correlation for continuous random variables by defining
a robust version as follows.
ρem (X; Y |U ) :=
inf
qXY U :qXY U =pXY U a.s.
(19)
ρm,q (X; Y |U ),
for continuous random variables X, Y, U, with pdf pXY U . We name ρem (X; Y |U ) as robust conditional maximal
correlation. Obviously, for discrete random variables case, robust conditional maximal correlation is consistent
with conditional maximal correlation. Moreover, if we take inf qXY U :qXY U =pXY U
a.s.
operation on each side of an
equality or inequality about qXY U , it usually does not change the equality or inequality. Hence in this paper, we
only consider conditional maximal correlations rather than their robust versions.
Lemma 3. (TV bound on maximal correlation). For any random variables X, Y, U with finite alphabets,
ρm,Q (X; Y |U ) ≥
where Pm =
min
x,y,u:P (x,y,u)>0
P (x, y|u), and δ =
ρm,P (X; Y |U ) −
1+
1+
where Qm =
min
x,y,u:Q(x,y,u)>0
4δ
Pm
4δ
Pm
(20)
,
max kPXY |U =u − QXY |U =u kT V .
u:P (u)>0
Remark 3. Lemma 3 implies
ρm,P (X; Y |U ) −
4δ
Pm
4δ
Pm
≤ ρm,Q (X; Y |U ) ≤
4δ
1+
Qm
ρm,P (X; Y |U ) +
4δ
,
Qm
(21)
Q(x, y|u).
Proof: Assume u achieves the supremum in (18), and f, g satisfying EP [f (X, U )|U = u] = 0, EP [g(Y, U )|U =
u] = 0, varP [f (X, U )|U = u] = 1, varP [g(Y, U )|U = u] = 1, achieves ρm,P (X; Y |U ). Then P (x|u)f 2 (x, u) ≤
P
2
x P (x|u)f (x, u) = 1 for any x, u, i.e.,
|f (x, u)| ≤ p
6
1
P (x|u)
(22)
for any x, u such that P (x|u) > 0. Furthermore, for any x, u such that P (x|u) > 0, we have P (x|u) ≥ P (x, y|u) ≥
Pm . Hence (22) implies
1
.
Pm
(23)
1
.
Pm
(24)
|f (x, u)| ≤ √
Similarly, we have
|g(y, u)| ≤ √
According to Property (7), the following inequalities hold.
|EQ [f (X, U )g(Y, U )|U ] − EP [f (X, U )g(Y, U )|U ]| ≤
(25)
2δ
2
kPX|U − QX|U kT V ≤ √
Pm
Pm
(26)
2
2δ
kPY |U − QY |U kT V ≤ √
Pm
Pm
(27)
|EQ [f (X, U )|U ]| ≤ √
|EQ [g(Y, U )|U ]| ≤ √
2δ
Pm
|EQ [f 2 (X, U )|U ] − 1| ≤
2δ
Pm
(28)
|EQ [g 2 (Y, U )|U ] − 1| ≤
2δ
.
Pm
(29)
and
Therefore, we have
EQ [f (X, U )g(Y, U )|U ] − EQ [f (X, U )|U ]EQ [g(Y, U )|U ]
q
ρm,Q (X; Y |U = u) ≥ q
EQ [f 2 (X, U )|U ] − E2Q [f (X, U )|U ] EQ [g 2 (Y, U )|U ] − E2Q [g(Y, U )|U ]
≥q
≥
=
EQ [f (X, U )g(Y, U )|U ] − P4δm
q
EQ [f 2 (X, U )|U ] − P4δm EQ [g 2 (Y, U )|U ] −
4δ
Pm
ρm,P (X; Y |U = u) − P4δm
q
q
1 + P4δm 1 + P4δm
ρm,P (X; Y |U = u) −
1+
4δ
Pm
(30)
(31)
(32)
4δ
Pm
.
(33)
Lemma 4. (Continuity and discontinuity). Assume X, Y, U have finite alphabets. Then given PU , ρm (X; Y |U ) is
continuous in PXY |U . Given PXY |U , ρm (X; Y |U ) is continuous on {PU : PU (u) > 0, ∀u ∈ U }. But in general,
ρm (X; Y |U ) is discontinuous in PXY U .
Proof: (21) implies for given PU , as
max kPXY |U =u − QXY |U =u kT V → 0, ρm,Q (X; Y |U ) → ρm,P (X; Y |U ).
u:P (u)>0
Hence for given PU , ρm,P (X; Y |U ) is continuous in PXY |U . Furthermore, since given PXY |U , ρm (X; Y |U ) =
sup
u:P (u)>0
λ2 (u), we have for given PXY |U , ρm (X; Y |U ) is continuous on {PU : PU (u) > 0, ∀u ∈ U}. But it is
worth noting that ρm (X; Y |U ) may be discontinuous at PU such that PU (u) = 0 for some u ∈ U. Therefore,
QXY U → PXY U in total variation sense does not necessarily imply ρm,Q (X; Y |U ) → ρm (X; Y |U ). That is, the
conditional maximal correlation may be discontinuous in probability distribution PXY U .
Furthermore, some other properties hold.
7
Lemma 5. (Concavity). Given PXY |U , ρm (X; Y |U ) is concave in PU .
Proof:
Fix PXY |U . Assume RU = λPU + (1 − λ) QU , λ ∈ (0, 1), then by Lemma 2, we have
ρm,R (X; Y |U ) =
=
sup
u:R(u)>0
sup
u:P (u)>0 or Q(u)>0
= max
(
(34)
ρm (X; Y |U = u)
sup
u:P (u)>0
(35)
ρm (X; Y |U = u)
ρm (X; Y |U = u),
sup
u:Q(u)>0
)
ρm (X; Y |U = u)
= max {ρm,P (X; Y |U ), ρm,Q (X; Y |U )} .
(36)
(37)
Hence ρm,R (X; Y |U ) ≥ λρm,P (X; Y |U ) + (1 − λ) ρm,Q (X; Y |U ), i.e., ρm (X; Y |U ) is concave in PU .
Lemma 6. For any random variables X, Y, Z, U , the following inequalities hold.
0 ≤ |ρ(X; Y |U )| ≤ θ(X; Y |U ) ≤ ρm (X; Y |U ) ≤ 1.
(38)
Moreover, ρm (X; Y |U ) = 0 if and only if X and Y are conditionally independent given U ; ρm (X; Y |U ) = 1 if
and only if X and Y have Gács-Körner common information given U.
Proof:
|E[cov(X, Y |U )]| = |E[(X − E[X|U ])(Y − E[Y |U ])]|
p
≤ E[(X − E[X|U ])2 ]E[(Y − E[Y |U ])2 ]
p
= E[var(X|U )]E[var(Y |U )],
(39)
(40)
(41)
where (40) follows from the Cauchy-Schwarz inequality. Hence
0 ≤ |ρ(X; Y |U )| ≤ 1
(42)
0 ≤ |ρ(X; Y |U )| ≤ θ(X; Y |U ) ≤ ρm (X; Y |U ) ≤ 1
(43)
which further implies
since both θ(X; Y |U ) and ρm (X; Y |U ) are conditional correlations for some variables.
If X and Y are conditionally independent given U , then for any functions f and g, f (X, U ) and g(Y, U ) are
also conditionally independent given U . This leads to ρm (X; Y |U ) = 0.
Conversely, if ρm (X; Y |U ) = 0, then
ρ(f (X, U ); g(Y, U )|U ) = 0
(44)
for any functions f and g. For any x, u, set f (X, U ) = 1{X = x, U = u} and g(Y, U ) = 1{Y = y, U = u}, then
E[cov(X, Y |U )] = P(X = x, Y = y|U = u) − P(X = x|U = u)P(Y = y|U = u)
8
(45)
= PXY |U (x, y|u) − PX|U (x|u)PY |U (y|u).
(46)
Hence (44) implies
PXY |U (x, y|u) = PX|U (x|u)PY |U (y|u).
(47)
This implies X and Y are conditionally independent given U . Therefore, ρm (X; Y |U ) = 0 if and only if X and
Y are conditionally independent given U.
Assume X and Y have Gács-Körner common information given U , i.e., f (X, U ) = g(Y, U ) with probability 1
for some functions f and g such that H(f (X, U )|U ) > 0. Then Evar(f (X, U )|U )Evar(g(Y, U )|U ) > 0, and
ρm (X; Y |U ) ≥ ρ(f (X, U ); g(Y, U )|U ) ≥ 1.
(48)
Combining this with ρm (X; Y |U ) ≤ 1, we have ρm (X; Y |U ) = 1.
Assume ρm (X; Y |U ) = 1, then f (X, U ) = g(Y, U ) with probability 1 for some functions f and g such that
Evar(f (X, U )|U )Evar(g(Y, U )|U ) > 0, or equivalently, H(f (X, U )|U ) > 0. This implies X and Y have GácsKörner common information given U . Therefore, ρm (X; Y |U ) = 1 if and only if X and Y have Gács-Körner
common information given U .
Lemma 7. For any random variables X, Y, Z, U , the following properties hold.
θ(X; Y Z|U ) ≥ θ(X; Y |U );
(49)
ρm (X; Y Z|U ) ≥ ρm (X; Y |U );
(50)
θ(X; Y |U ) =
=
ρm (X; Y |U ) = sup
f
= sup
f
s
E[var(E[X|Y U ]|U )]
E[var(X|U )]
s
1−
s
E[var(E[f (X, U )|Y U ]|U )]
E[var(f (X, U )|U )]
s
1−
E[var(X|Y U )]
;
E[var(X|U )]
E[var(f (X, U )|Y U )]
.
E[var(f (X, U )|U )]
(51)
(52)
In particular if U is degenerate, then the inequalities above reduce to
θ(X; Y Z) ≥ θ(X; Y );
(53)
ρm (X; Y Z) ≥ ρm (X; Y );
(54)
θ(X; Y ) =
=
s
s
var(E[X|Y ])
var(X)
1−
9
E[var(X|Y )]
;
var(X)
(55)
ρm (X; Y ) = sup
f
= sup
f
s
s
var(E[f (X)|Y ])
var(f (X))
1−
E[var(f (X)|Y )]
.
var(f (X))
(56)
Remark 4. Correlation ratio is also closely related to Minimum Mean Square Error (MMSE). The optimal MMSE
estimator is E[X|Y U ], hence the variance of the MMSE for estimating X given (Y, U ) is mmse(X|Y U ) =
E(X − E[X|Y U ])2 = E[var(X|Y U )] = E[var(X|U )](1 − θ2 (X; Y |U )).
Proof: According to definitions of conditional correlation ratio and conditional maximal correlation, (49) and
(50) can be proven easily.
In fact, we may, without loss of the generality, consider only such function g for which E[g(Y, U )|U = u] = 0, ∀u
and var(g(Y, U )|U = u) = 1, ∀u and suppose E[X] = 0, E[var(X|U )] = 1; for this case we have by the Cauchy-
Schwarz inequality
E[cov(X, g(Y, U )|U )] = E[(X − E[X|U ])(g(Y, U ) − E[g(Y, U )|U ])]
Therefore,
(57)
= E[E[(X − E[X|U ])|Y U ](g(Y, U ) − E[g(Y, U )|U ])]
(58)
= E[(E[X|Y U ] − E[X|U ])(g(Y, U ) − E[g(Y, U )|U ])]
p
≤ E[(E[X|Y U ] − E[X|U ])2 ]E[(g(Y, U ) − E[g(Y, U )|U ])2 ]
p
= E[var(E[X|Y U ]|U )]E[var(g(Y, U )|U )].
(59)
θ(X; Y |U ) = sup p
g
≤
s
E[cov(X, g(Y, U )|U )]
E[var(X|U )]E[var(g(Y, U )|U )]
E[var(E[X|Y U ]|U )]
.
E[var(X|U )]
(60)
(61)
(62)
(63)
It is easy to verify that equality holds if and only if g(Y, U ) = αE[X|Y U ] for some constant α > 0. Hence
s
E[var(E[X|Y U ]|U )]
θ(X; Y |U ) =
.
(64)
E[var(X|U )]
Furthermore, by law of total variance
var(Y ) = Evar(Y |X) + var(E(Y |X))
(65)
E[var(X|U )] = E[var(X|Y U )] + E[var(E[X|Y U ]|U )],
(66)
and the conditional version
we have
θ(X; Y |U ) =
=
s
s
E[var(E[X|Y U ]|U )]
E[var(X|U )]
1−
10
E[var(X|Y U )]
.
E[var(X|U )]
(67)
Furthermore, since ρm (X; Y |U ) = sup θ(f (X, U ); Y |U ), (52) follows straightforwardly from (67).
f
Lemma 8. (Correlation ratio equality). For any random variables X, Y, U,
1 − θ2 (X; Y Z|U ) = (1 − θ2 (X; Z|U ))(1 − θ2 (X; Y |ZU ));
(68)
1 − ρ2m (X; Y Z|U ) ≥ (1 − ρ2m (X; Z|U ))(1 − ρ2m (X; Y |ZU ));
(69)
θ(X; Y Z|U ) ≥ θ(X; Y |ZU );
(70)
ρm (X; Y Z|U ) ≥ ρm (X; Y |ZU ).
(71)
Remark 5. (71) is very similar to I(X; Y Z|U ) ≥ I(X; Y |ZU ). Furthermore, ρm (X; Y |U V ) ≥ ρm (X; Y |U ) or
ρm (X; Y |U V ) ≤ ρm (X; Y |U ) does not always hold. This is also similar to that I(X; Y |U V ) ≥ I(X; Y |U ) or
I(X; Y |U V ) ≤ I(X; Y |U ) does not always hold.
Proof: From (51), we have
1 − θ2 (X; Y Z|U ) =
1 − θ2 (X; Z|U ) =
and
1 − θ2 (X; Y |ZU ) =
E[var(X|Y ZU )]
,
E[var(X|U )]
(72)
E[var(X|ZU )]
,
E[var(X|U )]
(73)
E[var(X|Y ZU )]
.
E[var(X|ZU )]
(74)
Hence (68) follows immediately.
Suppose f achieves ρm (X; Y Z|U ), i.e., the supremum in (13), then
1 − ρ2m (X; Y Z|U ) = 1 − θ2 (f (X, U ); Y Z|U )
(75)
= (1 − θ2 (f (X, U ); Z|U ))(1 − θ2 (f (X, U ); Y |ZU ))
(76)
≥ (1 − ρ2m (X; Z|U ))(1 − ρ2m (X; Y |ZU )).
(77)
Furthermore, θ2 (X; Z|U ) ≥ 0, hence (70) follows immediately from (68).
Suppose f 0 achieves ρm (X; Y |ZU ), then
ρm (X; Y |ZU ) = θ(f 0 (X, U ); Y |ZU ) ≤ θ(f 0 (X, U ); Y Z|U ) ≤ ρm (X; Y Z|U ).
(78)
Lemma 9. For any PU XY V such that U → X → Y and X → Y → V , we have
ρm (U X; V Y ) = max{ρm (X; Y ), ρm (U ; V |XY )}.
(79)
Remark 6. A similar result can be found in [9, Eqn. (4)], where Beigi and Gohari only proved the equality above
as an inequality.
Proof: Beigi and Gobari [9, Eqn. (4)] have proven ρm (U X; V Y ) ≤ max{ρm (X; Y ), ρm (U ; V |XY )}. Hence
we only need to prove that ρm (U X; V Y ) ≥ max{ρm (X; Y ), ρm (U ; V |XY )}. According to the definition, ρm (U X; V Y ) ≥
11
ρm (X; Y ) is straightforward. From (71) of Lemma 8, we have ρm (U X; V Y ) ≥ ρm (U X; V |Y ) ≥ ρm (U ; V |XY ).
This completes the proof.
We also prove that conditioning reduces covariance gap as shown in the following lemma, the proof of which is
given in Appendix C.
Lemma 10. (Conditioning reduces covariance gap). For any random variables X, Y, Z, U,
i. e.,
p
Evar(X|ZU )Evar(Y |ZU ) − Ecov(X, Y |ZU ) ≤
p
p
Evar(X|Z)Evar(Y |Z) − Ecov(X, Y |Z),
(1 − θ2 (X; U |Z))(1 − θ2 (Y ; U |Z))(1 − ρ(X, Y |ZU )) ≤ 1 − ρ(X, Y |Z).
(80)
(81)
In particular, if Z is degenerate, then
i. e.,
p
Evar(X|U )Evar(Y |U ) − Ecov(X, Y |U ) ≤
p
var(X)var(Y ) − cov(X, Y ),
p
(1 − θ2 (X; U ))(1 − θ2 (Y ; U ))(1 − ρ(X, Y |U )) ≤ 1 − ρ(X, Y ).
(82)
(83)
Remark 7. The following two inequalities follows immediately.
and
p
p
(1 − ρ2m (X; U |Z))(1 − θ2 (Y ; U |Z))(1 − θ(X, Y |ZU )) ≤ 1 − θ(X, Y |Z),
(84)
(1 − ρ2m (X; U |Z))(1 − ρ2m (Y ; U |Z))(1 − ρm (X, Y |ZU )) ≤ 1 − ρm (X, Y |Z).
(85)
Furthermore, there are also some other remarkable properties.
Lemma 11. (Tensorization). Assume given U, (X n , Y n ) is a sequence of pairs of conditionally independent random
variables, then we have
ρm (X n ; Y n |U ) = sup ρm (Xi ; Yi |U ).
(86)
1≤i≤n
Proof: The unconditional version
ρm (X n ; Y n ) = sup ρm (Xi ; Yi ),
(87)
1≤i≤n
for a sequence of pairs of independent random variables (X n , Y n ) is proven in [2, Thm. 1]. Using this result and
Lemma 1, we have
ρm (X n ; Y n |U ) =
=
sup
u:P (u)>0
sup
ρm (X n ; Y n |U = u)
(88)
sup ρm (Xi ; Yi |U = u)
(89)
ρm (Xi ; Yi |U = u)
(90)
u:P (u)>0 1≤i≤n
= sup
sup
1≤i≤n u:P (u)>0
= sup ρm (Xi ; Yi |U ).
1≤i≤n
12
(91)
Lemma 12. (Gaussian case). For jointly Gaussian random variables X, Y, U , we have
ρm (X; Y ) = θ(X; Y ) = θ(Y ; X) = |ρ(X; Y )|,
(92)
ρm (X; Y |U ) = θ(X; Y |U ) = θ(Y ; X|U ) = |ρ(X; Y |U )|.
(93)
Proof: The unconditional version (92) is proven in [13, Sec. IV, Lem. 10.2]. On the other hand, given U =
u, (X, Y ) also follows jointly Gaussian distribution, and ρ(X; Y |U = u) = ρ(X; Y |U ) for different u. Hence
ρm (X; Y |U ) =
sup
u:P (u)>0
ρm (X; Y |U = u) =
sup
u:P (u)>0
|ρ(X; Y |U = u)| = |ρ(X; Y |U )|.
Furthermore, both θ(X; Y |U ) and θ(Y ; X|U ) are between ρm (X; Y |U ) and |ρ(X; Y |U )|. Hence (93) holds.
Lemma 13. (Data processing inequality). If random variables X, Y, Z, U form a Markov chain X → (Z, U ) → Y,
then
|ρ(X; Y |U )| ≤ θ(X; Z|U )θ(Y ; Z|U ),
(94)
θ(X; Y |U ) ≤ θ(X; Z|U )ρm (Y ; Z|U ),
(95)
ρm (X; Y |U ) ≤ ρm (X; Z|U )ρm (Y ; Z|U ).
(96)
Moreover, the equalities hold in (94)-(96), if (X, Z, U ) and (Y, Z, U ) have the same joint distribution. In particular
if U is degenerate, then
|ρ(X; Y )| ≤ θ(X; Z)θ(Y ; Z),
(97)
θ(X; Y ) ≤ θ(X; Z)ρm (Y ; Z),
(98)
ρm (X; Y ) ≤ ρm (X; Z)ρm (Y ; Z).
(99)
Proof: Consider that
E[cov(X, Y |U )] = E[(X − E[X|U ])(Y − E[Y |U ])]
(100)
= E[E[(X − E[X|U ])(Y − E[Y |U ])|ZU ]]
(101)
= E[E[X − E[X|U ]|ZU ]E[Y − E[Y |U ]|ZU ]]
(102)
= E[(E[X|ZU ] − E[X|U ])(E[Y |ZU ] − E[Y |U ])]
p
≤ E[(E[X|ZU ] − E[X|U ])2 ]E[(E[Y |ZU ] − E[Y |U ])2 ]
p
= E[var(E[X|ZU ]|U )]E[var(E[Y |ZU ]|U )]
(103)
(104)
(105)
where (102) follows by conditional independence, and (104) follows the Cauchy-Schwarz inequality. Hence
E[cov(X, Y |U )]
p
|ρ(X; Y |U )| = p
E[var(X|U )] E[var(Y |U )]
s
E[var(E[X|ZU ]|U )]E[var(E[Y |ZU ]|U )]
≤
E[var(X|U )]E[var(Y |U )]
13
(106)
(107)
(108)
= θ(X; Z|U )θ(Y ; Z|U ).
It is easy to verify the equalities hold if (X, Z, U ) and (Y, Z, U ) have the same joint distribution.
Similarly, (95) and (96) can be proven as well.
Furthermore, correlation ratio and maximal correlation are also related to rate-distortion theory.
Lemma 14. (Relationship to rate-distortion function) Let RX|U (D) denote the conditional rate distribution function
2
for source X given U with quadratic distortion measure d (x, x̂) = (x − x̂) . Then from rate-distortion theory, we
have
I(X; Y |U ) ≥ RX|U (E[var(X|Y U )])
(109)
= RX|U (E[var(X|U )](1 − θ2 (X; Y |U )))
(110)
≥ RX|U (E[var(X|U )](1 − ρ2 (X; Y |U ))).
(111)
From Shannon lower bound,
I(X; Y |U ) ≥ RX|U (E[var(X|Y U )])
≥ h(X|U ) −
1
log(2πeE[var(X|U )](1 − θ2 (X; Y |U ))).
2
(112)
(113)
If (X, U ) is jointly Gaussian, then
1
1
log+ (
)
2
2
1 − θ (X; Y |U )
1
1
≥ log+ (
).
2
2
1 − ρ (X; Y |U )
I(X; Y |U )≥
In particular if U is degenerate, then
I(X; Y ) ≥ RX (E[var(X|Y )])
= RX (var(X)(1 − θ2 (X; Y ))).
(114)
(115)
(116)
(117)
From Shannon lower bound,
I(X; Y ) ≥ RX (E[var(X|Y )])
≥ h(X) −
1
log(2πe(1 − θ2 (X; Y ))).
2
(118)
(119)
If X is Gaussian, then
1
1
log+ (
)
2
1 − θ2 (X; Y )
1
1
≥ log+ (
).
2
2
1 − ρ (X; Y )
I(X; Y ) ≥
(120)
(121)
From the properties above, it can be observed that maximal correlation or correlation ratio has many similar
properties as those of mutual information, such as invariance to one-to-one transform, chain rule (correlation ratio
equality), data processing inequality, etc. On the other hand, maximal correlation or correlation ratio also has
some different properties, such as for a sequence of pairs of independent random variables, the mutual information
14
between them is the sum of mutual information of all pairs of components (i.e., additivity); while the maximal
correlation is the maximum one of the maximal correlations of all pairs of components (i.e., tensorization).
C. Extension: Smooth Maximal Correlation
Next we extend maximal correlation to smooth version. Analogous extensions can be found in [15] and [16],
where Rényi divergence and generalized Brascamp-Lieb-like (GBLL) rate are extended to the corresponding smooth
versions.
Definition 4. For any random variables X and Y with alphabets X ⊆ R and Y ⊆ R, and ∈ (0, 1), the -smooth
(Pearson) correlation and the -smooth conditional (Pearson) correlation of X and Y given another random variable
U are respectively defined by
and
ρe (X; Y ) :=
ρe (X; Y |U ) :=
inf
QXY :kQXY −PXY kT V ≤
ρQ (X; Y ),
inf
QXY U :kQXY U −PXY U kT V ≤
ρQ (X; Y |U ).
(122)
(123)
Definition 5. For any random variables X and Y with alphabets X ⊆ R and Y, and ∈ (0, 1), the -smooth
correlation ratio and the -smooth conditional correlation ratio of X and Y given another random variable U are
respectively defined by
and
θe (X; Y ) :=
θe (X; Y |U ) :=
inf
QXY :kQXY −PXY kT V ≤
θQ (X; Y ),
inf
QXY U :kQXY U −PXY U kT V ≤
θQ (X; Y |U ).
(124)
(125)
Definition 6. For any random variables X and Y with alphabets X and Y, and ∈ (0, 1), the -smooth maximal
correlation and the -smooth conditional maximal correlation of X and Y given another random variable U are
respectively defined by
and
ρem (X; Y ) :=
ρem (X; Y |U ) :=
inf
QXY :kQXY −PXY kT V ≤
ρm,Q (X; Y ),
inf
QXY U :kQXY U −PXY U kT V ≤
ρm,Q (X; Y |U ).
(126)
(127)
According to definition, obviously we have
ρe (X; Y |U ) ≤ ρ(X; Y |U ),
(128)
ρem (X; Y |U ) ≤ ρm (X; Y |U ).
(130)
θe (X; Y |U ) ≤ θ(X; Y |U ),
(129)
Furthermore, note that adding inf QXY U :kQXY U −PXY U kT V ≤ operation before both sides of an equality or inequality
about PXY U does not change the equality or inequality. Hence some of above lemmas still hold for -smooth
version, e.g., Lemmas 1, 2, 6, and 7, and also (70) and (71) of Lemma 8.
15
III. G ENERALIZED C OMMON I NFORMATION : I NFORMATION -C ORRELATION F UNCTION
In this section, we generalize the existing common informations, and define β-approximate common information
(or approximate information-correlation function) and β-exact common information (or exact information-correlation
function), which measure how much information are approximately or exactly β-correlated between two variables.
Different from the existing common informations, β-common information is a function of conditional maximal
correlation β ∈ [0, 1], and hence it provides a soft-measure of common information.
As in the previous section, in this section we also assume all alphabets are general unless otherwise stated.
A. Definition
Suppose U is a common random variable extracted from X, Y , satisfying privacy constraint ρm (X; Y |U ) ≤ β,
then the β-private information corresponding to U should be H(XY |U ). We define the β-private information as
the maximum of such private informations over all possible U .
Definition 7. For sources X, Y , and β ∈ [0, 1], the β-approximate private information of X and Y is defined by
Bβ (X; Y ) =
sup
PU |XY :ρm (X;Y |U )≤β
H(XY |U ).
(131)
Common information is defined as Cβ (X; Y ) = H(XY ) − Bβ (X; Y ), which is equivalent to the following
definition.
Definition 8. For sources X, Y , and β ∈ [0, 1], the β-approximate common information (or approximate informationcorrelation function) of X and Y is defined by
Cβ (X; Y ) =
inf
PU |XY :ρm (X;Y |U )≤β
I(XY ; U ).
(132)
Similarly, exact common information can be generalized to β-exact common information as well.
Definition 9. For sources X, Y , and β ∈ [0, 1], the β-exact common information (rate) (or exact informationcorrelation function) of X and Y is defined by
Kβ (X; Y ) = lim
inf
n→∞ PUn |X n Y n :ρm (X n ;Y n |Un )≤β
1
H(Un ).
n
(133)
Furthermore, for β ∈ (0, 1], we also define
Cβ − (X; Y ) = lim Cα (X; Y ),
(134)
Kβ − (X; Y ) = lim Kα (X; Y ).
(135)
α↑β
α↑β
B. Properties
These two generalized common informations have the following properties.
Lemma 15. (a) For the infimum in (132), it suffices to consider the variable U with alphabet |U| ≤ |X ||Y| + 1.
16
(b) For any random variables X, Y, Cβ (X; Y ) and Kβ (X; Y ) are decreasing in β. Moreover,
Cβ (X; Y ) ≤ Kβ (X; Y ), for 0 ≤ β ≤ 1,
(136)
Cβ (X; Y ) = Kβ (X; Y ) = 0, for ρm (X; Y ) ≤ β ≤ 1,
(137)
C0 (X; Y ) = CW (X; Y ),
(138)
K0 (X; Y ) = KKLG (X; Y ),
(139)
C1− (X; Y ) = K1− (X; Y ) = CGK (X; Y ),
(140)
where KKLG (X; Y ) := limn→∞ inf PUn |X n Y n :X n →Un →Y n
1
n H(Un )
denotes the exact common information (rate)
proposed by Kumar, Li, and Gamal [12].
(c) If PU |X,Y achieves the infimum in (132), then ρm (X; Y |U ) ≤ ρm (X; Y |V ) for any V such that XY →
U →V.
Remark 8. For any random variables X, Y, Cβ (X; Y ) is decreasing in β, but it is not necessarily convex or concave;
see the Gaussian source case in the next subsection. Cβ (X; Y ) and Kβ (X; Y ) are discontinuous at β = 1, if there
is common information between the sources. Lemma 15 implies Gács-Körner common information, Wyner common
information and exact common information are extreme cases of β-approximate common information or β-exact
common information.
Proof: To show (a), we only need to show for any variable U , there always exists another variable U 0 such
that |U 0 | ≤ |X ||Y| + 1, ρm (X; Y |U 0 ) = ρm (X; Y |U ), and I(XY ; U 0 ) = I(XY ; U ). Suppose ρm (X; Y |U = u∗ ) =
ρm (X; Y |U ). According to Support Lemma [7], there exists a random variable U 0 with U 0 ⊆ U and |U 0 | ≤ |X ||Y|+1
such that
PU 0 (u∗ ) = PU (u∗ ),
(141)
H(XY |U 0 ) = H(XY |U ),
X
PXY =
PU 0 PXY |U 0 .
(142)
(143)
u0
(141) implies ρm (X; Y |U 0 ) = ρm (X; Y |U ). (143) implies H(XY ) is also preserved, and hence I(XY ; U ) =
I(XY ; U 0 ). This completes the proof of (a).
(b) (136) and (137) follow straightforwardly from the definitions. According to the definitions and Lemma 6
(ρm (X; Y |U ) = 0 if and only if X → U → Y ), we can easily obtain (138) and (139). Next we prove (140).
Consider
C1− (X; Y ) =
inf
PU |X,Y :ρm (X;Y |U )<1
I(XY ; U ).
(144)
Assume Gács-Körner common information is fGK (X, Y ). Set U = fGK (X, Y ), then we have
ρm (X; Y |U ) < 1,
(145)
I(XY ; U ) = H(fGK (X, Y )) = CGK (X; Y ).
(146)
17
3
H XY
H X |Y
H Y | X
I X ;Y
CW I XY ;W
H X |W
H Y | W
C I XY ;U
1
1
CGK =H V
2
2
Fig. 1. Illustration of the relationship among joint entropy, mutual information, Wyner common information, generalized common information,
Fig. 1.
Illustration of the relationship among joint entropy, mutual information, Wyner common information, generalized common information,
and Gács-Körner common information, where W, V and U are the Wyner, Gács-Körner, and β-common random variables, respectively,
and GK common information, where W, V and U are the Wyner, GK, and ρ-correlated common randomnesses, respectively, and Region 1
and Region 1 represents H(XY |U ) and Region 2 represents H(XY |V ). These terms satisfy CGK ≤ I(X; Y ) ≤ CW and CGK =
represents H (XY |U ) and Region 2 represents H (XY |V ). The inequalities CGK ≤ I (X; Y ) ≤ CW and CGK = C0 ≤ Cρ ≤ limρ↑1 Cρ =
C0 ≤ Cβ ≤ C1− (X; Y ) = CW , ∀0 ≤ β < 1.
CW , ∀0 ≤ ρ < 1 hold.
Hence by definition,
The total variation distance between two probability
C1− (X; Y ) measures
≤ CGK (X;PY and
). Q with common alphabet is defined
(147)by
On the other hand, for any U suchkP
that−ρQk
|Usup
) < 1,
|Pthe
(A)Gács-Körner
− Q(A)|, common information is determined (1)
m (X;
T V Y,
A∈F
by U , i.e., fGK (X, Y ) = g(U ) for some function g. Therefore, we have
where F is the σ-algebra of the probability space. The following properties of total variation distance hold.
I(XY ; U ) = I(XY ; U, fGK (X, Y )) ≥ H(fGK (X, Y )) = CGK (X; Y ).
(148)
Property 1. [13] Total variation distance satisfies:
Hence
1) If the support of P and Q is a countable set X , then
C1− (X; Y ) ≥ CGK (X; Y ).
1X
|P ({x}) − Q({x})|.
=
2
kP − QkT V
Combining (147) and (149) gives us
(149)
(2)
x∈X
C1− (X;
Y ) = range
CGK (X;
).
2) Let > 0 and let f (x) be a function with
bounded
of Ywidth
b > 0. Then
(150)
Similarly K1− (X; Y ) = CGK (X;
Y )Qk
can be<proven
(3)
EP f (X) − EQ f (X) < b,
kP −
=⇒as well.
TV
(c) Suppose PU |X,Y achieves the infimum in (132). If V satisfies both XY → U → V and XY → V → U , the
indicates
the Yexpectation
taken
with respect to the distribution P .
P (X;
wewhere
have E
ρm
Y |U ) =that
ρm (X;
|U V ) = ρmis
(X;
Y |V ).
3) Let
and →
QX
be two
distributions
common
channel
X PY |X XY
Y |X
|X . Then
If VPsatisfies
UP→
V but
doesjoint
not satisfy
XY → Vwith
→U
, then I(XY
; U) P
=YI(XY
; U V ) > I(XY ; V ).
Hence ρm (X; Y |U ) ≤ ρm (X; Y |V ), otherwise it contradicts with that PU |X,Y achieves the infimum in (132).
kPX PY |X − QX PY |X kT V = kPX − QX kT V .
Fig. 1 illustrates the relationship among joint entropy, mutual information, Gács-Körner common information,
(4)
Wyner common information, and generalized
information.
II. Mcommon
AXIMAL C ORRELATION
16. (Additivity
subadditivity).
Assume
(Xi , Yicorrelation
)ni=1 is a sequence
of pairs
of independent
random
InLemma
this section,
we defineand
several
correlations,
including
coefficient,
correlation
ratio, and
maximal
variables,
then
we study
have their properties.
correlation,
and
then
Cβ (X n ; Y n ) =
n
X
Cβ (Xi ; Yi ),
(151)
i=1
Definition 1. For any random variables X and Y with alphabets
X ⊆ R and Y ⊆ R, the (Pearson) correlation of
and
X and Y is defined by
ρ(X; Y ) =
September 19, 2016
√
Kβ (X
Kβ (X) n ; Y n ) ≤
i ; Yi ) ≤
cov(X,Y
0,
√
var(X)
var(Y )
18
n
X
Kβ (Xi ; Yi ).
, if var
(X) var (Y ) > 0,
i=1
(152)
(5)
if var (X) var (Y ) = 0.
DRAFT
Proof: For (151) it suffices to prove the n = 2 case, i.e.,
Cβ (X 2 ; Y 2 ) = Cβ (X1 ; Y1 ) + Cβ (X2 ; Y2 ).
(153)
ρm (X 2 ; Y 2 |U ) ≥ ρm (Xi ; Yi |U ), i = 1, 2,
(154)
I(X 2 Y 2 ; U ) ≥ I(X1 Y1 ; U ) + I(X2 Y2 ; U |X1 Y1 )
(155)
= I(X1 Y1 ; U ) + I(X2 Y2 ; U X1 Y1 )
(156)
≥ I(X1 Y1 ; U ) + I(X2 Y2 ; U ).
(157)
Observe for any PU |X 2 Y 2 ,
and
Hence we have
Cβ (X 2 ; Y 2 ) ≥ Cβ (X1 ; Y1 ) + Cβ (X2 ; Y2 ).
(158)
Moreover, if we choose PU |X 2 Y 2 = PU∗1 |X1 Y1 PU∗2 |X2 Y2 in Cβ (X 2 ; Y 2 ), where PU∗i |Xi Yi , i = 1, 2, is the distribu-
tion achieving Cβ (Xi ; Yi ), then we have
ρm (X 2 ; Y 2 |U ) = max ρm (Xi ; Yi |Ui ) ≤ β,
(159)
I(X 2 Y 2 ; U ) = I(X1 Y1 ; U1 ) + I(X2 Y2 ; U2 ) = Cβ (X1 ; Y1 ) + Cβ (X2 ; Y2 ).
(160)
i∈{1,2}
and
Therefore,
Cβ (X 2 ; Y 2 ) =
inf
PU |X 2 Y 2 :ρm (X 2 ;Y 2 |U )≤β
I(X 2 Y 2 ; U ) ≤ Cβ (X1 ; Y1 ) + Cβ (X2 ; Y2 ).
(161)
(158) and (161) implies (151) holds for n = 2.
Furthermore, the first inequality of (152) can be obtained directly from the definition of Kβ . The second inequality
of (152) can be obtained by restricting PU |X n Y n to the one with independent components (similar as the proof of
(161)).
For continuous sources, a lower bound on approximate common information is given in the following theorem.
Theorem 1. (Lower bound on Cβ (X; Y )). For any continuous sources (X, Y ) with correlation coefficient β0 , we
have
Cβ (X; Y )≥ h(XY ) −
for 0 ≤ β ≤ β0 , and Cβ (X; Y ) = 0 for β0 ≤ β ≤ 1.
1+β
1
log (2πe(1 − β0 ))2
2
1−β
(162)
Proof:
(163)
I(XY ; U ) = h(XY ) − h(XY |U )
1
log (2πe)2 det(ΣXY |U )
2
1
≥ h(XY ) − log (2πe)2 det(EU ΣXY |U )
2
≥ h(XY ) − EU
19
(164)
(165)
2
Cβ
1.5
1
0.5
0
0
0.2
0.4
0.6
0.8
1
β
Fig. 2. Information-correlation function for Gaussian sources in Theorem 2 with β0 = 0.9.
1
log (2πe)2 Evar(X|U )Evar(Y |U ) − (Ecov(X, Y |U ))2
2
1
= h(XY ) − log (2πe)2 Evar(X|U )Evar(Y |U )(1 − ρ2 (X; Y |U ))
2
1 − β0
1
2
2
2
) (1 − ρ (X; Y |U ))
≥ h(XY ) − log (2πe) (
2
1 − ρ(X, Y |U )
1
1 + ρ(X; Y |U )
= h(XY ) − log (2πe(1 − β0 ))2
2
1 − ρ(X; Y |U )
1
1+β
,
≥ h(XY ) − log (2πe(1 − β0 ))2
2
1−β
= h(XY ) −
(166)
(167)
(168)
(169)
(170)
where (165) follows from the function log(det(·)) is concave on the set of symmetric positive definite square
matrices [8, p.73], (168) follows from Lemma 10, and (170) follows from
ρ(X; Y |U ) ≤ β.
(171)
The equality holds in Theorem 1 if X, Y are jointly Gaussian. The proof is given in Appendix D.
Theorem 2. (Gaussian sources). For jointly Gaussian sources X, Y with correlation coefficient β0 ,
1
(G)
+ 1 + β0 1 + β
Cβ (X; Y ) = log
/
.
(172)
2
1 − β0 1 − β
1
(G)
(G)
+ 1 + β0
Remark 9. Specialized to the Wyner common information, CW (X; Y ) = C0 (X; Y ) = log
, which
2
1 − β0
was first given in [21].
For the doubly symmetric binary source, an upper bound on common information is given in the following
theorem.
20
Theorem 3. (Doubly symmetric
binary source (DSBS)).For doubly symmetric binary source (X, Y ) with crossover
probability p0 , i.e., PXY =
(B)
Cβ (X; Y
1
2
(1 − p0 )
1
2 p0
) ≤ 1 + H2 (p0 ) − H4
1
2
1
2 p0
1
2
(1 − p0 )
s
1 − p0 +
, we have
1 − 2p0 − β
1−β
!
1
,
2
1 − p0 −
s
1 − 2p0 − β
1−β
!
!
p0 p0
, ,
2 2
(173)
(B)
for 0 ≤ β < 1 − 2p0 , and Cβ (X; Y ) = 0 for β ≥ 1 − 2p0 , where H2 and H4 denote the binary and quaternary
entropy functions, respectively, i.e.,
H2 (p) = −p log p − (1 − p) log(1 − p),
(174)
H4 (a, b, c, d) = −a log a − b log b − c log c − d log d.
(175)
Proof: Assume p is a value such that 2pp̄ = p0 , p̄ := 1 − p. Then (X, Y ) can be expressed as
X = U ⊕ V ⊕ Z1 ,
(176)
Y = U ⊕ V ⊕ Z2 ,
(177)
where U ∼ Bern( 21 ), V
∼ Bern(α)
with 0 ≤ α ≤ 1, Z1 ∼ Bern(p), and Z2 ∼ Bern(p) are independent. Hence we
a pp̄
with a = αp2 + ᾱp̄2 , b = αp̄2 + ᾱp2 , and
have PV ⊕Z1 ,V ⊕Z2 =
pp̄ b
ρm (X, Y |U ) = ρm (V ⊕ Z1 , V ⊕ Z2 ).
By using the formula
ρ2m (X, Y
#
X P 2 (x, y)
−1
)≤
P (x)P (y)
x,y
"
(178)
(179)
for (X, Y ) with at least one of them being binary-valued, we have
(180)
ρm (X, Y ) = 1 − 2p0 .
(B)
Hence Cβ (X; Y ) = 0 for β ≥ 1 − 2p0 . Next we consider the case
(181)
β ≤ 1 − 2p0 .
To guarantee ρm (V ⊕ Z1 , V ⊕ Z2 ) ≤ β, we choose
1
a=
2
1
b=
2
1 − p0 +
1 − p0 −
s
s
1 − 2p0 − β
1−β
1 − 2p0 − β
1−β
This leads to the inequality (335). This completes the proof.
21
!
!
(182)
.
(183)
C. Relationship to Rate-Distortion Function
The approximate information-correlation function can be rewritten as
Cβ (X; Y ) =
inf
PU |X,Y :d(PU XY )≤β
(184)
I(XY ; U ).
where d(PU XY ) := ρm (X; Y |U ). This expression has a form similar to rate-distortion function, if we consider
maximal correlation as a special “distortion measure”. But it is worth nothing that maximal correlation is taken on
the distribution of X, Y , instead of on them itself.
Information-correlation function is also related to the rate-privacy function [20]
gβ (X; Y ) :=
sup
(185)
I(Y ; U ),
PU |Y :ρm (X;U )≤β
in which U can be thought of as the extracted information from Y under privacy constraint ρm (X; U ) ≤ β. But
there are three differences between gβ (X; Y ) and Cβ (X; Y ). 1) The privacy constraint in gβ (X; Y ) is a constraint
on unconditional maximal correlation, and moreover, this unconditional maximal correlation is that between the
remote source X and extracted information U , instead of between the sources. Hence gβ (X; Y ) is not symmetric
respect to X, Y . 2) In gβ (X; Y ), U is extracted from Y instead of both X, Y , hence X → Y → U is restricted in
gβ (X; Y ). 3) The optimization in Cβ (X; Y ) is infimum, while in gβ (X; Y ) is supremum.
IV. P RIVATE S OURCES S YNTHESIS
In order to provide an operational interpretation for information-correlation functions Cβ (X; Y ) and Kβ (X; Y ),
in this section, we consider private sources synthesis problem. We show that the information-correlation functions
correspond to the minimum achievable rates for the centralized setting version of this problem.
A. Problem Setup
Consider private sources synthesis problem shown in Fig. 3, where a simulator generates two source sequences
X and Y n from a common random variable M . X n and Y n are restricted to follow i.i.d. according to a target
Y
distribution
PXY .
n
Definition 10. A generator is defined by a pmf PM and a stochastic mapping PX n Y n |M : M 7→ X n × Y n .
Furthermore, Shannon’s zero-error source coding theorem states that, it is possible to compress a message M
(using a variable length coding) at rate R for sufficiently large n if R >
only if R ≥
1
n H(M ).
1
n H(M );
and conversely, it is possible
Hence we define the achievability of tuple (R, β) as follows.
Definition 11. The tuple (R, β) is approximately or exactly achievable if there exists a sequence of generators such
that
1) rate constraint:
lim sup
n→∞
1
H(M ) ≤ R;
n
(186)
2) privacy constraint:
ρm (X n ; Y n |M ) ≤ β, ∀n;
22
(187)
19
nR
M [2M
]
∈
19
M [2nRGenerator
] GeneratorX n X n
M
1 1
∈
XnXn
Generator
Generator
M [2nRGenerator
] Generator Y n Y n
M
2 2
∈
Yn Yn
n(X
Fig. Fig.
source
synthesis
problem:
(left)
setting;
(right)
distributed
setting.
this
problem
we constraint
assume
Fig.
3.3. 3. Private
Private
Private
source
source
synthesis
synthesis
problem:
problem:
1) privacy
1)centralized
privacy
constraint
constraint
ρm (X
ρm
; Y n ;|M
Y n) |M≤) ≤
ρ;
2)
ρ; source
2)In source
distribution
distribution
constraint
1
n
n
1) n→∞
rate
constraint
) in
≤ weak
R;
2)sense,
ρnm
Y sense.
|Msense.
) ≤ β; 3) source distribution constraint
n YX
n n−
n
nQ
n
n
n k=H(M
nconstraint
n=
n
n
lim
limn→∞
kP
QX
−sup
kT
0=
in 0weak
sense,
orprivacy
PX
ornP
Q=
in
strong
in; strong
n→∞
XkP
Ylim
YX
YV
YX
Yn
X nQ
YX
Y(X
nT V
lim kPX n Y n − QX n Y n kT V = 0 in approximate synthesis sense, or PX n Y n = QX n Y n in exact synthesis sense. For distributed setting,
n→∞
the M in the constraints is replaced with M1 M2 .
2) approximate
2) approximate
sources
sources
distribution
distribution
constraint:
constraint:
YY
3) approximate sources distribution constraint:
lim limPX nPYXnn−
Y n − PXYPXY = 0,= 0,
n→∞
n→∞
TV TV
(150)
(150)
or exact
or exact
sources
sources
distribution
distribution
constraint:
constraint:
Y
YY
nn
n=
n = PXY
PXXnP
P.XY
YY
X
Y−
n
lim kP
PXY
kT V. = 0,
n→∞
(151)
(151)
(188)
Definition
Definition
35. 35.
The
The
rate-correlation
rate-correlation
function
function
for for
approximate
approximate
private
private
source
source
synthesis
synthesis
is defined
is defined
by R
bySSR(ρ)
, ,
or exact
sources
distribution
constraint:
SS (ρ)
Y
inf {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable}.
is achievable}.
Similarly,
Similarly,
the the
rate-correlation
function
for exact
private
private
source
source
synthesis
synthesis
is defined
is defined
nY n =
PXrate-correlation
P
, ∀n.for exact
(189)
XYfunction
(E) (E)
, inf
, {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable}.
is achievable}.
by R
bySSR(ρ)
SS (ρ)
Definition 12. The rate-correlation function for approximate private sources synthesis is defined by RP SS (β) :=
Besides,
Besides,
we also
we also
consider
consider
distributed
distributed
setting
setting
as shown
as shown
in Fig.
in Fig.
3(b).3(b).
For For
this this
case,case,
the the
source
source
synthesis
synthesis
problem
problem
is is
inf {R : (R, β) is approximately achievable}. Similarly, the rate-correlation function for exact private sources synnamed
named
distributed
distributed
private
private
source
source
synthesis.
synthesis.
(E)
thesis is defined by RP SS (β) := inf {R : (R, β) is exactly achievable}.
nR nR n n
Definition
Definition
36. 36.
An An
(n, R)
(n, distributed
R) distributed
generator
generator
is defined
is defined
by two
by two
stochastic
stochastic
mappings:
mappings:
PX nP|M
7→ X
and and
X n:|M2 : 2 7→ X
nR nRwe
also
Furthermore,
consider
distributed
setting,
which
is
shown
in
Fig.
3
(b).
For
this
case,
the
source
synthesis
n
PY nP|M
7→
.Y n .
Y n:|M2 : 2 7→ Y
problem is named distributed private sources synthesis.
Definition
Definition
37. 37.
TheThe
tupletuple
(R, (R,
ρ) isρ) approximately
is approximately
or exactly
or exactly
achievable
achievable
for for
distributed
distributed
setting
setting
if there
if there
exists
exists
a a
Definition 13. A distributed generator is defined by a pmf PM and two stochastic mappings: PX n |M : M 7→ X n
sequence
sequence
of (n,
of R)
(n, distributed
R) distributed
generators
generators
suchsuch
that that
and PY n |M : M 7→ Y n .
1) privacy
1) privacy
constraint:
constraint:
(149);
(149);
2)
approximate
2) approximate
distribution
distribution
constraint:
(150),
(150),
ororexact
or
exact
source
source
distribution
distribution
constraint:
(151).
(151).if there exists a
Definition
14. source
Thesource
tuple
(R, β) constraint:
is approximately
exactly
achievable
for constraint:
distributed
setting
sequence
of
generators function
such
thatfor distributed
Definition
Definition
38.distributed
38.
TheThe
rate-correlation
rate-correlation
function
for distributed
approximate
approximate
private
private
or exact
or exact
source
source
synthesis
synthesis
is defined
is defined
(E) (E)
1)
rate
constraint:
(186);
by R
by R (ρ) (ρ)
, inf
, {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable}
is achievable}
and and
R R (ρ) (ρ)
, inf
, {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable},
is achievable},
respectively.
respectively.
DSSDSS
DSSDSS
2) privacy constraint: (187);
ThenThen
for distributed
for distributed
setting,
setting,
privacy
privacy
constraint
constraint
3) approximate source distribution constraint: (188), or exact source distribution constraint: (189).
ρm (X
ρmn(X
; Ynn;|M
Y n)|M
= )0= 0
(152)
(152)
Definition 15. The rate-correlation function for distributed approximate or exact private sources synthesis is defined
(E)
isbysatisfied
isDP
satisfied
immediately.
immediately.
Therefore,
Therefore,
R
SS (β) := inf {R : (R, β) is approximately achievable} and RDP SS (β) := inf {R : (R, β) is exactly achievable},
respectively.
For distributed setting, privacy constraint
September
September
19, 2016
19, 2016
ρm (X n ; Y n |M ) = 0
23
DRAFT
DRAFT
(190)
is satisfied immediately. Therefore,
RDP SS (β) = RDP SS (0),
(E)
(E)
RDP SS (β) = RDP SS (0).
(191)
(192)
We assume the synthesized sources have finite alphabets.
B. Main Result
1) Centralized Setting: For approximate private sources synthesis, we have the following theorems. The proof
of Theorem 4 is given in Appendix E.
Theorem 4. For approximate private sources synthesis,
RP SS (β) = Cβ (X; Y ).
(193)
Remark 10. From the proof we can see that using fixed-length coding is sufficient to achieve the rate-correlation
function RP SS (β).
Theorem 5. For exact private sources synthesis,
(E)
RP SS (β) = Kβ (X; Y ).
(194)
Proof: Achievability: Suppose R > Kβ (X; Y ). We will show that the rate R is achievable.
Input Process Generator: Generate input source M according to pmf PUn .
Source Generator: Upon m, the generator generate sources (X n , Y n ) according to PX n Y n |Un (xn , y n |m).
For such generator, the induced overall distribution is
PX n Y n M (xn , y n , m) := PX n Y n Un (xn , y n , m).
(195)
ρm (X n ; Y n |M ) ≤ β,
(196)
ρm (X n ; Y n |Un ) ≤ β.
(197)
This means
since
1
1
H(Un ) for some Un , R≥ (H(Un ) + 1) for n large enough. By the achievability
n
n
part of Shannon’s zero-error source coding theorem, it is possible to exactly generate (X n , Y n ) at rate at most
1
(E)
(H(Un ) + 1). Hence rate R is achievable and thus RP SS (β) ≤ Kβ (X; Y ).
n
Converse: Now suppose a rate R is achievable. Then there exists an (n, R)-generator that exactly generates
Since Kβ (X; Y ) = lim
n→∞
(X n , Y n ) such that
ρm (X n ; Y n |M ) ≤ β.
(198)
By the converse for Shannon’s zero-error source coding theorem,
lim
n→∞
1
H(M ) ≤ R.
n
24
(199)
Therefore,
R≥ lim
n→∞
1
1
H(M ) ≥ lim
inf n n
H(Un ) = Kβ (X; Y ).
n→∞
n
PUn |X n Y n :ρm (X ,Y |Un )≤β n
(200)
That is
(E)
RP SS (β) ≥ Kβ (X; Y ).
(201)
2) Distributed Setting: For distributed private sources synthesis, we have similar results.
Theorem 6. For distributed approximate private sources synthesis,
RDP SS (β) = C0 (X; Y ).
(202)
Remark 11. From the proof we can see that similar to centralized case, using fixed-length coding is also sufficient
to achieve the rate-correlation function RDP SS (β) for distributed case.
Proof: The theorem was essentially same to Wyner’s result [3]. In the following, we prove this theorem by
following similar steps to the proof of the centralized case.
Achievability: Consider the generator used for the centralized case (see Appendix E-A). Similar to the centralized
case, we can prove if R > C0 (X; Y ),
lim EC kPX n Y n − QX n Y n kT V = 0.
n→∞
(203)
Owing to the distributed setting, Markov chain X n → M → Y n holds. By Lemma 6, we have
ρm (X n ; Y n |M ) = 0.
(204)
RDP SS (β) ≤ C0 (X; Y ).
(205)
Hence
Converse: By slightly modified the proof of centralized case and combining with Markov chain X n → M → Y n ,
we can show that
RDP SS (β) ≥ C0 (X; Y ).
(206)
Theorem 7. For distributed exact private sources synthesis,
(E)
RDP SS (β) = K0 (X; Y ).
(207)
Proof: Achievability: Suppose R > K0 (X; Y ). We will show that the rate R is achievable.
Input Process Generator: Generate input source M according to PUn .
Source Generator: Upon m, the generator 1 generates source X n according to PX n |Un (xn |m), and the generator
2 generates source Y n according to PY n |Un (y n |m).
25
is
Similar to the centralized case, since ρm (X n ; Y n |Un ) = 0, i.e., X n → Un → Y n , the induced overall distribution
PX n Y n M (xn , y n , m) := PUn (m) PX n |Un (xn |m)PY n |Un (y n |m) = PX n Y n Un (xn , y n , m).
This means
PX n Y n (xn , y n ) =
n
Y
PXY (xi , yi ).
(208)
(209)
i=1
and
ρm (X n ; Y n |M ) = 0 ≤ β.
(210)
Hence the rate R is achievable, which further implies
(E)
RDP SS (β) ≤ K0 (X; Y ).
(211)
Converse: Suppose a rate R is achievable. Then there exists an (n, R)-generator that exactly generates (X n , Y n )
such that
ρm (X n ; Y n |M ) ≤ β.
(212)
Owing to the distributed setting, Markov chain X n → M → Y n holds naturally. By Lemma 6, we have
ρm (X n ; Y n |M ) = 0.
(213)
Furthermore, by the converse for Shannon’s zero-error source coding theorem,
1
H(M ) ≤ R.
n→∞ n
lim
(214)
Therefore,
R≥ lim
n→∞
1
1
H(M ) ≥ lim
inf n n
H(Un ) = K0 (X; Y ).
n→∞
n
n
n
PUn |X Y :ρm (X ,Y |Un )≤β n
(215)
That is
(E)
RDP SS (β) ≥ K0 (X; Y ).
(216)
V. C OMMON I NFORMATION E XTRACTION
In this section, we study another problem, common information extraction problem, which provides another
operational interpretation for information-correlation functions Cβ (X; Y ) and Kβ (X; Y ). Similar to private sources
synthesis problem, the information-correlation functions are proven to be the minimum achievable rates for the
centralized setting version of this problem as well.
26
23
nR
X n X n Extractor
Extractor MM
1 1 [2 ]
∈
1 1
XnXn
Yn Yn
23
nR
M [2
M ]
∈
Extractor
Extractor
nR
Y n Y n Extractor
2 2 [2 ]
Extractor MM
∈
2 2
Fig. Fig.
4.4.Common
information
extraction
problem:
(left) centralized
setting;
(right)
distributed
setting.
In
thisdistributed
problem
wesetting.
assume
constraint
Fig.
4.
Common
Common
information
information
extraction
extraction
problem:
problem:
(a) (a)
centralized
centralized
setting;
setting;
(b) (b)
distributed
setting.
1) 1)
criterion
1)ratecriterion
1 1
1
nn n n
n
n
n
n
n
n
inf
inf
lim
sup
ρm
(X
; Y(X;|M
Y; Y
) |M
≤|M
ρ;
)≤
ρ;ρm
2)(X
ρm (X
;or
Y strong
;|M
Y ) |M
≤
ρ..
) ≤ ρ..
limQsup
)2)≤
β,
∀n,
privacy
constraint: ρm (X ; Y |M ) ≤
H(M
) ,Y
≤XnnR;
2)
weak
privacy
constraint:
forn→∞
anyn→∞
ρ>
0,
ρem
m (X
QX,Y,M
:kQX:kQ
kT Vlim
→0sup
n
n ,M
n ,YX
nn,M
nT,M
V →0
X,Y,M
,M
,Y −P
X−P
,Y k
n→∞ n
β, ∀n. For distributed setting, the variable M in the constraints is replaced with M1 M2 .
TheThe
extractor
extractor
extracts
extracts
common
common
information
information
to satisfy
to satisfy
the the
privacy
privacy
constraint
constraint
measured
measured
by conditional
by conditional
maximal
maximal
A. Problem Setup
correlation.
correlation.
As a counterpart of private sources synthesis problem, we consider common information extraction problem
Definition
Definition
44. 44.
The
The
tupletuple
(R,
(R,
ρ) isρ)achievable
isextracts
achievable
if there
if there
exists
exists
a sequence
a sequence
of
of
R)
(n,two
extractors
R) extractors
such
that thatX n and Y n .
shown
in Fig.
4,
where
an
extractor
common
random
variable
M(n,
from
sourcesuch
sequences
criterion
criterion
1: weak
1: weak
privacy
privacy
constraint
constraint
X n and Y n are i.i.d. according to PXY .
inf inf
lim sup
lim sup
ρm (X
ρmn(X
; Ynn;|M,
Y n |M,
QX nQ,YXnn,M
) ,M
≤ )ρ,≤ ρ,
,Y n
nn,M
n ,M
n ,YX
nT,M
QX,Y,M
:kQX:kQ
−P
kT V →0
V →0
X,Y,M
n→∞
n→∞ PM |X n Y n : X n × Y n 7→ M.
X
,Y −P
Xby
,Y k
Definition 16. AnQextractor
isn ,Ydefined
ann,M
stochastic
mapping:
(179)
(179)
criterion
criterion
2: strong
2: strong
privacy
privacy
constraint
constraint
The extractor should extract an enough mount of common information to satisfy the privacy constraint measured
ρm (X
ρmn(X
; Ynn;|M
Y n)|M
≤ )ρ.≤ ρ.
(180)
(180)
by conditional maximal correlation.
Definition
Definition
45. 45.
TheThe
rate-correlation
rate-correlation
functions
functions
for for
weakly
weakly
common
common
information
information
extraction
extraction
problem
problem
and and
strongly
strongly
Definition 17. The tuple (R, β) is weakly or strongly achievable if there exists a sequence of extractors such that
(E) (E)
common
common
information
information
extraction
extraction
problem
problem
are defined
are defined
by Rby
RCIE
(ρ) (ρ)
, inf
, {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable}
is achievable}
and and
RCIE
RCIE
(ρ) (ρ)
, ,
CIE
1) rate constraint:
1
inf {R
inf :{R
(R,: (R,
ρ) is
ρ)achievable},
is achievable},
respectively.
respectively.
lim sup H(M ) ≤ R;
(217)
n→∞ n
Besides,
also
we also
consider
consider
distributed
distributed
information
extraction.
2a) Besides,
weakweprivacy
constraint:
for
any common
>common
0, it information
holds
that extraction.
nR nR
n
1
n : X
and
Definition
Definition
46. 46.
An (n,
An R)
(n, distributed
R) distributed
extractor
extractor
is ndefined
by
two
stochastic
mappings:
PM1P|XMn1 |X
:X
7→n →
7 2 21 and
ρeismdefined
(X
; Y nby|Mtwo
)≤
β, stochastic
∀n, mappings:
(218)
n
n
nR
nR
2
2
PM2P|YMn2 |Y
: Yn : Y
7→ 7→
2 2 suchsuch
that that
where ρem (X n ; Y n |M ) denotes -smooth conditional
1 1 maximal correlation; see (127);
H (M
H 1(M
M21)M≤2 )R≤ R
(181)
(181)
n
n
2b) or strong privacy constraint:
for some
for some
R1 , R12,. R2 .
ρm (X n ; Y n |M ) ≤ β, ∀n.
(219)
Definition
Definition
47. 47.
TheThe
tupletuple
(R, (R,
ρ) isρ)achievable
is achievable
for distributed
for distributed
setting
setting
if there
if there
exists
exists
a sequence
a sequence
of (n,
of R)
(n, distributed
R) distributed
Common information corresponds to the smallest information rate that makes the privacy constraint satisfied,
extractors
extractors
suchsuch
that that
hence the common information indeed represents a kind of “core” information.
criterion
criterion
1 1
Now we define the rate-correlation functions as follows.
inf inf
lim sup
lim sup
ρm (X
ρmn(X
; Ynn;|M
Y n1|M
, M12,,M
Q2X, nQ,YXnn,M
≤2 )ρ,≤ ρ, (182)
(182)
,Y 1n,M
,M21),M
n ,YX
nn,M
n ,YX
nn,M
QX,Y,M
QX,Y,M
:kQ :kQ
−PX−P
k,M
kweakly
T V2 →0
T V →0
n→∞
n→∞strongly common information extraction problems are
,Y1n
,M
,M
,Y1n
,M
,M
1 ,M2The
1 ,M2Xrate-correlation
Definition
18.
functions
and
21 ,M2
21for
criterion
criterion
2 R2CIE (β) := inf {R : (R, β) is weakly achievable} and R(E) (β) := inf {R : (R, β) is strongly achievable},
defined
by
CIE
respectively.
ρm (X
ρmn(X
; Ynn;|M
Y n1|M
, M12,)M≤2 )ρ.≤ ρ.
(183)
(183)
Furthermore, we also consider distributed common information extraction.
September
September
19, 2016
19, 2016
DRAFT
DRAFT
27
Definition 19. A distributed extractor is defined by two stochastic mappings PM1 |X n : X n 7→ M1 and PM2 |Y n :
Y n 7→ M2 .
Definition 20. The tuple (R, β) is achievable for distributed setting if there exists a sequence of distributed extractors
such that
1) rate constraint: (217);
2a) weak privacy constraint: for any > 0, it holds that
2b) or strong privacy constraint:
ρem (X n ; Y n |M1 , M2 ) ≤ β, ∀n,
(220)
ρm (X n ; Y n |M1 , M2 ) ≤ β, ∀n.
(221)
Definition 21. The rate-correlation functions for distributed weakly and strongly common information extrac(E)
tion problems are defined by RDCIE (β) := inf {R : (R, β) is weakly achievable} and RDCIE (β) := inf{R :
(R, β) is strongly achievable}, respectively.
We also assume the sources have finite alphabets.
B. Main Result
1) Centralized Setting: For weakly common information extraction, we have the following theorems. The proof
of Theorem 8 is given in Appendix F.
Theorem 8. For weakly common information extraction,
RCIE (β) = Cβ (X; Y ).
(222)
Remark 12. From the proof we can see that using fixed-length coding is sufficient to achieve the rate-correlation
function RCIE (β).
Theorem 9. For strongly common information extraction,
(E)
RCIE (β) = Kβ (X; Y ).
(223)
Proof: Achievability: Suppose R > Kβ (X; Y ). We will show that the rate R is achievable.
Extractor: Upon (xn , y n ), the extractor generates m according to PUn |X n Y n (m|xn , y n ).
For such extractor, the induced overall distribution is
PX n Y n M (xn , y n , m) = PX n Y n Un (xn , y n , m).
(224)
ρm (X n ; Y n |M ) = ρm (X n ; Y n |Un ) ≤ β.
(225)
Hence
28
1
1
H(Un ), R≥ (H(Un ) + 1) for n large enough. By the achievability part of Shannon’s
n
n
1
zero-error source coding theorem, it is possible to exactly generate (X n , Y n ) at rate at most (H(Un ) + 1). Hence
n
(E)
rate R is achievable and thus RCIE (β) ≤ Kβ (X; Y ).
Since Kβ (X; Y ) = lim
n→∞
Converse: Now suppose a rate R is achievable. Then there exists a sequence of extractors that generate M such
that
ρm (X n ; Y n |M ) ≤ β, ∀n.
(226)
By the converse for Shannon’s zero-error source coding theorem,
lim
n→∞
1
H(M ) ≤ R.
n
(227)
Therefore,
R≥ lim
n→∞
1
1
H(M ) ≥ lim
inf
H(Un ) = Kβ (X; Y ).
n→∞ PUn |X n Y n :ρm (X n ,Y n |Un )≤β n
n
(228)
That is
(E)
(229)
RCIE (β) ≥ Kβ (X; Y ).
2) Distributed Setting: For distributed common information extraction, we have similar results. The following
theorems hold for weakly and strongly common information extraction, respectively. The proof of Theorem 10 is
given in Appendix G.
Theorem 10. For distributed weakly common information extraction,
(D,LB)
Cβ
(D)
(D,U B)
(X; Y ) ≤ RDCIE (β) = Cβ (X; Y ) ≤ Cβ
(X; Y ),
(230)
where
(D,U B)
Cβ
(X; Y ) :=
inf
PU |X PV |Y :ρm (X,Y |U V )≤β
(D)
Cβ (X; Y ) := lim
inf
n→∞ PU |X n PV |Y n :ρm (X n ;Y n |U V )≤β
(D,LB)
Cβ
(X; Y ) :=
(231)
I(XY ; U V ),
inf
PT PU V |XY T :U T →X→Y,X→Y →V T,
ρm (U X;V Y |T )≤ρm (X;Y ),
ρm (X,Y |U V T )≤β
1
I(X n Y n ; U V ),
n
I(XY ; U V |T ).
(232)
(233)
Remark 13. From the proof we can see that similar to centralized case, using fixed-length coding is also sufficient
to achieve the rate-correlation function RDCIE (β) for distributed case.
Theorem 11. For distributed strongly common information extraction,
(E)
(D)
(234)
RDCIE (β) = Kβ (X; Y ),
where
(D)
Kβ (X n ; Y n ) := lim
inf
n→∞ PUn |X n PVn |Y n :ρm (X n ,Y n |Un Vn )≤β
(D)
1
H(Un Vn ).
n
Proof: Achievability: Suppose R > Kβ (X; Y ). We will show that the rate R is achievable.
29
(235)
Extractor: Upon (xn , Y n ), the extractor 1 generates m1 according to PUn |X n (m1 |xn ), and extractor 2 generates
m2 according to PVn |Y n (m2 |y n ).
For such extractor, the induced overall distribution is
PX n Y n M1 M2 (xn , Y n , m1 , m2 ) = PX n Y n Un Vn (xn , Y n , m1 , m2 ).
(236)
ρm (X n ; Y n |M1 M2 ) = ρm (X n ; Y n |Un Vn ) ≤ β.
(237)
Hence
1
1
H(Un Vn ), R≥ (H(Un Vn ) + 1) for n large enough. By the achievability part of Shann
n
1
non’s zero-error source coding theorem, it is possible to exactly generate (X n , Y n ) at rate at most (H(Un Vn ) + 1).
n
(E)
(D)
Hence rate R is achievable and thus RDCIE (β) ≤ Kβ (X; Y ).
Since Gβ (X; Y ) = lim
n→∞
Converse: Now suppose a rate R is achievable. Then there exists a sequence of extractors that generate (M1 , M2 )
such that
ρm (X n ; Y n |M1 M2 ) ≤ β, ∀n.
(238)
By the converse for Shannon’s zero-error source coding theorem,
lim
n→∞
1
H(M1 M2 ) ≤ R.
n
(239)
Therefore,
R≥ lim
n→∞
1
1
(D)
H(M1 M2 ) ≥ lim
inf
H(Un Vn ) = Kβ (X; Y ).
n→∞ pUn |X n pVn |Y n :ρm (X n ,Y n |Un Vn )≤β n
n
(240)
That is
(D)
(E)
RDCIE (β) ≥ Kβ (X; Y ).
(241)
VI. C ONCLUDING R EMARKS
In this paper, we unify and generalize Gács-Körner and Wyner common informations, and define a generalized
version of common information, (approximate) information-correlation function, by exploiting maximal correlation
as a commonness or privacy measure. The Gács-Körner common information and Wyner common information are
two special and extreme cases of our generalized definition. Furthermore, similarly exact information-correlation
function has been defined as well, which is a generalization of Gács-Körner common information and KumarLi-Gamal common information. We study the problems of common information extraction and private sources
synthesis, and show that these two information-correlation functions are equal to the optimal rates under given
correlation constraints in the centralized cases of these problems.
Our results have a sequence of applications:
•
Dependency measure: The generalized common informations defined by us provide a fresh look at dependency.
The more common information the sources share, the more dependent they are. To normalize the (approximate)
information-correlation function, we can define
Γβ (X; Y ) =
30
Cβ (X; Y )
,
H(X, Y )
(242)
or
Γβ (X; Y ) = 1 − 2−2Cβ (X;Y ) .
(243)
Furthermore, we define correlation-information function as the inverse function of information-correlation
function, i.e.,
βC (X; Y ) =
inf
PU |XY :I(XY ;U )≤C
ρm (X; Y |U ),
(244)
which represents the source dependency after extracting C-rate common information from X, Y . Obviously
βC (X; Y ) = ρm (X; Y ) when C = 0. Dependency measure can be further applied to feature extraction
and image classification. Furthermore, conditional maximal correlation can be also applied to measure the
dependency of distributed sources, which has been exploited to derive some converse results of distributed
communication; see our another work [14].
•
Game theory and correlation based secrecy: The common information extraction can be equivalently transformed into a zero-sum game problem. Consider two adversarial parties. One is Player A, and another
one is Players B and C. Players A and B share a source X, and Players A and C share another source
Y . Sources X, Y are correlated and memoryless. Players B and C cooperate to maximize the conditional
correlation ρ(f (X n , M ); g(Y n , M )|M ) (or ρQ (f (X n , M ); g(Y n , M )|M ) for some distribution QX n Y n M )
over all functions f, g, where M is a message received from Player A through a rate-limited channel, and
f (X n , M ) and g(Y n , M ) are the outputs of Players B and C respectively. Player A generates M from X n , Y n
and wants to minimize the optimal correlation induced by Players B and C (assume Player A does not know the
distribution Q Players A and B choose). Then our result on common information extraction can directly apply
to this case, and it implies the exact (or approximate) information-correlation function is equal to the minimum
rate needed for Player A to force B and C’s optimal strategy satisfying supf,g ρ(f (X n , M ); g(Y n , M )|M ) ≤ β
(or inf kQX n Y n M −PX n Y n M kT V ≤ supf,g ρQ (f (X n , M ); g(Y n , M )|M ) ≤ β for any > 0).
•
Privacy protection in data collection or data mining: In data collection or data mining, privacy protection of
users’ data is an important problem. To that end, we need first identify which part is common information and
which part is private information. Our result gives a better answer to this question and hence it can be directly
applied to privacy protection in data collection or data mining.
•
Privacy constrained source simulation: As stated in [11], the private sources simulation problem has natural
applications in numerous areas – from game-theoretic coordination in a network to control of a dynamical
system over a distributed network with privacy protection. Our results are expected to be exploited in many
future remote-controlled applications, such as drone-based delivery system, privacy-preserving navigation,
secure network service, etc.
A PPENDIX A
P ROOF OF E QUATION (2)
First we prove inf PU |XY :CGK (X;Y |U )=0 I(XY ; U ) ≤ CGK (X; Y ). Assume f ∗ , g ∗ achieve the supremum in (1),
then we claim that setting U = f ∗ (X) = g ∗ (Y ), it holds that CGK (X; Y |U ) = 0. We use contradiction to
31
prove this claim. Suppose CGK (X; Y |U ) > 0, i.e., there exists a pair of f 0 , g 0 such that f 0 (X, U ) = g 0 (Y, U ) and
H(f 0 (X, U ) |U ) > 0. Since U is a function of X and also a function of Y , we can express f 0 (X, U ) as f 00 (X) and
g 0 (Y, U ) as g 00 (Y ) for some functions f 00 and g 00 . Setting f (X) = (f ∗ (X) , f 00 (X)) and g (Y ) = (g ∗ (Y ) , g 00 (Y )),
we have f (X) = g (Y ) and
H (f (X)) = H (U ) + H(f 0 (X, U ) |U ) > H (U ) .
(245)
This contradicts with the assumption of f ∗ , g ∗ achieving the supremum in (1). Therefore, CGK (X; Y |U ) = 0. This
implies
inf
PU |XY :CGK (X;Y |U )=0
I(XY ; U ) ≤ H (U ) = CGK (X; Y ).
(246)
Next we prove inf PU |XY :CGK (X;Y |U )=0 I(XY ; U ) ≥ CGK (X; Y ). We also assume f ∗ , g ∗ achieve the supremum
in (1). Then we claim that for any U such that CGK (X; Y |U ) = 0, it holds that f ∗ (X) = g ∗ (Y ) = κ (U ) for
some function κ, i.e., U contains the common randomness of X, Y . Next we prove this claim.
Assume f 0 , g 0 achieve the supremum in (3). Then we have CGK (X; Y |U ) = 0 implies H(f 0 (X, U ) |U ) = 0,
which further implies f 0 (X, U ) is a function of U ; see [17, Problem 2.5]. Setting f (X, U ) = (f ∗ (X) , f 0 (X, U ))
and g (Y, U ) = (g ∗ (Y ) , g 0 (Y, U )), we have f (X, U ) = g (Y, U ) and
H(f 0 (X, U ) |U ) ≤ H(f (X, U ) |U ).
(247)
Owing to the optimality of f 0 , g 0 , the equality in (247) should hold. Therefore, H(f ∗ (X) |U, f 0 (X, U )) = 0. This
implies f ∗ (X) is a function of U and f 0 (X, U ). Combining it with that f 0 (X, U ) is a function of U , we have
f ∗ (X) is a function of U . Therefore, f ∗ (X) = g ∗ (Y ) = κ (U ) for some function κ.
Using the claim, we have
inf
PU |XY :CGK (X;Y |U )=0
I(XY ; U ) ≥
=
inf
I(XY ; κ (U ))
(248)
inf
I(XY ; f ∗ (X))
(249)
PU |XY :CGK (X;Y |U )=0
PU |XY :CGK (X;Y |U )=0
= H (f ∗ (X))
(250)
= CGK (X; Y ).
(251)
Combining these two cases above, we have inf PU |XY :CGK (X;Y |U )=0 I(XY ; U ) = CGK (X; Y ).
A PPENDIX B
P ROOF OF L EMMA 1
A proof for the unconditional version of the lemma can be found in [23]. Here we extend the proof to the
conditional version. To that end, we only consider finite valued random variables. For countably infinitely valued
or continuous random variables, the result can be proven similarly.
32
For finite valued random variables, we will show maximal correlation ρm (X; Y |U ) can also be characterized by
p(x, y, u)
p(x, y|u)
=p
.
the second largest singular value of the matrix Qu with entries Qu (x, y) := p
p(x|u)p(y|u)
p(x, u)p(y, u)
Without loss of generality, we can rewrite
ρm (X; Y |U ) = sup E[f (X, U )g(Y, U )],
(252)
f,g
where the maximization is taken over all f, g such that E[f (X, U )] = E[g(Y, U )] = 0, Evar(f (X, U )) = Evar(g(Y, U )) =
1. Observe that
X
E[f (X, U )g(Y, U )] =
p
p
(f (x, u) p(x, u))Qu (x, y)(g(y, u) p(y, u)),
(253)
x,y,u
Xp
X
p
p
p
p(x, u)Qu (x, y) = p(y, u),
Qu (x, y) p(y, u) = p(x, u),
x
(254)
y
and the conditions E[f (X, U )] = 0 and E[g(Y, U )] = 0 are respectively equivalent to requiring that (x, u) 7→
p
p
p
f (x, u) p(x, u) is orthogonal to (x, u) 7→ p(x, u) and that (y, u) 7→ g(y, u) p(y, u) is orthogonal to (y, u) 7→
n
X
p
p
p(y, u). By Singular Value Decomposition, Qu =
λu,i au,i bTu,i , where λu,1 = 1, au,1 = ( p(x, u))x , bu,1 =
i=1
p
( p(y, u))y . Therefore,
E[f (X, U )g(Y, U )] =
X
x,y,u
=
X
p
p
(f (x, u) p(x, u))Qu (x, y)(g(y, u) p(y, u))
fuT (
u
=
n
X
λu,i au,i bTu,i )gu
(255)
(256)
i=1
n
XX
λu,i cu,i du,i
(257)
u
≤
i=2
n
XX
λu,i
u
i=2
c2u,i + d2u,i
,
2
(258)
p
p
where fu := (f (x, u) p(x, u))x , gu := (g(y, u) p(y, u))y , cu,i := fuT au,i , du,i := guT bu,i , i ≥ 2. Furthermore,
X
u
n
X
i=2
n
X
i=2
Hence
X
kfu k2 =
u
kgu k2 = 1,
(259)
c2u,i = kfu k2 ,
(260)
d2u,i = kgu k2 .
(261)
n
XX
c2u,i = 1,
(262)
d2u,i = 1.
(263)
u
i=2
n
XX
u
i=2
33
Combining these with (252) and (258) gives us
ρm (X; Y |U ) ≤
sup
(264)
λu,2 .
u:P (u)>0
On the other hand, it is easy to verify that the upper bound
sup
λu,2 can be achieved by choosing
u:P (u)>0
fu =
and
gu =
Therefore,
au,2 , if u = u∗ ;
(265)
bu,2 , if u = u∗ ;
(266)
0,
otherwise.
0,
otherwise.
ρm (X; Y |U ) =
sup
(267)
λu,2 .
u:P (u)>0
A PPENDIX C
P ROOF OF L EMMA 10
By the law of total covariance, we have
Ecov(X, Y |Z) = Ecov(X, Y |ZU ) + EZ covU (E(X|ZU ), E(Y |ZU )).
(268)
Hence to prove Lemma 10, we only need to show
p
Evar(X|ZU )Evar(Y |ZU ) + EZ covU (E(X|ZU ), E(Y |ZU )) ≤
To prove this, we consider
p
Evar(X|Z)Evar(Y |Z).
(269)
Evar(X|ZU )Evar(Y |ZU )
= (Evar(X|Z) − EZ varU (E(X|ZU ))) (Evar(Y |Z) − EZ varU (E(Y |ZU )))
(270)
=Evar(X|Z)Evar(Y |Z) − Evar(X|Z)EZ varU (E(Y |ZU ))
− Evar(Y |Z)EZ varU (E(X|ZU )) + EZ varU (E(X|ZU ))EZ varU (E(Y |ZU ))
p
≤Evar(X|Z)Evar(Y |Z) − 2 Evar(X|Z)EZ varU (E(Y |ZU )) · Evar(Y |Z)EZ varU (E(X|ZU ))
+ EZ varU (E(X|ZU ))EZ varU (E(Y |ZU ))
p
2
p
=
Evar(X|Z)Evar(Y |Z) − EZ varU (E(X|ZU ))EZ varU (E(Y |ZU ))
(271)
(272)
(273)
where (270) follows from the law of total variance
Evar(X|Z) = EZ varU (X|ZU ) + EZ varU (E(X|ZU )).
(274)
Since EZ varU (X|ZU ) ≥ 0, from (274), we have
EZ varU (E(X|ZU )) ≤ Evar(X|Z).
34
(275)
Similarly, we have
EZ varU (E(Y |ZU )) ≤ Evar(Y |Z).
(276)
EZ varU (E(X|ZU ))EZ varU (E(Y |ZU )) ≤ Evar(X|Z)Evar(Y |Z).
(277)
Therefore,
Combining (273) and (277), we have
p
p
p
Evar(X|ZU )Evar(Y |ZU ) ≤ Evar(X|Z)Evar(Y |Z) − EZ varU (E(X|ZU ))EZ varU (E(Y |ZU )).
(278)
Furthermore, by Cauchy-Schwarz inequality, it holds that
Therefore,
|EZ covU (E(X|ZU ), E(Y |ZU ))| = |E [(E(X|ZU ) − E(X|Z)) (E(Y |ZU ) − E(Y |Z))] |
q
2
2
≤ E (E(X|ZU ) − E(X|Z)) · E (E(Y |ZU ) − E(Y |Z))
p
= EZ varU (E(X|ZU ))EZ varU (E(Y |ZU )).
p
p
Evar(X|Z)Evar(Y |Z) − |EZ covU (E(X|ZU ), E(Y |ZU ))|
p
≤ Evar(X|Z)Evar(Y |Z) − EZ covU (E(X|ZU ), E(Y |ZU )),
Evar(X|ZU )Evar(Y |ZU ) ≤
(279)
(280)
(281)
(282)
(283)
which implies (269). This completes the proof.
A PPENDIX D
P ROOF OF T HEOREM 2
From Theorem 1, the following inequality follows immediately.
1
1 + β0 1 + β
(G)
Cβ (X; Y )≥ log+
.
/
2
1 − β0 1 − β
(284)
On the other hand, (X, Y ) can be expressed as
X = αU +
Y = αU +
with
α=
and the covariance of (Z1 , Z2 )
where U ∼ N (0, 1). Hence we have
p
p
s
1 − α 2 Z1 ,
(285)
1 − α 2 Z2 ,
(286)
β0 − β
1−β
Σ(Z1 ,Z2 ) =
1
β
β
1
(287)
,
ρ(X, Y |U ) ≤ β
and
1
I(XY ; U ) = log+
2
35
1 + β0 1 + β
/
.
1 − β0 1 − β
(288)
(289)
(290)
Hence
(G)
Cβ (X; Y
1
)≤ log+
2
1
log+
2
Combining (284) and (291) gives us
(G)
Cβ (X; Y ) =
This completes the proof.
1 + β0 1 + β
/
.
1 − β0 1 − β
(291)
1 + β0 1 + β
/
.
1 − β0 1 − β
(292)
A PPENDIX E
P ROOF OF T HEOREM 4
A. Achievability
Codebook Generation: Suppose R > Cβ (X; Y ). Randomly and independently generate sequences un (m), m ∈
n
Y
[1 : 2nR ] with each according to
PU (ui ). The codebook C = {un (m), m ∈ [2nR ]}.
i=1
Input Process Generator: Generate input source M according to the uniform distribution over [2nR ].
n
Y
Source Generator: Upon m, the generator generates sources (X n , Y n ) according to
PXY |U (xi , yi |ui (m)).
i=1
For such generator, the induced overall distribution is
PX n Y n M (xn , y n , m) := 2−nR
n
Y
i=1
PXY |U (xi , yi |ui (m)).
(293)
According to soft-cover lemma [18], if R > I(XY ; U ), then
lim EC kPX n Y n −
n→∞
n
Y
i=1
(294)
PXY kT V = 0.
Given U n (m) = un , (X n , Y n ) is a conditionally independent sequence, i.e.,
n
n
PX n Y n |M (x , y |m) =
Hence according to Lemma 11, we get
n
Y
i=1
PXY |U (xi , yi |ui (m)).
ρm (X n ; Y n |M ) = sup ρm (Xi ; Yi |Ui (M )).
(295)
(296)
1≤i≤n
Furthermore, from Lemma 2, we have
ρm (Xi ; Yi |Ui (M )) =
sup
u:PUi (u)>0
λ2,PXY |U (u) ≤
sup
u:PU (u)>0
λ2,PXY |U (u) ≤ β.
(297)
Hence
ρm (X n ; Y n |M ) ≤ β.
(298)
RP SS (β) ≤ Cβ (X; Y ).
(299)
This implies
36
B. Converse
Assume there exists a sequence of distributed generators such that lim supn→∞ n1 H(M ) ≤ R, ρm (X n ; Y n |M ) ≤
Q
β, ∀n, and limn→∞ kPX n Y n − PXY kT V = 0. Consider that
n
1
1X
I(X n Y n ; M ) =
I(Xi Yi ; M |X i−1 Y i−1 )
n
n i=1
(300)
n
=
1X
H(Xi Yi |X i−1 Y i−1 ) − H(Xi Yi |M X i−1 Y i−1 )
n i=1
(301)
n
=
1X
HQ (Xi Yi ) − H(Xi Yi |M X i−1 Y i−1 )
n i=1
(302)
n
=
1X
H(Xi Yi ) − H(Xi Yi |M X i−1 Y i−1 )
n i=1
(303)
n
=
1X
I(Xi Yi ; M X i−1 Y i−1 )
n i=1
(304)
= I(XT YT ; M X T −1 Y T −1 |T )
(305)
= I(XT YT ; M X T −1 Y T −1 T )
(306)
≥ I(XT YT ; M T )
(307)
= I(XY ; V ),
(308)
where T is a time-sharing random variable uniformly distributed [1 : n] and independent of all other random
variables, and X := XT , Y := YT , V := M T . Combining the inequality above with
1
1
I(X n Y n ; M ) ≤ H(M ) ≤ R
n
n
(309)
I(XY ; V ) ≤ R.
(310)
gives us
On the other hand,
ρm (X n ; Y n |M )≥ sup ρm (Xi ; Yi |M )
(311)
i
= sup ρm (Xi ; Yi |M = m)
(312)
= sup ρm (XT ; YT |M = m, T = i)
(313)
= ρm (XT ; YT |M, T )
(314)
= ρm (X; Y |V ),
(315)
i,m
i,m
where (311) follows from the definition of maximal correlation, and (312) follows from Lemma 2.
Combining (310) with (315) gives us
R≥
inf
PU |X,Y :ρm (X;Y |V )≤β
I(X, Y ; V ) = Cβ (X; Y ).
37
(316)
Hence
(317)
RP SS (β) ≥ Cβ (X; Y ).
This completes the proof.
A PPENDIX F
P ROOF OF T HEOREM 8
A. Achievability
Codebook Generation: Suppose R > Cβ (X; Y ). Randomly and independently generate sequences un (m), m ∈
n
Y
[1 : 2nR ] with each according to
PU (ui ). The codebook C = {un (m), m ∈ [2nR ]}.
i=1
Extractor: Upon (X, Y n ), the extractor generates sources m using a likelihood encoder PM |X n Y n (m|xn , y n ) ∝
n
Y
PXY |U (xi , yi |ui (m)), where ∝indicates that appropriate normalization is required.
i=1
For such extractor, the induced overall distribution PX n Y n M is related to an ideal distribution
QX n Y n M (xn , y n , m) := 2−nR
n
Y
i=1
PXY |U (xi , yi |ui (m)).
(318)
According to soft-covering lemma [18], if R > I(XY ; U ), then
(319)
lim EC kPX n Y n − QX n Y n kT V = 0,
n→∞
where
n
Y
QX n Y n (xn , y n ) =
(320)
PXY (xi , yi ).
i=1
On the other hand, observe that PM |X n Y n = QM |X n Y n . Hence by Property 1, we further have
lim EC kPX n Y n M − QX n Y n M kT V = lim EC kPX n Y n − QX n Y n kT V = 0.
n→∞
n→∞
(321)
Given U n (m) = un , (X n Y n ) is an independently distributed sequence under distribution Q. That is
QX n Y n |M (xn , y n |m) =
n
Y
PXY |U (xi , yi |ui (m)).
(322)
ρm,Q (X n ; Y n |M ) = sup ρm,Q (Xi ; Yi |Ui (M )).
(323)
Hence according to Lemma 11, we get
i=1
1≤i≤n
Furthermore, from Lemma 2, we have
ρm,Q (Xi ; Yi |Ui (M )) =
sup
u:PUi (u)>0
λ2,PXY |U (u) ≤
sup
u:PU (u)>0
λ2,PXY |U (u) ≤ β.
(324)
Hence
ρm,Q (X n ; Y n |M ) ≤ β.
(325)
RCIE (β) ≤ Cβ (X; Y ).
(326)
This implies
38
B. Converse
Assume there exists a sequence of extractors such that
lim sup
n→∞
1
H(M ) ≤ R,
n
(327)
and
inf
QX n ,Y n ,M :kQX n ,Y n ,M −PX n ,Y n ,M kT V ≤n
ρm,Q (X n ; Y n |M ) ≤ β, ∀n,
(328)
for some n such that limn→∞ n = 0.
Assume QX n ,Y n ,M achieves the infimum in (328). Hence kQX n ,Y n ,M − PX n ,Y n ,M kT V → 0. Then by the
total-variation bound on entropy, we have
1
1
| HP (X n Y n M ) − HQ (X n Y n M )|
n
n
1
|X n × Y n × [2nR ]|
≤ 2kQX n ,Y n ,M − PX n ,Y n ,M kT V log
n
2kQX n ,Y n ,M − PX n ,Y n M kT V
= 2kQX n ,Y n ,M − PX n ,Y n ,M kT V log
2R |X ||Y|
2kQX n ,Y n ,M − PX n ,Y n ,M kT V
(329)
(330)
(331)
→ 0,
and similarly,
and
1
|X ||Y|
1
| HP (X n Y n ) − HQ (X n Y n )| ≤ 2kQX n ,Y n − PX n ,Y n kT V log
→ 0,
n
n
2kQX n ,Y n − PX n ,Y n kT V
1
2R
1
→ 0.
| HP (M ) − HQ (M )| ≤ 2kQM − PM kT V log
n
n
2kQM − PM kT V
Furthermore, observe
1
1
1
1
I(X n Y n ; M ) = H(X n Y n ) + H(M ) − H(X n Y n M ). Hence
n
n
n
n
1
1
IP (X n Y n ; M )≤ HP (M )
n
n
≤ R.
(332)
(333)
(334)
(335)
On the other hand, consider that
n
n
IP (X Y ; M ) =
=
n
X
i=1
n
X
IP (Xi Yi ; M |X i−1 Y i−1 )
(336)
IP (Xi Yi ; M X i−1 Y i−1 )
(337)
i=1
= nIP (XT YT ; M X T −1 Y T −1 |T )
(338)
= nIP (XT YT ; M X T −1 Y T −1 T )
(339)
≥ nIP (XT YT ; M T )
(340)
≥ nIQ (XT YT ; M T ) − nn
(341)
= nIQ (XY ; V ) − nn ,
(342)
39
where T is a time-sharing random variable uniformly distributed [1 : n] and independent of all other random
variables, and X := XT , Y := YT , V := M T . Combining the inequality above with (335) gives us
IQ (XY ; V ) ≤ R + n .
(343)
ρm,Q (X n ; Y n |M )≥ max ρm,Q (Xi ; Yi |M )
(344)
Furthermore,
i
= max ρm,Q (Xi ; Yi i, m|M = m)
(345)
= max ρm,Q (XT ; YT i, m|M = m, T = i)
(346)
= ρm,Q (XT ; YT |M, T )
(347)
= ρm,Q (X; Y |V ),
(348)
where (344) follows from the definition of maximal correlation, and (345) and (347) follow from Lemma 2.
Furthermore, (328) implies lim supn→∞ ρm,Q (X n ; Y n |M ) ≤ β. Hence
(349)
ρm,Q (X; Y |V ) ≤ β.
Combining (343) with (349) gives us
R≥
inf
PV |X,Y :ρm (X;Y |V )≤β
I(XY ; V ) − n = Cβ (X; Y ) − n .
(350)
Hence
(351)
RCIE (β) ≥ Cβ (X; Y ).
This completes the proof.
A PPENDIX G
P ROOF OF T HEOREM 10
A. Achievability
(D,U B)
For the achievability part, we only need to show the upper bound Cβ
equivalent to showing that (R, β) with R >
(D,U B)
Cβ
(X; Y
(X; Y ) is achievable. It is also
) is achievable. Next we use a random binning strategy,
OSRB (Output Statistics of Random Binning) [22] to prove this, instead of using soft-covering technique. This
is because the “soft-covering” lemma is not easily applicable to complicated network structures, but OSRB is.
Furthermore, it is worth noting that the random binning technique can be applied to prove the centralized setting
case as well. Next we give the proof by following the basic proof steps of [22].
Part (1) of the proof: We define two protocols, source coding side of the problem (Protocol A) and the main
problem (Protocol B). Fig. 5 illustrates how the source coding side of the problem can be used to prove the common
information extraction problem.
Protocol A (Source coding side of the problem). Let (X n , Y n , U n , V n ) be i.i.d and distributed according to
PXY PU |X PV |Y . Consider the following random binning (see the left diagram of Fig. 5): uniformly and independently assign two bin indices m1 ∈ [1 : 2nR1 ] and f1 ∈ [1 : 2nR̃1 ] to each sequence un ; and similarly, uniformly
40
Encoder 1:
P
U|X
PM F |U n PF | X n PU n M | X n F
1 1
1
1
1
U
Extractor 1: PF1 PU n M1 | X n F1
F1
Xn
PU | X
Un
M1
F1
Un
X
n
PU n | X n F
n
M1
Un
Vn
M2
Vn
U
1
Yn
PV |Y
Vn
M2
V
n
Yn
PV n |Y n F
2
F2
F2
Encoder 2: PV |Y PM 2 F2 |V n PF2 |Y n PV n M 2 |Y n F2
U
Extractor 2: PF2 PV n M 2 |Y n F2
(Left) Source coding side of the problem (Protocol A). We pass i.i.d. sources X n and Y n through virtual discrete
memoryless channels PU |X and PV |Y respectively to generate i.i.d. sequences U n and V n . We describe U n and V n through two
Fig. 5.
random bins Mi and Fi at rates Ri and R̃i , i = 1, 2, where Mi will serve as the message for the receiver i in the main problem,
while Fi will serve as the shared randomness. We use SW decoder for decoding. (Right) The common information extraction
problem assisted with the shared randomness (Protocol B). We pass the sources X n and Y n and the shared randomnesses F1
and F2 through the reverse encoders to generate sequences U n and V n . The joint distribution of X n , Y n , M1 , M2 , F1 , F2 of
protocol A is equal to that of protocol B in total variation sense.
and independently assign two bin indices m2 ∈ [1 : 2nR2 ] and f2 ∈ [1 : 2nR̃2 ] to each sequence v n . Furthermore,
we use Slepian-Wolf (SW) decoders to recover un , v n from (m1 , m2 , f1 , f2 ). Denote the outputs of the decoders
by ûn and v̂ n , respectively.
The pmf induced by the random binning, denoted by P , can be expressed as
P (xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n )
= P (xn , y n )P (un |xn )P (v n |y n )P (f1 |un )P (f2 |v n )P (m1 |un )P (m2 |v n )P SW (ûn , v̂ n |m1 , m2 , f1 , f2 )
(352)
= P (xn , y n )P (f1 , un |xn )P (f2 , v n |y n )P (m1 |un )P (m2 |v n )P SW (ûn , v̂ n |m1 , m2 , f1 , f2 )
(353)
= P (xn , y n )P (f1 |xn )P (f2 |y n )P (un |xn , f1 )P (v n |y n , f2 )P (m1 |un )P (m2 |v n )P SW (ûn , v̂ n |m1 , m2 , f1 , f2 ).
(354)
Protocol B (Common information extraction problem assisted with the shared randomness). In this protocol we
assume that the transmitters (extractors) and the receivers have access to the shared randomnesses F1 , F2 where Fi
is uniformly distributed over [1 : 2nR̃i ], i = 1, 2. Then, the protocol proceeds as follows (see also the right diagram
of Fig. 5):
•
The transmitter 1 generates U n according to the conditional pmf P (un |xn , f1 ) of protocol A; and the transmitter
2 generates V n according to the conditional pmf P (v n |y n , f2 ) of protocol A.
•
Next, knowing un , the transmitter 1 generates m1 according to the conditional pmf P (m1 |un ) of protocol A.
Similarly, the transmitter 2 generates m2 according to the conditional pmf P (m2 |v n ) of protocol A.
•
Finally, upon (m1 , m2 , f1 , f2 ), the receiver uses the Slepian-Wolf decoder P SW (ûn , v̂ n |m1 , m2 , f1 , f2 ) of
protocol A to obtain an estimate of (un , v n ).
41
The pmf induced by the protocol, denoted by Pe, can be expressed as
Pe(xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n )
= P (xn , y n )P U (f1 )P U (f2 )P (un |xn , f1 )P (v n |y n , f2 )P (m1 |un )P (m2 |v n )P SW (ûn , v̂ n |m1 , m2 , f1 , f2 ). (355)
Part (2a) of the proof (Sufficient conditions that make the induced pmfs approximately the same): Observe that
f1 is a bin index of un and f2 is a bin index of v n in protocol A. For the random binning in protocol A, [22, Thm.
1] says that if
R̃1 < H(U |XY )
(356)
R̃2 < H(V |XY )
(357)
R̃1 + R̃2 < H(U V |XY )
(358)
then P (xn , y n )P (f1 |xn )P (f2 |y n ) ≈ P (xn , y n )P U (f1 )P U (f2 ). Combining this with (354) and (355) gives us
Pe(xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n ) ≈ P (xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n ).
(359)
Part (2b) of the proof (Sufficient conditions that make the Slepian-Wolf decoders succeed): [22, Lem. 1] says that
if
R1 + R̃1 > H(U |V )
(360)
R2 + R̃2 > H(V |U )
(361)
R1 + R2 + R̃1 + R̃2 > H(U V )
(362)
then
P (xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n ) ≈ P (xn , y n , un , v n , f1 , f2 , m1 , m2 )1{ûn = un , v̂ n = v n }.
(363)
Using (359), (363) and the triangle inequality, we have
Pe(xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n ) ≈ P (xn , y n , un , v n , f1 , f2 , m1 , m2 )1{ûn = un , v̂ n = v n }.
(364)
Part (3) of the proof (Eliminating the shared randomness F1 , F2 ): (364) holds for the random pmfs induced by
random binning, by Property 1, which guarantees existence of a fixed binning such that (364) holds for the induced
non-random pmfs. (364) can be rewritten as
Pe(xn , y n , un , v n , f1 , f2 , m1 , m2 , ûn , v̂ n )
≈ P (f1 , f2 , m1 , m2 , ûn , v̂ n )P (xn , y n |un , v n )1{ûn = un , v̂ n = v n }.
(365)
From (365) we further have
Pe(xn , y n , f1 , f2 , m1 , m2 , ûn , v̂ n )
≈ P (f1 , f2 , m1 , m2 , ûn , v̂ n )PX n Y n |U n V n (xn , y n |ûn , v̂ n )
(366)
= P (f1 , f2 , m1 , m2 )1 {ûn = ûn (m1 , f1 ) , v̂ n = v̂ n (m2 , f2 )} PX n Y n |U n V n (xn , y n |ûn , v̂ n ),
(367)
42
where PX n Y n |U n V n =
Hence
Qn
i=1
PXY |U V , and ûn (m1 , f1 ) and v̂ n (m2 , f2 ) correspond to the Slepian-Wolf decoders.
Pe(xn , y n , f1 , f2 , m1 , m2 )
≈ Q(xn , y n , f1 , f2 , m1 , m2 )
(368)
:= P (f1 , f2 , m1 , m2 )PX n Y n |U n V n (xn , y n |ûn (m1 , f1 ) , v̂ n (m2 , f2 )).
(369)
Observe that under Q, given F1 F2 M1 M2 , X n Y n follows
QX n Y n |F1 F2 M1 M2 (xn , y n |f1 , f2 , m1 , m2 ) =
Hence by Lemma 11, we get
n
Y
i=1
PXY |U V (xi , yi |ûi (m1 , f1 ) , v̂i (m2 , f2 )).
ρm,Q (X n ; Y n |F1 F2 M1 M2 ) = sup ρm,Q (Xi ; Yi |Ûi (M1 , F1 ) , V̂i (M2 , F2 )).
(370)
(371)
1≤i≤n
On the other hand, from Lemma 2, we have
ρm,Q (Xi ; Yi |Ûi (M1 , F1 ) , V̂i (M2 , F2 )) =
≤
sup
u,v:PUi Vi (u,v)>0
sup
u,v:PU V (u,v)>0
λ2,PXY |U V (u, v)
λ2,PXY |U V (u, v)
(372)
(373)
(374)
≤ β.
Therefore,
ρm,Q (X n ; Y n |F1 F2 M1 M2 ) ≤ β.
(375)
By choosing F1 = f1 , F2 = f2 for arbitrary (f1 , f2 ), it holds that
ρm,Q (X n ; Y n |F1 = f1 , F2 = f2 , M1 , M2 ) ≤ β.
(376)
Finally, specifying P (m1 |xn , f1 ) as the encoder 1 and P (m2 |xn , f2 ) as the encoder 2 (which is equivalent to,
for encoder 1, generating random sequences un according to P (un |xn , f1 ) and then transmitting the bin index
m1 assigned to un , and for encoder 2, doing similar operations), and P SW (ûn , v̂ n |m1 , m2 , f1 , f2 ) as the decoder
results in a pair of encoder-decoder obeying the desired constraints:
ρm,Q (X n ; Y n |M1 M2 ) = ρm,Q (X n ; Y n |F1 = f1 , F2 = f2 , M1 , M2 ) ≤ β.
(377)
Observe that the common information extraction above only requires R1 + R2 > I(XY ; U V ) = IQ (XY ; U V ).
(D,U B)
This implies Cβ
(X; Y ) is achievable, which in turn implies
(D,U B)
RDCIE (β) ≤ Cβ
(D)
(X; Y ).
(378)
Furthermore, Cβ (X; Y ) = limn→∞ inf PU |X n PV |Y n :ρm (X n ;Y n |U V )≤β n1 I(X n Y n ; U V ) is also achievable, since
(D,U B)
it is a multiletter extension of Cβ
(X; Y ).
43
B. Converse
Assume there exists an extractor such that
inf
QX n ,Y n ,M1 ,M2 :kQX n ,Y n ,M1 ,M2 −PX n ,Y n ,M1 ,M2 kT V ≤n
ρm,Q (X n ; Y n |M1 , M2 ) ≤ β, ∀n,
(379)
for some n such that lim supn→∞ n = 0.
Set U = M1 , V = M2 and follow similar steps to Subsection F-B, then we have
(D)
Cβ (X; Y ) = lim
inf
n→∞ PU |X n PV |Y n :ρm (X n ;Y n |U V )≤β
1
I(X n Y n ; U V ) ≤ R.
n
(380)
Hence
(D)
(381)
RDCIE (β) ≥ Cβ (X; Y ).
(D)
Combining this with the achievability of Cβ (X; Y ) gives us
(D)
RDCIE (β) = Cβ (X; Y ).
(382)
(D)
(383)
Now we remain to show
(D,LB)
Cβ (X; Y ) ≥ Cβ
Consider PU |X n PV |Y n such that ρm (X n ; Y n |U V ) ≤ β and
(X; Y ).
n n
1
n I(X Y ; U V
hold.
) ≤ R. Then the following equations
U → XT → YT ,
(384)
XT → YT → V,
(385)
ρm (XT , YT |U V T ) ≤ ρm (X n ; Y n |U V ),
(386)
ρm (U XT ; V YT |T ) ≤ ρm (U X n ; V Y n ) = ρm (X n ; Y n ) = ρm (X; Y ),
(387)
and
IQ (X n Y n ; U V ) =
=
n
X
i=1
n
X
IQ (Xi Yi ; U V |X i−1 Y i−1 )
(388)
IQ (Xi Yi ; U V X i−1 Y i−1 )
(389)
i=1
= nIQ (XT YT ; U V X T −1 Y T −1 |T )
(390)
= nIQ (XT YT ; U V X T −1 Y T −1 T )
(391)
≥ nIQ (XT YT ; U V |T )
(392)
= nIQ (XY ; U V |T ),
(393)
where T is a time-sharing random variable uniformly distributed [1 : n] and independent of all other random
variables, and X := XT , Y := YT . Therefore,
(D)
(D,LB)
Cβ (X; Y ) ≥ Cβ
This completes the proof.
44
(X; Y ).
(394)
R EFERENCES
[1] P. Gács and J. Körner, “Common information is far less than mutual information,” Probl. Contr lnform. Theory vol. 2, no. 2, pp. 149-162,
1973.
[2] H. S. Witsenhausen, “On sequences of pairs of dependent random variables,” SIAM J. Appl. Math., vol. 28, no. 1, pp. 100-113, Jan. 1975.
[3] A. Wyner, “The common information of two dependent random variables,” IEEE Trans. lnf. Theory, vol. 21, no. 2, pp. 163-179, Mar.
1975.
[4] H. Gebelein, “Das statistische Problem der Korrelation als Variationsund Eigenwert-problem und sein Zusammenhang mit der Ausgleichungsrechnung,” Zeitschrift für angew. Math. und Mech. 21, pp. 364-379, 1941.
[5] H. O. Hirschfeld, “A connection between correlation and contingency,” Proc. Cambridge Philosophical Soc. 31, pp 520-524, 1935.
[6] A. Rényi, “On measures of dependence,” Acta Math. Hung., vol. 10, pp. 441-451, 1959.
[7] A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011.
[8] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2004.
[9] S. Beigi and A. Gohari, “Monotone measures for non-local correlations,” IEEE Trans. Inf. Theory, vol. 61, no. 9, pp. 5185-5208, 2015.
[10] C. T. Li and A. El Gamal, “Maximal correlation secrecy,” arXiv preprint arXiv:1412.5374, Oct. 2016.
[11] S. Kamath and V. Anantharam, “On non-interactive simulation of joint distributions,” IEEE Trans. lnf. Theory, vol. 62, no. 6, pp. 3419-3435,
Jun. 2016.
[12] G. R. Kumar, C. T. Li, and A. El Gamal, “Exact common information,” in Proc. IEEE Symp. Inf. Theory, Honolulu, HI, USA, Jun./Jul.
2014, pp. 161-165.
[13] Y. A. Rozanov, Stationary Random Processes. San Francisco, CA: Holden-Day, 1967.
[14] L. Yu, H. Li, and C. W. Chen, “Distortion Bounds for Transmitting Correlated Sources with Common Part over MAC,” in Proc. 54th Ann.
Allerton Conf. Commun., Contr., and Comput., Sep. 2016. [Online]. Available: https://arxiv.org/abs/1607.01345.
[15] J. Liu, P. Cuff, and S. Verdú, “Eγ -Resolvability,” IEEE Trans. lnf. Theory, vol. 63, no. 5, pp. 2629-2658, May 2017.
[16] J. Liu, T. A. Courtade, P. Cuff, and S. Verdú, “Smoothing Brascamp-Lieb inequalities and strong converses for CR generation,” in Proc.
IEEE Symp. Inf. Theory, Jul. 2016, pp. 1043-1047.
[17] T. M. Cover and J. A. Thomas, Elements of Information Theory, Wiley, New York, 1991.
[18] P. Cuff, “Distributed channel synthesis,” IEEE Trans. lnf. Theory, vol. 59, no. 11, pp. 7071-7096, 2013.
[19] C. Schieler, and P. Cuff, “The henchman problem: Measuring secrecy by the minimum distortion in a list,” IEEE Trans. lnf. Theory, vol.
62, no. 6, pp. 3436-3450, Jun. 2016.
[20] S. Asoodeh, M. Diaz, F. Alajaji, and T. Linder, “Information extraction under privacy constraints,” Information, vol. 7, no. 1, 2016.
[21] G. Xu, W Liu, and B. Chen, “Wyner’s common information for continuous random variables–A lossy source coding interpretation,” in
Proc. 45th Annu. Conf. CISS, 2011, pp. 16.
[22] M. Yassaee, M. Aref, and A. Gohari, “Achievability proof via output statistics of random binning,” IEEE Trans. Inf. Theory, vol. 60, pp.
6760–6786, Nov 2014.
[23] V. Anantharam, A. Gohari, S. Kamath, and C. Nair, “On maximal correlation, hypercontractivity, and the data processing inequality studied
by erkip and cover,” arXiv:1304.6133[cs.IT], 2013.
45
| 7cs.IT
|
arXiv:1610.03694v1 [math.ST] 12 Oct 2016
Local asymptotic normality property for
fractional Gaussian noise under
high-frequency observations
Alexandre Brouste∗ and Masaaki Fukasawa†
∗
†
Département de Mathématiques, Université du Maine
Graduate School of Engineering Science, Osaka University
1-3 Machikaneyama, Toyonaka, Osaka, 560-8531 JAPAN
[email protected]
October 13, 2016
Abstract
Local Asymptotic Normality (LAN) property for fractional Gaussian
noise under high-frequency observations is proved with a non-diagonal
rate matrix depending on the parameter to be estimated. In contrast to the
LAN families in the literature, non-diagonal rate matrices are inevitable.
1 Introduction
The theory of Local Asymptotic Normality (LAN) provides a powerful framework under which we are able to discuss asymptotic optimality of estimators.
When the LAN property holds true for a statistical experiment, minimax theorems [9, 17] can be applied and a lower bound for the variance of the estimators
can be derived. Beyond the classical IID setting [10], the LAN property (or Local
Asymptotic Mixed Normality property) has been proved for various statistical
models including linear processes [12], ergodic Markov chains [19], ergodic
diffusions [7, 16], diffusions under high-frequency observations [8] and several
Lévy process models [2, 14, 15]. The LAN property for fractional Gaussian
noise (fGn) was obtained in [3] under the large sample observation scheme. In
this setting, Maximum Likelihood (ML) and Whittle sequences of estimators
achieve optimality [5, 6, 18].
The statistical experiment of observing fGn under high-frequency scheme
has not been well understood in the literature, despite that high-frequency data
has attracted much attention recently due to their increasing availability. At
high-frequency, scaling effects from the variance and from the self-similarity of
the fGn are melting. The singularity of the joint estimation of diffusion coefficient and Hurst parameter was already noticed in [1, 11]. A weak LAN property
1
with a singular Fisher matrix was obtained in [13]. Due to this singularity, no
minimax theorem can be applied and in particular, it has been unclear whether
the ML estimator enjoys any kind of optimality property.
In this paper, we prove that the statistical experiment in fact enjoys the
LAN property. To discuss the difference from [13], let us be more precise in the
definition of the LAN property.
Definition 1. Let Θ ⊂ Rd . A family of measures {Pnθ , θ ∈ Θ} is called locally
asymptotically normal (LAN) at θ0 ∈ Θ if there exist nondegenerate d × d matrices
ϕn (θ0 ) and I(θ0 ) such that for any u ∈ Rd , the likelihood ratio
Zn (u) =
dPnθ0 +ϕn (θ0 )u
dPnθ0
admits the representation
1
Zn (u) = exp hu, ζn (θ0 )i − hI(θ0 )u, ui + rn (θ0 , u) ,
2
(1)
ζn (θ0 ) → N(0, I(θ0)), rn (θ0 , u) → 0
(2)
where
in law under
Pnθ0 .
This definition of the LAN property is equivalent to the one given in [10].
The matrix ϕn (θ0 ) is often called the rate matrix. Remark that the nondegeneracy of ϕn (θ0 ) and I(θ0 ) is essential in the following minimax theorem due to
Hajek [9] and Le Cam [17], which implies in particular the asymptotic efficiency
of the ML sequence of estimators in regular models.
Theorem 1. Let the family of measures {Pnθ , θ ∈ Θ}, Θ ⊂ Rd , be LAN at θ0 ∈ Θ for
nondegenerate matrices ϕn (θ0 ) and I(θ0 ). Then, for any sequence of estimators θ̂n ,
h
i Z
−1
l I(θ0 )−1/2 z φ(z)dz
lim inf lim inf sup Eθ l ϕn (θ0 ) (θ̂n − θ) ≥
δ→0
n→∞
Rd
|θ−θ0 |<δ
2
for any symmetric, nonnegative quasi-convex function l with lim|z|→∞ e−ǫ|z| l(z) = 0
for all ǫ > 0, where φ is the density of the d-dimensional standard normal distribution.
For the fGn under high-frequency observations, we consider the estimation
of the parameters (H, σ) ∈ Θ = (0, 1)×(0, ∞), where H is the Hurst parameter and
σ is the diffusion coefficient. It was shown in [13] that the family of measures
{Pn(H,σ) , (H, σ) ∈ Θ} satisfies both conditions (1) and (2) at any (H, σ) ∈ Θ for
1
√
ϕn (H, σ) = n log ∆n
0
0
√1
and I(H, σ) =
n
!
2
2/σ
,
2/σ 2/σ2
where n is the sample size and ∆n is the length of sampling intervals. Note that
I(H, σ) is singular. On the one hand, this result does not imply that the family
2
is LAN in the sense of Definition 1. On the second hand, Theorem 1 can not be
applied with this result.
To the best of our knowledge, for any LAN family in the literature, it is
always possible to take a diagonal rate matrix. A typical example is the LAN
property for the IID setting where
1
ϕn (θ0 ) = √ Id .
n
Here Id is the d×d identity matrix. However, both in Definition 1 and Theorem 1,
ϕn (θ0 ) is required only to be nondegenerate. In this paper, we prove the LAN
property for the statistical experiment of observing fGn under high-frequency
scheme for a certain class of non-diagonal matrices ϕn (θ0 ) depending on θ0 . In
particular, Theorem 1 can be applied. Non-diagonal rate matrices are inevitable
because the Fisher matrix is singular.
Basics for the fractional Brownian motion and the fGn are recalled in Section
2. The statistical experiment under high-frequency scheme is described and the
LAN property result is proved in Section 3. Efficient rates for the estimation
of H and σ are given in Section 4 giving the optimality of the ML sequence of
estimators.
2 Fractional Brownian motion and fractional Gaussian noise
Here we briefly review the basics of the fractional Brownian motion, fractional
Gaussian noise and their large sample theory. A centered Gaussian process BH
is called a fractional Brownian motion with Hurst parameter H if
1 2H
(|t| + |s|2H − |t − s|2H )
2
for all t, s ∈ R. Such a process exists and is continuous for any H ∈ (0, 1)
by Kolmogorov’s extension and continuity theorems. The process enjoys a
self-similarity property : for any ∆ > 0 and t ∈ R,
H
E[BH
t Bs ] =
H H
H
BH
t+∆ − Bt ∼ ∆ B1
in law. The spectral density of regular increments, that is, the function fH
characterized by
1
H
H
E[BH
|k + 1|2H − 2|k|2H + |k − 1|2H
1 (Bk+1 − Bk )] =
2 Z
π √
1
=
e −1kλ fH (λ)dλ, k ∈ Z
2π −π
is given by
fH (λ) = CH 2(1 − cos(λ))
X
k∈Z
|λ + 2kπ|−2H−1
3
with
CH =
Γ(2H + 1) sin(πH)
.
2π
For a fixed interval ∆ > 0, assume we observe the increments
H
H
H
H
H
H
Xn = σBH
,
σB
−
σB
,
σB
−
σB
,
.
.
.
,
σB
−
σB
∆
2∆
∆
3∆
2∆
n∆
(n−1)∆ ,
where (H, σ) is unknown. The random vector Xn is called fractional Gaussian
noise (fGn). Denote by Tn (H) the covariance matrix of
Xn
,
σ∆H
of which the distribution does not depend on σ and ∆ by the self-similarity
property. The (i, j) element of Tn (H) is given by
Z π √
1
e −1(i− j)λ fH (λ)dλ.
2π −π
Let us suppose ∆ = 1 for brevity. The log-likelihood ratio is then given by
n
log(2π) − n log σ
2
1
1
− log |Tn (H)| − 2 hXn , Tn (H)−1 Xn i.
2
2σ
ℓn (H, σ) = −
Let
√ 1
−1
hX
,
T
(H)
X
i
−
1
,
n
n
n
n
nσ2
1
1 1
−1
∂H log |Tn (H)| +
Bn = √
hX
,
∂
{T
(H)
}X
i
.
n H n
n
2 σ2
n 2
An =
The score function is then given by
!
√ !
−B√n n
∂H ℓn
.
=
∇ℓn =
∂σ ℓn
An nσ−1
Let Pn(H,σ) be the measure on Rn induced by Xn .
n
o
Theorem 2. The family of measures Pn(H,σ) ; (H, σ) ∈ (0, 1) × (0, ∞) is LAN at any
(H, σ) ∈ (0, 1) × (0, ∞) with
1
0
√n
ϕn (H, σ) =
√1
0
n
and
Rπ
2
1
R π|∂H log fH (λ)| dλ
I(H, σ) = 4π1 −π
∂ log f (λ)dλ
2πσ −π
H
H
1
2πσ
Rπ
∂
log
f
(λ)dλ
H
H
−π
.
2
σ2
In particular, (An , Bn ) converges in law to a centered normal distribution with covariance
Rπ
1
2
− 2π
∂H log fH (λ)dλ
−π
R
R
J(H) = 1 π
(3)
.
π
1
− 2π −π ∂H log fH (λ)dλ 4π
|∂ log fH (λ)|2 dλ
−π H
Proof. See [3].
4
3 The LAN property in high-frequency observation
Let Xn be again the observed fractional Gaussian noise
H
H
H
H
H
H
Xn = σBH
,
σB
−
σB
,
σB
−
σB
,
.
.
.
,
σB
−
σB
∆n
2∆n
∆n
3∆n
2∆n
n∆n
(n−1)∆n .
Here we consider high-frequency asymptotics, which means ∆n → 0 as n → ∞.
The parameters to be estimated are still H ∈ (0, 1) and σ > 0. As before, the
distribution of
1
Xn
σ∆H
n
is stationary centered Gaussian and does not depend on σ and ∆n by the selfsimilarity property. Its covariance matrix is Tn (H) defined in the previous
section. Denote by ℓn (H, σ) the log-likelihood ratio
n
log(2π) − nH log ∆n − n log σ
2
1
1
− log |Tn (H)| − 2 2H hXn , Tn (H)−1Xn i.
2
2σ ∆n
ℓn (H, σ) = −
Let
√
n
(
)
1
−1
hX
,
T
(H)
X
i
−
1
,
n
n
n
nσ2 ∆2H
n
(
)
1
1 1
−1
∂H log |Tn (H)| +
hXn , ∂H {Tn (H) }Xn i .
Bn = √
2 σ2 ∆2H
n 2
n
An =
We use the same notation as in the previous section because their distributions
are the same. In particular, we have that (An , Bn ) → (A, B) in law under Pn(H,σ) for
a nondegenerate Gaussian random variable (A, B) whose covariance is given
by (3). In this setting, the score function is given by
!
√ !
√
A n log√∆n − Bn n
∂ ℓ
∇ℓn = H n = n
.
∂σ ℓn
An nσ−1
From this expression, we clearly see that the leading terms of ∂H ℓn and ∂σ ℓn are
linearly dependent, which is exactly the reason why we have a singular Fisher
matrix when considering diagonal rate matrices as in [13]. Now, we state the
main result of the paper.
Theorem 3. Suppose infn n∆n > 0. Consider a matrix
αn
ϕn = ϕn (H, σ) =
βn
with the following properties :
1. |ϕn | = αn β̂n − α̂n βn , 0.
5
α̂n
β̂n
!
√
2. αn n → α for some α ≥ 0.
√
3. α̂n n → α̂ for some α̂ ≥ 0.
√
√
4. γn := αn n log ∆n + βn nσ−1 → γ for some γ ∈ R.
√
√
5. γ̂n := α̂n n log ∆n + β̂n nσ−1 → γ̂ for some γ̂ ∈ R.
6. αγ̂ − α̂γ , 0.
n
o
Then, the family Pn(H,σ) ; (H, σ) ∈ (0, 1) × (0, ∞) is LAN at any (H, σ) for ϕn (H, σ)
defined previously and
!
!
γ −α
γ
γ̂
I(H, σ) =
J(H)
,
γ̂ −α̂
−α −α̂
where J(H) is defined by (3).
Proof. Let θ0 = (H0 , σ0 ) and u ∈ R2 . Let B̄(θ0 , r) = {θ ∈ (0, 1) × (0, ∞); |θ − θ0| ≤ r}
for r > 0. By Taylor’s formula,
1
ℓn (θ0 + ϕn u) − ℓn (θ0 ) = hϕ∗n ∇ℓn (θ0 ), ui + hu, ϕ∗n ∇2 ℓn (θ0 )ϕn ui + rn ,
2
where
1
|ϕn u| |u|2 max{|ϕ∗n ∇3 ℓn (θ)ϕn |; θ ∈ B̄(θ0 , |ϕn u|)}.
6
Step 1). Let us show that ϕ∗n ∇ℓn (θ0 ) → N(0, I(θ0)). Note that
√ !
An γn − Bn αn √n
∗
.
ϕn ∇ℓn =
An γ̂n − Bn α̂n n
|rn | ≤
Therefore, it converges in law to
!
! !
Aγ − Bα
γ −α A
=
,
Aγ̂ − Bα̂
γ̂ −α̂ B
which is Gaussian.
Step 2). Here we compute
ϕ∗n ∇2 ℓn ϕn .
Let
Cn =
Dn =
1
nσ2 ∆2H
n
1
nσ2 ∆2H
n
hXn , Tn (H)−1 Xn i,
hXn , ∂H {Tn (H)−1 }Xn i,
(
)
1
1 1 2
2
−1
∂ log |Tn (H)| +
hXn , ∂H {Tn (H) }Xn i .
En =
n 2 H
2 σ2 ∆2H
n
6
Note that
∂H log |Tn (H)| = −tr(∂H {Tn (H)−1 }Tn (H))
and so,
∂2H log |Tn (H)| = −tr(∂2H Tn (H)−1 Tn (H)) + tr(Tn (H)−1∂H Tn (H)Tn (H)−1 ∂H Tn (H)).
Then, the (1, 1) element of ∇2 ℓn ϕn is
√
√
n
αn
n log ∆n ∂H An − n∂H Bn + βn
∂ H An
σ
√
= γn ∂H An − αn n∂H Bn
√
= γn n −2Cn log ∆n + Dn − αn n En − Dn log ∆n
√
and the (1, 2) element is
√
!
√
√
n
n
n
∂ H An + β n − 2 +
∂ σ An
αn
σ
σ
σ
√
√
n
n
n
Cn γn + Dn αn − βn 2 .
= −2
σ
σ
σ
It follows then that the (1, 1) element of ϕ∗n ∇2 ℓn ϕn is
−2Cn γ2n
√
+ 2Dn γn αn n − α2n nEn − β2n
√
n
.
σ2
From this, it is clear that the (2, 2) element of ϕ∗n ∇2 ℓn ϕn is
√
√
n
−2Cn γ̂2n + 2Dn γ̂n α̂n n − α̂2n nEn − β̂2n 2 .
σ
Also it is already not difficult to see that the (1, 2) element of ϕ∗n ∇2 ℓn ϕn is
√
√
√
n
−2Cn γn γ̂n + Dn (γ̂n αn n + γn α̂n n) − αn α̂n nEn − βn β̂n 2 .
σ
It is clear that nCn ∼ χ2 (n) and so, Cn → 1. By the same argument as the proof
of Lemma 3.5 of [3], we have
Z π
1
∂H log fH (λ)dλ = E[AB]
Dn → −
2π −π
and
En →
Therefore,
1
4π
Z
π
−π
|∂H log fH (λ)|2 dλ = E[B2 ].
ϕ∗n ∇2 ℓn ϕn
→−
2γ2 − 2γαE[AB] + α2 E[B2 ]
2γγ̂ − (γα̂ + γ̂α)E[AB] + αα̂E[B2 ]
7
!
2γγ̂ − (γα̂ + γ̂α)E[AB] + αα̂E[B2 ]
.
2γ̂2 − 2γ̂α̂E[AB] + α̂2 E[B2 ]
This coincides with −I(H, σ) because E[A2 ] = 2.
Step 3) It remains to show rn → 0. From the computation in Step 2, we deduce
that the tensor −ϕ∗n ∇3 ℓn ϕn consists of the vectors
√
2∇Cn γ2n − 2γn αn n∇Dn + αn n∇En ,
√
√
2∇Cn γn γ̂n − (γn α̂n n + γ̂n αn n)∇Dn + αn α̂n n∇En ,
√
2∇Cn γ̂2n − 2γ̂n α̂n n∇Dn + α̂2n n∇En .
By the same argument as the proof of Lemma 3.7 of [3], we have that there
exists ǫ > 0 such that ∇Cn , ∇Dn and ∇En are of Op (n1/2−ǫ | log ∆n |) uniformly in
θ ∈ B̄(θ0 , ǫ). On the other hand,
n(α2n + α̂2n ) → α2 + α̂2 ,
n
(β2 + β̂2n ) → σ2 (α2 + α̂2 ),
| log ∆n | n
p
which implies that |ϕn u| = O( | log |∆n |/n). Since infn n∆n > 0 by the assumption, we conclude that rn → 0.
Several examples can be elicited. A symmetric matrix for rate ϕn is
1
√n log ∆
ϕn = 1 n
√
n
√1
n
,
σ log ∆n
− √n
for which α = 0, α̂ = 1, γ = 1 + σ−1 and γ̂ = 0. Another example is that
√1
√1
n
n
ϕn = √σ
√σ γ̂ − log ∆n
γ
−
log
∆
n
n
n
for (γ, γ̂) with γ , γ̂, for which α = α̂ = 1.
Two other examples of rate matrix will be used in the following section,
namely
!
1
1
0
,
ϕn = √
n −σ log ∆n 1
which gives α = 1, α̂ = 0, γ = 0 and γ̂ = σ−1 and
1
ϕn = √
n
1
log ∆n
0
!
1
,
−σ log ∆n
which gives α = 0, α̂ = 1, γ = 1 and γ̂ = 0. Remark that all the examples are
non-diagonal rate matrix depending on the parameter σ.
8
4 Efficient rates of estimation and optimality of ML
estimtors
4.1 The efficient estimation rate for H
As the rate matrix, we can take
1
1
ϕn = √
n −σ log ∆n
!
0
,
1
which gives α = 1, α̂ = 0, γ = 0 and γ̂ = σ−1 . It is worth mentioning that the
rate matrix is not diagonal and depends on the parameter σ. By Theorem 1, the
LAN property implies that
!!#
"
!−1/2
E[B2 ]
−E[AB]/σ
−1 Ĥn − H
lim inf E(H,σ) l ϕn
≥ E l
N
−E[AB]/σ
2/σ2
n→∞
σ̂n − σ
for any loss function l satisfying the condition given in Theorem 1, where
N ∼ N(0, I2). Since
!
√
1
0
−1
,
ϕn = n
σ log ∆n 1
we obtain the asymptotic lower bound of
nE(H,σ) [(Ĥn − H)2 ]
by taking l(x, y) = x2 . This means that the efficient rate of estimation for H is
√
n when both
√ H and σ are unknown. Note that when σ is known, the efficient
rate for H is n log ∆n , which follows from, say, [13].
4.2 The efficient estimation rate for σ
As the rate matrix, we can take
1
ϕn = √
n
1
log ∆n
0
!
1
,
−σ log ∆n
which gives α = 0, α̂ = 1, γ = 1 and γ̂ = 0. It is worth repeating that the rate
matrix is not diagonal and depends on the parameter σ. Again by Theorem 1,
the LAN property implies that
!−1/2
!!#
"
2
−E[AB]
−1 Ĥn − H
N
lim inf E(H,σ) l ϕn
≥ E l
−E[AB] E[B2 ]
n→∞
σ̂n − σ
for any loss function l satisfying the condition given in Theorem 1, where
N ∼ N(0, I2). Since
!
√
−1
n −σ log ∆n
ϕ−1
=
−
,
1
n
0
σ
log ∆n
9
we obtain the asymptotic lower bound of
n
E(H,σ) [(σ̂n − σ)2 ]
(log ∆n )2
2
by
√ taking l(x, y) = y . This means that the efficient rate of estimation for σ is
n/| log ∆n | when both
√ H and σ are unknown. Note that when H is known, the
efficient rate for σ is n, which follows from, say, [13].
References
[1] Berzin, C. & Léon, J. (2008) Estimation in models driven by fractional Brownian
motion, Annales de l’Institut Henri Poincaré – Probabilités et Statistiques,
44, 191-213
[2] Clément, E. & Gloter, A. (2016) Local asymptotic mixed normality property
for discretely observed stochastic differential equations driven by stable Lévy
processes, Stochastic Process. Appl. 125 (2015), no. 6, 2316-2352.
[3] Cohen, S., Gamboa, F., Lacaux, C. & Loubes, J.-M. (2013) LAN property
for some fractional type Brownian motion, ALEA : Latin American Journal of
Probability and Mathematical Statistics, 10(1), 91–106.
[4] Coeurjolly, J.-F. & Istas, J. (2001) Cramer-Rao bounds for fractional Brownian
motions, Statistics & Probability Letters, 53, 435–447.
[5] Dahlhaus, R. (1989) Efficient parameter estimation for self-similar processes,
Annals of Statistics, 17(4), 1749–1766.
[6] Dahlhaus, R. (2006) Correction Efficient parameter estimation for self-similar
processes, Annals of Statistics, 34(2), 1045–1047.
[7] Gobet, E. (2001) Local asymptotic mixed normality property for elliptic diffusion:
a Malliavin calculus approach, Bernoulli, 7(6), 899–912.
[8] Gobet, E. (2002) LAN property for ergodic diffusions with discrete observations,
Annales de l’Institut Henri Poincaré, 38(5), 711–737.
[9] Hájek, J. (1972) Local asymptotic minimax and admissibility in estimation, Proc.
6th Berkeley Symp. Math. Statist. Prob., 1, 175-194.
[10] Ibragimov, I.A. & Has’minski, R.Z (1981).Statistical estimation: asymptotic
theory, Springer-Verlag.
[11] Istas, J. & Lang, G. (1997) Quadratic variations and estimation of the local
Hölder index of a Gaussian process, Annales de l’Institut Henri Poincar –
Probabilits et Statistiques, 33(4), 407–436.
[12] Kakizawa, Y. & Taniguchi, M (2000) Asymptotic theory of statistical inference
for time series, Springer-Verlag, New York.
10
[13] Kawai, R. (2013) Fisher Information for Fractional Brownian Motion Under
High-Frequency Discrete Sampling, Communications in Statistics - Theory
and Methods, 42(9), 1628-1636.
[14] Kawai, R. (2013) Local asymptotic normality property for Ornstein-Uhlenbeck
processes with jumps under discrete sampling, J. Theoret. Probab. 26, no. 4,
932-967.
[15] Kawai, R. & Masuda, H. (2013) Local asymptotic normality for normal inverse
Gaussian Lévy processes with high-frequency sampling, ESAIM Probab. Stat.
17, 13-32.
[16] Kutoyants, Y. (2004) Statistical inference for ergodic diffusion processes,
Springer-Verlag, London.
[17] Le Cam, L. (1972) Limits of experiments, Proc. 6th Berkeley Symp. Math.
Statist. Prob., 1, 245–261.
[18] Lieberman, O., Rosemarin, R. & Rousseau, J. (2012) Asymptotic theory for
maximum likelihood estimation of the memory parameter in stationary Gaussian
processes, Econometric theory, 28(02), 457–470.
[19] Roussas, G. (1972) Contiguity of probability measures. Cambridge University
Press.
[20] van der Vaart, A. (1998) Asymptotic Statistics, Cambridge University Press.
11
| 10math.ST
|
1
Modeling and Optimal Operation of Distributed
Battery Storage in Low Voltage Grids
arXiv:1603.06468v3 [cs.SY] 16 Mar 2017
Philipp Fortenbacher, Johanna L. Mathieu, and Göran Andersson
Abstract—Due to high power in-feed from photovoltaics, it
can be expected that more battery systems will be installed in
the distribution grid in near future to mitigate voltage violations
and thermal line and transformer overloading. In this paper, we
present a two-stage centralized model predictive control scheme
for distributed battery storage that consists of a scheduling entity
and a real-time control entity. To guarantee secure grid operation,
we solve a robust multi-period optimal power flow (OPF) for the
scheduling stage that minimizes battery degradation and maximizes photovoltaic utilization subject to grid constraints. The
real-time controller solves a real-time OPF taking into account
storage allocation profiles from the scheduler, a detailed battery
model, and real-time measurements. To reduce the computational
complexity of the controllers, we present a linearized OPF that
approximates the nonlinear AC-OPF into a linear programming
problem. Through a case study, we show, for two different
battery technologies, that we can substantially reduce battery
degradation when we incorporate a battery degradation model.
A further finding is that we can reduce battery losses by 30% by
using the detailed battery model in the real-time control stage.
Index Terms—optimal control, power systems, predictive control, energy storage
N OMENCLATURE
ηbat , ηin
dis
ch
ηbat
, ηbat
battery stack and inverter efficiency
battery stack discharging and charging efficiency
ηdis , ηch
total battery system discharging and charging efficiency
λ
SOS2 set
σ PV
standard deviation of the PV forecast error
Φ
discrete battery system dynamics matrix
a1 , a2 , a3
degradation plane parameter vectors of the
triangles from the convex hull
A
continuous battery system dynamics matrix
A1 , A2 , A3 , Az matrices to include battery degradation for
multiple battery systems and time steps
Aq
matrix to describe polygonal regions for
active and reactive power
bbat
battery system control input vector
B
battery system control input matrix for
multiple battery systems
B 1i , B 2i , bl
Matrices and vector specifying the PWA
network loss approximation
This work was, in part, financed by the Swiss Commission for Technology
and Innovation (CTI) 4/2013-10/2014
P. Fortenbacher and G. Andersson are with the Power Systems Laboratory, ETH Zurich, Switzerland (e-mail: [email protected], [email protected]).
J. L. Mathieu is with the Department of Electrical Engineering and
Computer Science, University of Michigan, USA (e-mail: [email protected]).
Br
Bv
Bq
Cg
C pv
cE
cE
cd
cn
cp
cr
cw
E
e
e(0)
emin , emax
E
H
i
ib
imax
b
Ibat
l
m
n
ng
nl
np
npv
ns
N
Pbat
pagg
bat
loss
Pbat
ploss
bat
+
−
Pbat,r
, Pbat,r
Pcell
branch flow matrix
linearized active and reactive power to
voltage matrix
matrix to describe polygonal regions for
active and reactive power
controllable generator to bus mapping matrix
PV generator to bus mapping matrix
battery capacity
battery capacity vector
battery degradation cost in e/kWh
feeder energy cost in e/MWh
generator energy cost vector
inverse charge recovery time in sec−1
parameter to describe the size of capacity
wells
state of energy of one battery system in
MWh
state of energy vector of ns battery systems
initial state of energy vector of ns battery
systems
storage allocation bounds from scheduler
state of energy evolution vector of ns battery systems
discrete battery control input matrix
nodal current vector in p.u.
branch current vector in p.u.
max branch current vector in p.u.
battery current in A
Luenberger observer gain
penalty factor for setpoint regions
number of buses
number of controllable generators
number of branches
number of triangles
number of PV units
number of battery systems
number of time steps in the scheduler horizon
active battery stack power
aggregated battery set point
total active power losses of one battery
system
total active power loss vector of ns battery
systems
active battery power range for linear power
loss approximation
active battery cell power
2
pd , q d
pgen , q gen
pv
pv
ppv
gen , q gen , p̂gen
s
ps,dis
gen ∈ pgen
s
ps,ch
gen ∈ pgen
ps,P
gen
ppl , pql
pld , q ld , p̂ld
pmin , pmax
pnet
R
smax
Sx, Su
T1 , T2
U
v
v min , v max
vs
Voc
wk ∈ W
xk ∈ X
xE
xE1
xE2
xset
X 0, X 1
agg
yset
z
zk ∈ Z
non-controllable nodal active and reactive
power load vectors
controllable active and reactive generator
power vectors
active and reactive power PV measurement
and prediction vectors
active discharging battery system grid
power vector
active charging battery system grid power
vector
supporting point vector for nonconvex battery system loss calculation
decision vectors of active network losses
active and reactive power load measurement and prediction vectors
min and max active generator power vectors
active feeder generator power
internal battery resistance
max apparent generator power vector
matrices to describe the storage evolution
sample times for scheduler and RT controller
decision vector for active battery power
nodal RMS voltage vector in p.u.
min and max nodal RMS voltage vectors
slack bus voltage vector
battery stack open-circuit voltage
box-constrained uncertainty set
decision vector for grid variables
dynamic state vector of one battery system
state of accessible energy well
state of non-accessible energy well
vector of helper decision variables to penalize deviations from the storage allocation band
battery system state vectors (initial and
after one time step) for ns battery systems
helper decision variable to penalize deviations from the aggregated battery set point
battery degradation variable for one battery
system
decision vector for battery degradation
variables
I. I NTRODUCTION
I
N the next few years, it can be expected that many
battery systems will be installed in the Low Voltage (LV)
distribution grid to cope with high in-feed from photovaltaics
(PV) [1] and other fluctuating energy sources also connected
to the distribution grid. In particular, battery systems can mitigate voltage violations and thermal line overloading, allowing
Distribution System Operators (DSOs) to defer line and transformer upgrades. Some recent papers [2], [3] developed decentralized battery control strategies to provide voltage support.
Decentralized control strategies have the advantage that they
rely only on local measurements. Thus, they do not require
communication infrastructure. However, addressing thermal
overloading requires coordination either via centralized or
distributed control, since the power flows are not observable
at a local level. In [4], a centralized predictive control scheme
is developed to avoid thermal line overloading, but the authors
do not consider voltage constraints. Furthermore, the proposed
control strategies in [2]–[4] do not consider battery degradation
or use detailed battery models.
Since investment costs for batteries are still high [5], a
battery’s expected lifetime greatly affects assessments of its
economic viability. Each control action applied to a battery
leads to charge capacity loss (i.e., degradation) [6], reducing
its lifetime. Some recent papers propose methods to include
battery degradation costs in economic cost functions [7], [8].
Another way to increase the economic viability of a battery is
to exploit its full capacity, enabling operation in low/high state
of charge regimes when the benefits outweigh the degradation
costs. However, simple battery models do not capture the
dynamics that are present when batteries operate in these
regimes. For example, operation in these regimes is only
possible at low charging/ discharging powers. It is possible to
model these dynamics with detailed battery models [9], [10].
The objective of this paper is to develop computationallytractable methods to control distributed batteries in distribution networks with high penetrations of PV to manage
network constraints such as voltage and thermal constraints.
To exploit the full potential of the network and to dispatch
the battery systems in an economic and efficient way, we
formulate multi-period AC Optimal Power Flow (AC-OPF)
problems. We use a linear approximation of the nonlinear
nonconvex AC-OPF from our previous work [11] that exploits
the radial structure of a LV network. The approximation
captures line losses and is linear in active and reactive power
injections. Linear approximations have been developed for
distribution networks [12], [13]; however, these neglect line
losses. The so-called DC power flow approximation and the
approximation in [14] are not applicable for distribution grids
since, in LV grids, active power flow is dominated by voltage
magnitude differences rather than voltage angle differences.
Nonlinear convex relaxations of the AC-OPF problem also
exist; however, the resulting semi definite [15], [16] and
second order cone [17], [18] programs are computationally
large and not suitable for computationally-limited controllers.
Linear programming approximations of the second order cone
programming problem have been developed, but the number
of the linear constraints is large [19]. Ref. [20] derives
an AC power flow approximation for distribution networks
that is linear between the complex voltage and the complex
power injections, but does not demonstrate how to use the
approximation within an Optimal Power Flow (OPF) problem.
Our contribution is the development of a control strategy
that leverages the linearized AC-OPF approximation to optimize distributed battery operation within a LV grid. We incorporate the linear AC-OPF into a two-stage Model Predictive
Control (MPC) control scheme that consists of a scheduler
and a real-time (RT) controller. The scheduler is a robust MPC
that solves a multi-period OPF minimizing battery degradation
and maximizing PV utilization subject to grid and storage
3
constraints. We use the concept of degradation maps allowing
us to describe the impact on degradation associated with
discrete control actions. By using a piecewise-affine (PWA)
representation of the map, we are able to use efficient Linear
Programming (LP) solvers. We link the planning domain with
the RT domain by computing storage allocation bounds that
are used by the RT controller, which solves an RT-OPF using
a detailed battery model and RT measurements of the network
and battery states to minimize the battery and network losses.
As a further contribution, we show that simple battery models
lead to high power errors within RT control applications.
This paper is organized as follows: Section II defines
the problem that we aim to solve. Section III presents the
battery models used in the proposed controllers. Section IV
reviews the linearized OPF formulation that is included in
our controllers. Section V describes the controller formulations
for the scheduler and RT controller. Section VI presents the
simulation results and a battery lifetime assessment. Finally,
Section VII presents the conclusions.
III. BATTERY M ODELING
In this section, we extend the linear battery models developed in our previous work [9]. The resulting models capture
the main relevant characteristics of battery systems and allow
SoE
Local Area Control
RT Control
Scheduler
(LAC)
Figure 1. Illustration of the problem environment. The overall control
objective is to maximize PV infeed while managing grid constraints and
minimizing battery degradation.
II. P ROBLEM D EFINITION
Existing inverter-level strategies to maximize PV infeed
include active power curtailment [21] and reactive power
control [22], which are used by DSOs in Germany [23]. In
contrast to OPF methods, no communication infrastructure is
needed. However, OPF methods enable optimal utilization of
the grid. We propose a method that combines distributionlevel OPF with a storage control strategy that could leverage
high bandwidth communication links already available at most
households.
Figure 1 illustrates the problem environment. Battery systems are used to mitigate voltage violations and thermal line
overloading that results from high PV infeed in LV networks,
enabling the DSO to defer equipment upgrades. The overall
control objective is to maximize PV infeed, while managing
grid constraints and minimizing battery degradation. PV maximization can be achieved by minimizing the net power pnet .
We assume local area control (LAC) entities compute optimal
allocation schedules for the storage devices in their area every
hour for a 24 hour horizon. We assume each area contains
a fraction of the total number of storage devices within
the full distribution network, for example, 10-200 devices
connected to a feeder. It uses the storage allocation schedule,
knowledge of each battery’s State of Energy (SoE), and RT
measurements of the active and reactive power consumption of
the loads (pld , qld ) and power production of the PV generators
pv
(ppv
gen , qgen ) to solve an RT-OPF to compute RT active and
s
reactive power battery setpoints (psgen , qgen
) every 10 seconds.
Figure 2 shows the relationships between the sections of the
paper. Different battery models are presented in Section III.
The linearized AC-OPF (Section IV) is incorporated into the
scheduler (Section V-B) and RT controller (Section V-C).
ppv
gen
pv
qgen
pld
qld
pnet
psgen
s
qgen
Li
fet
el
Mod Map Scheduler
c
i
s
(Sec.V-B)
Ba dation
ra
g
e
D
Battery
Modeling
(Sec.III)
im
eA
sse
ssm
en
t
Results
(Sec.VI)
Linearized
OPF (Sec.IV)
Ex
Det tende
aile d M
RT control
d L od
oss el
(Sec.V-C)
Mo
del
l
tro ce
n
o
c
an
RT rform
Pe
Figure 2. Relationships between the sections of the paper.
Pcell
ηbat
Ibat R
Voc
Pbat
ηin
ps,dis
gen
ps,ch
gen
Figure 3. Block diagram of a battery system consisting of a grid connected
inverter (right block) and battery stack (left block).
for tractable controller formulations. Note that these models
are used within our controllers, but not used to simulate
realistic battery systems. Recognizing that these models do
not capture all battery dynamics, we use the DUALFOIL
electrochemical battery model [24] to represent real batteries
in the RT simulations presented in Section VI.
A. Efficiency
In order to determine the cost optimal operation of multiple
battery systems, we first need a model that describes the
total efficiency of an individual battery system considering
both the battery and inverter losses. The models developed
in [9] only included the battery losses so we extend them to
include the inverter losses. Figure 3 shows a block diagram
4
Pcell
1
bat
ηbat
ηin
1
0.5
ηdis
ηch
0
−10
−5
Grid Power ps,ch
(kW)
gen
0
0
5
10
s,dis
Grid Power pgen (kW)
of a battery system. The crucial parameter that influences the
battery efficiency ηbat is the battery’s internal resistance R.
Using a Thevenin circuit equivalent, the battery power Pbat is
Pbat = Voc Ibat − RIbat 2
,
−
Pbat,r
(1)
where Voc is the open circuit voltage (OCV). The battery
efficiency for discharging currents Ibat > 0 is
1) Linear Basic Model: The first model is an integrator
model
Ė = −Pcell ,
(7)
where E denotes the time-varying SoE. As shown in Fig. 5,
Pcell is a nonlinear function of the battery power. To derive a
linear model, we linearize Pcell between 0 and rated powers
−
+
Pbat,r
, Pbat,r
by evaluating (5) at those points. By including
an averaged inverter efficiency η̄in , we obtain
s,dis
pgen
−
+
−1 −1
.
Ė ≈ [−ηbat (Pbat,r ) η̄in −ηbat (Pbat,r )η̄in ]
s,ch
pgen
{z
} |
{z
}
|
,
ch
ηbat
=
RIbat
Pcell
=1+
Pbat
Voc − RIbat
.
(3)
ch
For Voc >> RIbat , we can approximate ηbat
as follows:
ch
dis
ηbat
≈ ηbat
=1−
RIbat
Voc
.
|
(2)
where the internal cell power Pcell is referred to as the power
input. The battery efficiency for charging currents Ibat < 0 is
(4)
The battery efficiency can be expressed as a function of Pbat
by solving (1) for Ibat and substituting the resulting expression
into (4)
√
Voc − Voc 2 − 4RPbat
ηbat = 1 −
.
(5)
2Voc
Including the inverter efficiency ηin the total losses are
(
s,dis
−1
ηbat −1 ηin −1 − 1 ps,dis
gen = ηdis − 1 pgen
loss
Pbat =
s,ch
(ηbat ηin − 1)ps,ch
gen = (ηch − 1) pgen ,
(6)
s,ch
where ps,dis
gen ≥ 0 and pgen < 0 represent the discharging and
charging grid powers. The multiplication of ηbat and ηin leads
to a non-convex loss function, which is shown in Fig. 4 (blue
curve). The inverter efficiency is obtained from [25].
B. Dynamics
We also model fast battery dynamics. The models in [9]
expressed the state variable in terms of normalized charge,
and here we express it in terms of energy for convenience.
−ηch
−1
−ηdis
2
Voc Ibat − RIbat
RIbat
Pbat
=
=1−
Pcell
Voc Ibat
Voc
+
Pbat
Pbat,r
Figure 5. Linear approximation of the cell power Pcell with respect to the
battery power Pbat .
Figure 4. Battery system efficiencies and total power losses as a function of
the grid power. The plot is calculated for a 10kW battery system with R=
10mΩ,Voc = 42V, and a standard inverter efficiency curve from [25].
dis
ηbat
=
discharging
−1
Pcell = ηbat
Pbat
charging
Pcell = ηbat Pbat
Ploss
Efficiency η
Power Losses (kW)
2
{z
bT
bat
}
(8)
2) Linear Extended Model: The linear extended model
allows us to capture the rate capacity effect [26] or charge
relaxation effect [27]. This effect accounts for the ion diffusion
in the electrolyte and reduces the accessible battery capacity
for high battery powers. As we have presented in [9] we
capture this effect by modifying the KiBaM model [28]. The
model is
"
#
T s,dis
cr
− ccwr
pgen
bbat
1−cw
ẋE =
x
+
E
cr − cr
s,ch
pgen
0
cw
1−cw
|
{z
}
A
E
= xE1 + xE2
,
T
(9)
where xE = [xE1 , xE2 ] denotes the state vector for two
virtual wells that are interconnected. Energy can only be
withdrawn from the available well xE1 . The parameter cw
determines the size of the wells and cr is the inverse charge
recovery time.
3) Comparison of models: Figure 6 demonstrates the merit
of the linear extended model as compared to the linear basic
model, which does not capture the rate capacity effect. The
figure shows the relative error between the maximum power
that could be applied to the system as modeled with the
linear basic model as compared the linear extended model for
different states of energy and as a function of the sample time
used within discrete versions of the models. The linear basic
model overestimates the maximum power available at high/low
SoE regimes for short sample times (seconds to minutes);
however, it is as accurate as the linear extended model for
sample times greater than or equal to 45 min. This means
that for RT control applications, the linear extended model is
significantly more accurate than the linear basic model.
5
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Degradation (kWh/h)
State of Energy SoE (-)
0.8
Relative power error
1
1
Identified Degradation Map
PWA Convex Hull
0.01
0.005
0
10
-10
0
0
Power (kW)
0
0
15
30
45
Sample Time T (min)
60
Figure 6. Relative battery power error of the linear basic model as compared
to the linear extended model as a function of the sample time.
0
Figure 7. Illustration of the degradation map for a 10 kWh battery system
(cE = 10kWh). The red surface is the PWA convex hull (11) of the identified
map (blue surface).
Table I
DUALFOIL C ONFIGURATION
C. Degradation
Charge capacity loss in lithium ion (Li-ion) batteries is a
slow process and hard to model. Among the contributors to
capacity loss are two chemical side reactions that irreversibly
transform cyclable ions into solids during battery operation. In
[9], we presented a method to identify a stationary degradation
process on an arbitrary battery usage pattern. The method
produces a degradation map, where degradation is a function
of the battery power and SoE. Calendar life degradation is also
included in the degradation map. Specifically, when the battery
power is zero (i.e., the battery is resting), the degradation map
in Fig. 7 shows non-zero degradation, which is a function of
the SoE.
Unfortunately, degradation maps are in general nonconvex
[29], such that we cannot apply efficient convex solvers in
our optimization framework. As an extension to the work in
[9], we compute the convex hull of the degradation map from
[9] using Delaunay triangulation [30]. Figure 7 shows the
convex hull (red surface) of the identified degradation map
(blue surface), assuming a LiCoO2 graphite cell modeled with
the DUALFOIL electrochemical battery model [24] configured
as in Table I. The operating conditions used to determine the
map are detailed in [9]. The rate constants rsc and rsa are
tuned in such a way that after 4000 full cycles 80% of the
initial battery capacity remains. The map in [9] is normalized
with respect to the energy capacity cE allowing us to scale the
map for any battery size. By evaluating the plane parameters
a1 , a2 , a3 ∈ Rnp ×1 of np triangles from the convex hull, we
can define the following PWA mapping for the normalized
degradation rate z/cE :
z
Pbat
E
= max a1
+ a2
+ a3
cE
cE
cE
10
5
Energy (kWh)
.
(10)
By multiplying (10) by cE , we can express the degradation z
in absolute terms for any battery with energy capacity cE as
Pbat
z = max [a1 a2 a3 ] E .
(11)
cE
ambient conditions
cathode
anode
electrolyte
Acell
isothermal
LiCoO2
Graphite
LiPF6
1 m2
rsc
rsa
off
Vcut
1e-11
1e-11
3.0 – 4.2 V
cell area corresponds to 30 Ah
for the cut-off voltage window
side reaction rate constant cathode
side reaction rate constant anode
cut-off voltage window
IV. O PTIMAL P OWER F LOW FOR L OW VOLTAGE G RIDS
In our previous work [10], [11], we derived a linearized OPF
formulation. Based on the Forward Backward Sweep (FBS)
power flow method from [31], we linearly approximated voltage, power losses, and branch flow limits as linear functions of
the nodal reactive and active power injections. The following
assumptions were made.
• We assume three-phase balanced radial LV networks. It is
common to assume three-phase balanced LV networks in
Europe [32] since houses and PV inverters are connected
to three phases; however, this assumption is less valid
for North America. However, many studies make this
assumption e.g., [16], [33]–[35]. The power flow equations we use could also be extended to handle unbalanced
networks [31] and can be incorporated in the proposed
control framework. For example, [33] describes how one
might extend a related power flow method.
• We assume that LV networks have high resistance over
reactance ratios (R/X ≥ 2) [32].
• We assume that voltage angle differences in LV networks
are small (≤ 10◦ ).
We summarize the formulation here; for full details see [11].
First, we give the generic formulation, before we specialize
it to our scenario
in Section V. We
T define following decision
vector x = pgen , q gen , ppl , pql , v , where pgen ∈ Rng ×1 and
q gen ∈ Rng ×1 are the active and reactive generator powers,
where ng is the number of generators; ppl ∈ Rnl ×1 and pql ∈
Rnl ×1 are the approximate active power losses due to active
and reactive power injections (defined in [11]), where nl is
the number of lines; and v ∈ Rn×1 is the voltage vector,
where n is the number of buses. Assuming positive linear
6
pgen
pgen
smax
smax
φ
φ
qgen
qgen
a)
b)
Figure 8. Approximated reactive power capability areas a) circular-bounded
b) cos φ-bounded. The polygonal convex regions can be described with the
constraints (12h-l). [10]
costs cp for the active generator powers, the OPF problem
can be approximated as the following LP problem:
J ∗ = min cTp pgen
x
1T Cg pgen − 1Tppl − 1T pql = 1T pd
C g pgen
pd
(b) B v
− v = Bv
− vs
C g q gen
qd
(c) B 1l ppl − B 2l C g pgen ≥ B 2l pd + bl
(d) B 1l pql − B 2l C g q gen ≥ B 2l q d + bl
+ B r pd
+ B r pd ≤ B r C g pgen ≤ imax
(e) −imax
b
b
(f) v min ≤ v ≤ v max
(g) pmin ≤ pgen ≤ pmax
(h) pgen + Aq q gen ≤ smax
(i) pgen − Aq q gen ≤ smax
(j) pgen + Aq q gen ≥ −smax
(k) pgen − Aq q gen ≥ −smax
(l) −B q smax ≤ q gen ≤ B q smax ,
(12)
where pd ∈ Rn×1 and q d ∈ Rn×1 are the active and reactive
net load; C g ∈ Rn×ng maps generators to buses; and B v ,
B r , Aq , and B q are matrix parameters defined in [11] and
[10]. Constraint (12a) enforces power balance in the grid.
The voltage approximation is included in (12b), where each
element of v s ∈ Rnl ×1 is the slack bus voltage. The constraints
(12c,d) incorporate epigraph formulations that are piecewise
linear inner approximations of the power losses. We specify
the matrix and vector parameters B 1l , B 2l , and bl to obtain a
more compact form of the inequalities defined in [11]. Constraint (12e) includes branch flow limits, where ib ∈ Rnl ×1
are the branch flow currents. Constraints (12f,g) specify the
lower and upper bounds for the voltage (v min , v max ) and
active generator powers (pmin , pmax ). Constraints (12h-l) approximate the generators’ apparent power limits, where smax
is the generators’ maximum apparent power; specifically, we
define circular-bounded and cos φ-bounded active and reactive
power settings by approximating the circular area/segments
with convex sets [10] that describe the polygons depicted in
Fig 8.
s.t.
(a)
V. O PTIMAL BATTERY O PERATION
A. Control Structure
As shown in Fig. 9 our proposed control scheme consists
of two control entities working at different time scales. We
Local Area Control
Forecast
emin , emax
PV, Load
Scheduler
RT
pagg
bat
T1 = 1h
Control
p , q T2 = 10 s
d
e
d
Figure 9. Centralized control scheme consisting of a scheduler and an RT
controller. The scheduler uses robust MPC. The RT controller solves a RTOPF using the storage allocation from the scheduler and RT measurements.
do not allow for PV curtailment and so the scheduler has
to calculate future storage allocation bounds in terms of
energy emin , emax and aggregated power pagg
bat that are feasible
under the worst case PV forecast scenarios. The RT controller
tries to control the storage units within these bounds, while
minimizing network and storage losses and taking into account
RT measurements of the net load.
B. Scheduler
We solve a robust multi-period OPF subject to the worst
case PV predictions. Since we solve the multi-period problems
in a receding horizon fashion every hour with updated energy
levels, the scheduler acts as a robust MPC.
1) Incorporation of Distributed Storage: Due to its high
sample time T1 = 1 hour, the scheduler cannot capture the
detailed battery dynamics so we use the linear basic model (8).
The discrete version of (8) for ns battery systems is
e(k + 1) = I ns e(k) + T1 diag{−η −1
dis } diag{−η ch }
|
{z
}
B
ps,dis
gen
,
ps,ch
gen
| {z }
psgen
(13)
ns ×1
ns ×1
s,ch
<
0
∈
R
,
η
,
η
≥
0
∈
R
,
p
where ps,dis
dis
ch ∈
gen
gen
Rns ×1 are the total discharging and charging efficiencies of the
storage units, and I ns denotes the identity matrix of dimension
ns . The complete SoE evolution E = [e(1), ..., e(N − 1)]T
for a time horizon of length N is
ns
s
pgen (0)
I
B
0
..
E = ... e(0)+ ... . . .
.
,
|
I ns
{z }
Sx
B
|
···
{z
Su
B
}|
psgen (N − 1)
{z
}
U
(14)
where e(0) is the initial SoE vector.
2) Incorporation of Battery Degradation: In Section III-C
we calculated the PWA convex hull (11) of the degradation
map. To incorporate degradation into the planning problem,
we define helper decision variables Z = [z(0), . . . , z(N −
1) ∈ Rns ×1 ]T and include the epigraphs of (11) as constraints
in our problem. Specifically, we substitute E from (14) into
an extended version of (11) that includes ns battery systems
with energy capacities cE ∈ Rns ×1 and obtain a degradation
equation in terms of the control variable U as follows:
(A1 + A2 S u ) U + Az Z ≤ −A2 S x e0 − A3 cE
,
(15)
7
where A1 = [I N ns ⊗ a1 ][I N ⊗ [I N ns I N ns ]], A2 = [I N ns ⊗
a2 ], A3 = I N ⊗ [I ns ⊗ a3 ], and Az = [I N ns ⊗ −1np ×1 ].
The operator ⊗ denotes the Kronecker product.
3) Robust Scheduling: To solve a multi-period OPF we first
incorporate a temporally-coupled sequence of single period
OPF problems (12). We extend the decision vector to X =
[x0 , · · · , xN −1 ]T , where pgen,k = [pnet (k) psgen (k)]T ∈ xk
and pnet (k) (i.e., the net power into or out of the LAC) is
treated as a slack generator. Since we do not allow for PV
curtailment, we set
T
pd,k = −C pv (p̂pv
gen,k + w k 3σ pv,k ) + p̂ld,k
,
(16)
RT Control
emin , emax
pagg
bat
pd , q d
s,1
ps,1
e1
gen , qgen
Battery 1
..
MPC
.
s,n
s,n
Control pgen , qgen
en
Battery n
x1
Observer 1
..
.
xn
Observer n
Figure 10. Block diagram of the RT controller.
p̂pv
gen,k
where the PV and load forecasts are denoted by
and
p̂ld,k , and C pv ∈ Rn×npv maps the npv PV units to the
buses. The standard deviation σ pv,k of the PV forecast error
could be determined by using the results from [36]. We
define W = [w0 , . . . , wN −1 ]T , which is a box-constrained
uncertainty set where −1 ≤ W ≤ 1. Note that we do not
consider load forecast uncertainty, since absolute load forecast
errors are generally much smaller than absolute PV forecast
errors for a grid with high PV penetration. The scheduling
problem is
!
N
−1
X
T
min T1
cn pnet (k) + cd 1 z(k)
X,Z
k=0
(17)
s.t. (a) −S x e(0) ≤ S u U ≤ cE − S x e(0)
∀xk : max (12a-c,e), (12d,g-l),(15) ,
−1≤W ≤1
where cn and cd are cost parameters for the net power and the
battery degradation. With this formulation, the PV infeed will
be maximized since the optimizer minimizes pnet . The constraint set (17a) operates the storage devices at their minimum
and maximum bounds. To retrieve the robust counterpart of
(17), we eliminate wk and use |σ pv,k | within each constraint
[37], since X does not multiply the random variable [38].
4) Determination of Robust Storage Level Bounds: To link
the planning domain with the RT domain, the scheduler sends
robust allocation bounds for the RT controller
emin
=
min(e(0), e(1))
(18)
emax
=
max(e(0), e(1)) ,
(19)
where we evaluate e(1) with (14). The scheduler also sends
the aggregated battery setpoint pagg
bat to give the RT controller
the ability to operate the batteries at their individual optimal
operation points
T ∗s
pagg
bat = 1 pgen (1)
,
(20)
where p∗s
gen (1) is a portion of the optimal decision vector of
the robust problem (17).
C. Real-Time Control
As depicted in Fig. 10, the RT controller receives the storage allocation bounds and aggregated battery power setpoint.
pv
Using RT measurements of the PV injections ppv
gen , q gen and
loads pld , q ld to compute pd , q d , it sets the reactive and active
battery powers psgen , q sgen by solving an RT-OPF. In particular,
we solve an MPC problem for one step incorporating the
linear extended battery model (9), allowing us to estimate
the power available from each individual battery system and
utilize its full capacity potential. The decision vector is u0 =
agg
T
loss
[psgen q sgen λ ploss
bat yset xset ] , where λ and pbat are helper
variables to incorporate battery losses into the problem and
agg
yset
and xset are helper variables to penalize deviations from
the storage allocation bounds. The RT-OPF problem is
min
u0
1T pgen + 1T ploss
+
| {zbat}
| {z }
network losses
s.t.
battery losses
(12a) − (12l)
(a) X 1 =
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
(m)
agg
yset
+ 1T xset
{z
}
|
deviations from setpoint regions
ΦX 0 + Hpsgen
cw
0
≤ X1 ≤ 1
⊗
1 − cw
ns
− diag{m}CX 1 − I xset ≤ −emin m
diag{m}CX 1 − I ns xset ≤ emax m
I ns xset ≤ 0
agg
≤ −mpagg
−m1T psgen − yset
bat
agg
−yset ≤ 0
pd = −C pv ppv
gen + pld
q d = −C pv qpv
gen + q ld
−I ns psgen + I ns ⊗ ps,P
gen λ = 0
ns loss
+ I ns ⊗ f (ps,P
−I
gen ) λ = 0
ns pbat1×n
s
λ=1
I ⊗1
λ = [λ1 , . . . , λns ]T ∀λi : SOS2 ,
ns ×1
(21)
where (21a) is the discrete version of (9) for multiple battery
systems. Specifically, X 0 is the initial battery system state
and X 1 is the state after one time step; Φ is the battery
system dynamics matrix, H is the RT control input matrix,
and C maps the internal states of each battery system to the
SoE. Constraint (21b) avoids overspill of the battery capacity
wells. Constraints (21c-e) define two regions and enable us to
penalize deviations outside of the storage allocation bounds
emin , emax with the penalty factor m, where diag{m} is
a square matrix with m on the diagonal. We only penalize
the aggregated power setpoint in the discharging direction
with (21f,g), since overcharging can be dealt with in RT as
it will not result in line limit violations. This yields lower
storage utilization and thus results in lower battery degradation. Constraints (21h,i) incorporate the RT measurements
and (21j-m) represent nonconvex piecewise linear functions
of the battery losses using an efficient Special Ordered Set
(SOS) formulation [39]. In particular, we discretize (6) with
8
R14
Table II
S IMULATION PARAMETERS .
R15
R13
R12
R6
R3
R0
R1
R2
R4
R11
Storage units
Storage parameters
R17
R5
R10
R7
R16
18
s,max
= 10 kVar,
=10 kW, qgen
ps,max
gen
cE =20 kWh, ηdis = 0.91, ηch = 0.91
Voc = 42 V, R = 10 mΩ
efficiency curve from [25]
convexified degradation map from [9]
convexified degradation map from [29]
24 @ 1h sample time
100 e/MWh
400 e/kWh
P = 10
m = 1000
18
= 20 kW (λ = 1)
ppv,max
gen
1 year
18 @ 4kW (randomized profile)
European LV network [32]
vmax = 1.1,vmin = 0.9
R8
Inverter
Deg. model LiCoO2
Deg. model LiFePO4
Scheduler horizon N
Net energy cost cn
Degradation cost cd
Supporting points
Penalization factor
PV units
PV power
Simulation horizon
Households
Grid
Voltage Limits
R9
R18
P supporting points ps,P
gen and evaluate (6) for each point and
for each battery system. The sum of the elements in the set λi
has to be equal to 1, which is enforced in (21l), and at most two
elements in the set have to be consecutively non-zero (21m).
This is referred to as an SOS2 set. Constraint (21m) makes
the problem a Mixed Integer Linear Programming (MILP)
problem.
We also implemented an observer to estimate the internal
states xE of each individual battery system, since battery
management systems typically provide only the SoE. Since
the system (9) is detectable, [40] we can design following
Luenberger observer
1
Pbat
˙x̂E = (A − l[1 1])x̂E +
l
, (22)
0
E
Energy Level (MWh)
Figure 11. Modified Cigre test grid [32] with PV generators (20kW) and
battery storage units (10kVA/20kWh).
0.02
0.01
0
25
5
6
20
15
7
10
where l ∈ R2×1 is chosen such that (A − l[1 1]) is stable.
5
8
0
9
Storage ID
Time (h)
VI. S IMULATION R ESULTS
Through case studies we demonstrate the performance of
our centralized control scheme. The parameters are listed
in Table II. Figure 11 shows the units within the test grid.
We assess each control stage with simulations on different
timescales. To study the RT controller, we simulate its performance on a sunny day and use the DUALFOIL electrochemical battery model [41] in place of a real battery system. We
assume perfectly balanced cells, such that we can scale the
model to emulate any required power and energy capacity. To
study the impact of the scheduler on battery lifetimes, we need
to consider a long time horizon. However, to reduce simulation
times, we omit the RT stage and run our scheduler with the
simplified battery model and model the capacity loss with the
convexified degradation maps. To assess battery lifetimes, we
compute the capacity loss from the original degradation maps.
A. RT control performance
In Fig. 12 we show the RT evolution of the battery energy
levels. The gray boxes are the allocation bounds emin , emax
from the scheduler. The scheduler uses the linear basic model,
which does not include the rate capacity effect. However, the
RT controller uses the linear extended model and so, in high
state of energy regimes, it reduces the battery power to make
Figure 12. RT evolution of the battery energy levels. The gray boxes are the
storage allocation bounds from the scheduler, in which the RT controller can
operate. For high state of energy regimes, the RT controller has to reduce the
battery power to make the full capacity available.
the full capacity available, corresponding to a slower rise in
energy levels.
To reduce battery system losses, the RT controller dispatches the batteries to loss-optimal operating points, while
keeping them within the allocation bounds. This can be seen
in Fig. 13, where the RT controller sets the battery powers in
a rolling fashion, resulting in switched battery operation.
Figure 14 shows compliance with the voltage and line flow
constraints. Since we use a linear approximation of the power
flow in our controllers, we calculate the exact power flows
by running the full nonlinear AC power flow with all power
injection variables set to the setpoints calculated by the RT
controller. As already shown in [11], the linear approximation
leads to a solution that is always feasible. This we can see
from Fig. 14, since the maximum line capacity is not fully
utilized and the allowable voltage band (0.9 − 1.1 pu) is not
fully used. The mean absolute errors (MAEs) in voltage magnitudes/angles between the approximate power flow equations
and the AC power flow equations are 1.06 × 10−3 pu and
5.94 × 10−3 ◦ .
To illustrate the merit of the detailed power loss model,
9
0
Power (MW)
0.2
0.1 −0.05
0.5
1
1.5
2
0
−0.1
−0.2
−0.3
−0.4
5
10
15
Time (h)
20
pagg
bat
time (h)
0
0.25
1.05
1
0
5
10
15
20
25
Time (h)
line 1−2
line 2−3
line 3−4
line 6−7
line 4−12
line 14−15
0.5
0
−0.5
−1
0
5
10
15
20
25
10
15
20
25
0.025
0.02
0.15
0.015
0.1
0.01
0.05
0.005
0
RT−LP Method
RT−MILP Method
0
Figure 15. Battery and network losses for a sunny day. The bars show the
energy losses as a function of the RT control configuration. The curves show
power loss over the day for the different RT control configurations. With the
detailed loss model (RT-MILP) the system incurs almost 30% fewer losses.
Table III
L OSSES FOR DIFFERENT RT CONTROL CONFIGURATIONS
RT control
configuration
RT-MILP
RT-LP
Normalized Total Capacity
Voltage magnitude |v| (pu)
Branch flow sb/sb,max
node 0
node 6
node 12
node 15
node 16
node 17
node 18
5
0.2
Figure 13. RT dispatch for a one day simulation with a snapshot of the
battery discharging phase. The blue curve shows the aggregated battery power
signal from the scheduler. To reduce battery losses, the batteries are discharged
(positive powers) and charged (negative powers) in a switched way.
1.1
network losses
battery losses
RT−LP net losses
RT−LP bat losses
RT−MILP net losses
RT−MILP bat losses
Power losses (MW)
0.05
0.3
Energy Losses (MWh)
bat 1
bat 2
bat 3
bat 4
bat 5
bat 6
bat 7
bat 8
bat 9
bat 10
bat 11
bat 12
bat 13
bat 14
bat 15
bat 16
bat 17
bat 18
net
load
PV
0.4
battery
loss model
nonconvex
linear
network losses
(kWh)
141.8
141.2
battery losses
(kWh)
47.5
68.4
1
0.8
0.6
0.4
LiFePO4 w deg model
LiFePO4 w/o deg model
LiCoO2 w deg model
LiCoO2 w/o deg model
0.2
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan
Time (month)
Figure 16. Normalized total battery capacity over one year for different
battery technologies, with (w) and without (w/o) a degradation (deg) model.
Time (h)
Figure 14. Voltage magnitudes and normalized branch power flows for a one
day simulation with high PV infeed. The allowable voltage band is v min =
0.9 to v max = 1.1 pu.
we simulate the RT controller using the full model in (21)
(referred to as the RT-MILP configuration) and using a simpler
linear loss model, resulting in an LP problem (referred to as
the RT-LP configuration). As shown in Fig. 15 the battery
losses are 30% smaller, if we use the RT-MILP configuration.
Table III quantifies this finding. The loss reduction is mainly
due to the reduced battery losses. Network losses are almost
identical in both cases and are approximately two third of the
total losses. Table IV shows the computation time of each
stage. We use the CPLEX [42] optimization solver running
on an Intel Core i7 computer. The computation time for the
MILP problem is less than one second, which is reasonable
for the RT control stage. The MILP computation time for 200
devices is approximately 4 seconds and so the approach could
be implemented in real-time.
B. Lifetime Assessment
We next determine how the scheduler affects battery lifetimes, and compare two different battery technologies: LiCoO2 and LiFePO4. For LiCoO2, we take the degradation map
from [9]. For LiFePO4, we discretize the analytic degradation
function presented in [29] to obtain a map. Both maps are
convexified as described in Section III-C and are included
in the scheduler. We also simulate the scheduler without
using the degradation model, but constrain the minimum SoE
level to 5% of the capacity. In Fig. 16, we compare the
charge capacity loss over one year for the different battery
technologies. The scheduler without the degradation model
operates the batteries in low SoE regimes. Those regimes
result in high degradation, especially for the LiCoO2 system.
In Fig. 17, we also show the charge capacity loss for each
battery. Using degradation models, the scheduler balances
degradation across the batteries, contributing to longer battery
lifetimes (see Table V). We also extrapolated these results to
10
Table IV
P ROBLEM SIZES AND COMPUTATION TIMES FOR THE CONTROLLERS
problem class
number of constraints
number of decision variables
computation time ±1σ (sec)
Scheduler
LP
14328
4392
2.85 ± 0.26
LiFePO4 w deg model
LiFePO4 w/o deg model
LiCoO2 w deg model
LiCoO2 w/o deg model
initial capacity
0.025
Battery Capacity (MWh)
RT Controller
MILP (18 SOSs)
381
327
0.78 ± 0.12
0.02
0.015
0.01
0.005
ba
t
ba 1
t
ba 2
t
ba 3
t
ba 4
t
ba 5
t
ba 6
t
ba 7
t
ba 8
ba t 9
t
ba 10
t
ba 11
t
ba 12
t
ba 13
t
ba 14
t
ba 15
t
ba 16
t
ba 17
t1
8
0
Figure 17. Remaining battery capacities for each battery (bat) after a one
year simulation for different battery technologies, with (w) and without (w/o)
a degradation (deg) model. The blue line shows the initial battery capacities.
obtain estimates for the number of full cycles and lifetimes
assuming an end of life (EOL) criterion of 0.8, which means
that 80% of the initial capacity remains. We find that by using
a degradation model, we can prolong the lifetime by a factor
10 for the LiCoO2 system and by a factor 2 for the LiFePO4.
VII. CONCLUSION
In this paper, we present a novel two-stage centralized MPC
scheme for distributed battery storage to mitigate voltage and
line flow violations that are induced by high PV penetrations
in LV grids. To link the planning and real time domains,
our control scheme consists of a robust scheduler and a RT
controller. This division enables planning using course PV
and load forecasts on timescales of hours and RT control to
manage dynamics and forecast error on timescales of seconds.
Therefore, the scheduler uses simple battery models, while
the RT controller uses detailed models. We incorporate a
linearized AC-OPF into both the scheduler and RT controller
to reduce the computational complexity of the algorithms.
To guarantee secure grid operation, the scheduler solves a
robust multi-period OPF taking the worst-case PV forecast
into account. It produces robust feasible storage allocation
bounds for the RT controller, which maximizes PV utilization,
while keeping battery degradation to a minimum by solving
a single step OPF problem. We find that the control scheme
can substantially reduce both battery losses and degradation.
R EFERENCES
[1] “Battery energy storage for smart grid applications,” Eurobat, Tech.
Rep., 2013. [Online]. Available: https://http://www.eurobat.org/sites/
default/files/eurobat smartgrid publication may 2013 0.pdf
Table V
L IFETIME ASSESSMENT FOR DIFFERENT BATTERY TECHNOLOGIES
Battery
Technology
LiFePO4
LiCoO2
Configuration
w deg model
w/o deg model
w deg model
w/o deg model
Expected number
of full cycles
2816
1609
1652
166
Expected lifetime
(years)
4.4
2.64
2.63
0.29
[2] F. Marra, G. Yang, C. Træholt, J. Østergaard, and E. Larsen, “A
decentralized storage strategy for residential feeders with photovoltaics,”
IEEE Transactions on Smart Grid, vol. 5, no. 2, pp. 974–981, 2014.
[3] K. H. Chua, Y. S. Lim, P. Taylor, S. Morris, and J. Wong, “Energy
storage system for mitigating voltage unbalance on low-voltage networks
with photovoltaic systems,” IEEE Transactions on Power Delivery,
vol. 27, no. 4, pp. 1783–1790, 2012.
[4] T. Wang, H. Kamath, and S. Willard, “Control and optimization of gridtied photovoltaic storage systems using model predictive control,” IEEE
Transactions on Smart Grid, vol. 5, no. 2, pp. 1010–1017, 2014.
[5] B. Nykvist and M. n. Nilsson, “Rapidly falling costs of battery packs for
electric vehicles,” Nature Climate Change, vol. 5, no. 4, pp. 329–332,
2015.
[6] S. J. Moura, J. C. Forman, S. Bashash, J. L. Stein, and H. K. Fathy,
“Optimal control of film growth in lithium-ion battery packs via relay
switches,” IEEE Transactions on Industrial Electronics, vol. 58, no. 8,
pp. 3555–3566, 2011.
[7] M. Koller, T. Borsche, A. Ulbig, and G. Andersson, “Defining a
degradation cost function for optimal control of a battery energy storage
system,” in Proceedings of POWERTECH, 2013, pp. 1–6.
[8] A. Trippe, R. Arunachala, T. Massier, J. Jossen, and T. Hamacher,
“Charging optimization of battery electric vehicles including cycle
battery aging,” in Proceedings of ISGT EUROPE, Oct 2014.
[9] P. Fortenbacher, J. Mathieu, and G. Andersson, “Modeling, identification, and optimal control of batteries for power system applications,” in
Proceedings of the Power Systems Computation Conference, 2014.
[10] ——, “Optimal real-time control of multiple battery sets for power
system applications,” in Proceedings of POWERTECH, 2015.
[11] P. Fortenbacher, M. Zellner, and G. Andersson, “Optimal sizing and
placement of distributed storage in low voltage networks,” in 19th
Power Systems Computation Conference (PSCC), Genoa, Italy, 2016.
[Online]. Available: http://arxiv.org/abs/1512.01218
[12] L. Gan and S. H. Low, “Convex relaxations and linear approximation
for optimal power flow in multiphase radial networks,” Proceedings 2014 Power Systems Computation Conference, PSCC 2014, 2014.
[13] M. E. Baran and F. F. Wu, “Optimal sizing of capacitors placed on
a radial distribution system.” IEEE Transactions on Power Delivery,
vol. 4, no. 1, pp. 735–743, 1989.
[14] J. A. Taylor and F. S. Hover, “Linear relaxations for transmission system
planning,” IEEE Transactions on Power Systems, vol. 26, no. 4, pp.
2533–2538, 2011.
[15] D. K. Molzahn and I. A. Hiskens, “Moment-based relaxation of the
optimal power flow problem,” in Proceedings of Power Systems Computation Conference, 18-22 Aug. 2014.
[16] E. Dall’Anese, S. V. Dhople, and G. B. Giannakis, “Optimal dispatch
of photovoltaic inverters in residential distribution systems,” IEEE
Transactions on Sustainable Energy, vol. 5, no. 2, pp. 487–497, 2014.
[17] S. H. Low, “Convex relaxation of optimal power flow - part I: Formulations and equivalence,” IEEE Transactions on Control of Network
Systems, vol. 1, no. 1, pp. 15–27, March 2014.
[18] K. Christakou, D.-C. Tomozei, J.-Y. L. Boudec, and M. Paolone,
“AC OPF in radial distribution networks - parts i,ii,” 2015. [Online].
Available: http://arxiv.org/abs/1503.06809
[19] S. Mhanna, G. Verbic, and A. Chapman, “Tight LP approximations
for the optimal power flow problem,” 2016. [Online]. Available:
http://arxiv.org/abs/1603.00773
[20] S. Bolognani and S. Zampieri, “On the existence and linear approximation of the power flow solution in power distribution networks,” IEEE
Transactions on Power Systems, vol. 31, no. 1, pp. 163–172, 2016.
[21] R. Tonkoski, L. A. Lopes, and T. H. El-Fouly, “Coordinated active power
curtailment of grid connected PV inverters for overvoltage prevention,”
IEEE Trans on Sustainable Energy, vol. 2, no. 2, pp. 139–147, 2011.
[22] G. Kerber, R. Witzmann, and H. Sappl, “Voltage limitation by autonomous reactive power control of grid connected photovoltaic inverters,” in Compatibility and Power Electronics., May 2009, pp. 129–133.
11
[23] “VDE-AR-N 4105 power generation systems connected to the lowvoltage distribution network - technical minimum requirements for
the connection to and parallel operation with low-voltage distribution
networks,” VDE Application Rule, 2011.
[24] T. F. Fuller, M. Doyle, and J. Newman, “Simulation and optimization
of the dual lithium ion insertion cell,” Journal of The Electrochemical
Society, vol. 141, no. 1, pp. 1–10, 1994.
[25] “Efficiency and derating sunny boy / sunny tripower / sunny
mini central,” SMA, Tech. Rep., 2015. [Online]. Available: http:
//files.sma.de/dl/1348/WirkungDerat-TI-en-40.pdf
[26] M. Doyle and J. Newman, “Analysis of capacity-rate data for lithium
batteries using simplified models of the discharge process,” Journal of
Applied Electrochemistry, vol. 27, no. 7, pp. 846–856, 1997.
[27] T. F. Fuller, “Relaxation phenomena in lithium-ion-insertion cells,”
Journal of The Electrochemical Society, vol. 141, no. 4, p. 982, 1994.
[28] J. F. Manwell and J. G. McGowan, “Lead acid battery storage model
for hybrid energy systems,” Solar Energy, vol. 50, no. 5, pp. 399 – 405,
1993.
[29] J. C. Forman, S. J. Moura, J. L. Stein, and H. K. Fathy, “Optimal
experimental design for modeling battery degradation,” in Proceedings
of ASME Dynamic Systems and Control Conference, vol. 1, 2012, pp.
309–318.
[30] D. T. Lee and B. J. Schachter, “Two algorithms for constructing a delaunay triangulation,” International Journal of Computer & Information
Sciences, vol. 9, no. 3, pp. 219–242, 1980.
[31] J.-H. Teng, “A direct approach for distribution system load flow solutions,” IEEE Transactions on Power Delivery, vol. 18, no. 3, pp. 882–
887, July 2003.
[32] “Benchmark systems for network integration of renewable and distributed energy resources,” Cigre Task Force C6.04.02, Tech. Rep., 2014.
[Online]. Available: http://c6.cigre.org/Publications/Technical-Brochures
[33] E. Dall’Anese, S. V. Dhople, and G. B. Giannakis, “Photovoltaic
inverter controllers seeking AC optimal power flow solutions,” IEEE
Transactions on Power Systems, vol. 31, no. 4, pp. 2809–2823, 2016.
[34] L. Reyes-Chamorro, A. Bernstein, J.-Y. L. Boudec, and M. Paolone, “A
composable method for real-time control of active distribution networks
with explicit power setpoints. part ii: Implementation and validation,”
Electric Power Systems Research, vol. 125, pp. 265 – 280, 2015.
[35] F. Olivier, P. Aristidou, D. Ernst, and T. V. Cutsem, “Active management
of low-voltage networks for mitigating overvoltages due to photovoltaic
units,” IEEE Transactions on Smart Grid, vol. 7, no. 2, pp. 926–936,
March 2016.
[36] P. Bacher, H. Madsen, and H. A. Nielsen, “Online short-term solar power
forecasting,” Solar Energy, vol. 83, no. 10, pp. 1772 – 1783, 2009.
[37] A. Ben-Tal, L. El Ghaoui, and A. Nemirovski, Robust Optimization.
Princeton University Press, 2009.
[38] J. Löfberg, “Automatic robust convex programming,” Optimization
Methods and Software, vol. 27, no. 1, pp. 115–129, 2012.
[39] J. A. Tomlin, “Special ordered sets and an application to gas supply
operations planning,” Mathematical Programming, vol. 42, no. 1, pp.
69–84, 1988.
[40] J. O’Reilly, Observers for Linear Systems. Elsevier Science, 1983.
[41] M. Doyle, T. F. Fuller, and J. Newman, “Modeling of galvanostatic
charge and discharge of the lithium/polymer/insertion cell,” Journal of
The Electrochemical Society, vol. 140, no. 6, pp. 1526–1533, 1993.
[42] “IBM ILOG CPLEX user’s manual,” IBM, Tech. Rep., 2015.
| 3cs.SY
|
Edge-Caching Wireless Networks: Performance
Analysis and Optimization
Thang X. Vu, Member, IEEE, Symeon Chatzinotas, Senior Member, IEEE, and
arXiv:1705.05590v3 [cs.IT] 6 Feb 2018
Bjorn Ottersten, Fellow, IEEE
Abstract
Edge-caching has received much attention as an efficient technique to reduce delivery latency and
network congestion during peak-traffic times by bringing data closer to end users. Existing works usually
design caching algorithms separately from physical layer design. In this paper, we analyse edge-caching
wireless networks by taking into account the caching capability when designing the signal transmission.
Particularly, we investigate multi-layer caching where both base station (BS) and users are capable of
storing content data in their local cache and analyse the performance of edge-caching wireless networks
under two notable uncoded and coded caching strategies. Firstly, we calculate backhaul and access
throughputs of the two caching strategies for arbitrary values of cache size. The required backhaul
and access throughputs are derived as a function of the BS and user cache sizes. Secondly, closedform expressions for the system energy efficiency (EE) corresponding to the two caching methods are
derived. Based on the derived formulas, the system EE is maximized via precoding vectors design and
optimization while satisfying a predefined user request rate. Thirdly, two optimization problems are
proposed to minimize the content delivery time for the two caching strategies. Finally, numerical results
are presented to verify the effectiveness of the two caching methods.
Index terms— edge-caching, energy efficiency, beamforming, optimization.
I. I NTRODUCTION
Future wireless networks will have to address stringent requirements of delivering content
at high speed and low latency due to the proliferation of mobile devices and data-hungry
This research is supported, in part, by the ERC AGNOSTIC project under code R-AGR-3283 and the FNR CORE ProCAST
project.
The authors are with the Interdisciplinary Centre for Security, Reliability and Trust (SnT) – University of Luxembourg, 29
Avenue John F. Kennedy, L-1855 Luxembourg. E–Mail: {thang.vu,symeon.chatzinotas, bjorn.ottersten}@uni.lu.
Parts of this work have been presented at the IEEE SPAWC 2017 [29].
1
2
applications. It is predicted that by 2020, more than 70% of network traffic will be video
[1]. Although various network architectures have been proposed in order to boost the network
throughput and reduce transmission latency such as cloud radio access networks (C-RANs) [2–4]
and heterogeneous networks (HetNets), traffic congestion might occur during peak-traffic times.
A promising solution to reduce latency and network costs of content delivery is to bring the
content closer to end users via distributed storages through out the network, which is referred to
content placement or caching [5]. Caching usually consists of a placement phase and a delivery
phase. The former is executed during off-peak periods when the network resources are abundant.
In this phase, popular content is stored in the distributed caches. The later usually occurs during
peak-traffic times when the actual users’ requests are revealed. If the requested content is available
in the user’s local storage, it can be served immediately without being sent via the network. In
this manner, caching allows significant backhaul’s load reduction during peak-traffic times and
thus mitigating network congestion [5], [6].
Most research works on caching exploit historic user requested data to optimize either placement or delivery phases [5], [8], [9]. For a fixed content delivery strategy, the placement phase
is designed to maximize the local caching gain, which is proportional to the number of file
parts available in the local storage. This caching method stores the contents independently and
are known as uncoded caching. The caching gain can be further improved via multicasting a
combination of the requested files during the delivery phase, which is known as coded caching
[6], [7]. By carefully placing the files in the caches and designing the coded data, all users can
recover their desired content via a multicast stream. Rate-memory tradeoff is derived in [6],
which achieves a global caching gain on top of the local caching gain. This gain is inversely
proportional to the total cache memory. Similar rate-memory tradeoff is investigated in deviceto-device (D2D) networks [10] and secrecy constraints [11]. In [12], [13], the authors study the
tradeoff between the memory at edge nodes and the transmission latency measured in normalized
delivery time. The rate-memory tradeoff of multi-layer coded caching networks is studied in [14],
[15]. Note that the global gain brought by the coded caching comes at a price of coordination
since the data centre needs to know the number of users in order to construct the coded messages.
Recently, there have been numerous works addressing joint content caching and transmission
design for cache-assisted wireless networks. The main idea is to take into account the cached
content at the edge nodes when designing the link transmission to reduce the access and backhaul
costs. It is shown in [16] that transmit power and fronthaul bandwidth can be reduced via cache-
DRAFT
3
aware multicast beamforming design and power allocation. The impact of wireless backhaul on
the energy consumption was studied in [17]. The authors in [18] propose a joint optimization
of caching, routing and channel assignment via two sub-problems called restricted master and
pricing. The performance of caching wireless D2D networks are analysed in [19–22]. In [20],
the authors study D2D networks which allow the storage of files at either small base stations or
user terminals. Taking into account the wireless fading channels, a joint content replacement and
delivering scheme is proposed to reduce the system energy consumption. The throughput-outage
trade-off of the mmWave underlying D2D networks under a simplified grid topology is derived in
[21]. The stochastic performance of caching wireless networks is analysed in [23], in which the
nodes’ locations are modelled as a Poison point process (PPP). The average ergodic downlink rate
and outage probability are studied when cache capability is present at three tiers of base station
(BS), relay and D2D pairs. In [24], success delivery rate is studied in cluster-centric networks,
which group small base stations (SBSs) into disjoint clusters. In this work, the SBSs within one
cluster share a cache which is divided into two parts: one contains the most popular files, and
one comprises different files which are most popular locally. The authors in [25] study effects
of mobility on the caching wireless networks via a random-walk assumption of node mobilities.
In [26], a low-complexity greedy algorithm is proposed to minimize the content delivering delay
in cooperative caching C-RANs. Energy efficiency (EE) of cache-assisted networks are analysed
in [27], [28]. Focusing on the content placement phase in heterogeneous networks, the authors
in [27] study the trade-off between the expected backhaul rate and energy consumption. The
impact of caching is analysed in [28] via close-form expression of the approximated network
EE. We note that these works consider either only the uncoded caching method or the caching
at higher layers separated from the signal transmission.
In this paper, we investigate the performance of edge-caching wireless networks in which
multi-layer caches are available at either user or edge nodes. Our contributions are as follows:
•
Firstly, we investigate the performance of edge-caching networks under two notable uncoded
and inter-file coded caching strategies1 . In particular, we compute the required throughputs
on the backhaul and access links for both caching strategies with arbitrary cache sizes.
•
Secondly, we derive a closed-form expression for the system EE, which reveals insight
contributions of cache capability at the BS and users. Based on the derived formula, we
1
The inter-file coded caching is different from intra-file coded caching method.
DRAFT
4
maximize the system EE subject to a quality-of-service (QoS) constraint taking into account
the caching strategies. The maximum EE is obtained in closed-form for zero-forcing (ZF)
precoding and suboptimally solved via semi-definite relaxation (SDR) design. Our paper
differs from [27], [28] as following. We focus on the delivery phase, while [28] considers
the placement phase. We consider multi-layer cache and the two caching strategies, while
[27] only considers caching available at the BS with an uncoded caching algorithm.
•
Thirdly, we analyse and minimize the delivery time for the two caching strategies via two
formulated problems which jointly optimize the beamforming design and power allocation.
Our method is fundamentally different from [12] which studies the latency limit from
information-theoretic perspectives. Compared with [26], which studies only uncoded caching
at higher layers, we consider both caching strategies jointly with the signal transmission.
•
Finally, the analysed EE and delivery time are verified via selective numerical results. We
show an interesting result that the uncoded-caching is more energy-efficient only for the
small user cache sizes. This result is different from the common understanding that the coded
caching always outperforms the uncoded caching in terms of total backhaul’s throughput.
The rest of this paper is organised as follows. Section II presents the system model and
the caching strategies. Section III analyses the system energy efficiency. Section IV presents the
proposed EE maximization algorithms. Section V minimizes the delivery time. Section VI derives
the EE for general content popularity. Section VII shows numerical results. Finally, Section VIII
concludes the paper.
Notation: (.)H , (x)+ and Tr(.) denote the Hermitian transpose, max(0, x) and the trace(.)
function, respectively. bxc denotes the largest integer not exceeding x.
II. S YSTEM M ODEL
We consider the downlink edge-caching wireless network in which a data centre serves K
distributed users, denoted by K = {1, . . . , K}, via one BS, as depicted in Figure 1. This model
can also be applied in various practical scenarios in which the users can be replaced by various
cache-assisted edge nodes, e.g., edge nodes in fog radio access networks (F-RAN), small-cell
BSs in HetNet. The L-antenna BS, with L ≥ K, serves all users via wireless access networks
and connects to the data centre via an error-free, bandwidth-limited backhaul link. The wireless
transmissions are subjected to block Rayleigh fading channels, in which the channel fading
coefficients are fixed within a block and are mutually independent across the users. The block
DRAFT
5
Fig. 1: Multiple-layer cache-assisted wireless networks.
duration is assumed to be long enough for the users to be served the requested files. The data
centre contains N files of equal size of Q bits and is denoted by F = {F1 , . . . , FN }. In practice,
unequal size files can be divided into trunks of subfiles which have the same size.
A. Caching model
We consider multiple-layer caching networks in which both the BS and users are equipped
with a storage memory of size Mb and Mu files, with 0 ≤ Mb , Mu ≤ N , respectively. We
consider off-line caching, in which the content placement phase is executed during off-peak
times [6]. For robustness, we consider the completely distributed placement phase in which the
BS is unaware of user cache’s content. In particular, the BS stores
Mb Q
N
(non-overlapping) bits
of every file in its cache, which are randomly chosen2 . Similarly, each user stores
2
Mu Q
N
bits of
There may exists a better cache placement at the BS at the expense of coordination.
DRAFT
6
every file in its cache under the uncoded caching strategy3 . The placement phase at the user
caches under the coded caching is similar to [6]. The total number of bits stored at the BS and
user caches are respectively Mb Q and Mu Q bits, which satisfy the memory constraints.
At the beginning of the delivery phase, each user requests one file from the library. In order
to focus on the interplay between the EE and cache capabilities, we consider the worst case
in which the users tend to request different files and the content popularity follows a uniform
distribution [6]. The general case of content popularities, e.g., Zipf distribution, will be studied
in Section VI. Denote d1 , ..., dK as the file indices requested by user 1, ..., K, respectively. If the
requested bits (or subfile) is in its own cache, they can be served immediately. Otherwise, this
subfile is sent from the BS’s cache or the data centre through the backhaul link. We consider
two notable caching methods for the delivery phase: uncoded caching and coded caching.
1) Uncoded caching: This strategy sends parts of the requested files to each user independently. We note that the users do not know the cache content of each other. The advantage of
this method is the robustness and it does not require coordination. The total number of bits
transmitted through the backhaul link, Qunc,BH , and the access link, Qunc,AC , are given in the
following proposition.
Proposition 1: Under the uncoded caching strategy, the total number of bits transmitted
through the backhaul and access links are Qunc,BH = KQ 1 − MNu 1 − MNb and Qunc,AC =
KQ 1 − MNu , respectively.
Proof: See Appendix A.
2) Coded caching: In coded caching strategy, the data centre first intelligently encodes the
requested files and then sends them to the users. We note that this strategy requires the number
of users in order to construct the coded messages for all users.
u
c ∈ Z? and δ =
Proposition 2: Let m = b KM
N
KMu
N
− m with 0 ≤ δ < 1. Under the coded-
caching strategy, the throughput (in bits) on the access links is Qcod,AC = (1 − δ) Q(K−m)
+
m+1
m+1
Q(K−m)
Mb
Mb m+2 Q(K−m
δ Q(K−m−1)
,
and
the
backhaul
thoughtput
is
Q
=
(1−δ)
1
−
+δ
1
−
cod,BH
m+2
N
m+1
N
m+2
N 2N
Proof: We consider two cases: i) Mu ∈ {0, K
, K , . . . , (K−1)N
} and ii) Mu has arbitrary
K
value within (0, N ).
N 2N
Case 1: Mu ∈ {0, K
, K , . . . , (K−1)N
}
K
In this case, the user cache Mu is multiple times of
3
If
Mb Q
N
DRAFT
or
Mu Q
N
N
.
K
Denote m =
Mu K
N
∈ {0, 1, 2, . . . , K − 1}.
is not an integer, we round up this ratio to the closest integer and perform zero-padding to the last.
7
When m = 0, it is straightforward to see that Qcod,AC = QK and Qcod,BH = (1 − Mb /N )KQ
since there is no cache at the users. The computation for m ∈ {1, . . . , K − 1} is as follows.
Computation of Qcod,AC : We first calculate the total bits Qcod,AC need to be sent over the access
links under the coded-caching strategy. Let CC(F, K, m) denote the coded-caching algorithm that
the BS employs to serve K users. Each user is equipped with a cache of size
mN
,m
K
∈ Z? ,
and requests one file from the library F. CC(F, K, m) comprises of two phases: a placement
phase and a delivery phase. Due to space limitation, the details of CC(F, K, m) are omitted
here but can be found in [6, Sec.V]. We only present the essential information of CC(F, K, m)
m
which will be used in the next subsection. Each file Ff ∈ F is divided into CK
non-overlapped
subfiles. Then each file can be expressed as Ff = (Ff,T |T ⊂ K, |T | = m), where T is any
subset of K consisting of m different elements. During the delivery phase, the BS multicasts
XS = ⊕s∈S Ffs ,S\{s} to all users, where S ⊂ K with |S| = m + 1 and ⊕ denotes the XOR
operation. It has been shown in [6] that
m+1
Qcod,AC = CK
K −m
Q
=Q
(bits).
m
CK
1+m
Computation of Qcod,BH : Since the BS randomly stores parts of every file in its cache, the
probability that a bit in file Ff ∈ F is prefetched at the BS cache is p =
Mb
.
N
Now consider the
transmission of signal XS . Each bit in XS is the XORed of m + 1 bits from m + 1 different
files. If these bits are available at the BS cache, there is no need to send this XORed bit through
the backhaul. Otherwise, the data centre sends this XORed bit through the backhaul to the BS.
Because these m+1 files are independent, the probability that this XORed bit is not sent through
the backhaul is pm+1 . In other words, the probability that a XORed bit in XS is sent through
the backhaul is 1 − pm+1 . Since there are Qcod,AC XORed bits, the total bits sent through the
backhaul is
Qcod,BH = (1 − pm+1 )Qcod,AC
M m+1 Q(K − m)
b
= 1−
.
N
m+1
Case 2: 0 < Mu < N
This subsection calculates the throughput on the backhaul and access links for arbitrary values
N
of the BS and user cache size. Let m ∈ Z? and 0 < δ < 1 such as Mu = (m + δ) K
. For
every file Ff ∈ F, we divide it into two parts: Ff1 consisting of the first (1 − δ)Q bits and Ff2
DRAFT
8
consisting of the remaining δQ bits. Then the original library F is decomposed into two disjoint
sub-libraries F1 = {F11 , F21 , . . . , FN1 } and F2 = {F12 , F22 , . . . , FN2 }. Note that the file size in F1
and F2 is (1 − δ)Q and δQ bits, respectively.
Cahe placement phase: The placement phase in this case comprises of two steps. First, the
data centre applies the placement phase of CC(F1 , K, m) on F1 . After this step, each user cache
contains (1 − δ)Mu Q bits. Then, it applies CC(F2 , K, m + 1) on F2 . This steps results in δMu Q
bits on each user cache. In total, each user cache is prefetched with (1−δ)Mu Q+δMu Q = Mu Q
bits, which satisfies the memory constraint.
Delivery phase: We employ a time-splitting mechanism to serve the user requests. As a result,
the delivery phase consists of two consecutive steps. First, the delivery phase of CC(F1 , K, m)
bits. Then the delivery phase of
is applied for F1 . This will costs a throughput (1 − δ) Q(K−m)
m+1
CC(F2 , K, m + 1) is applied for F2 , which results in additional δ Q(K−m−1)
bits4 . Therefore, the
m+2
total throughput on the access links is
Qcod,AC = (1 − δ)Q
K −m
K −m−1
+ δQ
.
m+1
m+2
We observe that the probability that each XORed bit in F1 and F2 is stored at the BS cache
is q m+1 and q m+2 , respectively. Therefore, the backhaul throughput in this case is
K −m
Qcod,BH = (1 − δ) 1 − q m+1 Q
m+1
K
−
m−1
+ δ(1 − q m+2 )Q
.
m+2
Proposition 2 derives the aggregated throughput on the access links under the coded-caching
strategy for arbitrary values Mu ∈ [0, N ]. When δ = 0 and
KQ(1−Mu /N )
,
1+KMu /N
KMu
N
∈ Z? , the result is shorten as
which can also be found in [6]. Note that [6] derives the access link’s throughput
only for limited values of Mu such as
KMu
N
∈ Z. In other words, Proposition 2 generalizes the
result in [6] for arbitrary values of the user cache size.
B. Transmission model
This subsection describes the transmission of the requested files from the BS to users. Let
hk ∈ CL×1 denote the channel vector from the BS antennas to user k, which follows a circular4
This time-splitting mechanism can be seen as an implementation scheme to achieve the memory-sharing performance in [6].
DRAFT
9
symmetric complex Gaussian distribution hk ∼ CN (0, σh2k IK ), where σh2k is the parameter
accounting for the path loss from the BS antennas to user k. Perfect channel state information
(CSI) is assumed to be available at the BS. In practice, robust channel estimation can be achieved
through the transmission of pilot sequences. When a user requests a file, it first checks its own
cache. If the requested file is available in its cache, it can be served immediately. Otherwise,
the user sends the requested file’s index to the data centre. If the requested file is not at the BS
cache, it will be sent from the data centre via the backhaul. Then the BS transmits the requested
file to the user via the access links.
1) Signal transmission for uncoded caching strategy: The data stream for each user under
the uncoded caching method is transmitted independently. Denote Fd1 , . . . , FdK as the requested
files from user 1, . . . , K, respectively, and F̄d1 , . . . , F̄dK as parts of the requested files which are
not at the user cache. First, the BS modulates F̄dk in to corresponding modulated signal xk and
then sends the precoded signal through the access channels. Denote wk ∈ CL×1 as the precoding
P
H
vector for user k. The received signal at user k is given as yk = hH
k wk xk +
l6=k hk wl xl + nk ,
where nk is Gaussian noise with zero mean and variance σ 2 . The first term in yk is the desired
signal, and the second term is the inter-user interference. The signal-to-interference-plus-noise
ratio at user k is SINRk =
P
2
|hH
k wk |
H w |2 +σ 2 .
|h
l
l6=k
k
The information achievable rate of user k is
Runc,k = B log2 (1 + SINRk ) , 1 ≤ k ≤ K,
(1)
where B is the access links’ bandwidth.
The transmit power on the access links under the uncoded caching policy is Punc =
PK
k=1
k
wk k2 .
2) Signal transmission for coded caching strategy: Obviously, one can use the transmission
design derived for the uncoded caching to delivery the requested files in the coded-caching
method. However, since the coded caching strategy transmits a coded message to a group of
users all users during the delivery phase, using the orthogonal beams might result in resources
inefficiency. Thus, we employ physical-layer multicasting [30] to precode the data for the coded
caching strategy.
m+1
coded messages (of length
In the coded caching strategy, the BS will send CK
Q
m
CK
bits) in
total to the users, each of which is received by a subset of m + 1 users [6]. Denote by S ⊂ K
an arbitrary subset consisting of m + 1 users, and by S = {S | |S| = m + 1} all subsets of
DRAFT
10
m+1
m + 1 users. Obviously, |S| = CK
. For convenience, we denote XS as the coded message
targeted for the users in S. The received signal at user k ∈ S is given as yk = hH
k wS xS + nk ,
where wS is the beamforming vector for the users in S and xS is the modulated signal of XS .
The achievable rate for the users under physical-layer multicasting is
n
|hH wS |2 o
.
Rcod,S = min B log2 1 + k 2
k∈S
σ
(2)
The transmit power on the access links under the coded caching policy is Pcod =k wS k2 .
III. E NERGY- EFFICIENCY ANALYSIS
This section analyses the EE performance of the two caching strategies.
Definition 1 (Energy efficiency): The EE measured in bit/Joule is defined as:
EE =
KQ
,
EΣ
where KQ is the total requested bits from the K users and EΣ is the total energy consumption
for delivering these bits.
Since the cache placement phase in off-line caching occurs much less frequently (daily or
weekly) than the delivery phase, we assume the energy consumption in the placement phase is
negligible and thus EΣ is the energy cost in the delivery phase [6], [16].
A. EE analysis for uncoded caching strategy
The total energy cost under the uncoded caching policy is given as Eunc,Σ = Eunc,BH +Eunc,AC ,
where Eunc,BH and Eunc,AC are the energy cost on the backhaul and access links, respectively5 .
To compute the energy consumption on the access links, we note that each user requests
Qunc,AC
K
bits. The uncoded caching strategy sends these bits to each user independently via unicasting.
Since user k requests a file at rate Runc,k , it takes
Qunc,AC
KRunc,k
seconds to complete the transmission.
Therefore, the total energy consumed on the access links is calculated as
K
Eunc,AC =
5
Qunc,AC
Mu X k wk k2
Punc = Q(1 −
)
.
KRunc,k
N k=1 Runc,k
In practice, EΣ also includes a static energy consumption factor.
DRAFT
11
Sine the backhaul link provides enough capacity to serve the access network, the energy cost
on the backhaul is modelled as
Eunc,BH = ηQunc,BH
Mb
Mu
1−
,
= ηKQ 1 −
N
N
where η is a constant. In practices, η can be seen as the pricing factor used to trade energy for
transferred bits [16]. The actual value of η depends on the backhaul technology.
Therefore, the EE under the uncoded caching strategy is given as
K
EEunc =
1−
Mu
N
ηK 1 −
Mb
N
+
kwk k2
k=1 Runc,k
PK
.
(3)
It is observed from (3) that EEunc is jointly determined by the cache capacities Mu and Mb and
the transmitted power on the access links.
B. EE analysis for coded caching strategy
The energy cost on the backhaul link under the coded caching policy is given as Ecod,BH =
ηQcod,BH , where η is the pricing factor. In order to calculate the energy consumption on the
access links, Ecod,AC , we note that the BS multicasts the coded information XS to the users in
S. With the rate Rcod,S , it takes
the BS in this case is Ecod,AC =
Qcod,AC
seconds to send XS . The total energy consumed by
m+1
CK
Rcod,S
P
Pcod,S
Qcod,AC
m+1
S∈S Rcod,S . Therefore, the EE under the coded caching
CK
strategy is given as
EEcod =
KQ
KQ
=
Qcod,AC P
Ecod,Σ
ηQcod,BH + C m+1
S∈S
K
Pcod,S
Rcod,S
.
(4)
From Proposition 2 we obtain
u
1+ KM
N
KMu +1
!.
P kwS k2
1
+ m+1
(1− MNu ) η 1− MNb N
R
C
S∈S cod,S
K
Similarly, the EE under the coded-caching is determined by the BS and user storage capacity
EEcod =
and the transmit power on the access links.
C. Comparison between the two strategies
In general, the comparison between the two caching methods is complicated due to the
contributions of many system parameters. In some cases, e.g.,
KMu
N
∈ Z? , however, it is possible
to explicitly reveal each method’s performance. Assuming that all users are served at the same
rate, e.g., Runc,k = Rcod,S = γ, ∀k, S.
DRAFT
12
1) Free-cost backhaul link: This occurs when the BS cache is large enough to store all the
files, e.g., Mb = N or η = 0. All the requested files are available at either user cache or BS
cache. Consequently, we have:
EEunc =
K
1−
, EEcod =
Mu Punc
N
γ
KMu
N
Mu Pcod
N
γ
1+
1−
.
When the two methods use the same transmit power on the access links, i.e., Punc = Pcod , we
have EEunc > EEcod . In general, the coded caching strategy will achieve a higher EE than the
Pcod
1
uncoded caching method when Mu > Punc − K N .
2) Mb = 0: In this case, all the requested files which are not at the user cache will be sent
from the data centre, and thus
EEunc =
u
1 + KM
1
N
.
,
EE
=
cod
unc
1− MNu η+ PγK
1− MNu η + Pcod
γ
It is observed that the coded-caching strategy achieves higher EE than the uncoded caching
method for the same transmit power since
KMu
N
> 0 and
Punc
K
< Pcod .
IV. E NERGY-E FFICIENCY M AXIMIZATION IN EDGE - CACHING WIRELESS NETWORKS
We aim at maximizing the EE in edge-caching wireless networks under the two caching
strategies. The general optimization problem is formulated as follows:
Maximize EE
{ wk }K
k=1 ,w
s.t. QoS constraint,
(5)
where EE ∈ {EEunc , EEcod } and Rk ∈ {Runc,k , Rcod } which are given in Section II-B.
A. EE maximization for uncoded caching strategy
Let γk denote the QoS requirement of user k (bits per second). Without caching, it takes
tk =
Q
γk
seconds to send user k the requested file. However, since parts of the requested files
are available in the user cache, the BS needs to send only (1 − MNu )Q bits to user k. Therefore,
the rate requirement taking into account the user cache is γ̄k = (1 −
Mu
)Q/tk
N
= (1 −
Mu
)γk .
N
It is observed from (3) that for a given network topology, the BS and user cache memories are
DRAFT
13
fixed. Therefore, maximizing the EE is equivalent to minimizing the transmit power. Therefore,
the problem (5) is equivalent to the following problem:
K
X
k wk k2
|hH wk |2
Minimize
, s.t. P Hk 2
≥ ζk , ∀k,
Runc,k
|hk wl | +σ 2
{ wk ∈CL }K
k=1 k=1
(6)
l6=k
γ̄k
where the rate constraint is replaced by an equivalent SINR constraint ζk = 2 B − 1.
1) Cost minimization by Zero-Forcing precoding: In this subsection, we maximize the EE
based on the ZF design because of its low computational complexity. Since the direction of
the beamforming vectors are already defined by the ZF, only transmitting power on each beam
needs to be optimized. Let pk , 1 ≤ k ≤ K, denote the transmit power dedicated for user k. The
√
precoding vector for user k is given as wk = pk h̃k , where h̃k is the ZF beamforming vector
for user k, which is the k-th column of HH (HHH )−1 , with H = [h1 , . . . , hK ]T .
Theorem 1: Under the ZF design, the uncoded caching strategy achieves the maximal EE
given as
K
EEZF
unc =
1−
Mu
N
ηK 1 −
Mb
N
+
σ2
PK
2
k=1 ζk kh̃k k
.
γ̄k
2
Proof: By definition, |hH
l wk | = pk δlk , where δij is the Dirac delta function. Therefore,
the constraint in (6) becomes
pk
σ2
≥ ζk , ∀k. Consequently, the cost minimization problem is
formulated as follows:
Minimize
{ pk :pk ≥0}K
k=1
K
X
k=1
ak p k
log2 (1 + ak pk /σ 2 )
(7)
s.t. pk ≥ ζk σ 2 , ∀k,
where ak =k h̃k k2 .
ax
with a, b ≥ 0 in R+ . The derivative of f (x) is f 0 (x) =
Consider a function f (x) = log (1+bx)
2
a
bx
1 − log(1+bx)(1+bx)
> 0, ∀x > 0. Therefore, the objective function of (7) is a strictly
log (1+bx)
2
increasing function in its supports. Therefore, the optimal solution of (7) is achieved at p?k = ζk σ 2 ,
P
2
and the minimum transmit power is σ 2 K
k=1 ζk k h̃k k . Substituting this into EEunc , we obtain
the proof of Theorem 1.
2) Cost minimization by Semi-Definite Relaxation: In this subsection, we maximize the EE
by design the beamforming vectors and power allocation simultaneously. It is seen that (6) is a
DRAFT
14
NP-hard problem due to its non-convex objective functions as well as the constraints. Therefore,
we resort to solve a suboptimal solution of (6) by minimizing the upper bound of the objective
function. Since it requires Runc,k ≥ γ̄k to deliver the requested content to the users successfully,
we have
kwkk2
Runc,k
≤
kwkk2
.
γ̄k
Due to the difference of transmission time among the users, a user
who has received the requested file may not interfere the transmission of other users. Denote
Kt , {k | γ̄k ≤
Q
, ∀t
t
∈ [0, minQ
]} as the subset of active users at the time of interest. Then
k (γ̄k )
the resorted problem is stated as
Minimize
wk ∈CL
X k wk k2
, s.t.
γ̄
k
k∈K
P
t
2
|hH
k wk |
≥ ζk , ∀k.
2
2
|hH
k wl | +σ
(8)
k6=l∈Kt
L×L
. Since
We introduce new variables Xk = wk wkH ∈ CL×L and denote Ak = hk hH
k ∈ C
2
H
H
H
H
|hH
l wk | = hl wk wk hl = Tr(hl hl wk wk ) = Tr(Al Xk ), we can reformulate problem (8) as
Minimize
Xk ∈CL×L
X
Tr(Xk )
(9)
k∈Kt
X
s.t. Tr(Ak Xk) ≥ ζk
Tr(Al Xk )+ζk σ 2 , ∀k,
k6=l∈Kt
Xk 0, rank(Xk ) = 1, ∀k.
Problem (9) is still difficult to solve because the rank-one constraint is non-convex. Fortunately,
the objective function and the two first constraits are convex. Therefore, (9) can be effectively
solved by the SDR which is obtained by ignoring the rank one constraint. Since the SDR of
(9) is a convex optimization problem, it can be effectively solved by using, e.g., the primal-dual
interior point method [32]. Gaussian randomization procedure may be used to compensate the
ignorance of the rank-one constraint in the SDR solution [31]. It has been shown that SDR can
achieve a performance close to the optimal solution [31]. From the solution X?k of the SDR
of (9), we obtain the precoding vector wk? . Substituting wk? into (3) we obtain the EE of the
uncoded caching strategy under SDR design.
B. EE maximization for coded caching strategy
Given the QoS requirement γk , user k expects to receive the requested file in tk =
Q
.
γk
m+1
m
Since each user receives only CK−1
coded messages out of CK
, the active time for user k is
m
CK−1
m+1
CK
tk =
DRAFT
(m+1)Q
.
Kγk
Therefore, the required rate for user k is γ̄k = (
m
Q∗CK−1
)/( (m+1)Q
)
m
CK
Kγk
=
K−m
γ ,
m+1 k
15
where
m
Q∗CK−1
m
CK
is the number of coded bits sent to user k. Since the cache memories (4) are
constant, maximizing the EE is equivalent to minimizing
Pcod
,
Rcod,S
where Rcod,S is given in (2).
The optimization problem in this case is stated as
Minimize
wS ∈CL×1
k wS k2
,
Rcod,S
s.t. Rcod,S ≥ γ̄k , ∀k ∈ S.
(10)
We note that problem (10) optimizes the beamforming vector for only a subset of users in
S. Because
Pcod
Rcod,S
is not convex, we instead find a suboptimal solution of problem (10) by
minimizing the upper bound of
Pcod
,
Rcod,S
i.e.,
Pcod
Rcod,S
≤
Pcod
,
γ̄min,S
where γ̄min,S = mink∈S γ̄k . By
introducing a new variable X = wSH wS ∈ CL×L , the reformulated problem is given as
Minimize
X∈CL×L
Tr(X)
, s.t. X 0; rank(X) = 1;
γ̄min,S
Tr(Ak X) ≥ σ 2 (2
γ̄min,S
B
(11)
−1), ∀k ∈ S.
We observe that the objective function and the constraints of problem (11) are convex, except the
rank-one constraint. This suggests to solve problem (11) via SDR method by ignoring the rankone constraint. It is noted that the solution of SDR does not always satisfy the rank-one condition.
Thus, Gaussian randomization procedure might be used to obtain the approximated vector from
the SDR solution [31]. From the solution X? of problem (11), we obtain the precoding vector
wS? . Substituting wS? into (4), we obtain the EE for the coding caching strategy.
V. M INIMIZATION OF CONTENT DELIVERY TIME
In this section, we aim at minimizing the average time for delivering the requested files to
all users. In general, the delivery time is comprised of two parts caused by the backhaul and
access links. In practice, the backhaul capacity is usually much greater than the access capacity.
Therefore, we assume negligible delivery time on the backhaul link. It is also assumed that the
processing time at the BS is fixed and negligible. Therefore, the total delivery time is mainly
determined by the access links.
A. Minimization of delivery time for uncoded caching strategy
We would like to remind here that the uncoded caching strategy transmits independent data
streams to the users. Let tk be a time duration for the BS to transmit all the
Qunc,AC
K
requested
DRAFT
16
bits to user k. Since the BS serves user k with rate Runc,k , we have tk =
Qunc,AC
KRunc,k
seconds. The
average delivery time in the uncoded caching strategy is given as
τunc
K
Q(1 −
1 X
=
tk =
K k=1
K
K
Mu X
)
N
k=1
1
Runc,k
.
The minimization of τunc is formulated as
Q(1 −
Minimize
K
{ wk ∈CL }K
k=1
Mu
)
N
K
X
k=1
1
(12)
Runc,k
s.t. Runc,k ≥ γ̄k , ∀k;
K
X
k w k k 2 ≤ PΣ ,
k=1
where the first constraint is to satisfy the QoS requirement and PΣ is the total transmit power.
1) Zero-Forcing precoding design: Let h̃k be the ZF precoding vector for user k, which is
the k-th column of the ZF precoding matrix HH (HHH )−1 . The beamforming vector is parallel
√
to the ZF precoding vector as wk = pk h̃k , where pk is the power allocating to user k. Note
ZF
that under the ZF precoding, hH
l h̃k = δlk , we thus have Runc,k = log2 (1 +
pk
).
σ2
Therefore, the
problem (12) is equivalent to
Q(1 −
K
K
Mu X
)
N
1
log2 (1 + pk /σ 2 )
k=1
X
pk
pk k h̃k k2 ≤ PΣ .
s.t. 2 ≥ ζk , ∀k;
σ
k
Minimize
{ pk :pk ≥0}K
k=1
Proposition 3: Given the total power PΣ satisfying PΣ ≥ σ 2
PK
k=1 ζk
(13)
k h̃k k2 , the problem
(13) is convex and feasible.
Proof: We will show that PΣ ≥ σ 2
PK
k=1 ζk
k h̃k k2 is the necessary and sufficient conditions
of problem (13). It is straightforward to see that the constraints of (13) are convex. We will show
that the objective function is also convex. Indeed, consider the function f (x) = 1/ log2 (1 + ax)
in R+ with a > 0. The second-order derivative of f (x) is given as
a
,
log2 (1 + ax)(1 + ax)
a2
a2
00
f (x) =
+
.
log22 (1 + ax)(1 + ax)2 log2 (1 + ax)(1 + ax)2
f 0 (x) = −
It is verified that the second-order derivative is always positive, thus the objective function is
DRAFT
17
convex in its support. Consequently, this problem can effectively solved by efficient algorithms,
e.g., CVX [32].
Now assuming that the problem (13) is feasible. Then there exists a solution {p̄k }K
k=1 which
satisfies all the constraints. From the first constraint, it is straightforward to verify that PΣ ≥
P
2
σ2 K
k=1 ζk k h̃k k .
2) General beamforming design: Finding the optimal solution of the original problem (12)
is challenging because of the non-convex objective function. We instead propose to solve (12)
sub-optimally via minimizing the upper bound of τunc . Since
τunc ≤ max{t1 , . . . , tK } =
and Q(1 −
Mu
)
N
Q(1 − MNu )
,
min{Runc,1 , . . . , Runc,K }
is a positive constant, the suboptimal optimization of (12) is formulated as
Maximize min{Runc,1 , . . . , Runc,K }
{ wk ∈CL }K
k=1
s.t. Runc,k ≥ γ̄k , ∀k;
X
(14)
k w k k 2 ≤ PΣ .
k
By introducing an arbitrary positive variable x and resorting to SINR constraint, the above
problem is equivalent to
Maximize
x>0,{wk ∈CL }K
k=1
2
|hH
k wk |
≥ x, ∀k,
H
2
2
l6=k |hk wl | +σ
X
x ≥ ζk ;
k wk k2 ≤ PΣ ,
x, s.t. P
(15)
k
We introduce new variables Xk = wk wkH and remind that Ak = hk hH
k . The problem (15) is
equivalent to
Maximize
{ Xk ∈CL×L }K
k=1 ,x
x,
s.t. x ≥ ζk ;
X
Tr(Xk ) ≤ PΣ ;
(16)
k
Tr(Ak Xk ) − x
X
Tr(Ak Xl ) ≥ xσ 2 , ∀k;
l6=k
Xk 0; rank(Xk ) = 1.
It is observed that the third constraint is convex for a given x. Therefore, the SDR solution of
problem (16), which is obtained by ignoring the rank one constraint, can be solved via bisection.
DRAFT
18
TABLE I: A LGORITHM TO SOLVE (16)
1. Initialize AH , AL = ζ, and the accuracy .
2. AM = (AH + AL )/2.
3. Given AM , if (17) is feasible, then AL := AM .
Otherwise AH := AM .
4. Repeat step 2 and 3 until |AH − AL | ≤ .
The steps to solve are given in Table I.
find {Xk ∈ CL×L }K
k=1
s.t. Tr(Ak Xk ) − AM
(17)
X
Tr(Ak Xl ) + σ 2 ≥ 0, ∀k
l6=k
X
Tr(Xk ) ≤ PΣ ; Xk 0, ∀k.
k
B. Minimization of delivery time for coded caching strategy
The coded caching strategy multicasts the coded message XS to the users in S. Since each XS
Q
Qcod,AC P
1
contains Ccod,AC
m+1 bits, the delivery time under coded-caching strategy is τcod =
S∈S Rcod,S ,
C m+1
K
K
where Rcod,S is given in (2). Since the transmissions of XS are independent, the optimization
problem of τcod becomes minimizing the delivery time of each XS , as follows:
Minimize
wS
∈CL
1
Rcod,S
(18)
s.t. Rcod,S ≥ γ̄min,S ; k wS k2 ≤ PΣ .
By introducing new variables x > 0, X = wS wSH ∈ CL×L and using the equivalent SINR
constraint, the above optimization is equivalent to
Maximize x
x,X∈CL×L
(19)
s.t. Tr(Ak X) ≥ xσ 2 , ∀k ∈ S; X 0;
x ≥ 2γ̄min,S − 1; Tr(X) ≤ PΣ ; rank(X) = 1.
Similar to the previous subsection, we observe that the first constraint in (19) is convex for a
given x. Therefore, the above optimization can be solved via bisection and SDR by removing
DRAFT
19
TABLE II: A LGORITHM TO SOLVE (19)
1. Initialize AH , AL = 2γ̄min,S − 1, and the accuracy .
2. AM = (AH + AL )/2.
3. Given AM , if (20) is feasible, then AL := AM .
Otherwise AH := AM .
4. Repeat step 2 and 3 until |AH − AL | ≤ .
the rank one constraint. The steps to solve are given in Table II.
find X ∈ CL×L
(20)
s.t. Tr(Ak X) − AM σ 2 ≥ 0, ∀k ∈ S
Tr(X) ≤ PΣ ; X 0.
VI. N ON - UNIFORM FILE POPULARITY DISTRIBUTION
In most practical cases, the content popularity does not follow uniform distribution. In fact,
there are always some files which are more frequently requested than the others. In this section,
we consider arbitrary user content popularity and the uncoded caching strategy. Let pk =
PN
{qk,1 , . . . , qk,N } with
n=1 qk,n = 1 denote the content popularity of user k, where qk,n is
the probability of the n-th file being requested from user k.
The global file population at the BS is computed as follows:
qG,n
K
1 X
qk,n .
=
K k=1
(21)
We consider general cache memories in which the user caches’ size can be different. For
convenience, let M0 (files) denote the storage memory at the BS and Mk (files) denote the
storage memory at user k. In the placement phase, each user fills its cache based on the local
file popularity until full. Denote q̃k = Π(qk ) and q̃G = Π(qG ) as the sorted version in decreasing
order of qk and qG , respectively. Then user k stores the first nk = Mk files in q̃k . Similarly, the
BS stores the first nG = M0 files in q̃G .
In the delivery phase, the users send their requested file indices to the data centre.
Proposition 4: Let D = {d1 , . . . , dK } denote a set of file indices which are requested by the
DRAFT
20
users. The total throughput on the access links is given as
QAC (D) = Q
K
X
Ink (Πk (dk ))
(22)
InG (ΠG (dk )),
(23)
k=1
and the backhaul’s throughput is calculated as
QBH (D) = Q
K
X
k=1
where Πk (dk ) is the new position of file dk after sorted by Π(qk ), and In (i) = 1 if i > n and 0
otherwise.
The proof of Proposition 4 is straightforward followed by checking if the requested file is
available at the BS or user caches.
In this caching strategy, a user stores the whole file if it is cached. Therefore, the BS will
transmit only to a subset of users K̃(D) = {k | Πk (dk ) > nk } who do not cache the requested
files. In order to minimize the energy cost, the BS applies the signal transmission design as
follows:
Minimize
wk∈K̃(D) ∈CL
X k wk k2
,
R̃unc,k
(24)
k∈K̃(D)
s.t. R̃unc,k ≥ γ, ∀k ∈ K̃(D),
where R̃unc,k = B log2 1 +
P
2
|hH
k wk |
H w |2 +σ 2
|h
l
k6=l∈K̃(D)
k
.
The delivery time minimization problem is formulated as:
Minimize
wk∈K̃(D)
X
∈CL
k∈K̃(D)
Q
R̃unc,k
,
(25)
s.t. R̃unc,k ≥ γ, ∀k ∈ K̃(D).
The solution of problem (24) and (25) can be found by similar techniques in Section IV-A
and Section V-A, respectively.
VII. N UMERICAL RESULTS
This section presents numerical results to demonstrate the effectiveness of the studied caching
policies. The results are averaged over 500 channel realizations. For ease of presentation, the
uncoded caching under the general beamformer design using SDR in Section IV-A2 is named
DRAFT
21
300
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
Energy efficiency (Mbits/Joule)
250
200
150
6
100
4
2
0.1
50
0
0
0.1
0.15
0.2
0.2
0.3
0.4
0.5
0.6
Normalized user cache size
0.7
0.8
0.9
(a) Cost-free backhaul
70
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
Energy efficiency (Mbits/Joule)
60
50
40
30
20
10
0
DRAFT
0
0.1
0.2
0.3
0.4
0.5
0.6
Normalized user cache size
0.7
0.8
0.9
22
TABLE III: Simulation time in seconds, m = K − 1
K
Coded
4
8
0.197
0.204
UncodedSDR
0.384
1.131
UncodedZF
8.7e-5
10e-5
as SDR and the Zero-forcing design in Section IV-A1 is named as ZF in the figures. Unless
otherwise stated, the system setup is as follows: L = 10 antennas, K = 8 users, N = 1000 files,
B = 1 MHz, η = 10−6 bits/Joule [16], σh2k = 1, ∀k, Q = 10 Mb, γk = 2 Mbps , ∀k.
A. Energy efficiency performance
We first study the two caching strategies when the energy consumption on the backhaul is
negligible. This occurs when the BS cache is large enough to store all the files. In this case, the
EE only depends on the user cache size. Figure 2a presents the EE of the two caching strategies
as the function of the normalized user cache size (the user cache size Mu divided by the library
size N ). The EE is plotted based on the optimal precoding vectors obtained from Section IV.
It is shown that the uncoded caching under the SDR design achieves higher EE than the coded
caching when the normalized user cache is less than 0.2. This result suggests an important
guideline for using the uncoded caching since the user cache is usually small compared to the
library size in practice. When the user cache is capable of storing more than 20% of all the
files, it suggests to use the coded caching for larger system EE. It is also observed that the
uncoded caching under SDR design achieves higher EE than the ZF for all user cache size. This
is because the SDR design is more efficient than the ZF precoding.
Figure 2b compares the EE for various user cache size when Mb = 0.7N . In general, the coded
caching method is more efficient than the uncoded caching for most of user cache size values.
Increasing user cache capability results in larger relative gain of the coded-caching compared
with the uncoded method. The uncoded caching under SDR design achieves slightly better EE
than the ZF design at small user cache sizes, however, at an expense of higher computational
complexity as shown in Table III. From the practical point of view, ZF design is preferred in
this case because of its low complexity. When Mb increases, the SDR achieves significantly
higher EE than the ZF. Figure 2c presents the EE v.s. the user cache size when both BS and
user cache size are small. It is shown that the uncoded caching strategy with either SDR or ZF
DRAFT
23
design outperforms the coded caching scheme in the observed user cache sizes, which is in line
with the result in Figure 2b. Figure 2d compares the EE as a function of the BS cache size
when Mu = 0.5N . The result shows that the caching at the BS has more impacts on both the
caching strategies when the BS cache size is relatively large. It is shown that the coded-caching
outperforms the uncoded caching for all values Mb . It is also shown that the SDR design achieves
higher EE gain compared with the ZF as Mb increases.
Figure 3a presents the EE v.s. the normalized user cache size of the uncoded caching algorithm
under Zipf content popularity distribution, i.e., qk,n =
n−α
PN −α , ∀k.
i=1 i
It is observed that the SDR
design significantly surpasses the ZF design. In particular, at 40% library size of the user cache,
the SDR achieves almost 3 times EE higher than the ZF design. Greater Zipf exponent factor
results in higher EE for the both designs. This is because the content distribution in this case
is more centralized at some files. Figure 3b plots the EE v.s. the normalized BS cache size.
Similarly, the SDR design achieves higher EE than the ZF design. Also, the BS cache size has
smaller impacts on the system EE than the user cache size.
B. Delivery time performance
Figure 4 presents the delivery times of the two caching strategies as a function of the user
cache size with 8 users and transmit power equal to 10 dB. It is shown that the uncoded caching
strategy with both designs outperforms the coded counter part if the user cache is smaller
than 30% of the library. When the cache size is larger, the coded-caching method achieves
slightly smaller latency than the uncoded caching strategy. This important observation suggests
the optimal caching algorithm in practical systems depending on the memory availability at the
edge nodes. It is also shown that the delivery time of the uncoded caching strategy linearly
depends on the cache size. This can be seen from Proposition 1 that the network throughput in
the uncoded caching linearly depends on the cache size.
Figure 5 compares the delivery times of the two caching algorithms for various transmit
powers. Obviously, increasing the transmit power will significantly reduce the delivery times
in both strategies. When the user cache size is small (Fig. 5a), the uncoded caching strategies
deliveries the requested files faster than the coded caching method, which is in line with the
results in Fig. 4. When the user cache memory is capable of storing more content (Fig. 5b),
the coded caching strategy is more efficient than the uncoded caching. It is also observed that
DRAFT
24
60
SDR design
ZF design
Energy efficiency (Mbits/Joule)
50
40
α=1.5
30
α=1
20
10
0
0
0.1
0.2
0.3
0.4
Normalized user cache size
0.5
0.6
(a) Mb = 0.3N
SDR design
ZF design
Energy efficiency (Mbits/Joule)
20
16
α=1.5
12
8
α=1
4
DRAFT
0
0
0.1
0.2
0.3
0.4
Normalized BS cache size
0.5
0.6
25
1.6
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
Delivery time (seconds)
1.2
0.8
0.3
0.2
0.4
0.3
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Normalized user cache size
0.4
0.7
0.5
0.8
0.9
Fig. 4: Delivering time of the two caching methods v.s. the normalized user memory Mu . Average
transmit power is 10 dB.
the SDR design only outperforms the ZF design for small transmit power. This is because large
transmit power can supports optimal solution for both SDR and ZF designs.
Figure 6 plots the delivery times depending on the number of users K. For small K, the
uncoded caching strategy slightly outperforms the coded caching method. When K increases,
the coded caching tends to surpass the uncoded caching strategy. In this case, the total cache
size in the network is bigger in which the coded caching algorithm is more effective.
VIII. C ONCLUSIONS
We have analysed the performance of cache-assisted wireless networks under two notable
uncoded and coded caching strategies. First, we have expressed the energy efficiency metric in
closed-form expression for each caching strategy as a function of base station and user cache sizes
and the transmit power on the access links. Based on the derived closed-form, two optimization
DRAFT
26
0.8
Delivery time (seconds)
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
0.6
0.4
0.2
5
10
15
Average transmit power (dB)
20
25
(a) Mu = 0.3N
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
Delivery time (seconds)
0.3
DRAFT
0.2
0.1
5
10
15
Average transmit power (dB)
20
25
27
0.5
Coded caching
Uncoded caching, SDR
Uncoded caching, ZF
Delivery time (seconds)
0.4
0.3
0.2
0.1
2
4
6
Number of user
8
10
Fig. 6: Delivering time of the two caching methods v.s. the number of users. Average transmit
power is 10 dB, Mu = 0.4N .
problems have been formulated to maximize the system EE while satisfying a predefined user
rate requirement. Second, we have analysed the total delivery time for each caching strategy and
designed the beamforming vectors to minimize the total delivery time. It has been shown that
the uncoded caching algorithm achieves higher EE than the coded caching method only when
the user cache size is small and the BS cache is large enough.
Based on the studied work, several research directions can be extended. One is to consider
generic networks in which the data centre is serving multiple base stations. In this case, different
backhaul constraints for each BS should be taken into account when designing the caching
algorithms. Another direction is to consider the coded caching algorithm applied to non-uniform
content popularity. This requires a redesign of both cache placement and delivery phases in order
to take into consideration differences in user preferences.
DRAFT
28
A PPENDIX A
P ROOF OF P ROPOSITION 1
The proof can be found by similar techniques in [7, Sec. II]. When a user requests a file,
parts of the requested file are in the user cache. Since the users’ requests are independent, the
requested files can be either the same or different.
For any integer number m, 1 ≤ m ≤ N , there are N m ways to choose m elements out of the
set of size N , which can be further expressed as
N
m
=
m
X
N
am
l Cl ,
l=1
where ClN ,
N!
(N −l)!
m N
and am
l is a constant. In the above equation, al Cl is the number of choices
of m elements out of N which contains l different elements. By using the inductive method, we
can obtain:
1,
am
=
l
mam−1 + am−1 ,
l
l−1
if l = 1 or m
if 1 < l < m
For a choice comprising of l different values, the BS needs to send lQ(1 − Mu /N ) subfiles to
the users. Therefore, the average access throughput is calculated as
Qunc,AC
K
Mu
1 X
K N
lQal Cl 1 −
= K
N l=1
N
l
K
X
lQaK
Mu Y N − l + i
l
.
=
1−
N K−l
N i=1
N
l=1
(26)
It is observed that the library size N is usually very large compared to K, thus
N −l+i
N
' 1, ∀1 ≤
i ≤ l and
laK
l
N K−l
0, if l < K
'
.
K, if l = K
(27)
From (26) and (27) we obtain:
Qunc,AC
Mu
' KQ 1 −
.
N
(28)
To compute the backhaul throughput, we note that the BS randomly cache
file. Therefore, the probability that a bit is stored at the BS cache is
DRAFT
Mb
.
N
Mb
N
parts of every
Finally, since the BS
29
is the caching at the BS and users independent, we obtain Qunc,BH in Proposition 1.
R EFERENCES
[1] Cisco, “Cisco visual networking index: Global mobile data traffic forecast update 2016-2021,” 2017, white paper.
[2] T. X. Vu, H. D. Nguyen, T. Q. S. Quek, and S. Sun, “Adaptive cloud radio access networks: compression and optimization,”
IEEE Trans. Signal Process, vol. 65, no. 1, pp. 228–241, Jan. 2017.
[3] T. X. Tran and D. Pompili, “Dynamic Radio Cooperation for User-Centric Cloud-RAN With Computing Resource Sharing,”
IEEE Trans. Wireless Commun., vol. 16, no. 4, pp. 2379–2393, Apr. 2017.
[4] T. X. Tran, A. Hajisami, and D. Pompili, “QuaRo: A queue-aware robust coordinated transmission strategy for downlink
C-RANs,” in Proc. IEEE Int. Conf. Sensing, Commun. and Netw., London, 2016, pp. 1-9.
[5] S. Borst, V. Gupta, and A. Walid, “Distributed caching algorithms for content distribution networks,” in Proc. IEEE Int.
Conf. Comput. Commun., Mar. 2010, pp. 1–9.
[6] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2856–2867,
May 2014.
[7] T. X. Vu, S. Chatzinotas, and B. Ottersten, “Coded caching and storage allocation in heterogeneous networks,” in Proc.
IEEE Wireless Commun. Netw. Conf., San Francisco, CA, 2017, pp. 1–5.
[8] K. C. Almeroth and M. H. Ammar, “The use of multicast delivery to provide a scalable and interactive video-on-demand
service,” IEEE J. Sel. Areas Commun., vol. 14, no. 6, pp. 1110–1122.
[9] D. Christopoulos, S. Chatzinotas, and B. Ottersten, “Cellular-broadcast service convergence through caching for COMP
cloud RAN,” in Proc. IEEE Symp. Commun. Veh. Tech. in the Benelux, Luxembourg, 2015, pp. 1–6.
[10] M. Ji, G. Caire, and A. F. Molisch, “Fundamental limits of caching in wireless D2D networks,” IEEE Trans. Inf. Theory,
vol. 62, no. 2, pp. 849–869, Feb. 2016.
[11] A. Sengupta, R. Tandon, and T. C. Clancy, “Fundamental limits of caching with secure delivery,” IEEE Trans. Info.
Forensics and Security, vol. 10, no. 2, pp. 355–370, Feb. 2015.
[12] A. Sengupta, R. Tandon, and O. Simeone, “Cache aided wireless networks: Tradeoffs between storage and latency,” in
Proc. Annu. Conf. Info. Sci. Syst., Princeton, NJ, Mar. 2016, pp. 320–325.
[13] S. H. Park, O. Simeone, W. Lee, and S. Shamai, “Coded multicast fronthauling and edge caching for multi-connectivity
transmission in fog radio access networks,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun., Sapporo,
Japan, 2017, pp. 1-5.
[14] N. Karamchandani, U. Niesen, M. A. Maddah-Ali, and S. N. Diggavi, “Hierarchical coded caching,” IEEE Trans. Inf.
Theory, vol. 62, no. 6, pp. 3212–3229, Jun. 2016.
[15] L. Tang and A. Ramamoorthy, “Coded caching for networks with the resolvability property,” in Proc. IEEE Int. Symp. Inf.
Theory, Barcelona, Jul. 2016, pp. 420–424.
[16] M. Tao, E. Chen, H. Zhou, and W. Yu, “Content-centric sparse multicast beamforming for cache-enabled cloud RAN,”
IEEE Trans. Wireless Commun., vol. 15, no. 9, pp. 6118–6131, Sept. 2016.
[17] T. X. Vu, S. Chatzinotas, and B. Ottersten “Energy Minimization for Cache-assisted Content Delivery Networks with
Wireless Backhaul,” IEEE Wireless Commun. Lett., vol. pp, no. pp, pp. 1–1, 2018.
[18] A. Khreishah, J. Chakareski, and A. Gharaibeh, “Joint caching, routing, and channel assignment for collaborative small-cell
cellular networks,” IEEE J. Sel. Areas Commun., vol. 34, no. 8, pp. 2275–2284, IEEE Trans. Inf. Theory. 2016.
[19] L. Zhang, M. Xiao, G. Wu, and S. Li, “Efficient scheduling and power allocation for D2D-assisted wireless caching
networks,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2438–2452, Jun. 2016.
DRAFT
30
[20] M. Gregori, J. Gmez-Vilardeb, J. Matamoros, and D. Gndz, “Wireless content caching for small cell and D2D networks,”
IEEE J. Sel. Areas Commun., vol. 34, no. 5, pp. 1222–1234, May 2016.
[21] M. Ji, G. Caire, and A. F. Molisch, “Wireless device-to-device caching networks: Basic principles and system performance,”
IEEE J. Sel. Areas Commun., vol. 34, no. 1, pp. 176–189, Jan. 2016.
[22] M. Ji, G. Caire, and A. Molisch, “The throughput-outage tradeoff of wireless one-hop caching networks,” IEEE Trans.
Inf. Theory, vol. 61, no. 12, pp. 6833–6859, Dec. 2015.
[23] C. Yang, Y. Yao, Z. Chen, and B. Xia, “Analysis on cache-enabled wireless heterogeneous networks,” IEEE Trans. Wireless
Commun., vol. 15, no. 1, pp. 131–145, Jan. 2016.
[24] Z. Chen, J. Lee, T. Q. Quek, and M. Kountouris, “Cooperative caching and transmission design in cluster-centric small
cell networks,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 3401 – 3415, May 2016.
[25] G. Alfano, M. Garetto, and E. Leonardi, “Content-centric wireless networks with limited buffers: when mobility hurts,”
IEEE/ACM Trans. Netw., vol. 24, no. 1, pp. 299–311, Jan. 2016.
[26] T. X. Tran, F. Kazemi, E. Karimi, and D. Pompili, “Mobee: Mobility-aware energy-efficient coded Caching in cloud radio
access networks,” in Proc. IEEE Int. Conf. Mobile Ad-Hoc Sensor Syst. (MASS), Orlando, FL, 2017, pp. 461–465.
[27] F. Gabry, V. Bioglio, and I.Land, “On energy-efficient edge caching in heterogeneous networks,” IEEE J. Sel. Areas
Commun., vol. 34, no. 12, pp. 3288–3298, Dec. 2016.
[28] D. Liu and C. Yang, “Energy efficiency of downlink networks with caching at base stations,” IEEE J. Sel. Areas Commun.,
vol. 34, no. 4, pp. 907–922, Apr. 2016.
[29] T. X. Vu, S. Chatzinotas, and B. Ottersten, “Energy-efficient design for edge-caching wireless networks: When is codedcaching beneficial?” in Proc. IEEE Int. Workshop Signal Process. Wireless Commun., Sapporo, 2017, pp. 1–5.
[30] N. D. Sidiropoulos, T. N. Davidson, and Z.-Q. Luo, “Transmit beamforming for physical-layer multicasting,” IEEE Trans.
Signal Process, vol. 54, no. 6, pp. 2239–2251, Jun. 2006.
[31] Z.-Q. Luo, W. K. Ma, A. M. C. So, Y. Ye, and S. Zhang, “Semidefinite relaxation of quadratic optimization problems,”
IEEE Signal Process. Mag., vol. 27, no. 3, pp. 20–34, Mar. 2010.
[32] S. Boyd and L. Vandenberghe, Convex Optimization.
DRAFT
Cambridge Univ. Press, 2004.
| 7cs.IT
|
Tunable Sensitivity to Large Errors in Neural Network Training
Gil Keren
Sivan Sabato
Björn Schuller
arXiv:1611.07743v1 [stat.ML] 23 Nov 2016
Chair of Complex and Intelligent systems Department of Computer Science Chair of Complex and Intelligent systems
University of Passau
Ben-Gurion University of the Negev
University of Passau
Passau, Germany,
Beer Sheva, Israel
Passau, Germany
Machine Learning Group
[email protected]
Imperial College London, U.K.
Abstract
When humans learn a new concept, they might ignore examples that they cannot make sense of at first, and only later
focus on such examples, when they are more useful for learning. We propose incorporating this idea of tunable sensitivity
for hard examples in neural network learning, using a new
generalization of the cross-entropy gradient step, which can
be used in place of the gradient in any gradient-based training
method. The generalized gradient is parameterized by a value
that controls the sensitivity of the training process to harder
training examples. We tested our method on several benchmark datasets. We propose, and corroborate in our experiments, that the optimal level of sensitivity to hard example is
positively correlated with the depth of the network. Moreover,
the test prediction error obtained by our method is generally
lower than that of the vanilla cross-entropy gradient learner.
We therefore conclude that tunable sensitivity can be helpful
for neural network learning.
1
Introduction
In recent years, neural networks have become empirically
successful in a wide range of supervised learning applications, such as computer vision (Krizhevsky, Sutskever, and
Hinton 2012; Szegedy et al. 2015), speech recognition (Hinton et al. 2012), natural language processing (Sutskever,
Vinyals, and Le 2014) and computational paralinguistics
(Keren and Schuller 2016; Keren et al. 2016). Standard
implementations of training feed-forward neural networks
for classification are based on gradient-based stochastic optimization, usually optimizing the empirical cross-entropy
loss (Hinton 1989).
However, the cross-entropy is only a surrogate for the
true objective of supervised network training, which is in
most cases to reduce the probability of a prediction error
(or in some case BLEU score, word-error-rate, etc). When
optimizing using the cross-entropy loss, as we show below, the effect of training examples on the gradient is linear in the prediction bias, which is the difference between
the network-predicted class probabilities and the target class
probabilities. In particular, a wrong confident prediction induces a larger gradient than a similarly wrong, but less confident, prediction.
Copyright c 2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
In contrast, humans sometimes employ a different approach to learning: when learning new concepts, they might
ignore the examples they feel they do not understand, and
focus more on the examples that are more useful to them.
When improving proficiency regarding a familiar concept,
they might focus on the harder examples, as these can contain more relevant information for the advanced learner. We
make a first step towards incorporating this ability into neural network models, by proposing a learning algorithm with
a tunable sensitivity to easy and hard training examples.
Intuitions about human cognition have often inspired successful machine learning approaches (Bengio et al. 2009;
Cho, Courville, and Bengio 2015; Lake et al. 2016). In this
work we show that this can be the case also for tunable sensitivity.
Intuitively, the depth of the model should be positively
correlated with the optimal sensitivity to hard examples.
When the network is relatively shallow, its modeling capacity is limited. In this case, it might be better to reduce sensitivity to hard examples, since it is likely that these examples
cannot be modeled correctly by the network, and so adjusting the model according to these examples might only degrade overall prediction accuracy. On the other hand, when
the network is relatively deep, it has a high modeling capacity. In this case, it might be beneficial to allow more sensitivity to hard examples, thereby possibly improving the accuracy of the final learned model.
Our learning algorithm works by generalizing the crossentropy gradient, where the new function can be used instead
of the gradient in any gradient-based optimization method
for neural networks. Many such training methods have been
proposed, including, to name a few, Momentum (Polyak
1964), RMSProp (Tieleman and Hinton 2012), and Adam
(Kingma and Ba 2015). The proposed generalization is parameterized by a value k > 0, that controls the sensitivity
of the training process to hard examples, replacing the fixed
dependence of the cross-entropy gradient. When k = 1 the
proposed update rule is exactly the cross-entropy gradient.
Smaller values of k decrease the sensitivity during training
to hard examples, and larger values of k increase it.
We report experiments on several benchmark datasets.
These experiments show, matching our expectations, that in
almost all cases prediction error is improved using large values of k for deep networks, small values of k for shallow net-
works, and values close to the default k = 1 for networks of
medium depth. They further show that using a tunable sensitivity parameter generally improves the results of learning.
The paper is structured as follows: In Section 1.1 related
work is discussed. Section 2 presents our setting and notation. A framework for generalizing the loss gradient is developed in Section 3. Section 4 presents desired properties
of the generalization, and our specific choice is given in Section 5. Experiment results are presented in Section 6, and we
conclude in Section 7. Some of the analysis, and additional
experimental results, are deferred to the supplementary material due to lack of space.
1.1
Related Work
The challenge of choosing the best optimization objective
for neural network training is not a new one. In the past, the
quadratic loss was typically used with gradient-based learning in neural networks (Rumelhart, Hinton, and Williams
1988), but a line of studies demonstrated both theoretically
and empirically that the cross-entropy loss has preferable
properties over the quadratic-loss, such as better learning
speed (Levin and Fleisher 1988), better performance (Golik, Doetsch, and Ney 2013) and a more suitable shape of
the error surface (Glorot and Bengio 2010). Other cost functions have also been considered. For instance, a novel cost
function was proposed in (Silva et al. 2006), but it is not
clearly advantageous to cross-entropy. The authors of (Bahdanau et al. 2015) address this question in a different setting
of sequence prediction.
Our method allows controlling the sensitivity of the training process to examples with a large prediction bias. When
this sensitivity is low, the method can be seen as a form of
implicit outlier detection or noise reduction. Several previous works attempt to explicitly remove outliers or noise in
neural network training. In one work (Smith and Martinez
2011), data is preprocessed to detect label noise induced
from overlapping classes, and in another work (Jeatrakul,
Wong, and Fung 2010) the authors use an auxiliary neural
network to detect noisy examples. In contrast, our approach
requires a minimal modification on gradient-based training
algorithms for neural networks and allows emphasizing examples with a large prediction bias, instead of treating these
as noise.
The interplay between “easy” and “hard” examples during
neural network training has been addressed in the framework
of Curriculum Learning (Bengio et al. 2009). In this framework it is suggested that training could be more successful if
the network is first presented with easy examples, and harder
examples are gradually added to the training process. In another work (Kumar, Packer, and Koller 2010), the authors
define easy and hard examples based on the fit to the current model parameters. They propose a curriculum learning
algorithm in which a tunable parameter controls the proportions of easy and hard examples presented to a learner at
each phase. Our method is simpler than curriculum learning
approaches, in that the examples can be presented at random
order to the network. In addition, our method allows also a
heightened sensitivity to harder examples. In a more recent
work (Zaremba and Sutskever 2014), the authors indeed find
that a curriculum in which harder examples are presented in
early phases outperforms a curriculum that at first uses only
easy examples.
2
Setting and Notation
For any integer n, denote [n] = {1, . . . , n}. For a vector v,
its i’th coordinate is denoted v(i).
We consider a standard feed-forward multilayer neural
network (Svozil, Kvasnicka, and Pospichal 1997), where
the output layer is a softmax layer (Bridle 1990), with n
units, each representing a class. Let Θ denote the neural network parameters, and let zj (x; Θ) denote the value of output unit j when the network has parameters Θ, before the
applying the softmax function. Applying the softmax function, the probability assigned by the network to class j is
n
P
pj (x; Θ) := ezj /
ezi . The label predicted by the network
i=1
for example x is ŷ(x; Θ) = argmaxj∈[n] pj (x; Θ). We consider the task of supervised learning of Θ, using a labeled
training sample S = {(xi , yi )}m
i=1 , , where yi ∈ [n], by
m
P
optimizing the loss function: L(Θ) :=
`((xi , yi ); Θ). A
i=1
popular choice for ` is the cross-entropy cost function, defined by `((x, y); Θ) := − log py (x; Θ).
3
Generalizing the gradient
Our proposed method allows controlling the sensitivity of
the training procedure to examples on which the network
has large errors in prediction, by means of generalizing the
gradient. A naı̈ve alternative towards the same goal would
be using an exponential version of the cross-entropy loss:
` = −| log(py )k |, where py is the probability assigned to the
correct class and k is a hyperparameter controlling the sensitivity level. However, the derivative of this function with respect to py is an undesired term since it is not monotone in k
for a fixed py , resulting in lack of relevant meaning for small
or large values of k. The gradient resulting from the above
form is of a desired form only for k = 1, due to cancellation
of terms from the derivatives of l and the softmax function.
Another naı̈ve option would be to consider l = − log(pky ),
but this is only a scaled version of the cross-entropy loss and
amounts to a change in the learning rate.
In general, controlling the loss function alone is not sufficient for controlling the relative importance to the training
procedure of examples on which the network has large and
small errors in prediction. Indeed, when computing the gradients, the derivative of the loss function is being multiplied
by the derivative of the softmax function, and the latter is
a term that also contains the probabilities assigned by the
model to the different classes. Alternatively, controlling the
parameters updates themselves, as we describe below, is a
more direct way of achieving the desired effect.
Let (x, y) be a single labeled example in the training set,
and consider the partial derivative of `(Θ; (x, y)) with respect to some parameter θ in Θ. We have
n
∂`((x, y); Θ) X ∂` ∂zj
=
,
∂θ
∂zj ∂θ
j=1
where zj is the input to the softmax layer when the input
example is x, and the network parameters are Θ.
∂` ∂py
∂`
= ∂p
and
If ` is the cross-entropy loss, we have ∂z
j
y ∂zj
∂`
1
=− ,
∂py
py
∂py
py (1 − py ) j = y,
=
−py pj
j=
6 y.
∂zj
Hence
∂`
=
∂zj
pj − 1
pj
y=j
otherwise.
For given x, y, Θ, define the prediction bias of the network
for example x on class j, denoted by j , as the (signed) difference between the probability assigned by the network to
class j and the probability that should have been assigned,
based on the true label of this example. We get j = pj − 1
for j = y, and j = pj otherwise. Thus, for the crossentropy loss,
n
∂` X ∂zj
=
j .
(1)
∂θ
∂θ
j=1
In other words, when using the cross entropy loss, the effect
of any single training example on the gradient is linear in the
prediction bias of the current network on this example.
As discussed in Section 1, it is likely that in many cases,
the results of training could be improved if the effect of a
single example on the gradient is not linear in the prediction
bias. Therefore, we propose a generalization of the gradient
that allows non-linear dependence in .
For given x, y, Θ and for j ∈ {1, . . . , n}, define f :
[−1, 1]n → Rn , let = (1 , . . . , n ), and consider the fol∂`
lowing generalization of ∂θ
:
g(θ) :=
n
X
∂zj
j=1
∂θ
fj ().
(2)
Here fj is the j’th component of f . When f is the identity,
∂`
∂`
we have fj () ≡ ∂z
, and g(θ) = ∂θ
. However, we are now
j
at liberty to study other assignments for f .
We call the vector of values of g(θ) for θ in Θ a pseudogradient, and propose to use g in place of the gradient within
any gradient-based algorithm. In this way, optimization of
the cross-entropy loss is replaced by a different algorithm of
a similar form. However, as we show in Section 5.2, g is not
necessarily the gradient of any loss function.
4
Properties of f
Consider what types of functions are reasonable to use for
f instead of the identity. First, we expect f to be monotonic
non-decreasing, so that a larger prediction bias never results
in a smaller update. This is a reasonable requirement if we
cannot identify outliers, that is, training examples that have a
wrong label. We further expect f to be positive when j 6= y
and negative otherwise.
In addition to these natural properties, we introduce an
additional property that we wish to enforce. To motivate
this property, we consider the following simple example.
Assume a network with one hidden layer and a softmax
layer (see Figure 1), where the inputs to the softmax layer
are zj = hwj , hi + bj and the outputs of the hidden layer
are h(i) = hwi0 , xi + b0i , where x is the input vector, and
b0i , wi0 are the scalar bias and weight vector between the input layer and the hidden layer. Suppose that at some point
during training, hidden unit i is connected to all units j in
the softmax layer with the same positive weight a. In other
words, for all j ∈ [n], wj (i) = a. Now, suppose that the
training process encounters a training example (x, y), and
let l be some input coordinate.
softmax
a a a a a
hidden
.....
i
.....
w ' i (l)
input
...
l
.........
Figure 1: Illustrating the state of the network discussed in
above.
What is the change to the weight wi0 (l) that this training
example should cause? Clearly it need not change if x(l) =
0, so we consider the case x(l) 6= 0. Only the value h(i) is
directly affected by changing wi0 (l). From the definition of
pj (x; Θ), the predicted probabilities are fully determined by
the ratios ezj /ezj0 , or equivalently, by the differences zj −
zj 0 , for all j, j 0 ∈ [n]. Now, zj − zj 0 = hwj , hi + bj −
∂(z −z 0 )
j
j
= wj (i) − wj 0 (i) =
hwj 0 , hi + bj 0 . Therefore, ∂h(i)
a − a = 0, and therefore
∂(zj − zj 0 )
∂(zj − zj 0 ) ∂h(i)
=
= 0.
0
∂wi (l)
∂h(i) ∂wi0 (l)
We conclude that in the case of equal weights from unit i to
all output units, there is no reason to change the weight wi0 (l)
for any l. Moreover, preliminary experiments show that in
these cases it is desirable to keep the weight stationary, as
otherwise it can cause numerical instability due to explosion
or decay of weights.
Therefore, we would like to guarantee this behavior also
for our pseudo-gradients. Therefore, we require g(wi0 (l)) =
0 in this case. It follows that
n
X
∂zj
0 = g(wi0 (l)) =
f (j )
∂wi0 (l)
j=1
=
n
n
X
X
∂zj ∂h(i)
f
(
)
=
a · x(l) · f (j ).
j
∂h(i) ∂wi0 (l)
j=1
j=1
Dividing by a·x(l), we get the following desired property
for the function f , for any vector of prediction biases:
X
fy () = −
fj ().
(3)
1.0
0.8
j6=y
5
0.6
|fy (²)|
Note
P that this indeed holds for the cross-entropy loss, since
j∈[n] j = 0, and in the case of cross-entropy, f is the
identity.
k =0.25
k =0.5
k =1
k =2
k =4
0.4
Our choice of f
In the case of the cross-entropy, f is the identity, leading to
a linear dependence on . A natural generalization is to consider higher order polynomials. Combining this approach
with the requirement in Eq. (3), we get the following assignment for f , where k > 0 is a parameter.
j = y,
−|y |k
|y |k
fj () = P
(4)
k
ki · j otherwise.
0.2
0.0
0.0
0.2
0.4
²y
0.6
0.8
1.0
Figure 2: Size of fy () for different choices of k. Lines are
in the same order as in the legend.
i6=y
The expression
| |k
Py k
i
is a normalization term which makes
i6=y
sure Eq. (3) is satisfied. Setting k = 1, we get that g(θ) is the
gradient of the cross-entropy loss. Other values of k result in
different pseudo-gradients.
To illustrate the relationship between the value of k and
the effect of prediction biases of different sizes on the
pseudo-gradient, we plot fy () as a function of y for several values of k (see Figure 2). Note that absolute values of the pseudo-gradient are of little importance, since in
gradient-based algorithms, the gradient (or in our case, the
pseudo-gradient) is usually multiplied by a scalar learning
rate which can be tuned.
As the figure shows, when k is large, the pseudo-gradient
is more strongly affected by large prediction biases, comk
pared to small ones. This follows since |||0 |k is monotonic
increasing in k for > 0 . On the other hand, when using
k
a small positive k we get that |||0 |k tends to 1, therefore, the
pseudo-gradient in this case would be much less sensitive to
examples with large prediction biases. Thus, the choice of
f , parameterized by k, allows tuning the sensitivity of the
training process to large errors. We note that there could be
other reasonable choices for f which have similar desirable
properties. We leave the investigation of such other choices
to future work.
5.1
A Toy Example
To further motivate our choice of f , we describe a very simple example of a distribution and a neural network. Consider
a neural network with no hidden layers, and only one input
unit connected to two softmax units. Denoting the input by
x, the input to softmax unit i is zi = wi x + bi , where wi and
bi are the network weights and biases respectively.
It is not hard to see that the set of possible prediction
functions x 7→ ŷ(x; Θ) that can be represented by this network is exactly the set of threshold functions of the form
ŷ(x; Θ) = sign(x − t) or ŷ(x; Θ) = −sign(x − t).
For convenience assume the labels mapped to the two
softmax units are named {−1, +1}. Let α ∈ ( 21 , 1), and suppose that labeled examples are drawn independently at random from the following distribution D over R × {−1, +1}:
Examples are uniform in [−1, 1]; Labels of examples in
[0, α] are deterministically 1, and they are −1 for all other
examples. For this distribution, the prediction function with
the smallest prediction error that can be represented by the
network is x 7→ sign(x).
However, optimizing the cross-entropy loss on the distribution, or in the limit of a large training sample, would result
in a different threshold, leading to a larger prediction error
(for a detailed analysis see Appendix A in the supplementary material). Intuitively, this can be traced to the fact that
the examples in (α, 1] cannot be classified correctly by this
network when the threshold is close to 0, but they still affect
the optimal threshold for the cross-entropy loss.
Thus, for this simple case, there is motivation to move
away from optimizing the cross-entropy, to a different update rule that is less sensitive to large errors. This reduced
sensitivity is achieved by our update rule with k < 1. On
the other hand, larger values of k would result in higher sensitivity to large errors, thereby degrading the classification
accuracy even more.
We thus expect that when training the network using our
new update rule, the prediction error of the resulting network
should be monotonically increasing with k, hence values of
k which are smaller than 1 would give a smaller error. We
tested this hypothesis by training this simple network on a
synthetic dataset generated according to the distribution D
described above, with α = 0.95.
We generated 30,000 examples for each of the training,
validation and test datasets. The biases were initialized to
0 and the weights were initialized from a uniform distribution on (−0.1, 0.1). We used batch gradient descent with a
learning rate of 0.01 for optimization of the four parameters, where the gradient is replaced with the pseudo-gradient
Table 1: Experiment results for single-layer networks
DATASET
MNIST
MNIST
MNIST
SVHN
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
CIFAR-100
L AYER S IZE
400
800
1100
400
800
1100
400
800
1100
400
800
1100
M OMENTUM
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
S ELECTED k
0.5
0.5
0.5
0.25
0.125
0.25
0.25
0.125
0.25
0.25
0.25
0.125
Table 2: Toy example experiment results
k
Test error
Threshold
CE Loss
4
2
1
0.5
0.25
0.125
0.0625
8.36%
6.73%
4.90%
4.27%
4.04%
3.94%
3.61%
0.116
0.085
0.049
0.037
0.030
0.028
0.022
0.489
0.361
0.288
0.299
0.405
0.625
1.190
from Eq. (2), using the function f defined in Eq. (4). f is
parameterized by k, and we performed this experiment using values of k between 0.0625 and 4. After each epoch,
we computed the prediction error on the validation set, and
training was stopped after 3000 epochs in which this error
was not changed by more than 0.001%. The values of the
parameters at the end of training were used to compute the
misclassification rate on the test set.
Table 2 reports the results for these experiments, averaged
over 10 runs for each value of k. The results confirm our hypothesis regarding the behavior of the network for the different values of k, and further motivate the possible benefits of
using k 6= 1. Note that while the prediction error is monotonic in k in this experiment, the cross-entropy is not, again
demonstrating the fact that optimizing the cross-entropy is
not optimal in this case.
5.2
Non-existence of a Cost Function for f
It is natural to ask whether, with our choice of f in Eq. (4),
g(θ) is the gradient of another cost function, instead of the
cross-entropy. The following lemma demonstrates that this
is not the case.
Lemma 1. Assume f as in Eq. (4) with k 6= 1, and g(Θ) the
resulting pseudo-gradient. There exists a neural network for
which the g(Θ) is not a gradient of any cost function.
The proof of is lemma is left for the supplemental material. Note that the above lemma does not exclude the possi-
T EST E RROR
k=1
S ELECTED k
1.76%
1.74%
1.67%
1.65%
1.67%
1.65%
16.88%
16.16%
16.09%
15.64%
16.04%
15.53%
48.32%
47.06%
46.91%
46.01%
46.43%
45.84%
75.18%
74.41%
74.04%
73.78%
73.69%
73.11%
T EST C ROSS -E NTROPY L OSS
k=1
S ELECTED k
0.078
0.167
0.072
0.150
0.071
0.145
0.661
1.576
0.648
3.108
0.626
1.525
1.430
3.034
1.388
5.645
1.410
2.820
3.302
6.931
3.260
7.449
3.239
13.557
bility that a gradient-based algorithm that uses g instead of
the gradient still somehow optimizes some cost function.
6
Experiments
For our experiments, we used four classification benchmark
datasets from the field of computer vision: The MNIST
dataset (LeCun et al. 1998), the Street View House Numbers dataset (SVHN) (Netzer et al. 2011) and the CIFAR-10
and CIFAR-100 datasets (Krizhevsky and Hinton 2009). A
more detailed description of the datasets can be found in Appendix C.1 in the supplementary material.
The neural networks we experimented with are feedforward neural networks that contain one, three or five hidden layers of various layer sizes. For optimization, we used
stochastic gradient descent with momentum (Sutskever et al.
2013) with several values of momentum and a minibatch
size of 128 examples. For each value of k, we replaced
the gradient in the algorithm with the pseudo-gradient from
Eq. (2), using the function f defined in Eq. (4). For the multilayer experiments we also used Gradient-Clipping (Pascanu,
Mikolov, and Bengio 2013) with a threshold of 100. In the
hidden layers, biases were initialized to 0 and for the weights
we used the initialization scheme from (Glorot and Bengio
2010). Both biases and weights in the softmax layer were
initialized to 0.
In each experiment, we used cross-validation to select the
best value of k. The learning rate was optimized using crossvalidation for each value of k separately, as the size of the
pseudo-gradient can be significantly different between different values of k, as evident from Eq. (4). We compared
the test error between the models using the selected k and
k = 1, each with its best performing learning rate. Additional details about the experiment process can be found in
Appendix C.2 in the supplementary material.
We report the test error of each of the trained models for
MNIST, SVHN, CIFAR-10 and CIFAR-100 in Tables 1, 3
and 4 for networks with one, three and five layers respectively. Additional experiments are reported in Appendix C.1
in the supplementary material. We further report the crossentropy values using the selected k and the default k = 1.
Table 3: Experiment results for 3-layer networks
DATASET
MNIST
MNIST
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
L AYER S IZES
400
800
400
800
400
800
400
800
M OM ’
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
S ELECTED k
1
1
2
1
2
1
0.5
1
T EST E RROR
k=1
S ELECTED k
—
—
—
—
16.52%
16.52%
—
—
46.81%
46.63%
—
—
75.20%
74.95%
—
—
T EST CE L OSS
k = 1 S ELECTED k
—
—
—
—
1.604
0.968
—
—
3.023
2.121
—
—
3.378
4.511
—
—
Table 4: Experiment results for 5-layer networks
DATASET
MNIST
MNIST
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
MNIST
MNIST
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
L AYER S IZES
400
800
400
800
400
800
400
800
400
800
400
800
400
800
400
800
M OM ’
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.9
0.9
0.9
0.9
0.9
0.9
0.9
0.9
S ELECTED k
0.5
0.25
4
0.5
2
4
2
2
1
4
4
2
4
2
1
4
Several observations are evident from the experiment results. First, aligned with our hypothesis, the value of k selected by the cross-validation scheme was almost always
smaller than 1, for the shallow networks, larger than one
for the deep networks, and close to one for networks with
medium depth. Indeed, the capacity of network is positively
correlated with the optimal sensitivity to hard examples.
Second, for the shallow networks the cross-entropy loss
on the test set was always worse for the selected k than for
k = 1. This implies that indeed, by using a different value
of k we are not optimizing the cross-entropy loss, yet are
improving the success of optimizing the true prediction error. On the contrary, in the experiments with three and five
layers, the cross entropy is also improved by selecting the
larger k. This is an interesting phenomenon, which might be
explained by the fact that examples with a large prediction
bias have a high cross-entropy loss, and so focusing training
on these examples reduces the empirical cross-entropy loss,
and therefore also the true cross-entropy loss.
To summarize, our experiments show that overall, crossvalidating over the value of k usually yields improved results
over k = 1, and that, as expected, the optimal value of k
grows with the depth of the network.
T EST E RROR
k=1
S ELECTED k
1.71%
1.69%
1.61%
1.60%
17.41%
16.49%
17.07%
16.61%
48.05%
47.85%
44.21%
44.24%
75.69%
75.48%
74.10%
73.57%
—
—
1.58%
1.60
17.89%
16.54%
16.24%
15.73%
47.91%
47.57%
45.69%
44.11%
—
—
74.32%
74.62%
7
T EST CE L OSS
k = 1 S ELECTED k
0.113
0.224
0.118
0.390
1.436
0.708
1.343
2.604
2.017
1.962
4.610
1.677
3.611
3.228
4.650
4.439
—
—
0.098
0.060
1.284
0.718
1.647
0.998
2.202
1.648
3.316
2.171
—
—
3.872
3.432
Conclusions
Inspired by an intuition in human cognition, in this work we
proposed a generalization of the cross-entropy gradient step
in which a tunable parameter controls the sensitivity of the
training process to hard examples. Our experiments show
that, as we expected, the optimal level of sensitivity to hard
examples is positively correlated with the depth of the network. Moreover, the experiments demonstrate that selecting
the value of the sensitivity parameter using cross validation
leads overall to improved prediction error performance on a
variety of benchmark datasets.
The proposed approach is not limited to feed-forward
neural networks — it can be used in any gradient-based
training algorithm, and for any network architecture. In future work, we plan to study this method as a tool for improving training in other architectures, such as convolutional
networks and recurrent neural networks, as well as experimenting with different levels of sensitivity to hard examples
in different stages of the training procedure, and combining
the predictions of models with different levels of this sensitivity.
Acknowledgments
This work has been supported by the European Communitys
Seventh Framework Programme through the ERC Starting
Grant No. 338164 (iHEARu). Sivan Sabato was supported
in part by the Israel Science Foundation (grant No. 555/15).
References
Bahdanau, D.; Serdyuk, D.; Brakel, P.; Ke, N. R.;
Chorowski, J.; Courville, A.; and Bengio, Y. 2015. Task
loss estimation for sequence prediction. arXiv preprint
arXiv:1511.06456.
Bengio, Y.; Louradour, J.; Collobert, R.; and Weston, J.
2009. Curriculum learning. In Proc. of the 26th annual International Conference on Machine Learning (ICML), 41–
48. Montreal, Canada: ACM.
Bridle, J. S. 1990. Probabilistic interpretation of feedforward classification network outputs, with relationships to
statistical pattern recognition. In Neurocomputing. Springer.
227–236.
Cho, K.; Courville, A.; and Bengio, Y. 2015. Describing
multimedia content using attention-based encoder-decoder
networks. IEEE Transactions on Multimedia 17(11):1875–
1886.
Glorot, X., and Bengio, Y. 2010. Understanding the difficulty of training deep feedforward neural networks. In
Proc. of International Conference on Artificial Intelligence
and Statistics, 249–256.
Golik, P.; Doetsch, P.; and Ney, H. 2013. Cross-entropy vs.
squared error training: a theoretical and experimental comparison. In Proc. of INTERSPEECH, 1756–1760.
Hinton, G.; Deng, L.; Yu, D.; Dahl, G. E.; Mohamed, A.-r.;
Jaitly, N.; Senior, A.; Vanhoucke, V.; Nguyen, P.; Sainath,
T. N.; et al. 2012. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research
groups. Signal Processing Magazine, IEEE 29(6):82–97.
Hinton, G. E. 1989. Connectionist learning procedures. Artificial intelligence 40(1):185–234.
Jeatrakul, P.; Wong, K. W.; and Fung, C. C. 2010. Data
cleaning for classification using misclassification analysis.
Journal of Advanced Computational Intelligence and Intelligent Informatics 14(3):297–302.
Keren, G., and Schuller, B. 2016. Convolutional RNN: an
enhanced model for extracting features from sequential data.
In Proc. of 2016 International Joint Conference on Neural
Networks (IJCNN), 3412–3419.
Keren, G.; Deng, J.; Pohjalainen, J.; and Schuller, B. 2016.
Convolutional neural networks with data augmentation for
classifying speakers native language. In Proc. of INTERSPEECH, 2393–2397.
Kingma, D., and Ba, J. 2015. Adam: A method for stochastic optimization. In International Conference on Learning
Representations (ICLR).
Krizhevsky, A., and Hinton, G. 2009. Learning multiple
layers of features from tiny images.
Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012.
Imagenet classification with deep convolutional neural networks. In Proc. of Advances in Neural Information Processing Systems (NIPS), 1097–1105.
Kumar, M. P.; Packer, B.; and Koller, D. 2010. Self-paced
learning for latent variable models. In Proc. of Advances in
Neural Information Processing Systems (NIPS), 1189–1197.
Lake, B. M.; Ullman, T. D.; Tenenbaum, J. B.; and Gershman, S. J. 2016. Building machines that learn and think like
people. arXiv preprint arXiv:1604.00289.
LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998.
Gradient-based learning applied to document recognition.
Proceedings of the IEEE 86(11):2278–2324.
Levin, E., and Fleisher, M. 1988. Accelerated learning in
layered neural networks. Complex systems 2:625–640.
Netzer, Y.; Wang, T.; Coates, A.; Bissacco, A.; Wu, B.; and
Ng, A. Y. 2011. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning. Granada, Spain.
Pascanu, R.; Mikolov, T.; and Bengio, Y. 2013. On the difficulty of training recurrent neural networks. In Proceedings
of the 30th International Conference on Machine Learning
(ICML), 1310–1318.
Polyak, B. T. 1964. Some methods of speeding up the convergence of iteration methods. USSR Computational Mathematics and Mathematical Physics 4(5):1–17.
Rumelhart, D. E.; Hinton, G. E.; and Williams, R. J. 1988.
Learning representations by back-propagating errors. Cognitive modeling 5:3.
Silva, L. M.; De Sa, J. M.; Alexandre, L.; et al. 2006. New
developments of the Z-EDM algorithm. In Intelligent Systems Design and Applications, volume 1, 1067–1072.
Smith, M. R., and Martinez, T. 2011. Improving classification accuracy by identifying and removing instances that
should be misclassified. In The 2011 International Joint
Conference on Neural Networks (IJCNN), 2690–2697.
Sutskever, I.; Martens, J.; Dahl, G.; and Hinton, G. 2013.
On the importance of initialization and momentum in deep
learning. In Proc. of the 30th International Conference on
Machine Learning (ICML), 1139–1147.
Sutskever, I.; Vinyals, O.; and Le, Q. V. 2014. Sequence
to sequence learning with neural networks. In Advances in
neural information processing systems, 3104–3112.
Svozil, D.; Kvasnicka, V.; and Pospichal, J. 1997. Introduction to multi-layer feed-forward neural networks. Chemometrics and intelligent laboratory systems 39(1):43–62.
Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.;
Anguelov, D.; Erhan, D.; Vanhoucke, V.; and Rabinovich, A.
2015. Going deeper with convolutions. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Tieleman, T., and Hinton, G. 2012. Lecture 6.5-rmsprop:
Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning.
Zaremba, W., and Sutskever, I. 2014. Learning to execute.
arXiv preprint arXiv:1410.4615.
Tunable Sensitivity to Large Errors in Neural
Network Training
Supplementary material
A
Proof for Toy Example
Consider the neural network from the toy example in Section
5.1. In this network, there exists one classification threshold
such that examples above or below it are classified to different classes. We prove that for a large enough training set,
the value of the cross-entropy cost is not minimal when the
threshold is at 0.
Suppose that there is an assignment of network parameters that minimizes the cross-entropy which induces a
threshold at 0. The output of the softmax layer is determined
z0
uniquely by eez1 , or equivalently by z0 − z1 = x(w0 − w1 ) +
b0 − b1 . Therefore, we can assume without loss of generality
that w1 = b1 = 0. Denote w := w0 , b := b0 . If w = 0 in the
minimizing assignment, then all examples are classified as
members of the same class and in particular, the classification threshold is not zero. Therefore we may assume w 6= 0.
In this case, the classification threshold is −b
w . Since we assume a minimal solution at zero, the minimizing assignment
must have b = 0.
When the training set size approaches infinity, the crossentropy on the sample approaches the expected crossentropy on D. Let CE(w, b) be the expected cross-entropy
on D for network parameter values w, b. Then
Z 0
Z α
1
log(p0 (x))dx +
log(p1 (x))dx
CE(w, b) = −
2 −1
0
Z 1
+
log(p0 (x))dx .
α
And we have:
ewx+b
= wx + b − log(ewx+b + 1),
+1
1
log(p1 (x)) = log wx+b
= − log(ewx+b + 1).
e
+1
Therefore
∂CE(w, b)
=
∂b
Z
Z 0
1 ∂ 0
−
(wx + b)dx −
log(ewx+b + 1)dx
2 ∂b −1
−1
Z α
−
log(ewx+b + 1)dx
log(p0 (x)) = log
ewx+b
0
Z
1
Z
1
(wx + b)dx −
+
α
log(ewx+b + 1)dx
α
Z
1
log(ewx+b + 1)dx + 1 − α).
1
∂
= − (1 −
2
∂b
−1
Differentiating under the integral sign, we get
Z 1
∂CE(w, b)
1
ewx+b
=−
2−α−
wx+b + 1
∂b
2
−1 e
Since we assume the cross-entropy has a minimal solution
with b = 0, we have
∂CE(w, b = 0)
∂b
1
= 2 − α − (log(ew + 1) − log(e−w + 1)).
w
Therefore
ew + 1
= log(ew ) = w.
w(2 − α) = log −w
e +1
0 = −2
Since α 6= 1, it must be that w = 0. This contradicts our assumption, hence the cross-entropy does not have a minimal
solution with a threshold at 0.
B
Proof of Lemma 1
Proof. Consider a neural network with three units in the output layer, and at least one hidden layer. Let (x, y) be a labeled example, and suppose that there exists some cost func¯
tion `((x,
y); Θ), differentiable in Θ, such that for g as defined in Eq. (2) and f defined in Eq. (4) for some k > 0, we
∂ `¯
have g(θ) = ∂θ
for each parameter θ in Θ. We now show
that this is only possible if k = 1.
¯ for any two parameters θ1 , θ2 ,
Under the assumption on `,
¯
¯
∂
∂`
∂ 2 `¯
∂
∂`
=
=
,
∂θ2 ∂θ1
∂θ1 θ2
∂θ1 ∂θ2
hence
∂g(θ2 )
∂g(θ1 )
=
.
(5)
∂θ2
∂θ1
Recall our notations: h(i) is the output of unit i in the last
hidden layer before the softmax layer, wj (i) is the weight
between the hidden unit i in the last hidden layer, and unit j
in the softmax layer, zj is the input to unit j in the softmax
layer, and bj is the bias of unit j in the softmax layer.
Let (x, y) such that y = 1. From Eq. (5) and Eq. (2), we
have
n
n
X ∂zj
X ∂zj
∂
∂
fj () =
fj ().
∂w2 (1) j=1 ∂w1 (1)
∂w1 (1) j=1 ∂w2 (1)
Plugging in f as defined in Eq. (4), and using the fact that
= 0 for i 6= j, we get:
∂
∂z1
k
−
· |1 |
=
∂w2 (1) ∂w1 (1)
∂
∂z2
k
· k 2 k · |1 |k .
∂w1 (1) ∂w2 (1) 2 + 3
∂zj
∂wi (1)
Since y = 1, we have = (p1 − 1, p2 , p3 ). In addition,
∂zj
∂h(1)
∂wj (1) = h(1) and ∂wj (1) = 0 for j ∈ [2]. Therefore
−
∂
(1 − p1 )k =
∂w2 (1)
∂
pk2
k
(1 − p1 ) .
∂w1 (1) pk2 + pk3
(6)
Next, we evaluate each side of the equation separately,
using the following:
∀j 6= i,
∂pj
∂pj ∂zj
=
= h(1)pj (1 − pj ),
∂wj (1)
∂zj ∂wj (1)
∂pj
∂pj ∂zi
=
= −h(1)pi pj .
∂wi (1)
∂zi ∂wi (1)
For the LHS of Eq. (6), we have
−
∂
(1 − p1 )k = −k(1 − p1 )k−1 h(1)p1 p2 .
∂w2 (1)
For the RHS,
kh(1)p1 pk2 (1 − p1 )k
∂
pk2
k
(1
−
p
)
=
−
.
1
∂w1 (1) pk2 + pk3
pk2 + pk3
Hence Eq. (6) holds if and only if:
1=
pk−1
(1 − p1 )
2
.
pk2 + pk3
For k = 1, this equality holds since p1 + p2 + p3 = 1.
However, for any k 6= 1, there are values of p1 , p2 , p3 such
that this does not hold. We conclude that our choice of f
does not lead to a pseudo-gradient g which is the gradient of
any cost function.
C
C.1
Additional Experiment details and Results
datasets
The MNIST dataset (LeCun et al. 1998), consisting of
grayscale 28x28 pixel images of handwritten digits, with
10 classes, 60,000 training examples and 10,000 test examples, the Street View House Numbers dataset (SVHN)
(Netzer et al. 2011), consisting of RGB 32x32 pixel images of digits cropped from house numbers, with 10 classes
73,257 training examples and 26,032 test examples and the
CIFAR-10 and CIFAR-100 datasets (Krizhevsky and Hinton
2009), consisting of RGB 32x32 pixel images of 10/100 object classes, with 50,000 training examples and 10,000 test
examples. All datasets were linearly transformed such that
all features are in the interval [−1, 1].
C.2
Choosing the value of k
In each experiment, we used cross-validation to select the
best value of k. For networks with one hidden layer, k was
selected out of the values {4, 2, 1, 0.5, 0.25, 0.125, 0.0625}.
For networks with 3 or 5 hidden layers, k was selected
out of the values {4, 2, 1, 0.5, 0.25}, removing the smaller
values of k due to performance considerations (in preliminary experiments, these small values yielded poor results
for deep networks). The learning rate was optimized using
cross-validation for each value of k separately, as the size of
the pseudo-gradient can be significantly different between
different values of k, as evident from Eq. (4).
For each experiment configuration, defined by a dataset,
network architecture and momentum, we selected an initial
learning rate η, based on preliminary experiments on the
training set. Then the following procedure was carried out
for η/2, η, 2η, for every tested value of k:
1. Randomly split the training set into 5 equal parts,
S1 , . . . , S5 .
2. Run the iterative training procedure on S1 ∪ S2 ∪ S3 , until there is no improvement in test prediction error for 15
epochs on the early stopping set, S4 .
3. Select the network model that did the best on S4 .
4. Calculate the validation error of the selected model on the
validation set, S5 .
5. Repeat the process for t times after permuting the roles
of S1 , . . . , S5 . We set t = 10 for MNIST, and t = 7 for
CIFAR-10/100 and SVHN.
6. Let errk,η be the average of the t validation errors.
We then found argminη errk,η . If the minimum was found
with the minimal or the maximal η that we tried, we also
performed the above process using half the η or double the
η, respectively. This continued iteratively until there was no
need to add learning rates. At the end of this process we
selected (k ∗ , η ∗ ) = argmink,η errk,η , and retrained the network with parameters k ∗ , η ∗ on the training sample, using
one fifth of the sample as an early stopping set. We compared the test error of the resulting model to the test error of
a model retrained in the same way, except that we set k = 1
(leading to standard cross-entropy training), and the learning rate to η1∗ = argminη err1,η . The final learning rates in
the selected models were in the range [10−1 , 10] for MNIST,
and [10−4 , 1] for the other datasets.
C.3
Results
Additional experiment results with momentum values other
than 0.5 are reported in Table 5, Table 6.
Table 5: Experiment results for single-layer networks
DATASET
MNIST
MNIST
MNIST
MNIST
MNIST
MNIST
SVHN
SVHN
SVHN
SVHN
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-10
CIFAR-10
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
CIFAR-100
L AYER S IZE
400
800
1100
400
800
1100
400
800
1100
400
800
1100
400
800
1100
400
800
1100
400
800
1100
M OMENTUM
0
0
0
0.9
0.9
0.9
0
0
0
0.9
0.9
0.9
0
0
0
0.9
0.9
0.9
0.9
0.9
0.9
S ELECTED k
0.5
0.5
0.5
0.5
2
0.5
0.25
0.25
0.25
0.125
0.25
0.25
0.125
0.125
0.125
0.0625
0.125
0.125
0.25
0.125
0.25
T EST E RROR
k=1
S ELECTED k
1.71%
1.70%
1.66%
1.67%
1.64%
1.62%
1.75%
1.75%
1.71%
1.63%
1.74%
1.69%
16.84%
16.09%
16.19%
15.71%
15.97%
15.68%
16.65%
16.30%
16.15%
15.68%
15.85%
15.47%
48.15%
46.91%
46.92%
46.14%
46.63%
46.00%
48.19%
46.71%
47.09%
46.16%
46.71%
45.77%
74.96%
74.28%
74.12%
73.47%
73.47%
73.19%
T EST C ROSS -E NTROPY L OSS
k=1
S ELECTED k
0.0757
0.148
0.070
0.137
0.068
0.131
0.073
0.140
0.070
0.054
0.069
0.127
0.658
1.575
0.641
1.534
0.636
1.493
0.679
2.861
0.675
1.632
0.640
1.657
1.435
5.609
1.390
5.390
1.356
5.290
1.518
11.049
1.616
5.294
1.850
5.904
3.306
7.348
3.327
13.267
3.235
7.489
Table 6: Experiment results for 3-layer networks
DATASET
MNIST
MNIST
MNIST
MNIST
SVHN
SVHN
CIFAR-10
CIFAR-10
CIFAR-100
CIFAR-100
L AYER S IZES
400
800
400
800
400
800
400
800
400
800
M OM ’
0
0
0.9
0.9
0.9
0.9
0.9
0.9
0.9
0.9
S ELECTED k
1
1
1
0.5
1
2
2
2
0.25
0.5
T EST E RROR
k=1
S ELECTED k
—
—
—
—
—
—
1.60%
1.53%
—
—
16.14%
15.96%
47.52%
46.92%
45.27%
44.26%
74.97%
74.52%
74.48%
73.17%
T EST CE L OSS
k = 1 S ELECTED k
—
—
—
—
—
—
0.091
0.189
—
—
1.651
1.062
2.226
2.010
2.855
2.341
3.356
8.520
4.133
8.642
| 9cs.NE
|
arXiv:1803.04660v1 [cs.DM] 13 Mar 2018
Revisiting Radius, Diameter, and all Eccentricity
Computation in Graphs through Certificates∗
Feodor Dragan†
Michel Habib‡
Laurent Viennot§
March 14, 2018
Abstract
We introduce notions of certificates allowing to bound eccentricities in a graph. In particular, we revisit radius (minimum eccentricity) and diameter (maximum eccentricity) computation and explain the efficiency of practical radius and diameter algorithms by the existence
of small certificates for radius and diameter plus few additional properties. We show how
such computation is related to covering a graph with certain balls or complementary of balls.
We introduce several new algorithmic techniques related to eccentricity computation and propose algorithms for radius, diameter and all eccentricities with theoretical guarantees with
respect to certain graph parameters. This is complemented by experimental results on various real-world graphs showing that these parameters appear to be low in practice. We also
obtain refined results in the case where the input graph has low doubling dimension, has low
hyperbolicity, or is chordal.
1
Introduction
The radius and diameter of a graph are part of the basic global parameters that allow to apprehend
the structure of a practical graph. More broadly, the eccentricity of each node, defined as the
furthest distance from the node, is also of interest as a classical centrality measure [19]. It is
tightly related to radius which is the minimum eccentricity and to diameter which is the maximum
eccentricity. On the one hand, efficient computation of such parameters is still theoretically
challenging as truly sub-quadratic algorithms would improve the state of the art for other “hard
in P” related problems such as finding two orthogonal vectors in a set of vectors or testing if one
set in a collection is a hitting set for another collection [2]. A sub-quadratic diameter algorithm
would also refute the strong exponential time hypothesis (SETH) [25] and would improve the
state of the art of SAT solvers as noted for similar problems in [24]. On the other hand, a line
of practical algorithms has been proposed based on performing selected Breadth First Search
traversals (BFS) [23, 27, 28, 8, 5] allowing to compute the diameter of very large graphs [3].
However, such practical efficiency is still not well understood.
What are the structural properties that make practical graphs tractable? This paper answers
this question with the lens of certificate, that is a piece of information about a graph that allows
to compute its radius and diameter in truly sub-quadratic time. We propose a notion of certificate
tightly related to the class of algorithms based on one-to-all distance computations from selected
nodes. Existing practical algorithms fall into this category that we call one-to-all distance based
algorithms. Based on this approach, we propose algorithms with proven guarantees with respect
to several graph properties which appear to be generally met in practice.
Another intriguing question concerns the relationship between diameter and radius computations. The most advanced algorithms [28, 5] compute both parameters at the same time. Would
∗ Work
supported by ANR Distancia.
State University
‡ Paris Diderot University
§ Inria
† Kent
1
computing one parameter help for computing the other? We answer by the affirmative based on
the notion of certificate.
The paper is presented in the context of unweighted undirected graphs but all the notions and
algorithms extend to the weighted and/or directed cases.
1.1
Our contribution
We introduce the notion of certificate as a set of nodes such that the distances from these nodes
to all nodes (rather than all-to-all pairs) allow to deduce the value of the radius or the diameter
with certainty. Given a graph G with radius r, we define a radius certificate as a set L of nodes
such that any node of G is at distance at least r from a node of L. Given in addition a node
c with eccentricity r, the set L allows to certify that the radius rad(G) of G, i.e. the minimum
eccentricity, is indeed r: we can compute r with a BFS from c and certify that all nodes have
eccentricity r or more by checking that their distance to some node in L is at least r using |L|
BFS traversals. If L has size o(n), this opens the possibility of breaking the quadratic barrier
for radius computation if one can efficiently find a small certificate when there exists one. This
raises the problem of approximating the minimum certificate for radius. Interestingly, the size R
of the minimum radius certificate gives a lower bound on the complexity of one-to-all distance
based algorithms for radius: such an algorithm must perform at least R/2 one-to-all distance
computations. We also raise similar approximation problems for diameter and all eccentricity
computations.
p
6
8 ...
4
3
U
2
q-2
q-2
q-3
q-3
7
p
.
.
.
2
a
q+1
q+1
c
1
q+1
q-3
D
2
q+1
b
5
.
.
.
p
q+1
R
Figure 1: An example of graph BTp,q (for p ≥ 2 and q ≥ 6) with small certificates for radius (R),
diameter (D) and all eccentricities (R, U ). Its diameter is 4q − 2 (eccentricity of blue nodes). Its
radius is 2q + 1 (eccentricity of green nodes). Plain lines correspond to edges while a dashed line
with label ` corresponds to a path of length `.
We show that a radius certificate can be formally defined as a covering of the node set with
complementary of open balls of radius rad(G) (excluding nodes at distance rad(G)). We also
define a diameter certificate as a covering with balls B[x, diam(G) − e(x)] of radius diam(G) − e(x)
where diam(G) is the diameter of the graph and e(x) is the eccentricity of the center x of the
ball. Similarly, an all eccentricity certificate can be defined by combining two coverings as a pair
of lower/upper (see definitions in Section 3). Finding a minimum radius (or diameter) certificate
is shown to be equivalent to minimum set cover. It is thus NP-hard while O(log n)-approximation
2
(only) is doable in polynomial time. Compared to set cover, it has an additional difficulty: the
sets are not directly available and computing all of them would require quadratic time at least. It
should be noted that these notions of certificate are independent of any algorithm: it is a graph
property to have small or big certificates. As an example, for odd k, a k × k square grid has a
one-node diameter certificate (its center) and a radius certificate with four nodes (its corners).
Figure 1 presents an example of bow-tie shaped graph BTp,q for integral parameters p ≥ 2
and q ≥ 6 that has small certificates. The diameter certificate D contains three nodes a, b, c. The
central (green) node c has eccentricity 2q + 1. Note that its eccentricity is minimal (rad(BTp,q ) =
2q + 1) and c is called a center. Any node at distance r ≤ diam(BTp,q ) − (2q + 1) = 2q − 3 from c
has eccentricity at most r + 2q + 1 ≤ diam(BTp,q ) as it can reach any node v by following a path
of length r to c and then a path from c to v of length e(c) = 2q + 1 at most. In set-cover terms, c
covers the ball B[c, 2q − 3]. The rest of the graph is covered by the balls of radius q centered at a
and b, implying that D = {a, b, c} is a diameter certificate. The radius certificate R has five nodes
such that any node is at distance rad(BTp,q ) = 2q + 1 at least from one of them. In other words,
the complement of open balls of radius rad(BTp,q ) centered on them cover the whole graph.
We propose algorithms for radius, diameter and all-eccentricity certificate computation (as a
byproduct, our algorithms also provide radius, diameter and all eccentricities). They follow a
primal-dual approach that allows to obtain guarantees on the size of computed certificates and on
the number of BFS traversals performed with respect to graph parameters that seem to be low in
practical graphs. Our experiments on practical graphs from various sources show that these graphs
not only have small certificates but also small coverings with much reduced sets: we can still cover
the node set with few complementary of balls with increased radii (resp. decreased radii) compared
to radii required for a radius (resp. diameter) certificate. Such properties explain the efficiency of
a primal dual approach. Although our algorithms have some similarities with previous algorithms,
this primal-dual flavor was not noticed before. They have similar performances in practice but
provide significantly smaller certificates. Their proven guarantees also make them more robust.
In particular, our radius and diameter algorithms handle the graph BTp,q of Figure 1 with O(1)
BFS traversals while previous exact algorithms require Ω(p) BFS traversals.
Our experiments show a striking phenomenon concerning specifically lower-bounding eccentricities (as in radius computation) that we call “antipode sparsity”. Given a ranking of the nodes
(e.g., their ID order), we define the antipode of a node u as the node at furthest distance from u
having highest rank (the ranking is used for breaking ties among nodes at the same distance). We
say that a node is an antipode if it is the antipode of some other node. We observe that practical
graphs have very few antipodes for several rankings (i.e., large groups of nodes share the same
antipode): most the practical graphs tested (with up to hundred of thousands of nodes) have less
than 100 antipodes. Although our notion of antipode is reminiscent of the usage of antipodes on
the Earth, we see that it can significantly deviate from it. On the sphere, the antipode of a point
is the unique furthest point from it and the antipode of the antipode is the point itself. The same
situation can be met in graphs such as a cycle or a grid torus. However it appears to be much
different in practical graphs: the relation is highly asymmetric, most of the nodes have multiple
furthest nodes (i.e., nodes at furthest distance from them) while there are very few antipodes
overall. This situation is indeed highly favorable to one-to-all distance based algorithms as shown
by the following theorem summarizing our algorithmic results.
Theorem 1 Given a connected graph G having m edges and k antipodes overall (according to a
given ranking), it is possible to compute:
• its radius, a center and a radius certificate of size k at most,
• its diameter, a diametral node and a diameter certificate of size π1/3 at most where π1/3 is
the maximum packing size for open balls B(u, 13 (diam(G) − e(u))),
• all eccentricities, a lower certificate of size k at most and a minimum upper certificate UOP T ,
using O(1) BFS traversals per node of associated certificates (i.e., in O(km), O(π1/3 m) and
O(km + |UOP T | m) time respectively).
3
Concerning diameter (second item), we analyse a minimalist algorithm inspired by previous practical algorithms. A basic primal-dual argument implies that the maximum size π1 of a packing
for (closed) balls B[u, α(diam(G) − e(u))] for α = 1 is a lower bound of the minimum size of a
diameter certificate. The above theorem thus indeed proves that this basic approach approximates
minimum diameter certificate within a ratio of π1/3 /π1 . While the value π0.8 (with radii reduced
by a factor 0.8) appears to be generally small in practice, the π1/3 bound can be much higher than
π1 . However, this provides a first answer for the efficiency of practical diameter algorithms that
can be complemented with the following observation. A second property often met by practical
graphs is a high diameter to radius ratio diam(G)/ rad(G) (over 1.5 in our experiments) so that a
large part of the graph is included in any ball B[c, diam(G) − rad(G)] centered at a central node
c. Such a ball corresponds to the nodes covered by adding c to a diameter certificate in the associated set cover problem. We confirm this with a refinement of the parameter π1/3 in the above
theorem when combining radius and diameter computation where a center c of the graph is used
to initialize the basic diameter algorithm. We observe values for that refined parameter that are
generally within a small constant factor of π1 . This graph property associated with high diameter
to radius ratio and the discovery of a node with small eccentricity as part of diameter computation
is thus our second element for explaining the efficiency of practical diameter algorithms.
Concerning radius computation, practical algorithms tend to perform even faster than predicted by the first point of the above theorem (including the algorithm analyzed in the theorem).
This large diameter to radius ratio also allows us to give an intuition for this. Our radius algorithm iteratively selects a node with minimal eccentricity lower-bound (according to the radius
certificate computed so far) and adds its antipode to the candidate radius certificate. We can
show that the selected node is always in the intersection of all balls of radius rad(G) centered
at previous discovered antipodes. As antipodes tend to have high eccentricity to graph-radius
ratio (in the order of the diameter to radius ratio), this intersection quickly shrinks toward the
set of graph centers. As selecting a node with minimal lower-bound combined with discovering
high eccentricity nodes is a classical approach, this gives a second element for understanding the
efficiency of practical radius algorithms. Note that the idea of using antipodes systematically for
finding high eccentricity nodes is new.
Although we reuse classical algorithmic tools, our radius and all eccentricity algorithms rely on
a new algorithmic technique that we call minimum eccentricity selection which has its own interest.
It specifically leverages on antipode sparsity for enabling efficient selection of a node with minimum
eccentricity within a set maintained online. Its amortized complexity is low when the number of
antipodes is small. Interestingly, this technique also allows to design algorithms based on an oracle
giving access to all eccentricities. Such algorithm can then be efficiently implemented using our
technique as long as eccentricity values are used to iteratively select a node u such that f (u, e(u))
is minimal for a given computable function f satisfying some non-decreasing property. It is based
on the idea of using antipodes to enhance a (lower) certificate until an adequate node is found.
The technique also appears to be useful for optimizing diameter computation. We also introduce
a new technique for diameter computation that we call delegate certificate in order to obtain both
theoretical guarantees and efficient practical performances.
Finally, a surprising fact concerns the complexity of finding an optimum upper certificate
(a certificate with minimum size that provides a tight upper-bound of the eccentricity of each
node) as provided by our all eccentricity algorithm. Contrarily to radius and diameter certificates
(as discussed above), it appears to be tractable in polynomial time. In comparison, finding an
optimum lower certificate is also shown to be as hard as set-cover. Moreover, our all eccentricity
algorithm roughly performs one BFS traversal per node of the optimum upper certificate when
the number of antipodes is much smaller than the size of this upper certificate (as observed in
our experiments). Note that this is close to the best possible for an algorithm based on one-to-all
distance computations.
We additionally refine the performance analysis of our algorithms in particular cases when the
input graph 1) has bounded doubling dimension, 2) has small hyperbolicity, or 3) is a chordal
graph.
We believe that our certificate approach provides new insight on the efficiency of practical
4
algorithms for radius and diameter, allows to propose more robust practical algorithms with complexity guarantees, and significantly enhance the state of the art for all eccentricity computation.
Moreover, the new techniques proposed here could enable new types of radius and diameter algorithms with parametrized complexity.
1.2
Related work
The concept of certificate is somehow implicit in the method introduced in [27, 28] that consists
in maintaining lower and upper bounds on the eccentricity of each node. After each BFS traversal
these bounds are improved based on distances from the source of the traversal. The sources used
for the BFS traversals performed by the algorithm form what we call a certificate. Contrarily to
this approach, we distinguish nodes used for improving lower bounds (the lower certificate) from
those used for improving upper bounds (the upper certificate). Our definition of lower certificate
uses a looser lower-bounding inequality because of this distinction. The main approach proposed
for diameter computation [27] consists in alternating nodes with small lower bound and nodes with
large upper bound as BFS sources. This can be seen as a mix of our basic diameter algorithm
with a heuristic for finding nodes with small eccentricities.
The two-sweeps heuristic [23] performs only 2 traversals to provide a diameter estimate that
appears to be tight in practice. The idea is to use the last visited node in the first traversal to
start the second traversal. It thus introduces the idea of using what we call antipodes as tentative
diametral nodes. The technique was first introduced for trees [20] where it happens to be exact.
It was also shown to provide good approximation (up to a small constant) for chordal graphs and
various graph classes [14].
A four-sweeps heuristic is proposed in [8] and complemented with an exact diameter algorithm
called iFub. The four-sweep heuristic performs twice the two-sweep method, using a mid-point of
the longest path found in the first round as the starting point of the second one. The idea is that
mid-points of longest paths make good candidates for central nodes or at least nodes with small
eccentricity. The iFub method additionally inspects furthest nodes from the best candidate center
found with the four-sweep heuristic until exact value of the diameter can be inferred.
The exact-sum-sweep method computes (exactly) both radius and diameter while performing
few BFS traversals in practice [5]. It integrates many techniques proposed in previous practical
algorithm plus an heuristic based on sum of distances that boosts the discovery of nodes with
large eccentricity in an initial phase. It also handles the directed case in a very general manner.
The structure of random power law graphs is analyzed in [6] and the efficiency of practical
diameter and radius algorithms is discussed for that type of graphs. It is shown that random
power law graphs satisfy similar properties as those we insist on. The main argument proposed
for efficiency of practical algorithms resides in the fact that such graphs have few furthest nodes
(that is nodes that appear to be furthest from some other node). However, we observe much less
antipodes than furthest nodes in practice and some graphs do have a fairly high number of furthest
nodes. Our work provides a finer parameter and allows to extend such explanation to other types
of practical graphs such as road networks and grid like networks.
Packing and covering of hyperbolic graphs with balls is investigated in [9], although slightly
different problems are considered. It would be interesting to derive similar results in hyperbolic
graphs for the collections of balls (or complementary of balls) we consider here.
1.3
Structure of the paper
We introduce basic graph and set-cover terminology in Section 2. The notions of certificate for
radius, diameter, and all eccentricities are given in Section 3. We show how such notion can be
related to one-to-all distance based algorithms in Section 4. Section 5 is devoted to our radius
algorithm. We introduce in Section 6 the technique of minimum eccentricity selection which is the
core of this radius algorithm. We analyse a basic diameter algorithm and propose an optimization
based on radius computation and minimum eccentricity selection in Section 7. Computation of
5
all eccentricities is studied in Section 8. Theorem 1 is a consequence of the theorems proven in
Sections 5, 7 and 8. We present some experimental results in Section 9 about the measurement
on various practical graphs of the parameters involved in our theorems. Section 10 is devoted to
graphs with low doubling dimension: a refined algorithm for diameter computation is proposed and
our radius and diameter algorithms are analyzed in terms of radius and diameter approximation
respectively. Section 11 refines the analysis of our radius algorithm in the case of graphs with
low hyperbolicity. Finally, we study chordal graphs in Section 12: we show that centers form a
diameter certificate while diametral nodes form a radius certificate, this allows to derive a linear
time algorithm for computing all eccentricities of a bounded degree chordal graph.
2
Preliminaries
Given an undirected unweighted graph G we denote by V its set of nodes. Let d(u, v) be the
distance between two nodes u and v in G, that is the length of a shortest path from u to v.
The eccentricity e(u) of a node u is the maximum length of a shortest path from u, that is
e(u) = maxv∈V d(u, v). The furthest nodes of u are the nodes v at furthest distance from u, i.e.,
d(u, v) = e(u). Given a ranking r of the nodes, the antipode Antipoder (u) of a node u for r is
its furthest node with highest rank. Formally, Antipoder (u) = argmaxv∈V (d(u, v), r(v)) where
pairs are ordered lexicographically. A node is called a furthest node (resp. an antipode) if it is a
furthest node (resp. an antipode) of some other node. Given a set W ⊆ V , we let Antipoder (W ) =
{Antipoder (u) : u ∈ W } denote the set of antipodes from nodes in W . The diameter diam(G) =
maxu∈V e(u) of G is the maximum eccentricity in G and the radius rad(G) = minu∈V e(u) is the
minimum eccentricity in G. A diametral node b is a node with maximum eccentricity (e(b) =
diam(G)). A central node c (or simply center ) is a node with minimum eccentricity (e(c) =
rad(G)). We let B[u, r] = {v ∈ V | d(u, v) ≤ r} (resp. B(u, r) = {v ∈ V | d(u, v) < r}) denote the
(closed) ball (resp. open ball) with radius r centered at a node u. Similarly, we define its coball
of radius r as B(u, r) = {v ∈ V | d(u, v) ≥ r}, that is the complementary of B(u, r).
We restrict ourselves to algorithms based on one-to-all distance queries: we suppose that an
algorithm DistFrom for one-to-all distances is given (typically BFS or Dijkstra). It takes a graph
G and a node u as input and returns distances from u. More precisely, DistFrom(G, u) returns
a vector D such that D(v) = d(u, v) for all v ∈ V . In particular, e(u) can be obtained as the
maximum value in the vector and the antipode of u as the index with highest rank were this
value appears in D. We may measure the complexity of an algorithm by the number of one-toall distance queries it performs when its cost mainly comes from these operations. A one-to-all
distance based algorithm accesses the graph only through one-to-all distance queries and relies
solely on distances known from queries, triangle inequality, and non-negativeness of distances for
bounding unknown distances.
Given a collection S of subsets of V such that ∪S∈S S = V , a covering with S is a sub-collection
C ⊆ S of sets such that their union covers all V : V ⊆ ∪S∈C S. (A set S ∈ S is said to cover elements
in S.) Recall that the set-cover problem consists in finding a covering of minimum size. We define
a packing for S as a subset P ⊆ V such that any set of S contains at most one element in P . The
denomination comes from the fact that elements of P correspond to pairwise disjoint subsets of
the dual collection S ∗ = {{S ∈ S | u ∈ S} : u ∈ V }. A hitting set for S is a set P that intersects
all sets of S. (Equivalently, a hitting set can be defined as a covering for S ∗ but it may be more
convenient to consider a collection rather than its dual.) We let π(S) denote the maximum size
of a packing for S, and κ(S) denote the minimum size of a covering with S. As a covering must
cover each element of a packing with distinct sets, we obviously have π(S) ≤ κ(S) (weak duality).
We say that a collection R is restricted compared to S if there exists a one-to-one mapping f from
R to S such that S ⊆ f (S) for all sets S ∈ R. Note that this mapping then turns any covering
with R into a covering with S and we thus have κ(S) ≤ κ(R). Similarly, a packing for S is also
a packing for R and we have π(S) ≤ π(R). In other words, restricting the sets of a collection to
smaller subsets increases maximum packing size and minimum covering size.
6
3
Lower and upper certificates for eccentricities
Our notion of certificate is based on the fact that knowing all distances from a given node x allows
to derive some bounds on the eccentricities of other nodes:
∀u ∈ V, d(u, x) ≤ e(u) ≤ d(u, x) + e(x).
(1)
The first inequality derives directly from the eccentricity definition while the second one is a consequence of the triangle inequality. A possibly tighter lower-bound of max{d(u, x), e(u) − d(u, x)}
could be used as in [27] but this optimization does not allow to reduce drastically certificate size
(see Section 4).
We say that a set L (resp. U ) of nodes is a lower certificate (resp. an upper certificate) of G
when it is used to obtain lower bounds (resp. upper bounds) of eccentricities in G. Given the
distances from a node u to all nodes in L ∪ U and the eccentricities of nodes in U , we have the
following lower and upper bounds for the eccentricity of any node u (as a direct consequence of
Inequation 1):
U
e (u) = minx∈U d(u, x) + e(x)
eL (u) ≤ e(u) ≤ eU (u), where
eL (u) = maxx∈L d(u, x)
A lower (resp. upper) certificate L (resp. U ) is said to be tight when eL (u) = e(u) (resp.
eU (u) = e(u)) for all u ∈ V . An all-eccentricty certificate is defined as a pair L, U of a tight lower
certificate L and a tight upper certificate U .
Given a bound D and a node x, we have d(u, x) + e(x) ≤ D if and only if u ∈ B[x, D − e(x)].
Given an upper certificate U we thus have eU (u) ≤ D if and only if u ∈ ∪x∈U B[x, D − e(x)]. A
diameter certificate is a set U such that eU (u) ≤ diam(G) for all u ∈ V . Equivalently it can be
defined as a covering with {B[x, diam(G) − e(x)] : x ∈ V } using balls whose radius equals diam(G)
minus eccentricity of the center (and identifying a ball with its center).
Similarly, given a lower certificate L and a bound R, we obviously have eL (u) ≥ R for all
nodes u whose coball B(u, R) intersects L (i.e., there exists a node in L at distance R at least
from u). We thus define a radius certificate as a L such that eL (u) ≥ rad(G) for all u ∈ V or
equivalently as a hitting set L for the collection {B(u, rad(G)) : u ∈ V } of coballs of radius rad(G).
As x ∈ B(u, rad(G)) if and only if u ∈ B(x, rad(G)), the collection of coballs of radius rad(G) is
its own dual, and a radius certificate L can equivalently be defined as a covering for this collection.
Note that a tight lower certificate can equivalently be defined as a hitting set for the collection
{B(u, e(u)) : u ∈ V }. Similarly, a tight upper certificate can equivalently be defined as a covering
with the collection {{u ∈ V | d(u, x) ≤ e(u) − e(x)} : x ∈ V }.
Examples. A path with 2k+1 nodes has a radius certificate with two nodes (the two extremities)
and a diameter certificate with one node (its mid-point). More generally, any graph G such that
diam(G) = 2 rad(G) has a one node diameter certificate (a center). It can be shown that any tree
has a radius certificate of two nodes (two well chosen leaves) while its centers (at most two nodes)
form a diameter certificate. A square grid has a radius certificate with four nodes (the corners)
while its centers (at most four nodes) form a diameter certificate. As an extreme example of graph
with large certificates, consider a cycle C. Its only radius and diameter certificates are both the
whole set of its nodes. More generally, the whole set of nodes is the only diameter certificate of
any graph where all nodes have same eccentricity (when radius equals diameter).
Hardness of approximation. Similarly to [9], we note that set cover can easily be encoded
with ball cover: given a collection S of subsets of V of an instance of the set-cover problem,
consider the split graph where the sets S of S form a clique and the elements x ∈ V form a stable
set so that the nodes x and S are adjacent if and only if x ∈ S. Without loss of generality, we
may assume that no subset of S equals V (otherwise, the problem is trivial), no set is empty
(otherwise, we can remove it) and that there exists two elements such that no set contains both
7
of them (if needed, we add a singleton {z} to S where z is a new dummy element added to V ).
In this graph, sets and elements have eccentricity 2 and 3 respectively. Any minimum diameter
certificate is a covering with balls of radius 1 or 0 (if centered on a set or an element). One
can easily transform it into a covering with balls of radius 1 centered at nodes corresponding to
sets (only). It then corresponds to an optimal solution of the original set-cover problem. Now
consider the complementary graph which is also a split graph where elements form a clique while
sets form a stable set and where x and S are adjacent if and only if x ∈
/ S. Similarly, a minimum
radius certificate for this complementary graph corresponds to a covering with coballs of radius
2 centered at sets and is also an optimal solution to the original set-cover problem. For that
graph, finding a radius certificate is equivalent to finding a tight lower certificate. The hardness
of set-cover approximation [16] thus implies that computing a minimum diameter (resp. radius
or tight lower) certificate is NP-hard and that no polynomial time algorithm can approximate it
with a factor (1 − o(1)) log n unless P = N P . Surprisingly, we will see that finding a minimum
tight upper certificate can be done in polynomial time.
4
Lower bound for radius computation
We first show that the notion of radius certificate is related to the minimum number of queries a
one-to-all distance based algorithm must perform.
Theorem 2 Given a graph G, if a one-to-all distance based algorithm for radius queries a set L of
nodes then L∪Antipoder (L) is a radius certificate for any ranking r. Such a radius algorithm thus
requires at least 12 |LOP T | one-to-all distance queries where |LOP T | = κ({B(u, rad(G)) : u ∈ V }) is
the minimum size of a radius certificate.
Proof. Consider a one-to-all distance based algorithm and let L denote the set of nodes it queries
for one-to-all distances. A proof of correctness of the algorithm allows to conclude that all nodes
have eccentricity rad(G) at least based on triangle inequality and the distances known to the
algorithm. That is for each node u, there is a node v such that we can prove d(u, v) ≥ rad(G)
based on triangle inequality and distances from nodes in L. Consider a node v such that the
proof uses a minimum number of triangle inequalities. If neither u nor v is in L, the proof
must use a triangle inequality d(u, v) ≥ |d(u, x1 ) − d(v, x1 )| for some node x1 and a proof of
|d(u, x1 ) − d(v, x1 )| ≥ rad(G). In the case |d(u, x1 ) − d(v, x1 )| = d(u, x1 ) − d(v, x1 ), we would have
a shorter proof d(u, x1 ) ≥ rad(G) in contradiction with the choice of v. We thus consider only the
case |d(u, x1 )−d(v, x1 )| = d(v, x1 )−d(u, x1 ). We then have a proof of d(v, x1 ) ≥ rad(G)+d(u, x1 ).
Either x1 ∈ L or the proof uses a node x2 such that |d(v, x2 ) − d(x2 , x1 )| ≥ rad(G) + d(u, x1 ). The
choice of v again implies |d(v, x2 ) − d(x2 , x1 )| = d(v, x2 ) − d(x2 , x1 ) (otherwise x2 would provide a
shorter proof). By repeating this argument, we deduce that a shortest proof of d(u, v) ≥ rad(G)
uses a sequence x1 , . . . , xp of nodes such that d(v, xi ) ≥ rad(G) + d(u, x1 ) + d(x1 , x2 ) + · · · +
d(xi−1 , xi ) for i = 1..p and xp ∈ L. Consider the antipode a = Antipoder (xp ). We then
have d(a, xp ) ≥ d(v, xp ) ≥ rad(G) + d(u, x1 ) + · · · + d(xp−1 , xp ). By triangle inequality, we
have d(u, a) ≥ d(a, xp ) − d(u, xp ) and d(u, xp ) ≤ d(u, x1 ) + · · · + d(xp−1 , xp ). We thus have
d(u, a) ≥ rad(G). In all cases, L ∪ Antipoder (L) must contain a node at distance rad(G) or more
from u, it is thus a radius certificate.
Note that a similar result does not hold for diameter certificate. A one-to-all distance based
algorithm for diameter could query a set U of node such that for any pair u, v there is x ∈ U
satisfying d(u, x)+d(x, v) ≤ diam(G) which implies d(u, v) ≤ diam(G) by triangle inequality. Note
that checking that a set U has this property requires quadratic time in general (under SETH) even
if U has size O(log n) (see the reduction from SAT to diameter computation in [25]). Our diameter
certificate definition requires that for each node u a single node x allows to bound all distances
d(u, v) for v ∈ V using d(u, v) ≤ d(u, x) + e(x). The reason for this stronger requirement is to
enable sub-quadratic time verification that a certificate is indeed a certificate when it has o(n)
size.
8
5
Radius computation and certification
We now propose a radius algorithm with complexity parametrized by the number of antipodes in
the input graph. Similarly to previous algorithms [27, 5], it maintains lower bounds on eccentricities of all nodes and performs one-to-all distance queries from nodes with minimal lower bound as
a first ingredient. Similarly to the two-sweeps and four-sweeps heuristics [20, 23, 8], it performs
one-to-all distance queries from antipodes of previous query source as a second ingredient. Contrarily to these heuristics, it iterates until an exact solution is obtained (together with a radius
certificate).
The idea of the algorithm is to maintain both a set K of nodes with distinct antipodes and
a lower certificate L (initally empty). We iteratively select a node u with minimal lower-bound
eL (u) and perform a one-to-all distance query from u. As long as this bound is not tight (i.e.,
eL (u) < e(u)), we add Antipoder (u) to L and u to K while eccentricity lower-bounds are improved
accordingly. (The fact that the bound is not tight implies that no antipode of u could previously
be in L.) As soon as the bound is tight (i.e., eL (u) = e(u)), we then claim that u is a center (i.e.,
its eccentricity is minimal) and return e(u) as the radius and L as radius certificate. Algorithm 1
formally describes the whole method.
Note the primal-dual flavor of this algorithm as the set K (which has same size as L) is a
packing for {Antipode−1
r (u) : u ∈ V } which is a restricted collection of {B(u, rad(G)) : u ∈ V } for
which the computed certificate L is a covering.
Input: A graph G and a ranking r of its node set V .
Output: The radius rad(G) of G, a center c and a radius certificate L.
L := ∅ /* Lower certificate (tentative covering with {B(u, rad(G)) : u ∈ V })
Maintain eL (v) = maxx∈L d(v, x) (initially 0) for all v ∈ V .
K := ∅ /* Packing for {Antipode−1
r (u) : u ∈ V }.
Do
Select u ∈ V such that eL (u) is minimal.
Du := DistFrom(G, u) /* Distances from u.
e(u) := maxv∈V Du (v) /* Eccentricity of u.
If e(u) = eL (u) then
return e(u), u, and L
else
a := argmaxv∈V (Du (v), r(v)) /* Antipode of u for r.
Da := DistFrom(G, a) /* Distances from a.
K := K ∪ {u}
L := L ∪ {a}
For v ∈ V do eL (v) := max(eL (v), Da (v))
*/
*/
*/
*/
*/
*/
while minu∈V eL (u) < minu∈K e(u).
Return e(c), c and L where c = argminu∈K e(u).
Algorithm 1: Computing the radius, a center and a radius certificate.
Theorem 3 Given a graph G and a ranking r on its node set V , Algorithm 1 computes its radius
rad(G), a center c and a radius certificate L ⊆ Antipoder (V ) with 2 |L| + 1 = O(|Antipoder (V )|)
one-to-all distance queries.
Proof. We first prove the termination of Algorithm 1. Consider an iteration where we add the
antipode a of the selected node u to L. We cannot have a ∈ L as we would then have eL (u) = e(u)
which is the termination case. In other words, nodes added to K have distinct antipodes and
K is a packing for {Antipode−1
r (u) : u ∈ V }. As long as the do-while loop runs, each iteration
adds a new node to L. If ever we reach the point where L = Antipoder (V ), then the lowerbound of each node u ∈ V is tight: eL (u) = e(u). The next iteration must then terminate. The
complexity is straightforward: at most 2 |L| + 1 one-to-all distance queries are performed and
|L| ≤ |Antipoder (V )| as L ⊆ Antipoder (V ).
9
We now prove the correctness of Algorithm 1. Consider an iteration of the do-while loop. By
the choice of u, we then have eL (v) ≥ eL (u) for all v ∈ V . If the termination case eL (u) = e(u)
occurs, we have e(v) ≥ eL (v) ≥ eL (u) = e(u) for all v ∈ V . This ensures that u has minimum eccentricity (it is a center). We thus have rad(G) = e(u) and L is a radius certificate
as eL (v) ≥ rad(G) for all v ∈ V . Finally, if ever the condition for continuing the do-while
loop is false, we have minu∈V eL (u) ≥ minu∈K e(u). For c = argminu∈K e(u), we thus have
rad(G) = minu∈V e(u) ≥ minu∈V eL (u) ≥ e(c) ≥ rad(G). That is, L is a radius certificate and c is
a center.
In practice, we observe very fast convergence compared to |Antipoder (V )| (see Section 9).
We can give the following argument for that. The node u selected at each iteration satisfies
eL (u) = minv∈V eL (v) ≤ minv∈V e(v) ≤ rad(G). We thus have maxx∈L d(u, x) ≤ rad(G), that is
u ∈ ∩x∈L B[x, rad(G)]. It appears that the eccentricity of antipodes is generally large compared
to radius in practical graphs, and the set ∩x∈L B[x, rad(G)] tends to quickly shrink toward the set
of centers as we add antipodes to L.
6
Minimum eccentricity selection
The core of the above radius algorithm is a general technique depending on a user-defined function
f that we call minimum eccentricity selection (minES) for f . It is a procedure that returns a node
with minimum eccentricity with respect to f . Its amortized complexity is low in graphs with
few antipodes. More precisely, for a given graph G and a function f that maps a node v and
an estimation ` of e(v) to a value, it provides a function argminES returning a node u such that
f (u, e(u)) is minimum as long as f is non-decreasing, i.e., f (v, `) ≤ f (v, `0 ) for ` ≤ `0 for all v. A
similar function minES returns the value of f (u, e(u)) for such node u. The challenge here is to
avoid the computation of all eccentricities.
L := ∅ /* Lower certificate.
Maintain eL (v) = maxx∈L d(v, x) (initially 0) for all v ∈ V .
Function argminES(G, r, L, eL , f )
Repeat
u := argminv∈V f (v, eL (v))
Du := DistFrom(G, u) /* Distances from u.
e(u) := maxv∈V Du (v) /* Eccentricity of u.
If eL (u) = e(u) then
return u
else
a := argmaxv∈V (Du (v), r(v)) /* Antipode of u for r.
Da := DistFrom(G, a) /* Distances from a.
L := L ∪ {a}
For v ∈ V do eL (v) := max(eL (v), Da (v))
*/
*/
*/
*/
*/
Function minES(G, r, L, eL , f )
u := argminES(G, r, L, eL , f )
Return f (u, eL (u))
Algorithm 2: Minimum eccentricity selection with respect to function f .
We implement such a selection by maintaining lower bounds of all eccentricities as in Algorithm 1 and by using these lower bounds as estimates for true eccentricities. When the selection
procedure is called, a node u which is minimum according to lower bounds is considered. Such a
node is found by evaluating f (v, eL (v)) for all v ∈ V where eL (v) denotes the lower bound stored
for a node v. A one-to-all distance query from u is then performed. If its eccentricity happens
to be equal to its lower-bound eL (u) we claim that f (u, e(u)) is minimum and return that node.
Otherwise, the antipode of u is used to improve lower bounds before trying again. Algorithm 2
10
formally describes this.
Proposition 1 Given a graph G and a ranking r of its node set V , we consider a function f such
that f (v, `) can be evaluated for any v ∈ V and ` ≤ e(v). If f (v, .) is non-decreasing for all v ∈ V ,
i.e., f (v, `) ≤ f (v, `0 ) for ` ≤ `0 , function argminES of Algorithm 2 returns a node u such that
f (u, e(u)) is minimal and updates the lower certificate L such that eL (u) = e(u) and f (v, eL (v)) ≥
f (u, e(u)) for all v ∈ V . Moreover it can perform k computations of argmin f (u, e(u)) using
k + 2 |L0 | one-to-all distance queries and (k + 2 |L0 |)n calls to f where L0 ⊆ Antipoder (V ) denotes
the set of nodes added to L.
Note that (x + 2 |L0 |)n calls to f represent less than O(x + 2 |L0 |) one-to-all distance queries
with respect to time cost when f can be evaluated in constant time.
Proof. The correctness of the selection comes from the fact that f (v, .) is non-decreasing: if
eL (u) = e(u), we then have f (u, e(u)) = f (u, eL (u)) ≤ minv∈V f (v, eL (v)) ≤ minv∈V f (v, e(v)).
The case eL (u) < e(u) can only occur if the antipode of u was not in L and happens at most
|Antipoder (V )| times in total. In particular, each call to argminES terminates. If an algorithm
makes x calls to the argminES, the number of successful iterations where eL (u) = e(u) is precisely
x while the number of unsuccessful iterations is at most the number of nodes added to L. For each
such iteration we perform 2 one-to-all distance queries instead of 1. The total number of queries
is thus k + 2 |L0 |. In all cases, we perform n calls to f per iteration.
As an example of usage, Algorithm 1 for radius is equivalent to the following algorithm using
our minimum eccentricity selection for the basic function v, ` 7→ `.
L := ∅; eL (v) := 0 for all v ∈ V .
Function ecc(v, `): return `
c := argminES(G, r, L, eL , ecc)
return eL (c), c, L
As another example, the function f can be used to select a node with minimum eccentricity
in a set W of nodes when f (v, `) returns ` for v ∈ W and ∞ otherwise. One can easily check
that f (v, .) is non-decreasing for all v ∈ V . We use our minimum eccentricity selection as an
optimization for diameter computation and as a core tool for computing all eccentricities in the
next sections.
7
Diameter computation and certification
We now analyze a simple diameter algorithm. The main ingredient of the algorithm consists in
maintaining upper bounds of all eccentricities and performing one-to-all distance queries from
nodes with maximum upper bound. It thus follows the main line of previous practical algorithms [27, 5]. However, we present the algorithm with a more general primal-dual approach
which was not noticed before. Moreover, we introduce a new technique called delegate certificate:
after selecting a node u with maximal upper bound, it consists in performing a one-to-all distance
query from any node x such that (d(u, x) + e(x) = e(u). We call such a node x a tight upper certificate for u as we have e{x} = e(u). A possible choice for x is u itself in which case the algorithms
becomes a very basic version of [27]. However, we observe that choosing a node x with minimal
eccentricity offers much better performances in practice (see Section 9). Our complexity analysis
is independent of the choice of x, we thus present the algorithm in the most general manner.
The algorithm grows both a packing K and an upper certificate U until the upper bound eU (u)
on the eccentricity of any node u is at most the maximum eccentricity of nodes in K. As long
as this condition is not satisfied, a node u with maximal upper bound is selected and added to
K. We then choose a tight upper certificate x for u and add it to U . Note that we now have
eU (u) = e(u) ≤ maxv∈K e(v) and u cannot be selected again. This ensures that the termination
condition is reached at some point when U is a certificate that all nodes have eccentricity at most
11
that of a maximum eccentricity node in K which must thus be equal to diameter. See Algorithm 3
for a formal description.
We claim that the set K is a packing for the collection D1/3 = {B(u, 31 (diam(G) − e(u))) : u ∈ V }
of open balls. As it has same size as the certificate U returned by the algorithm in the end, this
allows to state the following theorem.
Theorem 4 Given a graph G, Algorithm 3 computes the diameter of G, a diametral node b and a
diameter certificate U of size π1/3 at most with O(π1/3 ) one-to-all distance queries where πα is the
maximum packing size for the collection of open balls Dα = {B(u, α(diam(G) − e(u))) : u ∈ V }
π
where π[1] is the
for α > 0. It approximates minimum diameter certificate within a factor π1/3
[1]
maximum packing size for the collection D[1] = {B[u, diam(G) − e(u)] : u ∈ V }.
1
2
Input: A graph G and a ranking r of its node set V .
Output: The diameter diam(G) of G, a diametral node b and a diameter certificate U .
U := ∅ /* Upper certificate (tentative covering with
{B[u, diam(G) − e(u)] : u ∈ V }).
Maintain eU (u) = minx∈U d(u, x) + e(x) (initially ∞) for all v ∈ V .
K := ∅ /* Packing for {B(u, 13 (diam(G) − e(u))) : u ∈ V }.
Do
Select u such that eU (u) is maximal.
Du := DistFrom(G, u) /* Distances from u.
e(u) := maxv∈V Du (v) /* Eccentricity of u.
K := K ∪ {u}
Select x such that d(u, x) + e(x) = e(u). /* Delegate certificate for u.
Dx := DistFrom(G, x) /* Distances from x.
e(x) := maxv∈V Dx (v) /* Eccentricity of x.
U := U ∪ {x}
For v ∈ V do eU (v) := min(eU (v), Dx (v) + e(x))
while maxu∈K e(u) < maxu∈V eU (u)
*/
*/
*/
*/
*/
*/
*/
Return e(b), b and U where b ∈ K satisfies e(b) = maxu∈K e(u).
Algorithm 3: Computing diameter and a diameter certificate. The basic version of the
algorithm consists in selecting x := u in Line 1. (The redundant one-to-all distance query in
Line 2 can then be omitted.)
Proof.We already argued the termination and the correctness of the algorithm above. We thus
show the packing property of K. Suppose for the sake of contradiction that K is not a packing for D1/3 . Consider the first iteration where a node v is added to K while some open ball
B(y, 13 (diam(G) − e(y))) in D1/3 contains both v and some node u ∈ K added previously. Let x
be the tight upper certificate for u that was added to U . By triangle inequality, we have d(x, v) ≤
d(x, u) + d(u, y) + d(y, v). The choice of x implies d(x, u) = e(u) − e(x) ≤ d(u, y) + e(y) − e(x).
Combining the two inequalities, we obtain d(x, v) ≤ 2d(u, y) + d(y, v) + e(y) − e(x). As u and
v are in B(y, 31 (diam(G) − e(y))), we have 2d(u, y) + d(y, v) < diam(G) − e(y) and finally get
d(x, v) < diam(G) − e(x) which implies eU (v) < diam(G). However, it is required that v has
maximal upper bound eU (v) = maxw∈V eU (w) when it is selected for being added to K, in contradiction with maxw∈V eU (w) ≥ maxw∈V e(w) = diam(G). We conclude that K must be a packing
for D1/3 . Both sizes of K and U are bounded by π1/3 . As any diameter certificate is a covering
π
for D[1] and has size π[1] at least, this guarantees that the size of U is within a factor π1/3
at most
[1]
from optimum.
This analysis can be complemented when we start Algorithm 3 with K := {u} and U := {c}
initially where c is a center of the graph computed with Algorithm 1. We reference this combination
as Algorithm 1+3 in the sequel. A similar proof then allows to show that K is a packing for
12
c
D1/3
(G) = {B(u, βu (diam(G) − e(u))) : u ∈ V } where βu = 1/3 for u 6= c and βc = 1. We obtain
the following corollary from Theorems 3 and 4.
Corollary 1 Given a graph G and a ranking r of its node set V , Algorithm 1+3 computes
c
the diameter of G, a diametral node b and a diameter certificate U of size π1/3
at most with
|U | + 2 |Antipoder (V )| + 1 one-to-all distance queries at most where c is a center of G returned
c
c
c
by Algorithm 1 and π1/3
= π(D1/3
) is the maximum packing size for the collection D1/3
=
1
{B(u, βu (diam(G) − e(u))) : u ∈ V } of open balls with radii factors βu = 3 for u 6= c and βc = 1
(for u = c).
Proof. In addition to the proof of Theorem 4, we just have to consider the case when a node
v would be added to K while having v ∈ B(c, diam(G) − e(c)). As c ∈ U , we then have
eU (v) ≤ d(c, v) + e(c) < diam(G). This would raise a contradiction as the choice of v relies
on eU (v) = maxw∈V eU (w) ≥ maxw∈V e(w) ≥ diam(G).
This explains efficiency of practical algorithms as we observe that coverings of small size ofc
ten exist for D1/3
(G) in practical graphs (see Section 9). As mentioned before, a further optimization consists in selecting a tight upper certificate x for u with minimal eccentricity. Using
a function f such that f (v, `) returns ` when Du (v) + ` ≤ e(u) and returns ∞ otherwise, it
can be obtained through our minimum eccentricity selection procedure by replacing Line 1 with
x := argminES(G, r, L, eL , f ). The algorithm is referenced as Algorithm 1+3’ in the sequel. This
optimization through the delegate certificate technique provides performances similar to previous
practical algorithms (see Section 9) while providing the above complexity guarantees (Corollary 1
also applies to this variant).
8
All eccentricities
We now present a novel algorithm for all eccentricities. It relies on minimum eccentricity selection
and properties of tight upper certificates.
8.1
Optimal tight upper certificate
We first characterize the minimum tight upper certificate of a graph G which is tightly related to
the notion of tight upper certificate.
Proposition 2 Given a graph G, being a tight upper certificate defines a binary relation which
is a partial order (u x stands for e(u) = d(u, x) + e(x)). Moreover, the set U of all maximum
elements of this partial order is the unique tight upper certificate of G with minimum size.
Proof. We first prove that the relation is a partial order. It is obviously reflexive as the
distance from a node to itself is zero, implying e(u) = d(u, u) + e(u). It is antisymmetric: if x
and y are both tight upper certificates one for the other, we then have e(x) = d(x, y) + e(y) and
e(y) = d(y, x) + e(x), and thus d(x, y) = 0. We finally show transitivity. Suppose that y is a tight
upper certificate for x and that z is an tight upper certificate for y, that is e(x) = d(x, y)+e(y) and
e(y) = d(y, z) + e(z). We thus have d(x, y) + d(y, z) + e(z) = e(x). As triangle inequality implies
e(x) ≤ d(x, z) + e(z), we obtain d(x, y) + d(y, z) ≤ d(x, z), and thus d(x, y) + d(y, z) = d(x, z) by
triangle inequality again. We finally get e(x) = d(x, z) + e(z) and z is a tight upper certificate for
x.
We now show that the set U of maximal elements for is the unique optimal tight upper certificate of G. For any non-maximal element u, we can build a chain u x1 x2 · · · where u has
a tight upper certificate x1 , if x1 is not in U , it has a tight upper certificate x2 , and so on. As the
partial order is finite, the chain must be finite and xk must be in U for some k. The transitivity
of implies that xk is a tight upper certificate for u implying eU (u) ≤ d(u, xk ) + e(xk ) = e(u).
13
This shows that U is a tight upper certificate. As each element x of U is the only tight upper
certificate for itself (as a maximal element), U is included in any tight upper certificate of G. In
particular, any minimum tight upper certificate must indeed equal U .
Note that U includes in particular all centers of the graph: a center c cannot have a tight
upper certificate x 6= c (otherwise we have e(x) = e(c) − d(c, x) < e(c) in contradiction with the
minimality of e(c)).
8.2
All eccentricity computation and certification
We now propose to compute all eccentricities of a graph as follows (see Algorithm 4 for a formal
description). We maintain both a lower certificate L and an upper certificate U . As long as
some node has untight upper bound, we select a node u with untight upper bound and minimal
eccentricity using our minimum eccentricity selection procedure which then additionally ensures
eL (u) = e(u). (We use for that purpose a function returning ∞ when the eccentricity value equals
the upper bound.) We claim that u is in U (see Lemma 1 bellow). We thus add u to the upper
certificate U and update upper bounds accordingly. When our minimum eccentricity selection
procedure detects that all nodes have tight upper bounds, lower bounds must be tight also. The
algorithm then terminates with the following guarantees.
Input: A graph G and a ranking r of V .
Output: All eccentricities, a tight lower certificate L of G and a tight upper certificate U
of G.
L := ∅ /* Lower certificate (tentative hitting set for {B(u, e(u)) : u ∈ V }) */
Maintain eL (v) = maxx∈L d(v, x) (initially 0) for all v ∈ V .
U := ∅ /* Upper certificate (maximal nodes for )
*/
Maintain eU (u) = minx∈U d(u, x) + e(x) (initially ∞) for all v ∈ V .
Function ecc-untight(v, `)
If ` < eU (v) then return ` else return ∞
While minES(L, eL , ecc-untight) < ∞ do
u := argminES(G, r, L, eL , ecc-untight)
Du := DistFrom(G, u) /* Distances from u.
*/
e(u) := maxv∈V Du (v) /* Eccentricity of u.
*/
U := U ∪ {u}
For v ∈ V do eU (v) := min(eU (v), Du (v) + e(u))
Return eU , L, U
Algorithm 4: Computing all eccentricities and tight lower/upper certificates.
Theorem 5 Given a graph G and a ranking r of its node set V , Algorithm 4 computes all eccentricities, a tight lower certificate L ⊆ Antipoder (V ) and the optimal tight upper certificate U
with U + 2 |L| one-to-all distance queries.
In practical graphs, we observe that the size of U is much larger than that of the computed
lower certificate L and the algorithm roughly costs U BFS traversals (see Section 9).
The correctness of Algorithm 4 mainly rely on the following lemma.
Lemma 1 Consider an upper certificate U and the set SU of nodes that do not have a tight upper
certificate in U . Any node v ∈ SU with minimum eccentricity (having e(v) = minu∈SU e(u)) is its
unique tight upper certificate (i.e., v ∈ U ).
Proof. For the sake of contradiction, suppose that v has a tight upper certificate x 6= v. As
e(v) = d(v, x) + e(x), we have e(x) < e(v) and x cannot be in SU by the minimality of e(v). It
thus has a tight upper certificate y ∈ U . But the transitivity of being a tight upper certificate
(Proposition 2) implies that y is also a tight upper certificate for v which is in contradiction with
14
type
comm
comm
game
geom
road
road
road
road
road
soc
soc
soc
soc
soc
synth
synth
synth
synth
synth
vlsi
web
web
web
name
n
m/n
d
w
diam
rad
D
c
π0.8
c
π1/3
nc /n
R
AID
F
Gnutella
skitter
FrozenSea
buddha
CAL-d
CAL-t
CAL-u
FLA-t
europe-t
Epinions
Hollywood
Slashdot
Twitter
dblp
bowtie500
grid1500-wd
grid500-10
pwlaw2.5
udg10
alue7065
BerkStan
Indochina
NotreDame
14149
1694616
753343
543652
1890815
1890815
1890815
1070376
18010173
32223
1069126
71307
68413
226413
505002
296680
250976
1000000
999888
34046
334857
3806327
53968
3.60
13.09
7.70
6.00
2.45
2.45
2.45
2.51
2.34
13.76
106.33
12.80
24.63
6.33
2.00
1.66
3.59
3.85
9.99
3.22
13.51
25.96
5.65
•
◦
◦
◦
◦
◦
◦
◦
•
•
◦
•
•
◦
◦
•
◦
◦
◦
◦
•
•
•
◦
◦
•
•
•
•
◦
•
•
◦
◦
◦
◦
◦
◦
•
◦
◦
◦
◦
◦
◦
◦
1.58
1.94
1.80
1.87
1.89
1.83
1.99
1.99
1.99
2
1.71
1.86
2.50
2
1.99
2.38
2
1.90
1.99
2
2.73
6.91
2.11
19
3
15
27
3
7
2
2
2
2
29
4
2
1
3
3
1
4
5
1
2
2
2
47
5
35
63
17
17
4
4
4
4
603
40
6
11
2001
12
5
24
17
5
7
4
5
1912
7
191
385
90
105
6
4
5
7
4183
65
8
20
4001
69
5
24
86
5
28
6
22
0.27
0.99
0.91
0.95
0.99
0.98
0.99
0.99
1
0.99
0.99
0.99
0.99
1
0.99
0.04
1
0.99
0.99
1
2e-03
0.99
0.01
4
3
7
14
3
5
3
2
2
2
3
4
4
5
5
4
3
2
4
3
3
3
2
10
6
384
897
11
13
7
2
7
34
12
31
6
5
5
4
19
5
4
16
2
23
6
388
897
11
13
11
2
20
335
135
4753
43
1507
6
4
47
10
4
17
2
Table 1: Diameter and radius certificate sizes (D, R) for various graphs, and related parameters.
v ∈ SU .
Proof.[of Theorem 5] We first prove U ⊆ U . As in Lemma 1, let SU denote the set of nodes
that do not have a tight upper certificate in U (SU = {v ∈ V | e(v) < eU (v)}). Now consider the
node u selected at some iteration of the while loop. We prove u ∈ U \ U . The correctness of
our minimum eccentricity selection (Proposition 1) implies that u has minimum eccentricity in
SU and is thus in U by Lemma 1. (Note that ecc-untight(v, .) is non-decreasing for all v ∈ V as
ecc-untight(v, `) = ` for ` < e(v) and ecc-untight(v, `) = ∞ for ` ≥ e(v).) Additionally, u ∈ SU
implies that u has untight upper bound and is not in U until we add it at that iteration.
The termination of the algorithm is guaranteed by the fact that U grows at each iteration. The algorithm ends when minimum eccentricity selection returns a node u such that
ecc-untight(u, e(u)) = ∞. Proposition 1 then ensures ecc-untight(v, eL (v)) = ∞ for all v. That
is eL (v) = eU (v) for all v and both bounds must equal e(v). This implies that L (resp. U ) is
then a tight lower (resp. upper) certificate of G. Moreover, U ⊆ U then implies U = U by
Proposition 2.
9
Experiments
We test social networks (Epinions, Hollywood, Slashdot, Twitter, dblp), computer networks
(Gnutella, Skitter), web graphs (BerkStan, IndoChina, NotreDame), road networks (CAL-t, CALd, CAL-u, FLA-t, europe-t), a 3D triangular mesh (buddha), and grid like graphs from VLSI
applications (alue7065) and from computer games (FrozenSea). The data is available from snap.
stanford.edu, webgraph.di.unimi.it, www.dis.uniroma1.it/challenge9, graphics.stanford.
edu, steinlib.zib.de and movingai.com. We also test synthetic inputs: bowtie500 is the graph
BT500,500 represented in Figure 1, grid500-10 is a 501 × 501 square grid with random deletion of
10% of the edges, grid1500-wd is a weighted directed graph obtained from a 1501 × 1501 square
grid where each edge is oriented randomly (with probability 1/2 for each direction) and assigned a
15
random weight uniformly in {0, 1, . . . , 9}, pwlaw2.5 is a random powerlaw graph where the number
of nodes of degree k is proportional to k −2.5 , udg10 is a random unit disk graph where field size is
parametrized to obtain average degree 10 roughly. Each graph is restricted to its largest (strongly)
connected component.
Table 1 summarizes our main practical observations. For each instance G, we show its type,
the number n of nodes in the largest (strongly) connected component, the average out-degree m/n,
whether it is directed (d) and weighted (w), and the diameter to radius ratio diam(G)
rad(G) . We then
show the size D of the diameter certificate computed by Algorithm 1+3’, bounds on maximum
c
c
packing sizes π0.8
and π1/3
(defined in Section 7), the proportion nc /n of nodes in the (in-)ball
of radius diam(G) − rad(G) centered at a center c, the size R of the radius certificate computed
by Algorithm 1, the number AID of antipodes for ID ranking and the number F of furthest
nodes. The two latter numbers were obtained by performing a traversal per node of the graph
(in quadratic time). A dash indicates a value that could not be obtained in less than few days of
computation.
The first observation is that diameter and radius certificates are extremely small for all instances
(less than 30 nodes for all of them). Several observations allow to explain this phenomenon. First,
all graphs have high diameter to radius ratio (over 1.5 for all of them). Note that this ratio is at
most 2 for an undirected graph (it is unbounded in general directed graphs). Undirected graphs
with ratio 2 have a one node diameter certificate: a center. This concerns two practical graphs
while several ones have ratio very close to 2. Coherently, the concentration of nodes around the
center is also high with respect to the diameter minus radius difference. This is measured by
the ratio nc /n where c is a center computed by our radius algorithm (e(c) = rad(G)) and nc
denotes the number of nodes in B[c, diam(G) − rad(G)] (in directed graphs we count the number
of nodes u such that d(u, c) ≤ diam(G) − rad(G)). It counts the proportion of nodes u such that
e{c} (u) ≤ diam(G) which appears to be very close to 1 for most of the graphs. Notable exceptions
are Gnutella, BerkStan and NotreDame. This may be explained by the low (compared to others)
diameter to radius ratio (1.58) of the first one and probably to the highly asymmetric nature of the
two others. Seeing diameter certification as a covering problem with balls B[x, diam(G) − e(x)],
there are thus few nodes that are not covered by a center c. Additionally, other nodes can be
c
c
covered using few balls with reduced radii: the columns π0.8
and π1/3
indicate the size of coverings
we could find using balls with radii reduced by a factor .8 and 1/3 respectively. These numbers
c
c
upper bound maximum packing sizes π0.8
and π1/3
of the associated collections of balls with
c
reduced radii. Our theoretical upper-bound of π1/3 thus explains fast diameter computation for
most of the graphs. A notable exception is BT500,500 alias bowtie500 which was tailored for making
c
former diameter algorithms slow (including Algorithm 1+3) and thus have large π1/3
value. Other
exceptions are Hollywood and Gnutella for which the diameter to radius ratios are not so high
either (compared to other graphs). However the parameter is still much smaller than the number
of nodes.
Concerning radius computation, we observe that most graphs have very few antipodes as
indicated by the AID column although the number F of furthest nodes can be much larger (as in
the Twitter graph). A notable exception is the buddha graph which is a triangulated 3D surface
and thus has more or less a sphere like topology (the arms form handles) that may explain why all
furthest nodes are antipodes. However the number of antipodes remains much smaller than the
number of nodes. ForzenSea also has a relatively large number of furthest nodes that are almost
all antipodes. This might come from the design of the graph as a map where players of a video
game evolve and should find dead ends.
10
Graphs with low doubling dimension
A graph is γ-doubling if every ball is included in the union of at most γ balls with half radius.
16
10.1
Exact diameter computation
In the case of γ-doubling graphs, the following theorem shows that we can obtain a diameter
certificate with linear size compared to the maximum packing size π(Dα ) for the collection Dα
(see Section 7). This complements Theorem 4 in the range 31 ≤ α < 1.
Theorem 6 Given a γ-doubling graph G and α < 1, the diameter diam(G), a diametral node p
1
and a diameter certificate U satisfying |U | ≤ πα γ O(1)+log 1−α log diam(G) can be computed with
2 |Antipode(V )| + |U | one-to-all distance queries where πα is the maximum size of a packing for
the collection Dα = {B(u, α(diam(G) − e(u))) : u ∈ V }.
Note that this implies that minimum diameter certificate can be approximated within a factor
1
1
γ O(1)+log(diam(G)−rad(G)) ππ[1]
when G is γ-doubling as Dα = D1 for α > 1 − r+1
where r =
diam(G) − diam(R) is the maximum radius of a ball in D1 (recall that π[1] lower bounds the size
of a minimum diameter certificate).
The above theorem is a consequence of Algorithm 5 which follows a primal-dual approach
by constructing both a packing K for Dα together with a covering U with D[1] such that |U | ≤
α
|K| γ O(1)+log 1−α log diam(G). Given a node u, let Su = {v ∈ V | ∃B ∈ Dα s.t. u, v ∈ B} denote
the set of nodes v that cannot be in packing for Dα containing u. The idea is to iteratively add
a node u to K with highest eccentricity according to eU and then to add sufficiently many nodes
to U so that any node v in Su gets an eccentricity upper bound eU (v) equal to diam(G) or less.
This will guarantee that no such node is added later to K and that K is a packing for Dα .
Input: A graph G and a parameter α with 0 < α < 1.
Output: The diameter diam(G) of G and a diameter certificate p, U .
K := ∅ /* Packing for {B(u, α(diam(G) − e(u))) : u ∈ V }.
U := ∅ /* Upper certificate.
Maintain eU (u) = minx∈U d(u, x) + e(x) (initially ∞) for all v ∈ V .
L := ∅ /* Lower certificate.
Maintain eL (v) = maxx∈L d(v, x) (initially 0) for all v ∈ V .
While maxp∈K∪P e(p) < maxu∈V eU (u) do
Select u such that eU (u) is maximal.
Du := DistFrom(u)
e(u) := maxw∈V Du (w) /* Eccentricity of u.
K := K ∪ {u}
U := U ∪ {u}
For w ∈ V do eU (w) := min(eU (w), Du (w) + e(u))
Function ecc-slack-far(v, `)
If eU (v) − ` > 1−α
2α Du (v) then return −Du (v) else return ∞
While minES(G, r, L, eL , ecc-slack-far) < ∞ do
v := argminES(G, r, L, eL , ecc-slack-far)
Dv := DistFrom(v)
e(v) := maxw∈V Dv (w) /* Eccentricity of v.
U := U ∪ {v}
For w ∈ V do eU (w) := min(eU (w), Dv (w) + e(v))
*/
*/
*/
*/
*/
p := argmaxp∈K∪P e(p)
Return e(p) and p, U .
Algorithm 5: Computing diameter and a diameter certificate assuming doubling property.
Our selection rule for adding nodes to U is based on comparing their eccentricity to their
distance to u. The rough idea is to add a node v as certificate in U when eU (v) > diam(G)
and diam(G) − e(v) = Ω(d(u, v)). Note that any node w ∈ B[v, diam(G) − e(v)] then satisfies
eU (w) ≤ diam(G) and the doubling property will allow us to bound the number of nodes added to
U . As the eccentricity of v is not known precisely until we perform a one-to-all distance query from
17
v, we use our minimum eccentricity selection technique using a lower certificate L. As diam(G)
is not known either, we select v such that eU (v) − e(v) = Ω(d(u, v)). This ensures that a ball of
radius Ω(d(u, v)) will then be covered, we thus prefer v such that d(u, v) is additionally maximal.
Proof.[of Theorem 6] The main arguments of the proof are the following.
The function ` 7→ ecc-slack-far(v, `) is non-decreasing as it returns −d(u, v) for ` <= eU (v) −
1−α
2α d(u, v) and ∞ otherwise. Note that the minimum eccentricity selection for ecc-slack-far thus
returns a node v such that eU (v) − e(v) ≥ 1−α
2α d(u, v) and d(u, v) is maximal.
1
We first prove that the inner loop performs at most γ O(1)+log 1−α log diam(G) iterations. This
bounds the number of
nodes added to U when one node is added to K and will thus ensures
1
|U | ≤ |K| γ O(1)+log 1−α log diam(G). Consider an iteration of the main loop where u is added
ρ
d(u, v) where
to K. Just after adding v to U in the inner loop, consider w s.t. d(v, w) ≤ 2+ρ
1−α
U
ρ = 2α . We then have e (w) ≤ d(v, w) + e(v) as v ∈ U and e(w) ≥ e(v) − d(v, w) by triangle
2ρ
2
inequality. This gives eU (w) − e(w) ≤ 2+ρ
d(u, v). As d(u, w) ≥ d(u, v) − d(v, w) ≥ 2+ρ
d(u, v),
ρ
U
we get e (w) − e(w) ≤ ρd(u, w) and nodes in B[v, 2+ρ d(u, v)] do not satisfy the condition of the
inner loop. The doubling property implies that the number of iterations where we select v such
1
O(1)+log 1−α
that d(u, v) > e(u)
. Then we may select v such that d(u, v) > e(u)
2 is at most γ
4 during
the same number of iterations at most, and so on until we eventually select u itself (the only
node v such d(u, v) = 0). The overall number of iterations of the inner loop is thus bounded by
1
γ O(1)+log 1−α log e(u).
For the sake of contradiction, suppose that K is not a packing and consider u, u0 ∈ K and x ∈ V
such that both u and u0 are in B(x, α(diam(G) − e(x))). Assume without loss of generality that u
was added to K before u0 . After the inner loop for u, we have eU (x)−e(x) ≤ 1−α
2α d(u, x). There thus
U
0
exists y ∈ U such that d(x, y) + e(y) ≤ e(x) + 1−α
d(u,
x).
We
thus
have
e
(u
) ≤ d(u0 , y) + e(y) ≤
2α
1−α
0
0
0
d(u , x) + d(x, y) + e(y) ≤ d(u , x) + e(x) + 2α d(u, x). As u, u ∈ B(x, α(diam(G) − e(x))), we
1−α
U
0
get eU (u0 ) < 1+α
2 diam(G) + 2 e(x). As e(x) ≤ diam(G), we obtain e (u ) < diam(G). This is
0
U
0
U
a contradiction since the choice of u implies e (u ) = maxv∈V e (v) ≥ diam(G).
10.2
Approximating radius and diameter
Interestingly, the following lemma links the gap between the bound provided by a lower/upper
certificate for a node u and the distance from the certificate to a tight lower/upper certificate for
u. Recall that a tight upper certificate for u is a node x such that e(u) = d(u, x) + e(x). We
similarly define a tight lower certificate for u as a node x such that e(u) = d(u, x) (equivalently, x
is a furthest node from u).
Lemma 2 Given a lower certificate L (resp. an upper certificate U ) and a node u, we have
e(u) − eL (u) ≤ d(x, L) (resp. eU (u) − e(u) ≤ 2d(x, U )) for any tight lower (resp. upper) certificate
x for u.
Proof. Consider a tight lower certificate x for a node u (e(u) = d(u, x)). Let y ∈ L be the
closest node to x in L (d(x, y) = d(x, L)). By triangle inequality, we have e(u) = d(u, x) ≤
d(u, y) + d(y, x) ≤ eL (u) + d(x, L).
Similarly, consider a tight upper certificate x for a node u (e(u) = d(u, x) + e(x)). Let
y ∈ U be the closest node to x in U (d(x, y) = d(x, U )). By triangle inequality, we have
e(u) = d(u, x) + e(x) ≥ d(u, y) − d(x, y) + e(y) − d(x, y) ≥ eU (u) − 2d(x, U ).
Now consider the choice of a node u with minimal eccentricity lower bound in Algorithm 1.
This choice implies eL (u) ≤ rad(G) and Lemma 2 then implies e(u) ≤ rad(G) + d(a, L) where a
is the antipode of u which is added to L (if u is not a center). As long as the selected node has
eccentricity greater than (1 + ε) rad(G), the nodes in L are ε rad(G) far apart and form a packing
for the collection of balls of radius ε rad(G)/2. Similarly, the choice of a node u with maximal
18
eccentricity upper bound in Algorithm 3 implies eU (u) ≥ diam(G) and Lemma 2 then implies
e(u) ≥ diam(G) − 2d(x, U ) where x is the tight upper certificate chosen for u that is added to U .
As long as the selected node has eccentricity less than (1 − ε) diam(G), the nodes in U form a
packing for the collection of balls of radius 2ε diam(G)/2. As the doubling property implies that
2
such packings have size bounded by γ dlog ε e+1 , we obtain the following approximation results for
radius and diameter.
Proposition 3 Given a γ-doubling graph G and ε > 0, Algorithm 1 (resp. Algorithm 3) provides
2
a node u with eccentricity (1+ε) rad(G) at most (resp. (1−ε) diam(G) at least) with O(γ dlog ε e+1 )
traversals.
Proof. As discussed above, the set L is a packing for balls of radius ε rad(G) until a node whose
eccentricity approximates the radius is found. We use the fact that packing size is bounded by
covering size. By the γ-doubling property, the whole graph can be covered by γ i balls of radius
2 rad(G)
as any ball of radius 2 rad(G) contains all nodes. For i ≥ log 2ε + 1 these balls have radius
2i
ε rad(G)/2 at most. Using that packing size is bounded by covering size, we can bound the number
of iterations of Algorithm 1 where chosen nodes u have eccentricity greater than (1 + ε) rad(G).
2
If we stop the algorithm after γ dlog ε e+1 iterations, the node u ∈ K with smallest eccentricity is
guaranteed to have eccentricity (1 + ε) rad(G) at most. The argument for diameter approximation
is similar.
11
Negatively curved graphs
In this section we analyze the behavior of our algorithms in the class of negatively curved graphs,
alias, δ-hyperbolic graphs.
Let (X, d) be a metric space and w ∈ X. The Gromov product of y, z ∈ X with respect to w
is defined to be
1
(y|z)w = (d(y, w) + d(z, w) − d(y, z)).
2
Let δ ≥ 0. A metric space (X, d) is said to be δ-hyperbolic [18] if
(x|y)w ≥ min{(x|z)w , (y|z)w } − δ
for all w, x, y, z ∈ X. Equivalently, (X, d) is δ-hyperbolic if for any four points u, v, x, y of X, the
two larger of the three distance sums d(u, v) + d(x, y), d(u, x) + d(v, y), d(u, y) + d(v, x) differ by
at most 2δ ≥ 0. A graph G = (V, E) is δ-hyperbolic if the associated shortest path metric space
(V, d) is δ-hyperbolic.
From the definition of a δ-hyperbolic graph, we immediately get the following simple but very
useful auxiliary lemma.
Lemma 3 Let G = (V, E) be a δ-hyperbolic graph. For every vertices c, v, x, y ∈ V , d(x, v) −
d(x, y) ≥ d(c, v) − d(y, c) − 2δ or d(y, v) − d(x, y) ≥ d(c, v) − d(x, c) − 2δ holds.
Proof. Assume, without loss of generality, that d(x, c) + d(v, y) ≤ d(y, c) + d(x, v). If also
d(c, v)+d(x, y) ≥ d(y, c)+d(x, v) then, by δ-hyperbolicity of G, d(c, v)+d(x, y)−d(y, c)−d(x, v) ≤
2δ, i.e., d(x, v) − d(x, y) ≥ d(c, v) − d(y, c) − 2δ. If d(c, v) + d(x, y) ≤ d(y, c) + d(x, v), then
d(x, v) − d(x, y) ≥ d(c, v) − d(y, c) ≥ d(c, v) − d(y, c) − 2δ.
As easy corollaries we get two useful results known also from [7]. Denote by F (s) := {v ∈ V :
d(s, v) = e(s)} the set of vertices furthest from s.
Corollary 2 For every δ-hyperbolic graph G, diam(G) ≥ 2 rad(G) − 4δ − 1.
19
Proof. Let x, y be vertices of G such that d(x, y) = diam(G). Let c be a middle vertex of any
shortest path connecting x with y. Apply Lemma 3 to c, v, x, y, where v ∈ F (c). Without loss of
generality, assume that d(x, v) − d(x, y) ≥ d(c, v) − d(y, c) − 2δ holds. Then, since d(x, y) ≥ d(x, v),
d(y, c) ≥ d(c, v) − 2δ. Hence, d(x, y) = d(x, c) + d(c, y) ≥ 2d(c, y) − 1 ≥ 2d(c, v) − 4δ − 1 =
2e(c) − 4δ − 1 ≥ 2 rad(G) − 4δ − 1.
Corollary 3 Let G = (V, E) be a δ-hyperbolic graph. For every vertices c, v ∈ V such that
v ∈ F (c), e(v) ≥ diam(G) − 2δ ≥ 2 rad(G) − 6δ − 1.
Proof. Apply Lemma 3 to c, v and vertices x, y such that d(x, y) = diam(G). Without loss of
generality, assume that d(x, v) − d(x, y) ≥ d(c, v) − d(y, c) − 2δ holds. Then, since d(c, v) ≥ d(c, y),
e(v) ≥ d(x, v) ≥ d(x, y) + d(c, v) − d(y, c) − 2δ ≥ d(x, y) − 2δ = diam(G) − 2δ.
We are ready to analyze Algorithm 1. Let ui and ai ∈ F (ui ) be the vertices picked in iteration
i of the do-while loop. Let Ki := {u1 , u2 , . . . , ui } and Li := {a1 , a2 , . . . , ai }. According to the
algorithm, u1 is picket arbitrarily (as initially L = ∅), a1 is the vertex furthest from u1 , u2 = a1
(as L1 = {a1 } is a singleton), a2 is a vertex most distant from u2 = a1 , u3 is a middle vertex of
a shortest (a1 , a2 )-path. By Corollary 3, we already have d(a1 , a2 ) ≥ diam(G) − 2δ. We can also
show that e(u3 ) ≤ rad(G) + 3δ.
Proposition 4 If G is a δ-hyperbolic graph, then d(a1 , a2 ) ≥ diam(G) − 2δ and e(u3 ) ≤ rad(G) +
3δ.
Proof. We need only to estimate the eccentricity of the vertex u3 . As u3 is a middle vertex of
a shortest (a1 , a2 )-path, min{d(c, a1 ), d(c, a2 )} ≥ b d(a12,a2 ) c ≥ b diam(G)
c − δ. Now, without loss
2
of generality, assume (see Lemma 3) that for vertices u3 , a3 , a2 , a1 ∈ V , d(a2 , a3 ) − d(a2 , a1 ) ≥
d(u3 , a3 ) − d(a1 , u3 ) − 2δ holds. Then, e(u3 ) = d(u3 , a3 ) ≤ d(a2 , a3 ) − d(a2 , a1 ) + d(a1 , u3 ) + 2δ =
d(a2 , a3 ) − (a2 , u3 ) + 2δ ≤ diam(G) − b diam(G)
c + δ + 2δ = d diam(G)
e + 3δ ≤ d 2 rad(G)
e + 3δ =
2
2
2
rad(G) + 3δ.
Thus, in δ-hyperbolic graphs, a vertex with eccentricity at most rad(G) + 3δ and a pair of
vertices that are at least diam(G) − 2δ apart from each other are found by Algorithm 1 in at most
3 iterations, i.e., in linear time. Note that a similar linear time algorithm was reported already in
[7]: in δ-hyperbolic graphs, a vertex with eccentricity at most rad(G) + 5δ and a pair of vertices
that are at least diam(G) − 2δ apart from each other can be found in linear time.
Next, we show that a vertex with eccentricity at most rad(G) + 2δ is found by Algorithm 1
in at most 2δ + 2 iterations. Consider iteration i ≥ 3 and let p0 , p00 be vertices of Li−1 with the
largest distance, i.e., d(p0 , p00 ) = max{d(x, y) : x, y ∈ Li−1 } =: diam(Li−1 ). If e(ui ) > rad(G) + 2δ
then, applying Lemma 3 to ui , ai , p0 , p00 ∈ V , we get d(p0 , ai ) − d(p0 , p00 ) ≥ d(ui , ai ) − d(p00 , ui ) − 2δ
or d(p00 , ai ) − d(p0 , p00 ) ≥ d(ui , ai ) − d(p0 , ui ) − 2δ. Hence, max{d(p0 , ai ) − d(p0 , p00 ), d(p00 , ai ) −
d(p0 , p00 )} ≥ e(ui ) − rad(G) − 2δ > 0 (as max{d(p00 , ui ), d(p0 , ui )} ≤ minu∈V eLi−1 (u) ≤ rad(G)).
That is, if e(ui ) > rad(G)+2δ then diam(Li ) > diam(Li−1 ). As diam(L2 ) = d(a1 , a2 ) ≥ diam(G)−
2δ, in at most 2δ +2 iterations of the while-loop of Algorithm 1 we will get diam(Li ) = diam(Li−1 )
(with i ≤ 2δ + 2) and hence e(ui ) ≤ rad(G) + 2δ must hold. Thus, we proved the following
proposition.
Proposition 5 If G is a δ-hyperbolic graph, then there is an index i ≤ 2δ + 2 such that e(ui ) ≤
rad(G) + 2δ. Furthermore, e(uj ) ≤ rad(G) + 2δ for all j ≥ 2δ + 2.
The second part of Proposition 5 says that all ui vertices generated by Algorithm 1 after
2δ + 1 iterations have eccentricity at most rad(G) + 2δ. Hence, in δ-hyperbolic graphs where the
set C 2δ (G) := {c ∈ V : e(c) ≤ rad(G) + 2δ} has cardinality bounded by some function g(δ),
20
depending only on δ, our algorithm will produce a vertex with eccentricity rad(G) (i.e., a central
vertex) in at most g(δ) + 2δ + 1 iterations.
Next we show that the set C 2δ (G) of a δ-hyperbolic graph has bounded diameter. Earlier, it
was known that diam(C(G)) ≤ 4δ + 1 and there exists a vertex c ∈ V such that d(v, c) ≤ 5δ + 1
for every v ∈ C(G) [7].
Proposition 6 If G is a δ-hyperbolic graph, then for every x, y ∈ C 2δ (G), d(x, y) ≤ 8δ + 1.
Furthermore, there is a vertex c ∈ V such that d(v, c) ≤ 6δ for every v ∈ C 2δ (G).
Proof. Let c be a middle vertex of any shortest path connecting x with y. Apply Lemma 3 to
c, v, x, y, where v ∈ F (c). Without loss of generality, assume that d(x, v) − d(x, y) ≥ d(c, v) −
d(y, c) − 2δ holds. Then, d(x, c) = d(x, y) − d(y, c) ≤ d(x, v) − d(c, v) + 2δ ≤ e(x) − e(c) + 2δ ≤
rad(G) + 2δ − rad(G) + 2δ = 4δ. Hence, d(x, y) = d(x, c) + d(y, c) ≤ 2d(x, c) + 1 ≤ 8δ + 1.
To prove the second assertion, consider a pair of vertices x, y ∈ V with d(x, y) = diam(G) and
a middle vertex c of any shortest (x, y)-path. Apply Lemma 3 to c, v, x, y, where v is an arbitrary
vertex from C 2δ (G). Without loss of generality, assume that d(x, v)−d(x, y) ≥ d(c, v)−d(y, c)−2δ
diam(G)
c ≥ b 2 rad(G)−4δ−1
c = rad(G) − 2δ
holds. We know also that d(x, c) ≥ b d(x,y)
2 c = b
2
2
(see Corollary 2). Hence, d(c, v) ≤ d(x, v) − d(x, y) + d(y, c) + 2δ = d(x, v) − d(x, c) + 2δ ≤
e(v) − rad(G) + 2δ + 2δ ≤ rad(G) + 2δ − rad(G) + 4δ = 6δ.
If the vertex degrees of a δ-hyperbolic graph are bounded by a constant ∆ then C 2δ (G) has at
most ∆O(δ) vertices. Summarizing, we have the following result.
Theorem 7 Let G = (V, E) be a δ-hyperbolic graph with m edges. Algorithm 1 finds
1. a vertex with eccentricity at most rad(G) + 3δ in at most O(m) time,
2. a vertex with eccentricity at most rad(G) + 2δ in at most O(δm) time,
3. a central vertex in at most O(m) time, if the vertex degrees and δ are bounded by constants.
Another linear time algorithm for finding a central vertex of a δ-hyperbolic graph with δ and
vertex degrees bounded by constants was proposed in [7].
12
Chordal graphs
Recall that F (s) = {v ∈ V : d(v, s) = e(s)} denotes the set of all vertices of G that are furthest
from s and C(G) = {c ∈ V : e(c) = rad(G)} denotes the set of all central vertices of G. The metric
interval I(u, v) between vertices u and v is defined by I(u, v) = {x ∈ V : d(u, x)+d(x, v) = d(u, v)},
i.e., it consists of all vertices of G laying on shortest paths between u and v.
In this section we analyze the behavior of our algorithms in the class of chordal graphs. Recall
that a graph G is chordal if every its induced cycle of length at least 4 has a chord. Chordal graphs
are interesting because a central vertex in them can be found in linear time [12] but finding the
diameter in linear time will refute the Orthogonal Vector Conjecture [13, 25].
First we give an example of an n-vertex chordal graph G on which Algorithm 1 will need n/2
iterations although G has a certificate for the radius consisting of only two vertices in L. Set
n = 2k and consider two sets of vertices X = {x1 , . . . , xk } and Y = {y1 , . . . , yk }. The vertex set of
G is X ∪ Y . Make X a clique and Y an independent set in G. Make every vertex xi adjacent to all
vertices yj with j ≤ i. Algorithm 1 may place vertices x1 , y2 , x2 , x3 , . . . , xk (in this order) into K
and vertices y2 , y1 , y3 , y4 , . . . , yk (in this order) into L. The central vertex xk will be determined
only when all Y -vertices are in L. On the other hand, (xk , {y1 , yk }) is a certificate for the radius
of G.
Note that the graph G constructed has vertices of large degrees (up-to n − 1). As every chordal
graph G has the hyperbolicity at most 1, it follows from Theorem 7 that our algorithm finds a
21
central vertex in linear time in every chordal graph with vertex degrees bounded by a constant.
It should be noted that there is a linear time algorithm that finds a central vertex of an arbitrary
chordal graph [12]; it uses additional metric properties of chordal graphs.
To analyze possible radius and diameter certificates in the class of chordal graphs, we will need
the following important lemma.
Lemma 4 ([10]) Let G be a chordal graph. Let x, y, v, u be vertices of G such that v ∈ I(x, y),
x ∈ I(v, u), and x and v are adjacent. Then d(u, y) ≥ d(u, x) + d(v, y). Furthermore, d(u, y) =
d(u, x) + d(v, y) if and only if there exist a neighbor x0 of x in I(x, u), a neighbor v 0 of v in I(v, y)
and a vertex w with N (w) ⊇ {x0 , x, v, v 0 }; in particular, x0 , v 0 and w lay on a common shortest
path of G between u and y.
Our analysis is based on the following propositions which are also of independent interest.
Recall that C 1 (G) := {v ∈ V : e(v) ≤ rad(G) + 1}.
Proposition 7 Let G = (V, E) be a chordal graph.
(i) If diam(G) < 2 rad(G) then, for every vertices s ∈ V and t ∈ F (s), there is a vertex
w ∈ I(s, t) ∩ C(G) such that t ∈ F (w).
(ii) If diam(G) = 2 rad(G) then, for every vertices s ∈ V and t ∈ F (s), there is a vertex
w ∈ I(s, t) ∩ C 1 (G) such that t ∈ F (w).
Proof. First we show that for every vertex x of G with e(x) = k > rad(G) there is a vertex y
such that e(y) = k − 1 and d(x, y) ≤ 2. This is true even in the case when diam(G) = 2 rad(G).
Consider a vertex y in G with e(y) = k −1 that is closest to x. Let z be any neighbor of y in I(y, x).
Necessarily, e(z) = k. Consider a vertex u ∈ F (z). Since d(y, u) ≤ e(y) = k − 1 = e(z) − 1 =
d(z, u) − 1 ≤ d(y, u), we have y ∈ I(z, u). Applying Lemma 4 to y ∈ I(z, u) and z ∈ I(y, x), we
get d(x, u) ≥ d(x, y) − 1 + d(y, u) = d(x, y) + k − 2. As d(x, u) ≤ e(x) = k, we conclude d(x, y) ≤ 2.
Next we claim that if diam(G) < 2 rad(G) then for every vertex x of G with e(x) = k > rad(G)
there is in fact a vertex z ∈ N (x) such that e(z) = k − 1. Furthermore, if diam(G) = 2 rad(G),
such a neighbor z with e(z) = k − 1 exists for every vertex x of G with e(x) = k > rad(G) + 1.
Assume, by way of contradiction, that no neighbor of x has eccentricity k − 1 and let y be an
arbitrary vertex of G with d(x, y) = 2 and e(y) = k − 1. Let also z be a vertex from N (x) ∩ N (y)
for which the set Sx (z) = {v ∈ F (x) : z ∈ I(x, v)} is largest. Necessarily, e(z) = k. As
before, consider a vertex u ∈ F (z). By Lemma 4, applied to y ∈ I(z, u) and z ∈ I(y, x), we get
d(x, u) ≥ d(y, u) + d(x, z) = k − 1 + 1 = k. As d(x, u) ≤ e(x) = k, we conclude d(x, u) = k. Hence,
by the second part of Lemma 4, there must exist a vertex w adjacent to y, z, x and at distance k −1
from u. As u ∈ F (x), w ∈ I(x, u), z ∈
/ I(x, u), by the maximality of |Sx (z)|, there must exist a
vertex u0 ∈ F (x) with w ∈
/ I(x, u0 ), z ∈ I(x, u0 ). We have d(z, u0 ) = k − 1, d(w, u0 ) = k, and hence
z ∈ I(w, u0 ) and w ∈ I(z, u). By Lemma 4, d(u, u0 ) ≥ d(u, w) + d(z, u0 ) = k − 1 + k − 1 = 2k − 2.
Hence, diam(G) ≥ d(u, u0 ) > 2 rad(G), if k > rad(G) + 1, and diam(G) ≥ d(u, u0 ) ≥ 2 rad(G), if
k = rad(G) + 1. These contradictions prove the claim.
Now we can conclude our proof. Consider arbitrary vertices s ∈ V and t ∈ F (s) and proceed
by the induction on k = e(s). If k = rad(G) then w = s and we are done. If k = rad(G) + 1 and
diam(G) = 2 rad(G) then again w = s and we are done. If k > rad(G) + 1 or k = rad(G) + 1 and
diam(G) < 2 rad(G) then a neighbor z of s with e(z) = k − 1 satisfies t ∈ F (z), and we can apply
the induction hypothesis.
A pair x, y of vertices is called a diametral pair of a graph G if d(x, y) = diam(G).
Proposition 8 The center C(G) of a chordal graph G is a diameter certificate of G (not necessarily a smallest one).
22
Proof. Additionally to Proposition 7(i), we need to mention only that in any graph G with
diam(G) = 2 rad(G), for every vertex v ∈ C(G) and every diametral pair of vertices x, y,
d(x, y) = d(x, c) + d(y, c) = d(x, c) + rad(G) = d(y, c) + rad(G) = 2 rad(G) holds.
Proposition 9 For every chordal graph G, the set C 1 (G) is a tight upper certificate of G (not
necessarily a smallest one).
Proof. The statement follows from Proposition 7 and the definition of a tight upper certificate.
So, it is interesting that if the center C(G) is known for a chordal graph G then its diameter
can be computed in O(|C(G)||E|) time. However, there is no way to bound the cardinality of
the set C(G) in an arbitrary chordal graph G. In fact C(G) may contain n − 2 vertices in some
chordal graphs. To construct such a graph G, take a complete graph Kn−2 on n − 2 vertices. Add
two new vertices u and v adjacent to all vertices of Kn−2 but not to each other. It is easy to see
that C(G) is exactly the vertices of Kn−2 .
Nevertheless, it is known that for every chordal graph G and any two vertices x, y from C(G),
d(x, y) ≤ 3 holds [11]. This suggest the following approach for computing the diameter of a chordal
graph G = (V, E).
- Use linear time algorithm from [12] to find a central vertex c of G.
- Set C3 := {x ∈ V : d(x, c) ≤ 3}. /* C(G) ⊆ C3 */
- Find a vertex p such that eC3 (p) is maximum.
- Report e(p).
The complexity of this approach is O(|C3 ||E|). As a consequence, we have that when the
vertex degrees are bounded in a chordal graph by a constant ∆ then its diameter can be computed
in linear time (as |C3 | is bounded by a constant ∆3 ). We are not aware if such a result was known
before. Note also that in general chordal graphs the cardinality of C3 cannot be bounded by a
constant since then the diameter of an arbitrary chordal graph can be computer in linear time
refuting the Orthogonal Vector Conjecture [13, 25].
From the proof of Proposition 7 it follows also that, for every vertex v with e(v) = rad(G) + 1,
d(v, C(G)) ≤ 2 holds. This suggest the following approach for computing the eccentricities of all
vertices of a chordal graph G = (V, E).
- Use linear time algorithm from [12] to find a central vertex c of G.
- Set C5 := {x ∈ V : d(x, c) ≤ 5}. /* C 1 (G) ⊆ C5 */
- For every vertex v ∈ V report e(v) = minc∈C5 d(v, c) + e(c).
The complexity of this approach is O(|C5 ||E|). As a consequence, we have that when the
vertex degrees are bounded in a chordal graph by a constant ∆ then the eccentricities of all its
vertices can be computed in linear time (as |C5 | is bounded by a constant ∆5 ). We are not aware
if such a result was known before.
Summarizing, we have the following result.
Theorem 8 Let G = (V, E) be a chordal graph with m edges and whose vertex degrees are bounded
by a constant. Then, eccentricities of all vertices of G can be computed in total O(m) time.
By Proposition 8, the center C(G) of a chordal graph G is a diameter certificate of G. Next
we will show that the set of all diametral vertices of a chordal graph G forms a radius certificate
of G.
Let Dk (G) := {v ∈ V : e(v) ≥ diam(G) − k}. It is known that for every vertex v of a chordal
graph G there is a vertex u ∈ F (v) with e(v) ≥ diam(G) − 1 [17]. Hence, the set D1 (G) contains
the output set L of Algorithm 1 and, therefore, gives already a radius certificate for a chordal
graph G. In fact, we can prove a stronger result.
Proposition 10 For an arbitrary graph G with diam(G) ≥ 2rad(G) − 1, any diametral pair of
c.
vertices x, y forms a minimum radius certificate of G. Futhermore, rad(G) = b d(x,y)+1
2
23
Proof. Let x, y be an arbitrary diametral pair of G, e.g., d(x, y) = diam(G). If there is a vertex
z ∈ V with max{d(z, x), d(z, y)} ≤ rad(G) − 1 then diam(G) = d(x, y) ≤ d(z, x) + d(z, y) ≤
2rad(G) − 2, and a contradiction arises.
Proposition 11 For every chordal graph G, the set D(G) := D0 (G) is a radius certificate of G
(not necessarily a smallest one).
Proof. By Proposition 10 and the fact that in any chordal graph G, diam(G) ≥ 2rad(G) − 2 [11],
we need to consider only the case when diam(G) = 2rad(G) − 2.
Assume that there is a vertex u in G such that d(u, t) ≤ rad(G) − 1 for every vertex t ∈ D(G).
Denote by S the set of all such vertices u. Denote by S 0 those vertices from S that have the
minimum eccentricity. Finally, denote by S 00 those vertices u from S 0 that have the smallest
number of vertices in F (u).
Consider a vertex u ∈ S 00 , a vertex v ∈ F (u) and a neighbor w of u on a shortest path from
u to v. Consider also an arbitrary vertex x ∈ D(G) and an arbitrary vertex y ∈ F (x) ⊂ D(G).
Since d(x, y) = 2rad(G) − 2, we have d(x, u) = d(y, u) = rad(G) − 1 and hence u ∈ I(x, y). We
claim that d(w, x) ≤ rad(G) − 1 as well. Assume that d(w, x) > rad(G) − 1. Then, u ∈ I(x, w)
and w ∈ I(u, v). By Lemma 4, d(x, v) ≥ d(x, u) + d(w, v) ≥ rad(G) − 1 + e(u) − 1 ≥ 2rad(G) − 2,
i.e., d(x, v) = 2rad(G) − 2 (hence v must belong to D(G)) and d(u, v) = e(u) = rad(G). The
latter contradicts with the choice of u (as u ∈ S). Thus, d(w, x) ≤ rad(G) − 1 for every vertex
x ∈ D(G), i.e., w ∈ S.
As u ∈ S 0 , e(u) ≤ e(w). First assume that e(u) = e(w), i.e., w ∈ S 0 . Since v ∈ F (u)
and v ∈
/ F (w) (note that d(w, v) = d(u, v) − 1 = e(u) − 1 ≤ e(w) − 1), by the choice of u (as
u ∈ S 00 ), there must exist a vertex t ∈ V such that t ∈ F (w) and t ∈
/ F (u). We necessarily have
d(t, u) = e(w) − 1 ≥ rad(G) − 1 and d(v, w) = e(u) − 1 ≥ rad(G) − 1. One can apply now Lemma
4 to u ∈ I(w, t) and w ∈ I(u, v) and get d(v, t) ≥ d(v, w) + d(t, u) ≥ 2rad(G) − 2 = diam(G).
That is, both v and t must be in D(G), contradicting again with the choice of u (as u ∈ S and
d(u, v) = e(u) ≥ rad(G)).
Assume now that e(u) < e(w) and consider an arbitrary vertex t ∈ F (w). Necessarily, u ∈
I(w, t). Hence again one can apply Lemma 4 to u ∈ I(w, t) and w ∈ I(u, v) and get d(v, t) ≥
d(v, w) + d(t, u) = e(u) − 1 + e(w) − 1 ≥ 2rad(G) − 1 > diam(G), and a contradiction arises.
Contradictions obtained prove that for every vertex u ∈ V there is a vertex t ∈ D(G) such
that d(u, t) ≥ rad(G), i.e., D(G) is a radius certificate of G.
13
Conclusion
A rough description for radius and diameter algorithms could be: take a node u not yet covered
by a certificate C and add a node x to C such that u is now covered and possibly many other
nodes are also covered. Selecting a node x covering u which is at maximum distance from u tends
to provide a maximal covering set in both radius and diameter cases. Although this is not the
classical greedy set-cover heuristic it is not surprising after all that it works well in practice.
References
[1] Amir Abboud, Fabrizio Grandoni, Virginia Vassilevska Williams: Subcubic Equivalences Between Graph Centrality Problems, APSP and Diameter. SODA 2015, pp..1681–1697.
[2] A. Abboud, J. Wang, V. Vassilevska Williams, Approximation and Fixed Parameter Subquadratic Algorithms for Radius and Diameter in Sparse Graphs, In Proceedings of the
Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, pp..
377–391, Arlington, VA, USA, January 10-12, 2016.
24
[3] L. Backstrom, P. Boldi, M. Rosa, J. Ugander, S. Vigna: Four degrees of separation. WebSci,
pp..33–42, 2012.
[4] M. Borassi, P. Crescenzi, M. Habib, Into the square - on the complexity of quadratic-time
solvable problems, Electr. Notes Theor. Comput. Sci. 322: 51-67, 2016.
[5] Michele Borassi, Pierluigi Crescenzi, Michel Habib, Walter A. Kosters, Andrea Marino, Frank
W. Takes: Fast diameter and radius BFS-based computation in (weakly connected) real-world
graphs: With an application to the six degrees of separation games. Theor. Comput. Sci. 586:
59-80 (2015).
[6] Michele Borassi, Pierluigi Crescenzi, Luca Trevisan: An Axiomatic and an Average-Case
Analysis of Algorithms and Heuristics for Metric Properties of Graphs. SODA 2017: 920-939.
[7] V. Chepoi, F. F. Dragan, B. Estellon, M. Habib, and Y. Vaxès, Diameters, centers, and approximating trees of δ-hyperbolic geodesic spaces and graphs, Symposuium on Computational
Geometry (SoCG’08), ACM, New York, 2008, pp. 59–68.
[8] Pierluigi Crescenzi, Roberto Grossi, Michel Habib, Leonardo Lanzi, Andrea Marino, On computing the diameter of real-world undirected graphs, Theoret. Comput. Sci. 514 (2013) 84–95.
[9] Victor Chepoi, Bertrand Estellon: Packing and Covering delta-Hyperbolic Spaces by Balls.
APPROX-RANDOM 2007: 59-73.
[10] V. Chepoi, Some d-convexity properties in triangulated graphs, in Mathematical Research,
vol. 87, Ştiinţa, Chişinău, 1986, pp. 164–177 (Russian).
[11] V. Chepoi, Centers of triangulated graphs, Math. Notes 43 (1988), 143–151.
[12] V.D. Chepoi and F.F. Dragan, A Linear-Time Algorithm for Finding a Central Vertex of a
Chordal Graph, ESA 1994: 159-170.
[13] D.G. Corneil, F.F. Dragan, M. Habib, C. Paul, Diameter determination on restricted graph
families, Discrete Applied Mathematics 113(2001), 143–166.
[14] Derek G. Corneil, Feodor F. Dragan and Ekkehard Köhler, On the power of BFS to determine
a graph’s diameter, Networks vol. 42, 4, pp. 209–222, 2003.
[15] Derek G. Corneil, Jérémie Dusart, Michel Habib, Antoine Mamcarz, Fabien de Montgolfier,
A tie-break model for graph search. Discrete Applied Mathematics 199: 89-100 (2016).
[16] Irit Dinur and David Steurer, Analytical Approach to Parallel Repetition, Symposium on
Theory of Computing (STOC’14), ACM, 2014, pp. 624–633.
[17] Feodor F. Dragan, Falk Nicolai, Andreas Brandstädt: LexBFS-Orderings and Power of
Graphs. WG 1996: 166-180
[18] M. Gromov, Hyperbolic groups, Essays in group theory, Math. Sci. Res. Inst. Publ., vol. 8,
Springer, New York, 1987, pp. 75–263.
[19] P. Hage and F. Harary. Eccentricity and centrality in networks. Social Networks 17, pp. 57–63,
1995.
[20] G. Y. Handler, Minimax Location of a Facility in an Undirected Tree Graph, Transportation
Science 7, pp. 287–293, 1973.
[21] Camille Jordan, Sur les assemblages de lignes, Journal für reine und angewandte Mathematik
70, pp. 185–190, 1869.
[22] Jure Leskovec and Andrej Krevl, SNAP Datasets: Stanford Large Network Dataset Collection,
http://snap.stanford.edu/data, June 2014.
25
[23] Clémence Magnien, Matthieu Latapy, Michel Habib, Fast computation of empirically tight
bounds for the diameter of massive graphs, ACM J. Exp. Algorithmics 13 (2009) 1–7.
[24] Mihai Patrascu, Ryan Williams: On the Possibility of Faster SAT Algorithms. SODA 2010:
1065-1075.
[25] L. Roditty, V. Vassilevska Williams, Fast approximation algorithms for the diameter and
radius of sparse graphs, In Proceedings of the 45th annual ACM symposium on Symposium
on theory of computing, STOC’13, pp..515–524, New York, NY, USA, 2013. ACM.
[26] V. Vassilevska Williams, Hardness of Easy Problems: Basing Hardness on Popular Conjectures such as the Strong Exponential Time Hypothesis, In Proceedings of the 10th International Symposium on Parameterized and Exact Computation (IPEC 2015), pp. 17–29,
September 16-18, 2015, Patras, Greece, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik
2015.
[27] Frank W. Takes, Walter A. Kosters, Determining the diameter of small world networks,
in: Proceedings of the 20th ACM Conference on Information and Knowledge Management,
CIKM, 2011, pp. 1191–1196.
[28] Frank W. Takes, Walter A. Kosters, Computing the eccentricity distribution of large graphs,
Algorithms 6 (2013) 100–118.
26
| 8cs.DS
|
Use of Signed Permutations in Cryptography
arXiv:1612.05605v3 [cs.CR] 15 Jun 2017
Iharantsoa Vero RAHARINIRINA
[email protected]
Department of Mathematics and computer science,
Faculty of Sciences, BP 906 Antananarivo 101, Madagascar
June 16, 2017
Abstract
In this paper we consider cryptographic applications of the arithmetic on the hyperoctahedral
group. On an appropriate subgroup of the latter, we particularly propose to construct public key
cryptosystems based on the discrete logarithm. The fact that the group of signed permutations
has rich properties provides fast and easy implementation and makes these systems resistant to
attacks like the Pohlig-Hellman algorithm. The only negative point is that storing and transmitting
permutations need large memory. Using together the hyperoctahedral enumeration system and
what is called subexceedant functions, we define a one-to-one correspondance between natural
numbers and signed permutations with which we label the message units.
Keywords : Signed Permutation, Public Key System, Discrete Logarithm Problem, Hyperoctahedral Enumeration System, Subexceedant Function
1 Introduction
The term cryptography refers to the study of techniques providing information security. Several mathematical objects have been used in this field to label the so-called message units. These objects are
often integers. They may also be points or vectors on some curves. For this work, we shall use signed
permutations.
Let us denote by:
• [n] the set {1, · · · , n},
• [±n] the set {−n, · · · , −1, 1, · · · , n},
• Sn the symmetric group of degree n.
Definition 1.1. A bijection π : [±n] −→ [±n] satisfying π(−i) = −π(i) for all i ∈ [±n] is called
”signed permutation”.
We can also write a signed permutation π in the form
1
2
...
n
π=
with σ ∈ Sn and εi ∈ {±1} .
ε1 σ1 ε2 σ2 . . . εn σn
1
Under the ordinary composition of mappings, all signed permutations of the elements of [n] form a
group Bn called hyperoctahedral group of rank n. We write π k = π
· · ◦ π} for an integer k and
| ◦ ·{z
k-times
π ∈ Bn and when we multiply permutations, the leftmost permutation acts first. For example,
1
2 3
4
1
2 3 4
1
2 3 4
.
=
◦
3 −4 1 −2
3 −2 4 1
1 −3 4 2
Before the 1970’s, to cipher or decipher a message, two users of cryptographic system must safely
exchange information (the private key) which is not known by anyone else. New keys could be periodically distributed so as to keep the enemy guessing. In 1976, W. Diffie and M. Hellman [1] discovered an entirely different type of cryptosystem for which all of the necessary information to send an
enciphered message is publicly available without enabling anyone to read the secret message. With
this kind of system called public key, it is possible for two parties to initiate secret communications
without exchanging any preliminary information or ever having had any prior contact.
The security of public key cryptosystems is based on the hardness of some mathematical problems. As a public key cryptosystem consists of a private key (the deciphering key) that is kept secret
and a public key (the enciphering key) which is accessible to the public, then the straightforward way
to break the system is to draw the private key from the public key. Therefore, the required computation
cost is equivalent to solving these difficult mathematical problems. For instance for the two particularly important examples of public key cryptosystems : RSA and Diffie-Hellman, both are connected
with fundamental questions in number theory which are : difficulty of factoring a large composite
integer whose prime factors are not known in advance and intractability of discrete logarithm in the
multiplicative group of a large finite field, respectively.
Definition 1.2. The discrete logarithm problem in the finite group G to the base g ∈ G is the problem
: given y ∈ G, of finding an integer x such that g x = y, provided that such an integer exists (in other
words, provided that y is in the subgroup generated by g).
If we really want our random element y of G to have a discrete logarithm, g must be a generator
of G.
Example 1. Let G = (Z/19Z)∗ be the multiplicative group of integers modulo 19. The successive
powers of 2 reduced mod 19 are : 2, 4, 8, 16, 13, 7, 14, 9, 18, 17, 15, 11, 3, 6, 12, 5, 10, 1. Let g be the
generator 2, then the discrete logarithm of 9 to the base 2 is 8.
In many ways, the group of signed permutations is analogous to the multiplicative group of a
finite field. For this purpose, cryptosystems based on the latter can be translated to systems using the
hyperoctahedral group.
This work addresses public key cryptosystems based on the group of signed permutations and
related to the discrete logarithm problem. We shall illustrate this in section 3 by an hyperoctahedral
group analogue for the Diffie-Hellman key exchange system then by describing two hyperoctahedral
group public key cryptosystems for transmitting information. We shall also explain why there is no
subexponential time algorithms to break the proposed systems. Before introducing the cryptosystems
themselves, we give in section 2 a method to define a bijection between integers and signed permutations. This time, we first convert the natural number in hyperoctahedral system and then we use the
result to compute the corresponding signed permutation by means of subexceedant function.
2
2 Integer representation of signed permutations
2.1 Converting from one base to another
Definition 2.1. Hyperoctahedral number system is a system that expresses all natural number n of N
in the form :
k(n)
n=
X
i=0
di .Bi , where k(n) ∈ N, di ∈ {0, 1, 2, · · · , 2i + 1} and Bi = 2i i! .
(1)
This definition is motivated by the fact that Bi is the cardinal of the hyperoctahedral group Bi .
To denote the non-negative integer n of the equation (1) in the hyperoctahedral system, we use by
convention the (k + 1)-digits representation
dk dk−1 · · · d2 d1 d0 .
Theorem 2.2. Every positive integer has an unique representation in the hyperoctahedral system.
Before giving the proof of this theorem, these are some properties of the hyperoctahedral system.
Lemma 2.3. If n = dk · · · d1 d0 is a number in hyperoctahedral system, then
0 6 n 6 Bk+1 − 1
Proof. Since 0 6 di 6 2i + 1, we have
06
k
X
i=0
di Bi 6
k
X
(2i + 1)Bi .
i=0
Recall that Bi = 2i i! ,
(2i + 1)Bi = (2(i + 1 − 1) + 1)Bi
= (2(i + 1) − 1)Bi
= 2(i + 1)Bi − Bi
= Bi+1 − Bi .
Therefore,
k
k
X
X
(Bi+1 − Bi ) = Bk+1 − B0 = Bk+1 − 1 .
(2i + 1)Bi =
i=0
i=0
Lemma 2.4. Let n = dk · · · d1 d0 be a number in hyperoctahedral system, then
dk Bk 6 n < (dk + 1)Bk .
Proof. Let m = dk−1 · · · d1 d0 . From lemma 2.3, 0 6 m < Bk . We also have n = dk Bk + m, hence
dk Bk 6 n < Bk + dk Bk = (1 + dk )Bk .
3
Now, we can proceed to the demonstration of theorem 2.2 which states the one-to-oneness between
non negative integers and hyperoctahedral base numbers.
Proof. Let an · · · a1 a0 with an 6= 0 and bm · · · b1 b0 , bm 6= 0 be two representations of a positive
integer N in hyperoctahedral system. First, bm ≥ 1 and an ≥ 1 imply
Bm ≤ bm Bm ≤ bm · · · b1 b0 and Bn ≤ an · · · a1 a0 .
The relation n = m is immediate because if n < m then
Bm ≥ Bn+1 > an · · · a1 a0
by lemma 2.3 so
Similarly, if m < n then
bm · · · b1 b0 > an · · · a 1 a 0 .
an · · · a1 a0 > bm · · · b1 b0 .
We get a contradiction. Consequently n = m and ai = bi for all i ∈ {0, 1, . . . , n} by induction and
by the unicity of the expression N = an Bn + rn with rn = an−1 · · · a1 a0 < Bn .
To express a positive integer n in the hyperoctahedral system, one proceeds with the following
manner. Start by dividing n by 2 and let d0 be the rest r0 of the expression
n = r0 + (2)q0 .
Divide q0 by 4, and let d1 be the rest r1 of the expression
q0 = r1 + (4)q1 .
Continue the procedure by dividing qi−1 by 2(i + 1) and taking di := ri of the expression
qi−1 = ri + 2(i + 1)qi
until ql = 0 for some l ∈ N. In this way, we obtain n = dl : dl−1 : · · · : d1 : d0 and we also have
n = d0 + 2 (d1 + 4 · (d2 + 2(3) · (d3 + · · · ))) .
Now let n = dk−1 : dk−2 : · · · : d1 : d0 be a number in hyperoctahedral system. By definition 2.1,
one way to convert n to the usual decimal system is to calculate
dk−1 2k−1 (k − 1)! + · · · + d1 .2 + d0 .
In practice, one can use this algorithm :
Input :
Output :
An integer dk−1 : dk−2 : · · · : d1 : d0 in hyperoctahedral system.
An integer d in decimal system.
1. initiate the value of d :
2. for i from k − 1 to 1 do :
3. return d
4
d ← dk−1
d ← d.2.i + di−1
2.2 A bijection between the subexceedant functions and permutations
Definition 2.5. A subexceedant function f on [n] is a map f : [n] −→ [n] such that
1 6 f (i) 6 i for all i ∈ [n] .
We will denote by Fn the set of all subexceedant functions on [n], and we will represent a subexceedant function f over [n] by the word f (1)f (2) · · · f (n).
Example 2. These are the sets Fn for n = 1, 2, 3 :
F1 = {1}
F2 = {11, 12}
F3 = {111, 112, 113, 121, 122, 123}.
It is easy to verify that card Fn = n!, since from each subexceedant f over [n − 1], one can
obtain n distinct subexceedant functions over [n] by adding any integer i ∈ [n] at the end of the word
representing f .
We will give a bijection between Sn and Fn . Let be the map
φ : Fn −→ Sn
.
f 7−→ σf = (1 f (1))(2 f (2)) · · · (n f (n))
Notice that there is an abuse of notation in the definition of σf . Indeed, if f (i) = i, then the cycle
(i f (i)) = (i) does not really denote a transposition but simply the identity permutation.
Lemma 2.6. The map φ is a bijection from Fn onto Sn .
Proof. Since Sn and Fn both have cardinality n!, it suffices to prove that f is injective. Let f and g
be two subexceedant functions on [n]. Assume that φ(f ) = φ(g) i.e. σf = σg . So we have :
(1 f (1))(2 f (2)) · · · (n f (n)) = (1 g(1))(2 g(2)) · · · (n g(n)).
(2)
Since σf = σg , then in particular σf (n) = σg (n). By definition σf (n) = f (n) and σg (n) = g(n),
so f (n) = g(n). Let us multiply both members of equation (2) on the right by the permutation
(n f (n)) = (n g(n)), we obtain :
(1 f (1))(2 f (2)) · · · (n − 1
f (n − 1)) = (1 g(1))(2 g(2)) · · · (n − 1
g(n − 1)).
Now, if we apply the same process to this equation, we obtain f (n − 1) = g(n − 1). By iterating, we
conclude that f (i) = g(i) for all integers i ∈ [n] and then f = g.
Let σ be a permutation of the symmetric group Sn and f be the inverse image of σ by φ. Then f
can be constructed as below :
1. Set f (n) = σ(n).
2. Multiply σ on the right by (n σ(n)) (this operation consists in exchanging the image of n and
the image of σ −1 (n)), we obtain a new permutation σ1 having n as a fixed point. Thus σ1 can
be considered as a permutation of Sn−1 . Then set f (n − 1) = σ1 (n − 1).
3. In order to obtain f (n − 2), apply now the same process to the permutation σ1 by multiplying
σ1 by (n − 1 σ1 (n − 1)). Iteration determines f (i) for all integers i of [n].
5
2.3 Mapping hyperoctahedral base numbers to signed permutations
Let dn−1 dn−2 · · · d1 d0 be a n-digits number in hyperoctahedral system. That is di ∈ {0, 1, 2, · · · , 2i+
1} for i = 0, · · · , n − 1.
Writing di = 2qi + ri where ri ∈ {0, 1} and qi ∈ {0, . . . , i}, gives
• the subexceedant function f = f (1) · · · f (n) with f (i) = 1 + qi−1 , i = 1, . . . , n
• and the sequence (ε1 , . . . , εn ) with εi = (−1)ri−1 for i = 1, . . . , n.
So to each integer dn−1 dn−2 · · · d1 d0 , we associate the signed permutation
1
2
...
n
ε1 σ1 ε2 σ2 . . . εn σn
where σ = σ1 · · · σn is the permutation associated to the subexceedant function f by the map
φ : Fn −→ Sn
.
f 7−→ σf = (1f (1))(2f (2)) · · · (nf (n))
1
2
...
n
be a signed permutation. As φ is a bijection, from π we
Now, let π =
ε1 σ1 ε2 σ2 . . . εn σn
have
(
0 if εi = 1
−1
for i = 1, . . . , n.
f = f (1) · · · f (n) = φ (σ) and ri−1 =
1 if εi = −1
The digits di = 2(f (i+1)−1)+ri , i = 0, . . . , n−1 form the hyperoctahedral number dn−1 dn−2 · · · d1 d0 .
It is easy to verify that di ∈ {0, . . . , 2i + 1}. We have
1 6 f (i) 6 i
for i = 1, . . . , n
0 6 f (i + 1) 6 i + 1
0 6 2(f (i + 1) − 1) 6 2i
0 6 di 6 2i + 1
for i = 0, . . . , n − 1
for i = 0, . . . , n − 1
because 0 6 ri 6 1.
3 Some cryptosystems based on hyperoctahedral group
We suppose that we are using message units with signed permutations as equivalents (according to
section 2) in some publicly known large hyperoctahedral group Bn .
3.1 Analog of the Diffie-Helman key exchange
The Diffie-Hellman key exchange [1] which was the first public key cryptosystem originally used the
multiplicative group of a finite field. It can be adapted for signed permutations group as follow.
Suppose that two users Alice and Bob want to agree upon a secret key, a random element of Bn ,
which they will use to encrypt their subsequent messages to one another. They first choose a large
integer n for the hyperoctahedral group Bn and select a signed permutation β ∈ Bn to serve as their
”base” and make them public. Alice selects a random integer 0 < a < Bn , which she keeps secret,
and computes β a ∈ Bn which she makes public. Bob does the same. He selects a random integer
0 < b < Bn , and transmits β b to Alice over a public channel. Alice and Bob agree on the secret key
β ab by computing (β b )a and (β a )b respectively.
6
3.2 Analog of the ElGamal system
Suppose user ”A” requires sending a message µ to user ”B”. As in the key exchange system above,
we start with a fixed publicly known hyperoctahedral group Bn for a large integer n and a signed
permutation β ∈ Bn (preferably, but not necessarily, a generator) as base. Each user selects a random
integer 0 < u < |β| (instead of u we take a for user ”A” and b for user ”B”) which is kept secret, and
computes β u that he publishes as a public key. With the public key β b , user ”A” encrypt the message
µ and sends the pair of signed permutations (m1 = β a , m2 = µ.(β b )a ) to user ”B”. To recover µ, B
computes (mb1 )−1 = ((β a )b )−1 and multiplies m2 on the right by the result.
3.3 Analog of the Massey-Omura system for message transmission.
We assume the same setup as in the previous subsection. Suppose that Alice wants to send Bob a
message µ. She chooses a random integer c satisfying 0 < c < Bn and g.c.d.(c, Bn ) = 1, and
transmits µc to Bob. Next, Bob chooses a random integer d with the same properties, and transmits
′
(µc )d to Alice. Then Alice transmits (µcd )c = µd back to Bob, where c′ c ≡ 1 mod Bn . Finally,
′
Bob computes (µd )d where d′ d ≡ 1 mod Bn .
3.4 Security
For the Diffie-Hellman’s key exchange above, a third party knows only the elements β , β a and β b
of Bn which are public knowledge. Obtaining β ab knowing only β a and β b is as hard as taking the
discrete logarithm a from β and β a (or b knowing β and β b ). Then an unauthorized third party must
solve the discrete logarithm problem to the base β ∈ Bn to determine the key.
Both ElGamal’s [5] and Massey-Omura’s [4] cryptosystems are essentially variants of DiffieHellman’s key exchange system. Therefore, breaking either of the systems above requires the solution
of the discrete logarithm problem for hyperoctahedral group.
Definition 3.1. The hyperoctahedral discrete logarithm problem to the base β ∈ Bn is the problem,
given π ∈ Bn , of finding an integer x such that π = β x if such x exists.
Many improvements are made for solving the discrete logarithm problem in finite group. In hyperoctahedral group based cryptosystems of the sort discussed above, one does not work with the entire
group Bn , but rather with cyclic subgroups : the group < β > in the Diffie-Hellman system and
ElGamal system and the group < µ > in the Massey-Omura system where we designate by
• < π >= {π i | i = 1, . . . , |π|} the cyclic subgroup of Bn generated by π ∈ Bn ,
• |π| the order of the signed permutation π.
In other words, one works in the subgroup generated by the base of the discrete logarithm in each
proposed system.
Choice of a suitable base The order of the cyclic subgroup of Bn has an important role to avoid an
easy solution of the discrete logarithm. Recall that we can also represent a signed permutation π in
the form
1
2
...
n
π=
with σ ∈ Sn and εi ∈ {±1} .
ε1 σ(1) ε2 σ(2) . . . εn σ(n)
As stated in the work of Victor Reiner [2], a signed permutation decomposes uniquely into a product
of commuting cycles just as permutations do.
7
Example 3. The disjoint cycle form of
1 2
3 4
5
6 7
3 6 −2 7 −5 −1 4
1
3 2
6
3 −2 6 −1
4 7
7 4
5
−5
is
.
Definition 3.2. The order of a signed permutation π is the smallest positive integer m such that
π m = ι where ι denotes the identity permutation.
i1
i2 . . . iℓ−1
iℓ
An ℓ-cycle C =
where εi ∈ {±1} has order ℓ. The proof of
ε1 i2 ε2 i3 . . . εℓ−1 iℓ εℓ i1
the following theorem can be found in [6].
Theorem 3.3. The order of a permutation written in disjoint cycle form is the least common multiple
of the lengths of the cycles.
Let us now assume that we have β ∈ Bn as base of the discrete logarithm. The order of < β >
can be very large for a large value of n, since the disjoint cycles of β can be selected in such a way
that their least common multiple be very large. Therefore brute force search is inefficient to solve
the discrete logarithm problem to the base β ∈ Bn . < β > is a cyclic group of finite order so it is
commutative. Thus, in order to avoid an easy solution to the discrete logarithm problem using the
√
techniques that apply to any finite abelian group (which take approximately p operations, where p
is the largest prime dividing the order of the group), it is important for the order of the commutative
group < β > to be non smooth, that is, divisible by a large prime.
Definition 3.4. Let n be a positive real number. we say that n is smooth if all of the prime factors of
n are small.
The method of Pohlig-Hellman [3] can efficiently computes the discrete logarithm in a group G if
its order is B-smooth for a reasonably small B.
Definition 3.5. Let B be a positive real number. An integer is said to be B-smooth if it is not divisible
by any prime greater than B.
The following theorem allows to generate the base β so that the order of the subgroup < β > of
Bn have arbitrary smoothness.
Theorem 3.6. Let p be a prime. For a large integer n, a cyclic subgroup G of Bn which its order is
p-smooth and is not (p − 1)-smooth can be constructed.
Proof. Let n be a large integer and p ≤ n a prime. There exists a p-cycle γp ∈ Bn . Let γp , C1 , C2 , . . . , Ck ∈
Bn be disjoint cycles of length respectively p, l1 , l2 , . . . , lk with
1 ≤ li ≤ p for i = 1, . . . , k and
k
X
i=1
li ≤ n − p .
Let us now consider the signed permutation π = γp C1 C2 . . . Ck ∈ Bn of order
|π| = lcm(p, l1 , l2 , . . . , lk ) .
8
Let q be a prime such that q|lmc(p, l1 , l2 , . . . , lk ). We show that q ≤ p. Let us suppose that q > p. We
obtain li < p < q which implies q ∤ |π| so q ≤ p. We have just shown that every prime q dividing the
order of π is smaller than p, that is, |π| is p-smooth. The order | < π > | = |π| of the cyclic subgroup
< π > of Bn generated by π is p-smooth but it is not (p − 1)-smooth because p | | < π > | and
p > p − 1.
This theorem provides high flexibility in selecting a subgroup of Bn on which cryptosystem resists
to attacks by Silver-Pohlig-Hellman’s algorithm.
The cryptosystems that we have proposed are easy to implement by applying optimized method
for exponentiation. Moreover, the multiplication on Bn which is the composition of mappings can be
performed in time O(n). However, as we must work in a very large hyperoctahedral group, the need
of large memory from the point of view implementation requires improvements.
References
[1] W. Diffie and M. E. Hellman, ”New directions in cryptography”, IEEE Transactions on Information Theory IT-22 (1976), 644-654.
[2] Reiner, Victor, ”Signed Permutation Statistics and Cycle Type”, Europ. J. Combinatorics 14
(1993), 569-579.
[3] S.C. Pohlig and M.E. Hellman, ”An improved algorithm for computing logarithms over GF (p)
and its cryptographic significance”, IEEE Transactions on Information Theory 24 (1978), 106110.
[4] P. K. S. Wah and M. Z. Wang. Realization and application of the Massey-Omura lock. Proc.
International Zurich Seminar, pages 175–182, 1984.
[5] T. ElGamal, ”A public key cryptosystem and a signature scheme based on discrete logarithms”,
IEEE Transactions on Information Theory IT-31 (1985), 469-472.
[6] J.J. Rotman, ”Advanced Modern Algebra”, Prentice Hall, 1st edition, New Jersey, (2002).
9
| 7cs.IT
|
1
Theory and Practice of Logic Programming
arXiv:cs/0312027v1 [cs.PL] 15 Dec 2003
PROGRAMMING PEARL
An Open Ended Tree
HENK VANDECASTEELE
PharmaDM
Ambachtenlaan 54D
B-3001 Leuven, Belgium
[email protected]
GERDA JANSSENS
Katholieke Universiteit Leuven,
Department of Computer Science
Celestijnenlaan 200A
B-3001 Leuven, Belgium
[email protected]
Abstract
An open ended list is a well known data structure in Prolog programs. It is frequently
used to represent a value changing over time, while this value is referred to from several
places in the data structure of the application. A weak point in this technique is that the
time complexity is linear in the number of updates to the value represented by the open
ended list. In this programming pearl we present a variant of the open ended list, namely
an open ended tree, with an update and access time complexity logarithmic in the number
of updates to the value.
1 Introduction
Many applications in Logic Programming deal with variables of which the content changes over time. In this programming pearl these variables are called the
application variables. An example of such an application is a Constraint Logic Programming Finite Domain solver (CLP(FD)) (Dincbas, Van Hentenryck, Simonis,
Aggoun, Graf and Berthier. 1988). In such a solver the application variables are
the finite domain variables. The solver changes the domains of the finite domain
variables and also the set of constraints associated with the finite domain variables.
Another example is found in a fix-point process that computes subsequent approximations for an entity before a final value – the fix-point – is reached. It is often
the case that several entities depend on each other. The application variables are
the entities for which a fix-point has to be computed. The application variables are
updated and used in an interleaved way. The number of application variables is not
known in advance.
The problem is to find a representation for such application variables that are
updated and accessed in an interleaved and unpredictable way. Several solutions
exist to tackle this problem:
• Replacement
2
Henk Vandecasteele, Gerda Janssens
In this solution every occurrence of the application variable in the data structure must be replaced on every change of the content of the application variable. For simple applications this might be feasible, but in a complex application as a CLP Finite Domain solver this is not reasonable, since each finite
domain variable can occur in numerous constraints and the solver has to traverse all these constraints at each change in the domain of a finite domain
variable.
• Threading
Threading a pair of arguments, containing the current values of the application
variables, is usually seen as the most obvious solution. The first argument then
contains the incoming state, the current values at the time of the call. The
second argument contains the set of values as the resulting state of the call.
All occurrences of the application variables in other data structures simply
refer to the values in these states (e.g. by some numbering scheme). Although
preferable from logical point of view, it can be problematic from efficiency
point of view. When dealing with a large number of application variables, the
access and update is at least logarithmic in the number of application variables
(e.g. when stored in a balanced tree). In the case of a demanding application
with many values to be maintained, such as a Finite Domain solver, this extra
time complexity is significant.
If the application variables are known beforehand, the programmer can thread
a corresponding number of pairs through the program and avoid the search
for the value. In our examples, the number of application variables is different
for each use of the program.
• Open ended lists
The use of open ended lists avoids dependency on the number of application
variables in the application and also avoids replacing each occurrence of them
whenever the value changes. The rule is that the last element before the
open end is the current value of the application variable. Whenever the same
application variable occurs in a data structure of the application, the same
open ended list is referred to. Whenever the value of the application variable
changes, the end of the list is instantiated to a new list with the new value as
first element and with a new open end. In this way all the other occurrences
of the same application variable can see the change. When storing an already
existing application variable one can use a list that consists only of the last
element and the open end. (e.g. [a,b,c,d|Var] can be replaced by [d|Var]). The
target applications in this pearl do not often allow this replacement.
As every application variable is represented by a separate open ended list, update and access times do not depend on the number of application variables.
Every update to an application variable adds one element to the open ended
list: the length of the list is equal to the number of updates to the application variable. From the observations that in order to add a value one has to
instantiate the tail and that the current value is the last element in the open
ended list, we can conclude that the update and access time is linear in the
number of updates to the application variable. From data complexity point
Programming pearl
3
of view, open ended lists have a serious disadvantage compared to threading:
the technique will keep alive all values that are in the list. This means that
the garbage collector will never be able to collect any of the values used in
the past.
• Assert and retract
Assert and retract can be used to store the changing content over time. This
method has a time complexity for update and access which is independent
of the number of application variables and the number of updates. One may
have to deal with a high constant factor in most Prolog systems. Every lookup
requires the creation of a new instance of the value. When the values of
the application variables contain logic variables, this method will not work
because every lookup will return an instance with fresh variables. When large
values are used, it may lead to a lot of overhead due to the creation and the
garbage collection of the instances. In all other solutions discussed here, no
new instances are created when accessing the current value. When dynamic
predicates are used, old values will not be restored on backtracking as it is
the case for open ended lists or threading. Assigning a new value is now
destructive and old values can be removed if the Prolog system has a garbage
collector for dynamic code.
• Non portable solutions
Some Prolog systems provide their own solution based on backtrackable destructive assignment (Holzbauer 1992, Le Houitouze 1990), for example as
attributed variables (Swe 2000) and meta variables (Eur 1998). Using these
features is probably efficient but unfortunately not portable. The update and
access times for these solutions are O(1). Also data complexity is optimal in
these solutions: old values that are not kept alive by some choice point can
be collected by the heap garbage collector.
In this programming pearl we present the open ended tree as an alternative
data structure for representing application variables that are updated and accessed
in an interleaved and unpredictable way. The open ended tree is an ISO-compatible
solution and has an update and access complexity which is logarithmic in the number of updates to the application variable at hand. The data complexity is equivalent
to the data complexity of the open ended list solution.
In Section 2 we explain how an open ended tree is used to represent an application
variable. Section 3 gives the Prolog predicates for accessing and updating the value
of an application variable. Section 4 shows some efficiency results and presents some
variants that can be used to tune the application at hand.
2 Open ended trees
In an open ended tree, the current value of the application variable is found in the
rightmost leaf of the tree: it is the last instantiated node that would be encountered when the tree were traversed in a depth-first left-to-right way. The tree is
constructed such that the number of steps for finding this rightmost leaf is loga-
4
Henk Vandecasteele, Gerda Janssens
1
_C
2
_B
_D
Fig. 1. An open ended tree with 1 collector node and a tree of depth 1
rithmic in the number of nodes in the tree. This number of nodes is the same as
the number of updates.
The main issue is the shape of the tree. Since the number of updates is not known
in advance and the nodes of the tree can not be rearranged, a balanced tree is out
of the question. The solution is to create a sequence of binary trees, where each tree
is one level deeper than the previous tree. This sequence of trees could be stored
in an open ended list, but for simplicity of the lookup-procedure the binary tree
structure is reused.
The nodes that are used to build the sequence of trees are called the collector
nodes. The right child of a collector node is – if already created – again a collector
node. A new collector node can only be created if the left child of the parent has
reached depth N in all its branches. The left child of the newly created collector
node is restricted to depth N+1. Each node in the open ended tree contains two
children and a data field. A child can be a free variable or again a node. Such a free
variable can be instantiated later with a node, as is done in open ended lists. The
root node is a collector node, whose left child’s depth is limited to 1. An empty
open ended tree is represented by a free variable.
Example 2.1
Suppose the open ended tree O is used to represent an application variable whose
subsequent values are 1, 2, 3, . . ., 10. These values will be added one after another
to O. Firstly, “1” is added and O gets bound to
tree( A,1, B),
a collector node with free variables as children. The left child becomes a binary
tree of depth 1 when “2” is added:
tree(tree( C,2, D),1, B).
This open ended tree is shown in Figure 1. Note that collector nodes are put in bold
in the text. When adding the third value “3”, a new collector node is created as
the right child of the root collector node:
tree(tree( C,2, D),1,tree( E,3, F)).
Adding “4”, a tree of depth 2 is started:
tree(tree( C,2, D),1, tree(tree( G,4, H),3, F))).
After “5” and “6” have been added, the tree of depth 2 is completed as shown in
Figure 2:
tree(tree( C,2, D), 1, tree(tree(tree( I,5, J),4,tree( K,6, L)),3, F)) .
Adding all values up to “10” gives the following open ended tree:
5
Programming pearl
1
_C
2
_D
3
_F
6
_L
4
_I
5
_J
_K
Fig. 2. An open ended tree with 2 collector nodes and a tree of depth 1 and depth 2
tree(tree( C,2, D), 1, tree(tree(tree( I,5, J),4,tree( K,6, L)), 3,
tree(tree(tree(tree( T,10, U),9, S),8, Q),7, O))) .
The use of a set of trees to reduce the access time to some data structure is
not new, e.g. Fibonacci heaps (Fredman, Tarjan 1987). However the shape and the
properties of open ended trees are quite different: open ended trees are binary and
have a different shape, and no reorganisation of the trees is ever needed to assure
logarithmic time complexity for the operations we need.
We can prove that the update and access to the data structure are logarithmic
PN −1
in the number of updates. A tree of depth N contains maximum i=0 2i = 2N − 1
nodes. Since the collector node contains a value as well, we have 2N values in a
tree of depth N and its corresponding collector node. After the tree of depth N has
PN
been completed, the data structure contains i=1 2i = 2N +1 − 2 elements. After
the tree of depth N-1 is completed and the tree of depth N is under construction
the data structure contains M nodes where
(2N − 2) < M ≤ (2N +1 − 2)
(1)
M is also the number of updates. From (1) we deduce that N = ceil(log2(M +2))−1.
Then finding the last tree takes N steps and finding the rightmost leaf in this tree
takes at most N steps as well. This results in a time complexity of O(log(M )).
The main disadvantage of this approach compared to an open ended list is its
space consumption: it uses twice as much memory in most Prolog implementations.
An open ended list uses one ./2 term for each element in the list. A ./2 term needs
two heap cells, whereas an open ended tree has per element one tree/3 term which
takes 4 heap cells.
3 Code for the open ended tree
In this section we give the Prolog predicates that define the two operations on an
open ended tree that represents an application variable:
6
Henk Vandecasteele, Gerda Janssens
• lookup(ApplVar, Value) unifies Value with the current value of the application
variable represented by ApplVar.
• insert(ApplVar, New) stores New as the (updated) current value of the application variable represented by ApplVar.
Subsequent calls of insert(T, X) and lookup(T, Y) will always unify the two variables X and Y.
% lookup(T,V) finds the current value V in the rightmost leaf of the tree T
lookup(tree(Left, El, Right), Value):( var(Right) →
( var(Left) → El = Value ; lookup(Left, Value) )
;
lookup(Right, Value)
).
% insert(T,V) stores the current value V in a new node in the tree T as the
% rightmost leaf
insert(Tree, Value):( nonvar(Tree) → insert1(Tree, Value, 1)
;
Tree = tree( , Value, ) % the first collector node
).
% insert1(T,V,D) first finds the last collector node in T and meanwhile computes
% the depth D of the tree in the left branch of T. Next it inserts the value V
% in the left branch of the collector node. In case this tree is full, insert1 creates
% a new collector node in the right child of the node in variable T.
insert1(tree(Left, , Right), Value, Depth):( var(Right) →
insert2(Left, Value, Depth, Right)
;
Depthplus1 is Depth + 1,
insert1(Right, Value, Depthplus1)
).
% insert2(T,V,D,R) inserts V in the tree T, unless this would make the depth of T
% larger than D. In the latter case a node is created in the variable R. This variable
% R is known to be the next subtree to be instantiated, it could be a collector node.
insert2(tree(Left, El, Right), Value, Depth, Back):( var(El) → El = Value
;
( Depth == 1 → Back = tree( , Value, )
;
Depthmin1 is Depth - 1,
( var(Right) →
insert2(Left, Value, Depthmin1, Right)
;
insert2(Right, Value, Depthmin1, Back)
)
)
).
7
Programming pearl
# updates
10
100
1000
10000
ilProlog
updating
lookup
list
tree
list
tree
0.07
0.53
4.88
49.86
0.16
0.28
0.39
0.52
0.07
0.51
4.85
48.82
0.08
0.15
0.20
0.27
SICStus
updating
lookup
list
tree
list
tree
0.18
1.50
14.67
147.1
0.32
0.62
0.90
1.29
0.19
1.82
18.11
181.2
0.22
0.41
0.53
0.82
Depth
of tree
6
12
18
26
Table 1. Comparison of the execution times.
4 Efficiency results and optimisations
4.1 Measuring efficiency
Our benchmarks measure the difference in efficiency between open ended lists and
open ended trees, which are both portable solutions that can be used in the same
kind of circumstances.
Four experiments were done, each starting from a data structure with already
several updates. In the first experiment we started with a data structure with
already 10 updates; the subsequent experiments had a data structure with already
100, 1000 and 100000 updates. Then, in each of these experiments, the time to
update the data structure, and the time to access the current value was measured.
Each of the operations (update and access) was repeated 100000 times (In case
of update, the update was undone by backtracking to prevent the data structure
from growing). Each of the experiments was done on an implementation with open
ended lists and open ended trees. The computation was performed on a Pentium III
666 Mhz, running Linux 2.2.20, both with ilProlog(version 0.9.6) (Vandecasteele,
Demoen, Janssens 2000) and SICStus(version 3.9.0) (Swe 2000). The times are
reported in seconds and do not include the time for setting up the benchmark.
From the experiments we can see that with only 10 updates the overhead is
larger than the benefit of using open ended trees: the disadvantage is rather small
for lookup, but considerable for update. From 100 updates on, the overhead is
compensated. Furthermore, the timings for the open ended tree exhibit the expected
logarithmic behaviour. The timings for the open ended list show linear behaviour.
4.2 Variants
• When using open ended lists, one can always replace the list by some tail
of the list, as long as the tail contains at least one element (e.g [1,2,3,4|V]
can be replaced by [4|V]). Although this does not change the time complexity
of the resulting program, still a considerable speed-up can be realised. The
same technique can be used with open ended trees, after a modification of
the code above. This modification consists of storing the depth of the tree
at each collector node. When this information is available at the collector
8
Henk Vandecasteele, Gerda Janssens
#updates
10
100
1000
10000
100000
Starting depth = 1
updating
lookup
0.16
0.28
0.39
0.52
0.66
0.08
0.15
0.20
0.27
0.34
Starting depth = 10
updating
lookup
0.27
0.32
0.31
0.42
0.57
0.11
0.15
0.12
0.20
0.29
Table 2. Starting the sequence of trees with a larger depth
node, the root node can always be replaced by one of the lower collector
nodes. Optionally one can choose to put the depth in each node, such that
computing depth while inserting can be avoided.
• When memory consumption is an issue, all leaves of the tree can be replaced
by a smaller term1 , e.g. leaf(value), or even simply the value if it is known to
be nonvar. The leaves of an open ended tree are all the nodes that occur at
the maximal depth in the trees. In ilProlog a tree/3 term takes 4 heap cells,
whereas a leaf/1 term takes only 2. As on average half of the values are leaves
and we gain 2 cells per leaf, the gain will be 1 heap cell per value. If the value
is stored directly in the leaf, 2 heap cells per value can be gained, and we
have almost the same memory consumption as with open ended lists. These
observations are confirmed by our experiments.
• When known in advance that application variables will have many updates,
one can start with larger trees in the sequence (e.g depth 10). This speeds
up the updates/lookups as can be seen in Table 2. The timings are obtained
with ilProlog.
References
Blockeel, H., Demoen, B., Dehaspe, L., Janssens, G., Ramon, J. and Vandecasteele, H.
(2000). Executing query packs in ILP., Proceedings of the 10th International Conference
in Inductive Logic Programming, volume 1866 of Lecture Notes in Artificial Intelligence,
London, UK, pp. 60–77, Springer.
Dincbas, M., Van Hentenryck, P., Simonis, H., Aggoun, A., Graf, T. and Berthier., F. (1988).
The constraint logic programming language CHIP., Proceedings of the International
Conference on Fifth Generation Computer Systems, Tokyo, pp. 693–702.
Eur (1998). ECLiPSe 3.7 User Manual.
Fredman, M. L. and Tarjan, R. E. (July 1987). Fibonacci Heaps and Their Uses in Improved
Network Optimization Algorithms Journal of the Association for Computing Machinery,
Vol. 34, No. 3, pp. 596–615.
Holzbauer, C. (1992). Meta-structures vs. Attributed Variables in the Context of Extensible Unification. Proceedings of the Fourth International Symposium on Programming
1
Thanks to a referee mentioning this
Programming pearl
9
Language Implementation and Logic Programming, volume 631 of Lecture Notes in Computer Science, Leuven, Belgium, pp. 260–268, Springer.
Le Houitouze, S. (1990). A New Data Structure for Implementing Extensions to Prolog.
Proceedings of the Second International Symposium on Programming Language Implementation and Logic Programming, volume 456 of Lecture Notes in Computer Science,
Linköping, Sweden, pp. 136–150, Springer.
Swe (2000). SICSTUS Prolog 3.8.5 User’s Manual.
Vandecasteele, H., Demoen, B. and Janssens, G. (July 2000). Compiling Large Disjunctions,
Technical report CW295,
http://www.cs.kuleuven.ac.be/publicaties/rapporten/CW2000.html Leuven.
| 6cs.PL
|
Optimal localist and distributed coding of spatiotemporal
spike patterns through STDP and coincidence detection
Timothée Masquelier12∗and Saeed Reza Kheradpisheh3
1
2
Centre de Recherche Cerveau et Cognition, UMR5549 CNRS - Université Toulouse 3, Toulouse,
France.
Instituto de Microelectrónica de Sevilla (IMSE-CNM), CSIC, Universidad de Sevilla, Sevilla, Spain.
arXiv:1803.00447v1 [cs.NE] 1 Mar 2018
3
Department of Computer Science, School of Mathematics, Statistics, and Computer Science,
University of Tehran, Tehran, Iran.
studies, in which higher thresholds led to localist
coding only.
Repeating spatiotemporal spike patterns exist and
Taken together these results suggest that coincarry information. Here we investigate how a sin- cidence detection and STDP are powerful mechagle neuron can optimally signal the presence of one nisms, compatible with distributed coding.
given pattern (localist coding), or of either one of
several patterns (distributed coding, i.e. the neu- Keywords: neural coding, coincidence detection,
ron’s response is ambiguous but the identity of the leaky integrate-and-fire neuron, multi-neuron spike
pattern could be inferred from the response of mul- sequence, spatiotemporal spike pattern, unsupertiple neurons). Intuitively, we should connect the vised learning, spike-timing-dependent plasticity
detector neuron to the neurons that fire during (STDP)
the patterns, or during subsections of them. Using a threshold-free leaky integrate-and-fire (LIF)
neuron with time constant τ , non-plastic unitary 1 Introduction
synapses and homogeneous Poisson inputs, we deElectrophysiologists report the existence of repeatrived analytically the signal-to-noise ratio (SNR)
ing spike sequences involving multiple cells, also
of the resulting pattern detector, even in the prescalled “spatiotemporal spike patterns”, with preence of jitter. In most cases, this SNR turned out
cision in the millisecond range, both in vitro and
to be optimal for relatively short τ (at most a few
in vivo, lasting from a few tens of ms to sevtens of ms). Thus long patterns are optimally deeral seconds, and these typically carry informatected by coincidence detectors working at a shorter
tion [Tiesinga et al., 2008]. How this information is
timescale, although these ignore most of the patextracted by downstream neurons is unclear. Can
terns. When increasing the number of patterns,
it be done by neurons only one synapse away from
the SNR decreases slowly, and remains acceptable
the recorded neurons? Or are multiple integration
for tens of independent patterns.
steps needed? Can it be done by simple coinciNext, we wondered if spike-timing-dependent
dence detector neurons, or should other temporal
plasticity (STDP) could enable a neuron to reach
features, such as spike ranks [Thorpe and Gautrais,
the theoretical optimum. We simulated a LIF
1998], be taken into account? Here we wondered
equipped with STDP, and repeatedly exposed it
how far we can go with the simplest scenario: the
to multiple input spike patterns. The LIF proreadout is done by simple coincidence detector neugressively became selective to every repeating patrons only one synapse away from the neurons intern with no supervision, even when the patterns
volved in the repeating patterns. We demonstrate
were embedded in Poisson activity. Furthermore,
that this approach can lead to very robust pattern
using certain STDP parameters, the resulting patdetectors, provided that the membrane time contern detectors were optimal. Tens of independent
stants are relatively short, possibly much shorter
patterns could be learned by a single neuron with
than the patterns’ durations.
a low adaptive threshold, in contrast with previous
In addition, some theoretical studies by us
∗ e-mail: [email protected]
and others have shown that neurons equipped
Abstract
1
3 A theoretical optimum
2
4
Afferents
10
with STDP can become selective to arbitrary
repeating spike patterns, even without supervision [Gilson et al., 2011, Humble et al., 2012, Hunzinger et al., 2012, Kasabov et al., 2013, Klampfl
and Maass, 2013, Krunglevicius, 2015, Masquelier,
1
2017, Masquelier et al., 2008, 2009, Nessler et al.,
0
2013, Sun et al., 2016, Yger et al., 2015]. Using
numerical simulations, we show here that the re200
sulting detectors can be close to the theoretical optimum, and that, surprisingly, a single neuron can
robustly learn up to ∼ 40 independent patterns.
This was not clear from previous studies, in which
neurons only learned one pattern (localist cod0
ing) [Gilson et al., 2011, Humble et al., 2012, Hun0
zinger et al., 2012, Kasabov et al., 2013, Klampfl
and Maass, 2013, Krunglevicius, 2015, Masquelier,
2017, Masquelier et al., 2008, 2009, Nessler et al., Fig. 1:
2013, Sun et al., 2016], or two patterns [Yger et al.,
2015]. This shows that STDP and coincidence detection are compatible with distributed coding.
0.6
t(s)
Potential
t
2
Formal description of the problem
We assessed the problem of detecting one or several spatiotemporal spike patterns with a single LIF
neuron. Intuitively, one should connect the neurons
that are active during the patterns (or during subsections of them) to the LIF neuron. That way, the
LIF will tend to be more activated by the patterns
than by some other inputs. More formally, we note
P the number of spike patterns, and assume that
they all have the same duration L. We note N the
number of neurons involved. For each pattern, we
chose a subsection with duration ∆t ≤ L, and we
connect the LIF to the M neurons that emit at least
one spike during at least one of these subsections
(see Fig. 1).
We hypothesize that all afferent neurons fire according to an homogeneous Poisson process with
rate f , both inside and outside the patterns. That
is the patterns corresponds to some realizations of
the Poisson process, which can be repeated (this
is sometimes referred to a “frozen noise”). At
each repetition a random time lag (jitter) is added
to each spike, drawn from a uniform distribution
over [−T, T ] (a normal distribution is more often used, but it would not allow analytical treatment [Masquelier, 2017]).
We also assume that synapses are instantaneous
0.6
L
t(s)
(Top) P = 2 repeating spike patterns (colored rectangles) with duration L, embedded
in Poisson noise. The LIF is connected to
the neurons that fire in some subsections of
the patterns with duration ∆t ≤ L (these
emit red spikes) (Bottom) The LIF potential peaks for patterns, and we want to optimize the SNR.
(i.e., excitatory postsynaptic currents are made of
Diracs), which facilitates the analytic calculations.
For now we ignore the LIF threshold, and we
want to optimize its signal-to-noise ratio (SNR),
defined as:
SN R =
Vmax − V noise
,
σnoise
(1)
where Vmax is the maximal potential reached during
the pattern presentations, V noise is the mean value
for the potential with Poisson input (noise period),
and σnoise is its standard deviation. Obviously, a
higher SNR means a larger difference between the
LIF membrane potential during the noise periods
and its maximum value, which occurs during the
selected ∆t window of each pattern. Therefore, the
higher the SNR the lower the probability of missing
patterns, and of false alarms.
3
3.1
A theoretical optimum
Deriving the SNR analytically
Here we are to find the optimum SNR of the LIF for
P patterns. To this end we should first calculate
3 A theoretical optimum
3
the SNR analytically. In this section, we assume
non-plastic unitary synaptic weights. That is an
afferent can be either connected (w = 1) or disconnected (w = 0). In the Appendix we estimate the
cost of this constraint on the SNR.
Since the LIF has instantaneous synapses and
firing is Poisson, during the noise periods and outside the ∆t
p windows we have: V noise = τ f M and
σnoise = τ f M/2 [Burkitt, 2006], where τ is the
membrane’s time constant and M is the number of
connected input neurons (with unitary weights).
To compute Vmax , it is convenient to introduce
the reduced variable:
vmax =
Vmax − V noise
V
∞
− V noise
,
(2)
In Section 3.2 we justify this last approximation
through numerical simulations, and we also show
that this mean SNR scenario is not much different
from the SNR of particular Poisson realizations.
The last step to compute hSN Ri in Equation 5
is to calculate hM i and hri. Here, we consider independent patterns, i.e. with chance-level overlap.
The probability that a given afferent fires at least
once in a given pattern subsection of length ∆t is
p = 1 − e−f ∆t , since firing is Poisson with rate
λ = f ∆t. Because an afferent can be involved
in any number of patterns, from the exclusioninclusion principle [Berndt and Brualdi, 1980], it
follows that the number of selected afferents M is
on average:
∞
where V = τ r is the mean potential of the steady
regime that would be reached if ∆t was infinite,
and r is the effective input spike rate during the
∆t window [Masquelier, 2017], that is the rate of
input spikes received by the LIF neuron from the
M afferents with synapses of weight one.
vmax can be calculated by exact integration of the
LIF differential equation [Masquelier, 2017]. Here
we omit the derivation and present the final equation:
∆t
vmax = min 1,
2T
τ
−
log 1 − e− max(∆t,2T )/τ + e−|∆t−2T |/τ .
2T
(3)
Using the definition of vmax in Equation 2, we
can rewrite the SNR equation as:
SN R = vmax
V
∞
− V noise
.
σnoise
(4)
Obviously, different Poisson pattern realizations
will lead to different values for M and r that con∞
sequently affect each of the terms V , V noise and
σnoise . Here we want to compute the expected value
of the SNR across different Poisson pattern realizations:
* ∞
+
V − V noise
hSN Ri = vmax
σnoise
p
r − fM
(5)
√
= vmax 2τ /f
M
p
hri − f hM i
p
≈ vmax 2τ /f
.
hM i
hM i = N
P
X
P
(−1)i+1 (1 − e−f ∆t )i .
i
i=1
(6)
And finally, the expected effective input spike
rate during the ∆t window is the expected total
number of spikes, f N ∆t, divided by ∆t, thus:
hri = f N.
We note that the SNR scales with
3.2
(7)
√
N.
Numerical validations
We first checked if the approximation we made
in Equation 5 is reasonable. As can be seen on
Figure 2, the average SNR across different Poisson patterns is very close to the SNR corresponding to the average-case scenario, i.e. M = hM i
and r = hri (as defined by Equations 6 and 7 respectively). Note that this Figure was done with
relatively small values for the parameters P , ∆t
and f (respectively 1, 2ms, and 1Hz). Our simulations indicate that when increasing these parameter values, the approximation becomes even better
(data not shown). Additionally, the left plot in Figure 2 shows that the SNR does not change much for
different Poisson pattern realizations and that the
average SNR well represents the SNR distribution
even for the worst and best cases.
Next, we verified the complete SNR formula
(Eq. 5), which also includes vmax , through numerical simulations. We used a clock-based approach,
and integrated the LIF equation using the forward
Euler method with a 0.1ms time bin. We used
3 A theoretical optimum
4
Fig. 2: Numerical validation of the averaging approximation. (Left) M × r plane. The black dots
correspond to different realizations of a Poisson pattern (a jitter was added to better visualize
√
density, given that both M and r are discrete). The background color shows (r − f M )/ M .
The red cross corresponds√to the average-case scenario M = hM i and r = hri. (Right) The
distribution of (r − f M )/ M values across Poisson realizations. The vertical blue solid
p line
shows its average. The vertical red dotted line shows our approximation, (hri − f hM i)/ hM i,
which matches very well the true average. Parameters: P = 1, N = 104 , ∆t = 2ms, f = 1Hz.
Simulations
Theory
80
70
60
SNR
50
40
30
20
P = 1 and P = 5 patterns, and performed 100
simulations with different random Poisson patterns
of duration L = 20ms, involving N = 104 neurons
with rate f = 5Hz. We chose ∆t = L = 20ms,
i.e. the LIF was connected to all the afferents that
emitted at least once during one of the patterns. In
order to estimate Vmax , each pattern was presented
1000 times, every 400ms. The maximal jitter was
T = 5ms. Between pattern presentations, the afferents fired according to a Poisson process, still with
rate f = 5Hz, which allowed to estimate V noise
and σnoise . We could thus compute the SNR from
Equation 1 (and its standard deviation across the
100 simulations), which, as can be seen on Figure 3,
matches very well the theoretical values, for P = 1
and 5.
10
3.3
0
1
5
Optimizing the SNR
We now want to optimize the SNR given by Equation 5. We consider that f and T are external variables, and that we have the freedom to choose τ
Fig. 3: Numerical validation of the theoretical SNR
and ∆t. We also add the constraint τ f M ≥ 10
values, for P = 1 and 5 patterns. Error bars
(large number of synaptic inputs), so that the disshow ±1 s.d.
tribution of V is approximately Gaussian [Burkitt,
P
4 Simulations show that STDP can be close-to-optimal
2006]. Otherwise, it would be positively skewed,
thus a high SNR would not guarantee a low false
alarm rate. We assume that L is sufficiently large so
that an upper bound for ∆t is not needed. We used
the Matlab R2017a Optimization Toolbox (MathWorks Inc., Natick, MA, USA) to compute the optimum numerically.
Figure 4 illustrates the results with P = 2. One
can make the following observations (similar to
our previous paper which was limited to P = 1
[Masquelier, 2017]):
• Unless f and T are both high, the optimal τ
and ∆t have the same order of magnitude (see
Figure 4 middle). Unless T is high (>10ms), or
f is low (<1Hz), then these timescales should
be relatively small (at most a few tens of ms;
see Figure 4 left). This means that even long
patterns (hundreds of ms or more) are optimally detected by a coincidence detector working at a shorter timescale, and which thus ignores most of the patterns. One could have
thought that using τ ∼ L, to integrate all
the spikes from the pattern would be the best
strategy. Instead, it is more optimal to use
subpatterns as signatures for the whole patterns. This could explain the apparent paradox between typical ecological stimulus durations (hundreds of ms or above) and the neuronal integration timescales (at most a few tens
of ms).
• The constraint τ f M ≥ 10 imposes larger τ
when both f and T are small. In the other
cases, it is naturally satisfied. (see Figure 4
left)
• Unsurprisingly, the optimal SNR decreases
with T (see Figure 4 right). What is more
surprising, is that it also decreases with f . In
other words, sparse activity is preferable. We
will come back to this point in the discussion.
What is the biological range for T , which corresponds to the spike time precision? Millisecond
precision in cortex has been reported [Havenith
et al., 2011, Kayser et al., 2010, Panzeri and Diamond, 2010]. We are aware that other studies
found poorer precision, but this could be due to
uncontrolled variable or the use of inappropriate
reference times [Masquelier, 2013].
5
We now focus, as an example, on the point on
the middle of the T × f plane – T = 3.2ms and
f = 3.2Hz – and vary P (Fig. 5). When increasing
P , the optimal τ and ∆t decrease. Unsurprisingly,
the resulting SNR also decreases, but only slowly.
It thus remains acceptable for several tens of independent patterns (e.g. SNR ∼ 7 for P = 40).
4
Simulations show that STDP can be
close-to-optimal
Next we investigated, through numerical simulations, if STDP could turn a LIF neuron into an
optimal multi-pattern detector. More specifically,
since STDP does not adjust τ , we set it to the
optimal value and investigated whether STDP
could learn all the patterns with an optimal
∆t. We used a clock-based approach, and the
forward Euler method with a 0.1ms time bin. The
Matlab R2017a code for these simulations will be
made available in ModelDB [Hines et al., 2004]
at https://senselab.med.yale.edu/modeldb/
once this paper is accepted in a peer-reviewed
journal.
4.1
Input spikes
The setup we used was similar to the one of our previous studies [Gilson et al., 2011, Masquelier, 2017,
Masquelier et al., 2008, 2009]. We used N = 104
afferents, which is in the biological range. Between
pattern presentations, the input spikes were generated randomly with a homogeneous Poisson process with rate f . The P spike patterns with duration L = 100ms were generated only once using
the same Poisson process (frozen noise). The pattern presentations occurred every 400ms (in previous studies, we demonstrated that irregular intervals did not matter [Gilson et al., 2011, Masquelier
et al., 2008, 2009], so here regular intervals were
used for simplicity). The P patterns were presented
alternatively, over and over again. Figure 6 shows
an example with P = 2 patterns. At each pattern
presentation, all the spike times were shifted independently by some random jitters uniformly distributed over [−T, T ].
4 Simulations show that STDP can be close-to-optimal
SNR
t/
1s
100
6
100
100
80
10ms
1
1ms
0.1
0.0001 0.001 0.01
10
0.5
1
f (Hz)
f (Hz)
10
f (Hz)
1
100ms
10
60
40
1
20
0.1
0.0001 0.001 0.01
0.1ms
0.1
T (s)
0
0.1
0.1
0.0001 0.001 0.01
T (s)
0.1
0
T (s)
Fig. 4: Optimal parameters for P = 2, as a function of f and T . (Left) Optimal τ (note the logarithmic
colormap). (Middle) Optimal ∆t, divided by τ . (Right) Resulting SNR.
100
0.02
t
SNR
s
50
4.3
0.01
0
10
20
30
40
0
50
P
Fig. 5: Optimal τ and ∆t (for f = 3.2Hz, T =
3.2ms) and resulting SNR as a function of
P.
4.2
value for θ0 could lead to the optimum. We thus
performed and exhaustive search, using a geometric
progression with a ratio of 2.5%.
A LIF neuron with adaptive
threshold
We simulated a LIF neuron connected to all of
the N afferents with plastic synaptic weights wi ∈
[0, 1]. The LIF had instantaneous synapses (i.e., excitatory postsynaptic currents are made of Diracs),
and an adaptive threshold. This adaptive threshold was increased by a fixed amount (1.8θ0 ) at each
postsynaptic spike, and then exponentially decayed
towards its baseline value θ0 with a time constant
τθ = 80ms. This is a simple, yet good model of
cortical cells, in the sense that it predicts very well
the spikes elicited by a given input current [Gerstner and Naud, 2009, Kobayashi et al., 2009]. Here,
such an adaptive threshold is crucial to encourage
the neuron to learn multiple patterns, as opposed to
fire multiple successive spikes to the same pattern.
Since the theory developed in the previous sections
ignored the LIF threshold, we did not know which
Synaptic plasticity
Initial synaptic weights were all equal. Their value
was computed so that V noise = θ + σnoise (leading
to an initial firing rate of about 4Hz, see Fig. 6
top). They then evolved in [0, 1] with all-to-all
spike STDP. Yet, we only modeled the Long Term
Potentiation part of STDP, ignoring its Long Term
Depression (LTD) term. As in Song et al. [2000], we
used a trace of presynaptic spikes at each synapse
i, Aipre , which was incremented by δApre at each
presynaptic spike, and then exponentially decayed
towards 0 with a time constant τpre = 20ms. At
each postsynaptic spike this trace was used for LTP
at each synapse: wi → wi + wi (1 − wi )Aipre .
Here LTD was modeled by a simple homeostatic mechanism. At each postsynaptic spike, all
synapses were depressed: wi → wi + wi (1 − wi )wout
where wout < 0 is a fixed parameter [Kempter
et al., 1999].
Note that for both LTP and LTD we used the
multiplicative term wi (1 − wi ), in contrast with
purely additive STDP, with which the ∆w is independent of the current weight value [Kempter et al.,
1999, Song et al., 2000]. This multiplicative term
ensures that the weights remain in the range [0,1],
and the weight dependence creates a soft bound effect: when a weight approaches a bound, weight
changes tend toward zero. Here it was found to
increase performance (convergence time and stability), in line with our previous studies [Kheradpisheh et al., 2016, 2018, Masquelier and Thorpe,
Potential
2400.2
After convergence
80.2
During learning
0.2
t (s)
t (s)
t (s)
2402.2
82.2
2.2
0
t
t (s)
Pattern 1
Pattern 1
0.1
10
4
0
t (s)
Pattern 1
0.1
10
4
10
4
1
0
opt
t (s)
0.1
10
4
1
4
1
10
1
4
1
10
1
0
0
0
t
opt
t (s)
t (s)
Pattern 2
t (s)
Pattern 2
Pattern 2
0.1
0.1
0.1
Fig. 6: Unsupervised STDP-based pattern learning. The neuron becomes selective to P = 2 patterns. (Top) Initial state. On the left,
we plotted the neuron’s potential as a function of time. Colored rectangles indicate pattern presentations. Next, we plotted
the two spike patterns, coloring the spikes as a function of the corresponding synaptic weights: blue for low weight, purple for
intermediate weight, and red for high weight. Initial weights were uniform. (Middle) During learning. Selectivity progressively
emerges. (Bottom) After convergence. STDP has concentrated the weights on the afferents which fire at least once in at least one
of the pattern subsections, located at the beginning of each pattern, and whose duration roughly matches the optimal ∆t (shown
in green). This results in one postsynaptic spike each time either one of the two pattern is presented. Elsewhere both V noise and
σnoise are low, so the SNR is high. In addition V noise roughly matches the optimal value (in green). We also indicate the the
optimal value for Vmax (in green). However, the potential never reaches it, because the adaptive threshold is reached before.
0
400
0
400
0
Initial state
Afferent
Afferent
Afferent
Potential
Potential
Afferent
Afferent
Afferent
400
4 Simulations show that STDP can be close-to-optimal
7
5 Discussion
8
0.45
1 pattern
2 patterns
5 patterns
10 patterns
20 patterns
40 patterns
0.4
Convergence index
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
2000
4000
6000
8000
10000 12000
t (s)
Fig. 7: Convergence index as a function of time and
number of patterns, for an example of optimal simulation. The convergence index
is defined as the mean distance between
the full precision weights, and their binary
quantization (0 if w < 0.5, and 1 otherwise).
2007, Mozafari et al., 2017].
The ratio between LTP and LTD, that is between
δApre and wout is crucial: the higher, the more
synapses reach the upper bound (1) after convergence. Here we chose to keep δApre = 0.1 and to
systematically vary wout , using again a geometric
progression with a ratio of 2.5%.
4.4
Results
For each θ0 × wout point, 100 simulations were performed with different random pattern realizations,
and we computed the proportion of “optimal” ones
(see below), and reported it in Table 1. After
12,000s of simulated time, the synaptic weights had
converged by saturation. That is synapses were either completely depressed (w = 0), or maximally
reinforced (w = 1). A simulation was considered
optimal if all the patterns were learned, and in
an optimal way, that is if all patterns exhibited
a subsection in which all spikes corresponded to
maximally reinforced synapses (w = 1), and whose
duration roughly matched the theoretical optimal
∆t. In practice, we used the total number of reinforced synapses as a proxy of the mean subsection
duration (since there is a non-ambiguous mapping
between the two variables, given by Equation 6),
and checked if this number matched the theoretical
optimal M (Eq. 6) with a 5% margin. Note that
the learned subsections typically corresponded to
the beginning of the patterns, because STDP tracks
back through them [Gilson et al., 2011, Masquelier
et al., 2008, 2009], but this is irrelevant here since
all the subsections are equivalent for the theory.
Figure 6 shows an optimal simulation with P = 2
patterns.
As can be seen on Table 1, the proportion of optimal simulations decreases with P , as expected. But
more surprisingly, several tens of patterns can be
optimally learned with reasonably high probability.
With P = 40 the probability of optimal simulations
is only 58%, but the average number of learned patterns is high: 39.5! This means that nearly all patterns are learned in all simulations, yet sometimes
in a suboptimal manner. Finally, Figure 7 shows
that convergence time increases with P .
5
Discussion
The fact that STDP can generate selectivity to
any repeating spike pattern in an unsupervised
manner is a remarkable, yet well documented
fact [Gilson et al., 2011, Humble et al., 2012, Hunzinger et al., 2012, Kasabov et al., 2013, Klampfl
and Maass, 2013, Krunglevicius, 2015, Masquelier,
2017, Masquelier et al., 2008, 2009, Nessler et al.,
2013, Sun et al., 2016, Yger et al., 2015]. Here we
have shown that, in addition, a single neuron can
become optimally selective to several tens of independent patterns. This shows that STDP and coincidence detection are compatible with distributed
coding. Yet one issue with having one neuron selective to multiple patterns is stability. If one of
the learned pattern does not occur for a long period during which the other patterns occur many
times, causing postsynaptic spikes, the unseen pattern will tend to be forgotten. This is not an issue
with localist coding: if the learned pattern does
not occur, the threshold is hardly ever reached so
the weights are not modified, and the pattern is
retained indefinitely, even if STDP is “on” all the
time.
Another issue with distributed coding is how the
readout is done, that is how the identity of the
5 Discussion
9
Tab. 1: Performance as a function of the number of patterns P . The first four lines are computed from
the theoretical optimum. The next two lines are the optimal values found through exhaustive
search (see text). The last four lines are performance indicators, estimated during the last 100
presentations of each pattern. hPlearned i is the mean number of “learned patterns”, that is by
convention patterns which elicit at least one postsynaptic spike. The following line is the mean
hit rate for those patterns. The subsequent line gives the false alarm rate, but we never observed
any here. Finally P(opt) is the proportion of optimal cases.
P
∆topt (ms)
τ opt (ms)
M opt
SN Ropt
θ0
wout
hPlearned i
Hit rate (%)
False alarms (Hz)
P(opt) (%)
5
11
8.9
1600
31
190
−6.2 10−3
5
98.9
0
100
10
8.1
6.8
2300
20
140
−6.3 10−3
10
98.6
0
100
stimulus can be inferred from multiple neuron responses, given that each response is ambiguous?
This is out of the scope of the current paper, but
we suspect that STDP could again help. As shown
in this study, each neuron equipped with STDP
can learn to fire to multiple independent stimuli.
Let’s suppose that stimuli are shown one at a time.
When stimulus A is shown, all the neurons that
learned this stimuli (among others) will fire synchronously. Let us call S this set of neurons. A
downstream neuron equipped with STDP could
easily become selective to this synchronous volley
of spikes from neurons in S [Brette, 2012]. With
an appropriate threshold, this neuron would fire if
and only if all the neurons in S have fired. Does
that necessarily mean that A is there? Yes, if the
intersection of the sets of stimuli learned by neurons in S only contains A. In the general case, the
intersection is likely to be small, much smaller than
the typical sets of stimuli learned by S neurons, so
much of the ambiguity should be resolved.
Here the STDP rule we used always led to binary
weights after learning. That is an afferent could be
either selected or discarded. We thus could use our
SNR calculations derived with binary weights, and
checked that the selected set was optimal given the
binary weight constraint. Further calculations in
the Appendix suggest that removing such a con-
20
5.7
5.6
3100
12
110
−6.5 10−3
20
97.9
0
100
40
3.7
5.1
3800
6.7
92
−6.7 10−3
39.5
96.5
0
58
straint could lead to a modest increase in SNR,
of about 10%. More research is needed to see if
a multiplicative STDP rule, which does not converge towards binary weights [Gütig et al., 2003,
van Rossum et al., 2000], could lead to the optimal
graded weights.
Our theoretical study suggests that synchrony
could be an important part of the neural
code [Stanley, 2013], that it is computationally efficient [Brette, 2012, Gütig and Sompolinsky, 2006],
and that coincidence detection could be the main
function of neurons [Abeles, 1982, König et al.,
1996]. In line with this proposal, neurons in vivo
appear to be mainly fluctuation-driven, not meandriven [Brette, 2012, 2015, Rossant et al., 2011].
This is the case in particular in the balanced
regime [Brette, 2015], which appears to be the
prevalent regime in the brain [Denève and Machens,
2016]. Several other points suggest that coincidence
detection is the main function of neurons. Firstly,
strong feedforward inhibitory circuits throughout
the central nervous system often shorten the neurons’ effective integration windows [Bruno, 2011].
Secondly, the effective integration time constant in
dendrites might be one order of magnitude shorter
than the soma’s one [König et al., 1996]. Finally, recent experiments indicate that a neuron’s threshold
quickly adapts to recent potential values [Fontaine
5 Discussion
et al., 2014, Mensi et al., 2016, Platkiewicz and
Brette, 2011], so that only a sudden potential increase can trigger a postsynaptic spike. This enhances coincidence detection.
Our results show that, somewhat surprisingly,
lower firing rates lead to better signal-to-ratio. It is
worth mentioning that mean firing rates are probably largely overestimated in the electrophysiological literature, because extracellular recordings – by
far the most popular technique – are totally blind
to cells that do not fire at all [Thorpe, 2011]. Even
a cell that fire only a handful of spikes will be ignored, because spike sorting algorithms need tens
of spikes from a given cell before they can create
a new cluster corresponding to that cell. Furthermore, experimentalists tend to search for stimuli
that elicit strong responses, and, when they can
move the electrode(s), tend to look for most responsive cells, introducing strong selection biases.
Mean firing rates, averaged across time and cells,
are largely unknown, but they could be smaller
than 1 Hz [Shoham et al., 2006]. It seems like
coding is sparse: neurons only fire when they need
to signal an important event, and that every spike
matters [Wolfe et al., 2010].
Acknowledgments
10
should put strong weights on the synapses corresponding to the most recent pattern spikes, since
these weights will increase Vmax more than Vnoise .
Conversely, very old pattern spikes that fall outside
the integration window (if any) should be associated to nil weights: any positive value would only
increase Vnoise , not Vmax . But between those two
extremes, it might be a good idea to use intermediate weight values.
To check this intuition, we used numerical optimizations using a simplified setup. We used a single
pattern (P = 1), that was repeated in the absence
of jitter (T = 0). We divided the pattern into n
different periods ∆t1 , ...∆tn (in reverse chronological order), each one corresponding to a different
synaptic weight w1 , ...wn (see Figure 8 – left for
an example with n = 2). More specifically: the
M1 afferents that fire in the ∆t1 window are connected with weight w1 . The M2 afferents that fire
in the ∆t2 window, but not in the ∆t1 one, are
connected with weight w2 . More generally, the Mi
afferents that fire in the ∆ti window, but not in the
∆t1 ...∆ti−1 ones, are connected with weight wi .
With this simple set up, the SNR can be computed analytically. For example, if n = 2 (Fig. 8 –
left), we have:
hM1 i = N (1 − e−f ∆t1 ),
(8)
This research received funding from the European
Research Council under the European Union’s 7th
hM2 i = N (1 − e−f ∆t2 )e−f ∆t1 .
(9)
Framework Program (FP/2007-2013) / ERC Grant
The asymptotic steady regimes for the two time
Agreement n.323711 (M4 project). We thank Milad
windows are:
Mozafari for smart implementation hints, and Jean
Pierre Jaffrezou for his excellent copy editing.
hV1∞ i = τ f w1 N,
(10)
Appendix
hV2∞ i = τ f w2 N + (w1 − w2 ) hM1 i .
Graded weights
(11)
Let’s call Vi the potential at the end of window
In this paper, we assumed unitary (or binary) ∆ti , and Vn+1 = Vnoise . Then Vmax = V1 can be
synaptic weights: all connected afferents had the computed iteratively:
same synaptic weight1 . This constraint strongly
V2 = (1 − e−∆t2 /τ )(V2∞ − V3 ),
(12)
simplified the analytical calculations. But could the
SNR be even higher if we removed this constraint,
and by how much? Intuitively, when one wants to
V1 = (1 − e−∆t1 /τ )(V1∞ − V2 ).
(13)
detect a spike pattern that has just occurred, one
Furthermore [Burkitt, 2006],
1 Numerical simulations with STDP used graded
weights during learning, but not after convergence.
Vnoise = τ f (w1 M1 + w2 M2 ),
(14)
5 Discussion
11
and:
q
τ f (w12 M1 + w22 M2 )/2.
0.8
(15)
exp(t/ )
f=1Hz
f=5Hz
f=10Hz
So we have everything we need to compute the
SNR.
Equations 8 – 15 can be generalized to n > 2:
hMi i = N (1 − e
−f ∆ti
−f
)e
i−1
P
j=1
0.6
V2
w
σnoise =
1
V1
V max
0.4
∆tj
,
0.2
(16)
V noise
0
t2
hVi∞ i = τ f wi N +
t1
-0.06
-0.04
-0.02
0
t (s)
i−1
X
(wj − wi ) Mj
(17) Fig. 8: Optimization with graded weights. (Left)
Didactic example with n = 2 weight values: w1 for all the afferents that fire in the
and Vmax = V1 can be computed iteratively from
∆t1 window, and w2 < w1 for all the afferVn+1 = Vnoise using:
ents that fire in the ∆t2 window but not in
−∆ti−1 /τ
∞
the ∆t1 one. V1∞ and V2∞ are the asympVi−1 = (1 − e
)(Vi−1 − Vi ).
(18)
totic potentials for the two periods. Vmax
can be computed from those two values (see
Furthermore [Burkitt, 2006],
text). (Right) Numerical optimization of
X
the weights with n = 70. With small f ,
Vnoise = τ f
w i Mi ,
(19)
the optimal solution appears to be close to
and:
et/τ .
q X
σnoise = τ f
wi2 Mi /2.
(20)
j=1
So the SNR can be computed whatever n, and,
importantly, it is differentiable with respect to the
wi . We can thus use efficient numerical methods to
optimize these weights. Since scaling the weights
does not change the SNR, we imposed w1 = 1.
Figure 8 – right gives an example with n = 70. Here
the ∆ti were all equal to 5τ /n, and we optimized
the corresponding wi . We chose τ = 10ms, and
f = 1, 5, and 10Hz. The gain w.r.t. binary weights
for the SNR were modest: 10.5%, 9.6% and 8.9%
respectively. As f tends towards 0, the optimal
weights appears to converge towards et/τ (even if
we could not prove it).
References
Abeles, M. (1982). Role of the cortical neuron: integrator or coincidence detector? Isr J Med Sci.,
18(1):83–92.
Berndt, B. C. and Brualdi, R. A. (1980). Introductory Combinatorics. The American Mathematical Monthly, 87(6):498.
Brette, R. (2012).
Computing with neural synchrony.
PLoS computational biology,
8(6):e1002561.
Brette, R. (2015).
Philosophy of the Spike:
Rate-Based vs. Spike-Based Theories of the
Brain.
Frontiers in Systems Neuroscience,
9(November):1–14.
Bruno, R. M. (2011). Synchrony in sensation. Current Opinion in Neurobiology, 21(5):701–708.
Burkitt, A. N. (2006). A review of the integrateand-fire neuron model: I. Homogeneous synaptic
input. Biological Cybernetics, 95(1):1–19.
Denève, S. and Machens, C. K. (2016). Efficient
codes and balanced networks. Nature Neuroscience, 19(3):375–382.
Fontaine, B., Peña, J. L., and Brette, R. (2014).
Spike-Threshold Adaptation Predicted by Membrane Potential Dynamics In Vivo. PLoS Computational Biology, 10(4):e1003560.
5 Discussion
Gerstner, W. and Naud, R. (2009). How good are
neuron models? Science, 326:379–380.
12
Kempter, R., Gerstner, W., and van Hemmen, J. L.
(1999). Hebbian learning and spiking neurons.
Phys Rev E, 59(4):4498–4514.
Gilson, M., Masquelier, T., and Hugues, E. (2011).
STDP allows fast rate-modulated coding with Kheradpisheh, S. R., Ganjtabesh, M., and MasquePoisson-like spike trains. PLoS computational bilier, T. (2016). Bio-inspired unsupervised learnology, 7(10):e1002231.
ing of visual features leads to robust invariant object recognition. Neurocomputing, 205:382–392.
Gütig, R., Aharonov, R., Rotter, S., and Sompolinsky, H. (2003). Learning input correlations Kheradpisheh, S. R., Ganjtabesh, M., Thorpe,
through nonlinear temporally asymmetric HebS. J., and Masquelier, T. (2018). STDP-based
bian plasticity. J Neurosci, 23(9):3697–3714.
spiking deep convolutional neural networks for
object recognition. Neural Networks, 99:56–67.
Gütig, R. and Sompolinsky, H. (2006). The tempotron: a neuron that learns spike timing-based Klampfl, S. and Maass, W. (2013). Emergence of
decisions. Nat Neurosci, 9(3):420–428.
Dynamic Memory Traces in Cortical Microcircuit Models through STDP. Journal of NeuroHavenith, M. N., Yu, S., Biederlack, J., Chen, N.science, 33(28):11515–11529.
H., Singer, W., and Nikolic, D. (2011). Synchrony makes neurons fire in sequence, and stim- Kobayashi, R., Tsubo, Y., and Shinomoto, S.
ulus properties determine who is ahead. J Neu(2009). Made-to-order spiking neuron model
rosci, 31(23):8570–8584.
equipped with a multi-timescale adaptive threshold. Frontiers in computational neuroscience,
Hines, M. L., Morse, T., Migliore, M., Carnevale,
3(July):9.
N. T., and Shepherd, G. M. (2004). ModelDB:
A Database to Support Computational Neuro- König, P., Engel, A. K., and Singer, W. (1996).
science. Journal of computational neuroscience,
Integrator or coincidence detector? The role of
17(1):7–11.
the cortical neuron revisited. Trends Neurosci,
19(4):130–7.
Humble, J., Denham, S., and Wennekers, T. (2012).
Spatio-temporal pattern recognizers using spik- Krunglevicius, D. (2015).
Competitive STDP
ing neurons and spike-timing-dependent plasticLearning of Overlapping Spatial Patterns. Neuity. Frontiers in computational neuroscience,
ral Computation, 27(8):1673–1685.
6(October):84.
Masquelier, T. (2013). Neural variability, or lack
Hunzinger, J. F., Chan, V. H., and Froemke,
thereof.
Frontiers in Computational NeuroR. C. (2012).
Learning complex temporal
science, 7:1–7.
patterns with resource-dependent spike timingdependent plasticity. Journal of Neurophysiol- Masquelier, T. (2017). STDP allows close-tooptimal spatiotemporal spike pattern detection
ogy, 108(2):551–566.
by single coincidence detector neurons. NeuroKasabov, N., Dhoble, K., Nuntalid, N., and Inscience.
diveri, G. (2013). Dynamic evolving spiking
neural networks for on-line spatio- and spectro- Masquelier, T., Guyonneau, R., and Thorpe, S. J.
(2008). Spike timing dependent plasticity finds
temporal pattern recognition. Neural networks,
the start of repeating patterns in continuous
41(1995):188–201.
spike trains. PLoS ONE, 3(1):e1377.
Kayser, C., Logothetis, N. K., and Panzeri, S.
(2010). Millisecond encoding precision of audi- Masquelier, T., Guyonneau, R., and Thorpe, S. J.
tory cortex neurons. Proc Natl Acad Sci U S A,
(2009). Competitive STDP-Based Spike Pattern
107(39):16976–16981.
Learning. Neural Comput, 21(5):1259–1276.
5 Discussion
13
Masquelier, T. and Thorpe, S. J. (2007). Unsu- Sun, H., Sourina, O., and Huang, G.-B. (2016).
pervised learning of visual features through spike
Learning Polychronous Neuronal Groups Ustiming dependent plasticity. PLoS Comput Biol,
ing Joint Weight-Delay Spike-Timing-Dependent
3(2):e31.
Plasticity. Neural Computation, 28(10):2181–
2212.
Mensi, S., Hagens, O., Gerstner, W., and Pozzorini,
C. (2016). Enhanced Sensitivity to Rapid Input Thorpe, S. J. (2011). Grandmother Cells and
Fluctuations by Nonlinear Threshold Dynamics
Distributed Representations. Visual population
in Neocortical Pyramidal Neurons. PLOS Comcodes-Toward a common multivariate framework
putational Biology, 12(2):e1004761.
for cell recording and functional imaging, pages
23–51.
Mozafari, M., Kheradpisheh, S. R., Masquelier,
T., Nowzari-Dalini, A., and Ganjtabesh, M. Thorpe, S. J. and Gautrais, J. (1998). Rank Order
(2017). First-spike based visual categorization
Coding. In Bower, J. M., editor, Computational
using reward-modulated STDP. arXiv.
Neuroscience : Trends in Research, pages 113–
118. New York: Plenum Press.
Nessler, B., Pfeiffer, M., Buesing, L., and Maass,
W. (2013). Bayesian Computation Emerges in Tiesinga, P., Fellous, J.-M., and Sejnowski, T. J.
Generic Cortical Microcircuits through Spike(2008). Regulation of spike timing in visual corTiming-Dependent Plasticity. PLoS Computatical circuits. Nat Rev Neurosci, 9(2):97–107.
tional Biology, 9(4):e1003037.
van Rossum, M. C., Bi, G. Q., and Turrigiano,
Panzeri, S. and Diamond, M. E. (2010). InforG. G. (2000). Stable Hebbian learning from
mation carried by population spike times in the
spike timing-dependent plasticity. J Neurosci,
whisker sensory cortex can be decoded without
20(23):8812–8821.
knowledge of stimulus time. Frontiers in SynapWolfe, J., Houweling, A. R., and Brecht, M. (2010).
tic Neuroscience, 2(17):1–14.
Sparse and powerful cortical spikes. Curr Opin
Platkiewicz, J. and Brette, R. (2011). Impact
Neurobiol, 20(3):306–312.
of Fast Sodium Channel Inactivation on Spike
Threshold Dynamics and Synaptic Integration. Yger, P., Stimberg, M., and Brette, R. (2015). Fast
Learning with Weak Synaptic Plasticity. Journal
PLoS Computational Biology, 7(5).
of Neuroscience, 35(39):13351–13362.
Rossant, C., Leijon, S., Magnusson, A. K., and
Brette, R. (2011). Sensitivity of noisy neurons to
coincident inputs. The Journal of neuroscience
: the official journal of the Society for Neuroscience, 31(47):17193–206.
Shoham, S., O’Connor, D. H., and Segev, R.
(2006). How silent is the brain: is there a
”dark matter” problem in neuroscience? Journal of comparative physiology. A, Neuroethology, sensory, neural, and behavioral physiology,
192(8):777–84.
Song, S., Miller, K. D., and Abbott, L. F. (2000).
Competitive hebbian learning through spiketiming-dependent synaptic plasticity. Nat Neurosci, 3(9):919–926.
Stanley, G. B. (2013). Reading and writing the neural code. Nature Neuroscience, 16(3):259–263.
| 9cs.NE
|
A Digital Hardware Fast Algorithm and FPGA-based Prototype for
a Novel 16-point Approximate DCT for Image Compression
Applications
arXiv:1702.01805v1 [cs.MM] 6 Feb 2017
F. M. Bayer∗
R. J. Cintra†
A. Edirisuriya‡
A. Madanayake‡
Abstract
The discrete cosine transform (DCT) is the key step in many image and video coding standards.
The 8-point DCT is an important special case, possessing several low-complexity approximations widely
investigated. However, 16-point DCT transform has energy compaction advantages. In this sense, this
paper presents a new 16-point DCT approximation with null multiplicative complexity. The proposed
transform matrix is orthogonal and contains only zeros and ones. The proposed transform outperforms
the well-know Walsh-Hadamard transform and the current state-of-the-art 16-point approximation. A fast
algorithm for the proposed transform is also introduced. This fast algorithm is experimentally validated
using hardware implementations that are physically realized and verified on a 40 nm CMOS Xilinx
Virtex-6 XC6VLX240T FPGA chip for a maximum clock rate of 342 MHz. Rapid prototypes on FPGA
for 8-bit input word size shows significant improvement in compressed image quality by up to 1-2 dB at
the cost of only eight adders compared to the state-of-art 16-point DCT approximation algorithm in the
literature [S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A novel transform for image compression.
In Proceedings of the 53rd IEEE International Midwest Symposium on Circuits and Systems (MWSCAS),
2010].
Keywords
DCT Approximation, Fast algorithms, FPGA
1
Introduction
The discrete cosine transform (DCT) [1, 12, 38] is a pivotal tool in digital signal processing, whose popularity
is mainly due to its good energy compaction properties. In fact, the DCT is a robust approximation for the
optimal Karhunen-Loève transform when first-order Markov signals, such as images, are considered [12, 30, 38].
Indeed, the DCT has found application in several image and video coding schemes [5, 12], such as JPEG [36],
MPEG-1 [39], MPEG-2 [22], H.261 [23], H.263 [24], and H.264 [32, 46, 51].
Through the decades signal processing literature has been populated with efficient methods for the DCT
computation, collectively known as fast algorithms. This can be observed in several works with efficient
hardware and software implementations, including [2, 3, 13, 14, 17, 21, 31, 44]. Methods such as the Arai DCT
algorithm [2] can greatly reduced the number of arithmetic operations required for the DCT evaluation.
∗ F.
M. Bayer is with the Departamento de Estatı́stica, Universidade Federal de Santa Maria. E-mail: [email protected]
J. Cintra is with the Signal Processing Group, Departamento de Estatı́stica, Universidade Federal de Pernambuco. E-mail:
[email protected]
‡ A. Edirisuriya and A. Madanayake are with the ECE, The University of Akron, Akron, OH, USA
† R.
1
Indeed, current algorithms for the exact DCT are mature and further complexity reductions are very difficult
to achieve. Nevertheless, demands for real-time video processing and transmission are increasing [27, 42].
Therefore, complexity reductions for the DCT must be obtained using different methods.
One possibility is the development of approximate DCT algorithms. Approximate transforms aim
at demanding very low complexity while offering a close estimate of the exact calculation. In general,
the elements of approximate transform matrices require only {0, ±1/2, ±1, ±2} [15]. This implies null
multiplicative complexity; only addition and bit shifting operations are usually required. While not computing
the DCT exactly, such approximations can provide meaningful estimations at low-complexity requirements.
In particular, 8-point DCT approximations have been attracting signal processing community attention.
This particular blocklength is widely adopted in several image and video coding standards, such as JPEG
and MPEG family [5, 30, 36]. Prominent 8-point DCT approximations include the signed discrete cosine [20],
the level 1 approximation by Lengwehasatit-Ortega [29], the Bouguezel-Ahmad-Swamy (BAS) series of
algorithms [8–11], and the DCT round-off approximations [4, 15]. However, transforms with blocklength
greater than eight has several advantages such as better energy compaction and reduced quantization error [16].
In [16], an adapted version of the 16-point Chen’s fast DCT algorithm [13, 38] is suggested for video
encoding. Chen’s algorithm requires multiplicative constants cos(kπ/32), k = 1, 2, . . . , 15, which can be
approximated by fixed precision quantities [16, Sec. 5]. Indeed, dyadic rational were employed [12], resulting
in a non-orthogonal transform [16, Sec. 5]. The International Telecommunication Union fosters image blocks
of 16×16 pixels [47] instead of the 4×4 and 8×8 pixel blocks required by the H.264/MPEG-4 AVC standard
for video compression [33]. The main reason for such recommendation is the improved coding gains [28]. It is
clear that for such large transform blocklengths, minimizing the computational complexity becomes a central
issue [16].
In this context, the main goal of this paper is to advance 16-point approximate DCT architectures. First,
we introduce a new low-complexity 16-point DCT approximation. The proposed transform is sought to be
orthogonal and to possess null multiplicative complexity. Second, we propose an efficient fast algorithm for
the new transform. Third, we introduce hardware implementations for the proposed transform as well as for
the 16-point DCT approximate method introduced by Bouguezel-Ahmad-Swamy (BAS-2010) in [10]. Both
methods are demonstrated to be suitable for image compression.
The paper unfolds as follows. In Section 2, the new proposed transform is introduced and mathematically
analyzed. Error metrics are considered to assess its proximity to the exact DCT matrix. In Section 3, a fast
algorithm for the proposed transform is derived and its computational complexity is compared with existing
methods. An image compression simulation is described in Section 4, indicating the adequateness of the
introduced transform. In Section 5, FPGA-based hardware implementations for both the proposed transform
and the BAS-2010 approximation are detailed and analyzed. Conclusions and final remarks are given in
Section 6.
2
16-point DCT Approximation
In this section, a new 16-point multiplication-free transform is presented. The proposed matrix transform
T was obtained by judiciously replacing each floating point of the 16-point DCT matrix for 0, 1, or −1.
Substitutions were computationally performed in such a way that: (i) the resulting matrix could satisfy the
2
following orthogonality-like property:
T × T> = diagonal matrix,
(ii) DCT symmetries could be preserved, and (iii) the resulting matrix could offer good energy compaction
properties [20]. Among the several possible outcomes, we isolated the following matrix:
T=
1
1
1
1
1
1
1
1
1
1
1
0
1
1
0
1
1
1
1
1
1
1
0
0
−1
−1
−1
−1
−1
−1
−1
−1
1
1
1
1
−1
−1
−1
−1
−1
−1
0
1
1
1
1
0
1
1
0
0
−1
−1
−1
1
1
1
1
1
−1
−1
−1
−1
1
1
0
−1
−1
−1
1
1
1
−1
−1
−1
−1
0
1
1
1
0
−1
−1
−1
1
1
1
−1
−1
0
1
1
1
−1
−1
1
1
−1
−1
1
1
0
−1
−1
0
1
1
−1
−1
1
1
1
1
−1
−1
1
0
−1
−1
1
1
−1
−1
1
1
0
−1
1
−1
−1
1
1
0
−1
1
1
−1
−1
1
1
−1
0
1
1
−1
−1
1
1
−1
0
1
−1
0
1
−1
−1
1
1
−1
1
0
−1
1
−1
−1
1
−1
−1
1
0
−1
1
−1
−1
1
1
−1
0
1
−1
1
1
−1
1
1
−1
1
−1
0
1
−1
1
−1
0
0
−1
1
−1
−1
1
−1
1
−1
−1
1
−1
1
1
−1
1
−1
−1
1
−1
1
−1
1
0
−1
1
−1
1
0
1
−1
1
−1
1
−1
0
0
−1
1
−1
1
−1
1
−1
1
1
−1
1
−1
1
−1
1
−1
1
−1
1
0
1
−1
0
−1
.
(1)
Above matrix furnishes a DCT approximation given by
Ĉ = D · T,
where
D = diag
1 √1
1
1 √1
1
√1
√1
√
√
4 , 14 , 2 3 , 14 , 4 , 14 , 2 3 , 14 ,
1 √1
1
1 √1
1
√1
√1
√
√
4 , 14 , 2 3 , 14 , 4 , 14 , 2 3 , 14
,
and diag(·) returns the block diagonal concatenation of its arguments.
The proposed transform Ĉ is orthogonal and requires no multiplications or bit shifting operations. Only
additions are required for the computation of the proposed DCT approximation. Moreover, the scaling
matrix D may not introduce any additional computational overhead in the context of image compression. In
fact, the scalar multiplications of D can be merged into the quantization step [8, 9, 11, 15, 29]. Therefore, in
this sense, the approximation Ĉ has the same low computational complexity of T.
Now we aim at comparing the proposed transformation with existing low-complexity approximations for the
16-point DCT. Although there is a reduced number of 16-point transforms with null multiplicative complexity
in signal processing literature, we could separate two orthogonal transformations for comparison: (i) the wellknown Walsh-Hadamard transform (WHT) [41, p. 1087] and (ii) the 16-point BAS-2010 approximation [10].
The WHT is selected for its simplicity of implementation [19, p. 472]. The BAS-2010 method considered
since it is the most recent method for DCT approximation for 16-point long data.
A classical reference in this field is the signed DCT (SDCT) [20], which became a standard for comparison
when considering 8-point DCT approximations. However, for 16-point data, the signed DCT is not orthogonal
and its inverse transformation requires several additions and multiplications [10]. Thus, we could not consider
3
SDCT for any meaningful comparison.
According to the methodology employed in [20] and supported by [15], we can assess how adequate
the proposed approximation is. For such analysis, each row of a 16×16 approximation matrix A can be
interpreted as the coefficients of a FIR filter. Therefore, the following filters are defined:
hm [n] = am,n ,
m = 0, 1, . . . , 15,
where am,n is the (m + 1, n + 1)-th entry of A.
Thus, the transfer functions associated to hm [n], m = 0, 1, . . . , 15, can computed by the discrete-time
Fourier transform defined over ω ∈ [0, π] [34]:
Hm (ω; A) =
15
X
hm [n] exp(−j n ω),
m = 0, 1, . . . , 15,
n=0
where j =
√
−1.
Spectral data Hm (ω; A) can be employed to define a figure of merit for assessing DCT approximations.
Indeed, we can measure the distance between Hm (ω; C) and Hm (ω; A), where C is the exact DCT matrix. We
adopted the squared magnitude as a distance measure function. Thus, we obtained the following mathematical
expression:
2
Dm (ω; A) , Hm (ω; C) − Hm (ω; A) ,
m = 0, 1, . . . , 15,
Note that Dm (ω; A) is an energy-related error measure. For each row m at any angular frequency ω ∈ [0, π]
in radians per sample, above expression quantifies how discrepant a given approximation matrix A is from
the DCT matrix.
Fig. 1 shows the plots for Dm (ω; A), m = 1, 2, . . . , 15, where A is either the WHT, the BAS-2010
approximation, or the proposed transform T. The particular plot for m = 0 was suppressed, since all
considered transforms could match the DCT exactly.
The error energy departing from the actual DCT can be obtained by integrating Dm (ω; A) over ω ∈
[0, π] [15]:
Z
m (A) =
π
Dm (ω; A)dω,
0
Table 1 summarizes the obtained values of m (A), m = 0, 1, . . . , 15. These quantities were computed by
numerical quadrature methods [37].
3
Fast Algorithm
As defined in (1), transformation matrix T requires 208 additions, which a significant number of operations. In
the following, we present a factorization of T obtained by means of butterfly-based methods in a decimationin-frequency structure [7]. For notational purposes, we denote In as the identity matrix of order n, Īn as the
opposite diagonal identity matrix of order n, and the butterfly matrix as
"
Bn ,
In/2
Īn/2
Īn/2
−In/2
4
#
.
8
14
2
4
D3(ω, A)
4
6
D2(ω, A)
6 8 10
D1(ω, A)
6 8 10
4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
0
2
2
0
π 4
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
0
π 4
π 2
ω
3π 4
π
D6(ω, A)
6 8 10
8
D5(ω, A)
4
6
0
D9(ω, A)
4
6
2
0
0
8
D12(ω, A)
4
6
0
0
0
0
0.0
2
2
0.5
D13(ω, A)
1.0 1.5
D14(ω, A)
4
6
D15(ω, A)
4 6 8 10
2.0
8
2.5
0
2
5
D11(ω, A)
5
10
D10(ω, A)
10
15
15
0
1
2
D7(ω, A)
2
3
D8(ω, A)
4
6
4
8
8
5
10
0
0
2
2
4
D4(ω, A)
5
10
0
14
15
0
Figure 1: Plots of Dm (ω; A) for m = 1, 2, . . . , 15 and ω ∈ [0, π], considering the proposed transform (solid
curve), the BAS-2010 transform (dashed curve), and the WHT (dotted curve).
5
Table 1: Error energy m (A) for selected DCT approximatinons.
m
Proposed
BAS-2010
WHT
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
0.00
0.78
0.20
0.78
0.50
0.95
0.22
0.87
0.00
0.79
0.19
0.72
0.46
0.70
0.22
0.70
0.00
0.61
6.34
1.52
6.28
4.31
8.60
2.69
0.00
6.57
6.69
8.47
5.84
1.60
0.75
6.72
0.00
6.78
6.09
5.20
5.70
5.44
6.60
4.81
5.50
6.57
10.62
8.06
7.29
1.60
5.97
6.42
Total
8.08
66.99
92.65
We maintain that T can be decomposed into less complex matrix terms according to the following
factorization:
T = P × diag B2 , B̄2 , E, O × diag(B4 , I12 ) × diag(B8 , I8 ) × B16 ,
where the required matrices are described below:
"
B̄2 = B2 × Ī2 =
1
1
−1
1
#
,
0
1
1
1
−1
E=
1
−1
−1
0
1
,
1
0
0 −1
1 −1
1
1
0
1
1
1
1
1
−1
0
−1
O=
1
−1
1
−1
−1
−1
−1
0
1
1
1 −1
−1
−1
1
1
−1
0
1 −1
−1
1
1
1
,
1
0
1
1
1
−1
1
1
0 −1
−1
1
1 −1
1
0 −1
1 −1
1 −1
1 −1
0 −1
1
−1
1 −1
6
(2)
and matrix P is a permutation matrix given by
h
P=
e1
e2
e9
e4
e5
e6
e13
e8
e3
e10
e7
e11
e15
e12
e14
e16
i
,
where ej is a 16-point column vector with one in position j and zero elsewhere.
Matrix E corresponds to the even-odd part, whereas matrix O is linked to the odd part of the proposed
transformation [6, p. 71]. A row permuted version of matrix E was already reported in literature in the
derivation of the 8-point DCT approximation described in [15, Fig. 1].
On the other hand, matrix O does not seem to be reported. Without any further consideration, matrix O
requires 48 additions. The locations of zero elements in (2) is such that a decimation-in-frequency operation
by means of a butterfly structure is prevented. In order to obtain the required symmetry, we propose the
following manipulation:
O0 = O − S,
(3)
where
0
0
0
1
0
0
0
0
S=
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
−1
0
0
1
0
0
0
0
0
0
0
0
−1
0
0
0
0
0
1
.
0
0
0
0
The resulting matrix O0 can factorized according to:
1
−1
0
−1
0
O =
0
0
0
0
0
0
1
0
1
0 −1
0
0
0
0
1
0
0
0
× (I4 ⊗ B2 ),
0 −1
0
0
0 −1
0 −1
0
0
1
−1
0
0
1
0
0
1
0 −1
0
0
1
0
1
1
0
0
0
0
−1
0 −1
0
0
−1
1
−1
0
0
1
(4)
where ⊗ denotes the Kronecker product. The additive complexity of matrix O0 is 20 additions.
Above mathematical description can be given a flow diagram, which is useful for subsequent hardware
implementation. Fig. 2(a) depicts the general structure of proposed fast algorithm. Block A and Block B
represent the operations associated to matrix E and O, respectively. The structure of Block A is disclosed in
Fig. 2(b). Fig. 3(a) details the inner structure of Block B as described in (3). Fig. 3(b) exhibits Block C
according to (4).
7
x0
X0
x1
X8
x2
X4
x3
X12
x4
X2
x5
X6
Block A
x6
X10
x7
X14
x8
X1
x9
X3
x10
X5
x11
X7
Block B
x12
X9
x13
X11
x14
X13
x15
X15
(a) Full diagram
(b) Block A
Figure 2: Flow diagram of the fast algorithm for the proposed transform.
The proposed algorithm requires only 72 additions. Bit shifting and multiplication operations are absent.
Arithmetic complexity comparisons with selected 16-point transforms are summarized in Table 2.
4
Application to Image Compression
This section presents the application of the proposed transform to image compression. We produce evidence
that it outperforms the other transforms in consideration. For this analysis, we used the methodology
described in [20], supported in [8–11], and extended in [15].
A set of 45 512×512 8-bit greyscale images obtained from a standard public image bank [48] was considered.
We adapted the JPEG compression technique [36] for the 16×16 matrix case. Each image was divided into
Table 2: Arithmetic complexity analysis.
Operation
Proposed BAS-2010 [10] WHT [18]
Addition
Bit shifting
Multiplication
72
0
0
64
8
0
64
0
0
Total
72
72
64
8
Block C
(a) Block B
(b) Block C
Figure 3: Flow diagram of the inner structure of Block B.
16×16 sub-blocks, which were submitted to the two-dimensional (2-D) transform procedure associated to the
DCT matrix, the BAS-2010 [10] matrix, the WHT [18] matrix, and the proposed matrix Ĉ. A 16×16 image
block K has its 2-D transform mathematically expressed by [45]:
A · K · A> ,
where A is a considered transformation.
This computation furnished 256 approximate transform domain coefficients for each sub-block. A hard
thresholding step was applied, where only the r initial coefficients were retained, being the remaining ones set
to zero. Coefficients were ordered according to the usual zig-zag scheme extended to 16×16 image blocks [35].
We adopted r ∈ {2, 4, . . . , 254, 256}. The inverse procedure was then applied to reconstruct the processed
data and image quality was assessed.
Image degradation was evaluated using three different quality measures: (i) the peak signal-to-noise
ratio (PSNR), (ii) the mean square error (MSE), and (iii) the universal quality index (UQI) [49]. The
PSNR and MSE were selected due to their wide application as figures of merit in image processing. The
UQI is considered an improvement over PSNR and MSE as a tool for image quality assessment [49]. The
UQI includes luminance, contrast, and structure characteristic in its definition. Another possible metric is
the structural-similarity-based image quality assessment (SSIM) [50]. Being a variation of the UQI, SSIM
results were not very different from the measurements offered by the UQI for the considered images. Indeed,
whenever a difference was present, it was in the order of 10−4 only. Therefore, SSIM results are not presented
here.
Moreover, in contrast with the JPEG image compression simulations described in [8–11], we considered
the average measures from all images instead of the results derived from selected images. In fact, average
calculations may furnish more robust results [26].
Fig. 4 shows the resulting quality measures. The proposed transform could outperform both the BAS-2010
transform and the WHT in all compression rates according to all considered quality measures. Fig. 4(b)
shows that the proposed transform outperformed the BAS-2010 transform in ≈ 1 dB and the WHT in ≈ 8 dB,
which corresponds to ≈ 26% and ≈ 630% gains, respectively. At the same time, Fig. 4(b) shows that the
9
results of the proposed transform are at most 2 dB way from when DCT results at compression ratios superior
to 85% (r < 40).
In order to convey a qualitative analysis, Figures 5 and 6 show two standard images compressed according
to the considered transforms. The associate differences with respect to the original uncompressed images are
also displayed. For better visualization, difference images were scaled by a factor of two. This procedure is
routine and described in further detail in [40, p. 273]. The images compressed with the proposed transform
are visually more similar to the images compressed with DCT than the others. As expected, the WHT
exhibits a poor performance.
5
FPGA-based Hardware Implementation
In this section, the proposed DCT approximation and the BAS-2010 algorithm [10] were physically implemented
on a field programmable gate array (FPGA) device. We employed the 40 nm CMOS Xilinx Virtex-6
XC6VLX240T FPGA for algorithm evaluation and comparison. Beforehand it is expected that the proposed
algorithm exhibit modestly higher hardware demands. This is due to the fact that it requires 72 additions,
whereas the BAS-2010 algorithms demands 64 additions.
We furnished circuit performance using metrics of (i) area (A) based on the quantity of required elementary
programmable logic blocks (slices), the number of look-up tables (LUTs), and the flip-flops count, (ii) the
speed, using the critical path delay (T ), and (iii) the dynamic power consumption.
The number of occupied slices furnished an estimate of the on-chip silicon real estate requirement, whereas
number of LUTs and flip-flops are the main logic resources available in a slice. In Xilinx FPGAs, a LUT
is employed as a combinational function generator that can implement a given boolean function and a
flip-flop is utilized as a 1-bit register. The critical path delay corresponds to the delay associated with the
longest combinational path and directly governs the operating frequency of the hardware. The total power
consumption of the hardware design constitutes of static and dynamic components. Static power consumption
in FPGAs is dominated by the leakage power of the logic fabric and the configuration static RAM. Thus, is
mostly design independent. On the other hand, the dynamic power consumption, which accounts for the
dynamic power dissipation associated with clocks, logic blocks, and signals, provides a metric for the power
efficiency of a given design [43]. The respective results are shown in Tables 3, 4, and 5, where the metrics
corresponding to each design were measured for several choices of finite precision using input word length
W ∈ {4, 8, 12, 16}.
From Table 3, it is observed that the proposed design consumes ≈ 10% more LUTs hardware resources
than [10]. For W = 8 the proposed design shows a ≈ 20% increase in the number of slices consumed (and 10%
more LUTs) while gaining 1-2 dB of improvement in PSNR compared to [10]. The increase in area shown
by the proposed design has led to an increase in the critical path delay, area-time (AT ), area-time-squared
(AT 2 ) metrics and to a higher power consumption as indicated in Tables 4 and 5. Of particular interest is
the case for W = 8 input word size, where the proposed algorithm and hardware design shows a 5.7% and
10% increase in the critical path delay and dynamic power consumption, respectively, when compared to the
algorithm in [10].
In this paper, we define a metric consisting of the product of error figures and the AT value:
(error figure) × (area-time product).
10
10
60
PSNR difference (dB)
2
4
6
8
Proposed Transform
BAS−2010
WHT
20
0
Average PSNR (dB)
30
40
50
DCT
Proposed Transform
BAS−2010
WHT
0
50
100
150
200
250
0
50
100
r
200
250
(b) PSNR difference relative to DCT (dB)
Proposed Transform
BAS−2010
WHT
0
0
MSE difference
100
200
300
DCT
Proposed Transform
BAS−2010
WHT
400
(a) Average PSNR
Average MSE
200
400
600
150
r
0
50
100
150
200
250
0
50
100
r
150
100
150
200
0.1
UQI difference
0.2 0.3 0.4
0.5
Proposed Transform
BAS−2010
WHT
0.0
Average UQI
0.4
0.6
0.8
0.0
0.2
DCT
Proposed Transform
BAS−2010
WHT
50
250
(d) MSE difference relative to DCT
1.0
(c) Average MSE
0
200
r
250
0
r
50
100
150
200
r
(e) Average UQI
(f) UQI difference relative to DCT
Figure 4: Quality measures for several compression ratios.
11
250
(a) DCT
(b) proposed
(c) BAS-2010
(d) WHT
(e) DCT
(f) proposed
(g) BAS-2010
(h) WHT
Figure 5: Compressed images (a–d) and difference images (e–h) using the DCT, the proposed transform, the
BAS-2010 approximation, and the WHT for the Lena image, considering r = 100.
The considered error figure can be the 1/PSNR, MSE, 1/UQI, or the total error energy, as given in
Table 1. This metric aims at combining both the mathematical and the hardware aspects of the resulting
implementation. The total error energy has the advantage of being image independent, being adopted in
the combined metric. Considering the proposed architecture and [10], the obtained values for the combined
metric are shown in Table 6.
Although the proposed DCT approximation consumes more resources than [10], a much better approximation for the exact DCT is achieved (see Table 1). This leads to superior compressed image quality (see
Fig. 4). Indeed, the choice of algorithm is always a compromise between its mathematical properties, such as
DCT proximity, energy error, and resulting image quality; and the related hardware aspects, such as area,
speed, and power consumption. This implies our proposed algorithm is a better choice over [10] when picture
quality is of higher importance.
6
Conclusion
This paper introduced a new 16-point DCT approximation. The proposed transform requires no multiplication
or bit shifting operations, is orthogonal, and its matrix elements are only {−1, 0, 1}. Using spectral analysis
methods described in [15, 20], we demonstrated that the proposed transform outperforms the WHT and the
BAS-2010 as an approximation for the 16-point DCT. The proposed transform was considered into standard
image compression methods. The resulting images were assessed for quality by means of PSNR, MSE, and
UQI. According to these metrics, the proposed transform could outperform the WHT and the BAS-2010
12
(a) DCT
(b) proposed
(c) BAS-2010
(d) WHT
(e) DCT
(f) proposed
(g) BAS-2010
(h) WHT
Figure 6: Compressed images (a–d) and difference images (e–h) using the DCT, the proposed transform, the
BAS-2010 approximation, and the WHT for the Airplane (F-16) image, considering r = 40.
Table 3: Area utilization for FPGA implementation.
Input
Area
word
Input word
BAS-2010 [10]
Proposed
length
Registers
LUTs
Slices
Registers
LUTs
Slices
4
403
543
172
597
524
178
8
828
704
241
956
909
290
12
1128
958
317
1253
1316
384
16
1432
1243
390
1597
1676
491
Table 4: Speed, AT , and AT 2 metrics for FPGA implementation.
Speed (MHz)
AT (Slices · µs)
AT 2 (Slices · µs2 · 10−3 )
length
BAS-2010 [10]
Proposed
BAS-2010 [10]
Proposed
BAS-2010 [10]
Proposed
4
369.13
342.9
0.466
0.519
1.26
1.51
8
361.92
342.114
0.666
0.848
1.84
2.48
12
363.63
336.813
0.872
1.140
2.40
3.38
16
341.29
338.18
1.143
1.452
3.35
4.29
13
Table 5: Dynamic power consumption for FPGA implementation.
Input
Dynamic Power (mW)
word
BAS-2010 [10]
Proposed
length
Clocks
Logic
Signals
Total
Clocks
Logic
Signals
Total
4
0.033
0.022
0.030
0.085
0.040
0.020
0.029
0.089
8
0.041
0.016
0.034
0.091
0.041
0.023
0.037
0.101
12
0.059
0.022
0.050
0.131
0.054
0.033
0.054
0.141
16
0.069
0.028
0.070
0.167
0.065
0.042
0.077
0.184
Table 6: Comparison of the cost associated with each design
Input
Combined metric
word length
BAS-2010 [10]
Proposed
4
31.22
4.19
8
44.62
6.85
12
58.42
9.21
16
76.57
11.73
approximation at any compression ratio. We also derived an efficient fast algorithm for the proposed matrix,
which required 72 additions.
This algorithm was implemented in hardware and compared with a state-of-the-art 16-point DCT
approximation [10]. FPGA-based rapid prototypes were designed, simulated, physically implemented, and
tested for 4-, 8-, 12-, and 16-bit input data word sizes. A typical application having 8-bit input image data
could be subject to 16-point DCT approximations at a real-time rate of 342 · 106 transforms per second, for
FPGA clock frequency of 342 MHz, leading to a pixel rate of 5.488 · 109 pixels/second. Both proposed and
BAS-2010 algorithms were realized on FPGA and tested and hardware metrics including area, power, critical
path delay, and area-time complexity. Additionally, an extensive investigation of relative performance in both
subjective mode as well as objective picture quality metrics using average PSNR, average MSE, and average
UQI was produced. The proposed DCT approximation algorithm improves on the state-of-art algorithm
in [10] by 1-2 dB for PSNR at the cost of only eight extra adders.
Video coding using motion partitions larger than 8×8 pixels is investigated in [16, 47] with satisfactory
application in H.264/AVC standard for video compression. In this perspective, the new proposed approximation
transform is a candidate technique to image and video coding with block size equal to 16×16. This blocklength
is of particular importance in the emerging H.265 reconfigurable video codec standard [25].
Acknowledgments
This work was partially supported by CNPq and FACEPE (Brazil); and by the College of Engineering at the
University of Akron, Akron, OH, USA.
14
References
[1] N. Ahmed, T. Natarajan, and K. R. Rao. Discrete cosine transform. IEEE Transactions on Computers,
C-23(1):90–93, Jan. 1974.
[2] Y. Arai, T. Agui, and M. Nakajima. A fast DCT-SQ scheme for images. Transactions of the IEICE, E-71(11):1095–
1097, 1988.
[3] H. L. P. Arjuna Madanayake, R. J. Cintra, D. Onen, V. S. Dimitrov, and L. T. Bruton. Algebraic integer based
8×8 2-D DCT architecture for digital video processing. In Proceedings of the IEEE International Symposium on
Circuits and Systems (ISCAS), pages 1247–1250, May 2011.
[4] F. M. Bayer and R. J. Cintra. Image compression via a fast DCT approximation. IEEE Latin America
Transactions, 8(6):708–713, Dec. 2010.
[5] V. Bhaskaran and K. Konstantinides. Image and Video Compression Standards. Kluwer Academic Publishers,
Boston, 1997.
[6] G. Bi and Y. Zeng. Transforms and Fast Algorithms for Signal Analysis and Representations. Birkhäuser, 2004.
[7] R. E. Blahut. Fast Algorithms for Digital Signal Processing. Addison-Wesley, 1985.
[8] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. Low-complexity 8×8 transform for image compression.
Electronics Letters, 44(21):1249–1250, Sept. 2008.
[9] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A fast 8×8 transform for image compression. In Proceedings
of the 2009 Internation Conference on Microelectronics, 2009.
[10] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A novel transform for image compression. In Proceedings of
the 53rd IEEE International Midwest Symposium on Circuits and Systems (MWSCAS), 2010.
[11] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy. A low-complexity parametric transform for image compression.
In Proceedings of the 2011 IEEE International Symposium on Circuits and Systems, 2011.
[12] V. Britanak, P. Yip, and K. R. Rao. Discrete Cosine and Sine Transforms. Academic Press, 2007.
[13] W.-H. Chen, C. H. Smith, and S. C. Fralick. A fast computational algorithm for the discrete cosine transform.
IEEE Transactions on Communications, COM-25(9):1004–1009, Sept. 1977.
[14] N. I. Cho and S. U. Lee. Fast algorithm and implementation of 2-D discrete cosine transform. IEEE Transactions
on Circuits and Systems, 38(3):297–305, Mar. 1991.
[15] R. J. Cintra and F. M. Bayer. A DCT approximation for image compression. IEEE Signal Processing Letters,
18(10):579–582, Oct. 2011.
[16] T. Davies, K. R. Andersson, R. Sjöberg, T. Wiegand, D. Marpe, K. Ugur, J. Ridge, M. Karczewicz, P. Chen,
G. Martin-Cocher, K. McCann, W.-J. Han, G. Bjontegaard, and A. Fuldseth. Joint collaborative team on
video coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC JTC1/SC29/WG11: Suggestion for a test model.
JCTVC-A033, International Telecommunication Union, Dresden, DE, Apr. 2010.
[17] V. S. Dimitrov, K. Wahid, and G. A. Jullien. Multiplication-free 8 × 8 DCT architecture using algebraic integer
encoding. Electronics Letters, 40(20):1310–1311, 2004.
[18] B. J. Fino. Relations between Haar and Walsh/Hadamard transforms. Proceedings of the IEEE, 60(5):647–648,
May 1972.
[19] R. C. Gonzalez and R. E. Woods. Digital Image Processing. Prentice Hall, Upper Saddle River, NJ, 2002.
[20] T. I. Haweel. A new square wave transform based on the DCT. Signal Processing, 82:2309–2319, 2001.
15
[21] H. S. Hou. A fast recursive algorithm for computing the discrete cosine transform. IEEE Transactions on
Acoustic, Signal, and Speech Processing, 6(10):1455–1461, 1987.
[22] International Organisation for Standardisation. Generic coding of moving pictures and associated audio information
– Part 2: Video. ISO/IEC JTC1/SC29/WG11 - coding of moving pictures and audio, ISO, 1994.
[23] International Telecommunication Union. ITU-T recommendation H.261 version 1: Video codec for audiovisual
services at p × 64 kbits. Technical report, ITU-T, 1990.
[24] International Telecommunication Union. ITU-T recommendation H.263 version 1: Video coding for low bit rate
communication. Technical report, ITU-T, 1995.
[25] H. Kalva. The H.264 Video Coding Standard. IEEE Multimedia, 13(4):86–90, Oct. 2006.
[26] S. M. Kay. Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory, volume 1 of Prentice
Hall Signal Processing Series. Prentice Hall, Upper Saddle River, NJ, 1993.
[27] W.-K. Kuo and K.-W. Wu. Traffic prediction and QoS transmission of real-time live VBR videos in WLANs.
ACM Transactions on Multimedia Computing, Communications and Applications, 7(4):36:1–36:21, Dec. 2011.
[28] K. H. Lee, E. A. J. H. Park, W. J. Han, and J. H. Min. Technical considerations for ad hoc group on new
challenges in video coding standardization. MPEG Doc. M15580, Hannover, Germany, July 2008.
[29] K. Lengwehasatit and A. Ortega. Scalable variable complexity approximate forward DCT. IEEE Transactions
on Circuits and Systems for Video Technology, 14(11):1236–1248, Nov. 2004.
[30] J. Liang and T. D. Tran. Fast multiplierless approximations of the DCT with the lifting scheme. IEEE
Transactions on Signal Processing, 49(12):3032–3044, Dec. 2001.
[31] C. Loeffler, A. Ligtenberg, and G. Moschytz. Practical fast 1D DCT algorithms with 11 multiplications. In
Proceedings of the International Conference on Acoustics, Speech, and Signal Processing, pages 988–991, 1989.
[32] A. Luthra, G. J. Sullivan, and T. Wiegand. Introduction to the special issue on the H.264/AVC video coding
standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):557–559, July 2003.
[33] H. S. Malvar, A. Hallapuro, M. Karczewicz, and L. Kerofsky. Low-complexity transform and quantization in
H.264/AVC. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):598–603, July 2003.
[34] A. V. Oppenheim and R. W. Schafer. Discrete-Time Signal Processing. Prentice Hall, 3 edition, 2009.
[35] I.-M. Pao and M.-T. Sun. Approximation of calculations for forward discrete cosine transform. IEEE Transactions
on Circuits and Systems for Video Technology, 8(3):264–268, June 1998.
[36] W. B. Pennebaker and J. L. Mitchell. JPEG Still Image Data Compression Standard. Van Nostrand Reinhold,
New York, NY, 1992.
[37] R. Piessens, E. deDoncker-Kapenga, C. Uberhuber, and D. Kahaner. Quadpack: a Subroutine Package for
Automatic Integration. Springer-Verlag, 1983.
[38] K. R. Rao and P. Yip. Discrete Cosine Transform: Algorithms, Advantages, Applications. Academic Press, San
Diego, CA, 1990.
[39] N. Roma and L. Sousa. Efficient hybrid DCT-domain algorithm for video spatial downscaling. EURASIP Journal
on Advances in Signal Processing, 2007(2):30–30, 2007.
[40] D. Salomon. Data Compression. Springer, 3 edition, 2004.
[41] D. Salomon. The Computer Graphics Manual, volume 1. Springer-Verlag, London, UK, 2011.
[42] S. Saponara. Real-time and low-power processing of 3D direct/inverse discrete cosine transform for low-complexity
video codec. Journal of Real-Time Image Processing, 7:43–53, 2012. 10.1007/s11554-010-0174-5.
16
[43] M. Shafique and J. Henkel. Background and related work. In Hardware/Software Architectures for Low-Power
Embedded Multimedia Systems. Springer New York, 2011.
[44] N. Suehiro and M. Hateri. Fast algorithms for the DFT and other sinusoidal transforms. IEEE Transactions on
Acoustic, Signal, and Speech Processing, 34(6):642–644, 1986.
[45] T. Suzuki and M. Ikehara. Integer DCT based on direct-lifting of DCT-IDCT for lossless-to-lossy image coding.
IEEE Transactions on Image Processing, 19(11):2958–2965, Nov. 2010.
[46] J. V. Team. Recommendation H.264 and ISO/IEC 14 496–10 AVC: Draft ITU-T recommendation and final draft
international standard of joint video specification. Technical report, ITU-T, 2003.
[47] Telecommunication Standardization Sector. Video coding using extended block sizes. COM 16-C 123-E,
International Telecommunication Union, Jan. 2009.
[48] The USC-SIPI image database. http://sipi.usc.edu/database, 2011. University of Southern California, Signal and
Image Processing Institute.
[49] Z. Wang and A. Bovik. A universal image quality index. IEEE Signal Processing Letters, 9(3):81–84, 2002.
[50] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to
structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, Apr. 2004.
[51] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the H.264/AVC video coding standard.
IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560–576, July 2003.
17
| 8cs.DS
|
Monomial ideals of weighted oriented graphs
Yuriko Pitones∗, Enrique Reyes†, and Jonathan Toledo‡
arXiv:1710.03785v1 [math.AC] 10 Oct 2017
Departamento de Matemáticas
Centro de Investigación y de Estudios Avanzados del Instituto Politécnico Nacional
Apartado Postal 14–740, Ciudad de México
07000 México
Abstract
Let I = I(D) be the edge ideal of a weighted oriented graph D. We determine the
irredundant irreducible decomposition of I. Also, we characterize the associated
primes and the unmixed property of I. Furthermore, we give a combinatorial characterization for the unmixed property of I, when D is bipartite, D is a whisker or
D is a cycle. Finally, we study the Cohen-Macaulay property of I.
Keywords: Weighted oriented graphs, unmixed property, irreducible decomposition, CohenMacaulay property.
1
Introduction
A weighted oriented graph is a triplet D = (V (D), E(D), w), where V (D) is a finite set,
E(D) ⊆ V (D) × V (D) and w is a function w : V (D) → N. The vertex set of D and the
edge set of D are V (D) and E(D), respectively. Some times for short we denote these sets
by V and E respectively. The weight of x ∈ V is w(x). If e = (x, y) ∈ E, then x is the
tail of e and y is the head of e. The underlying graph of D is the simple graph G whose
vertex set is V and whose edge set is {{x, y}|(x, y) ∈ E}. If V (D) = {x1 , . . . , xn }, then we
consider the polynomial ring R = K[x1 , . . . , xn ] in n variables over a field K. In this paper
w(x )
we introduce and study the edge ideal of D given by I(D) = (xi xj j : (xi , xj ) ∈ E(D))
in R, (see Definition 3.1).
In Section 2 we study the vertex covers of D. In particular we introduce the notion
of strong vertex cover (Definition 2.6) and we prove that a minimal vertex cover is
strong. In Section 3 we characterize the irredundant irreducible decomposition of I(D).
In particular we show that the minimal monomial irreducible ideals of I(D) are associated
with the strong vertex covers of D. In Section 4 we give the following characterization
of the unmixed property of I(D).
I(D) is unmixed
G is unmixed and D has the minimal-strong property
All strong vertex
covers have the
same cardinality
All minimal vertex
covers have the
same cardinality
∗ [email protected]
† [email protected]
‡ [email protected]
All strong vertex
covers are minimals
Furthermore, if D is bipartite, D is a whisker or D is a cycle, we give an effective
(combinatorial) characterization of the unmixed property. Finally in Section 5 we study
the Cohen-Macaulayness of I(D). In particular we characterize the Cohen-Macaulayness
when D is a path or D is complete. Also, we give an example where this property depend
of the characteristic of the field K.
2
Weighted oriented graphs and their vertex covers
In this section we define the weighted oriented graphs and we study their vertex covers.
Furthermore, we define the strong vertex covers and we characterize when V (D) is a
strong vertex cover of D. In this paper we denote the set {x ∈ V | w(x) 6= 1} by V + .
Definition 2.1. A vertex cover C of D is a subset of V , such that if (x, y) ∈ E, then
x ∈ C or y ∈ C. A vertex cover C of D is minimal if each proper subset of C is not a
vertex cover of D.
+
(x) =
Definition 2.2. Let x be a vertex of a weighted oriented graph D, the sets ND
−
{y : (x, y) ∈ E(D)} and ND (x) = {y : (y, x) ∈ E(D)} are called the out-neighbourhood
and the in-neighbourhood of x, respectively. Furthermore, the neighbourhood of x is the
+
−
set ND (x) = ND
(x) ∪ ND
(x).
Definition 2.3. Let C be a vertex cover of a weighted oriented graph D, we define
+
L1 (C) = {x ∈ C | ND
(x) ∩ C c 6= ∅},
−
(x) ∩ C c 6= ∅} and
L2 (C) = {x ∈ C | x ∈
/ L1 (C) and ND
L3 (C) = C \ (L1 (C) ∪ L2 (C)),
where C c is the complement of C, i.e. C c = V \ C.
Proposition 2.4. If C is a vertex cover of D, then
L3 (C) = {x ∈ C | ND (x) ⊂ C}.
+
−
(x) ⊆ C, since x ∈
/ L1 (C). Furthermore ND
(x) ⊆ C, since
Proof. If x ∈ L3 (C), then ND
x∈
/ L2 (C). Hence ND (x) ⊂ C, since x ∈
/ ND (x). Now, if x ∈ C and ND (x) ⊂ C, then
x∈
/ L1 (C) ∪ L2 (C). Therefore x ∈ L3 (C).
Proposition 2.5. If C is a vertex cover of D, then L3 (C) = ∅ if and only if C is a
minimal vertex cover of D.
Proof. ⇒) If x ∈ C, then by Proposition 2.4 we have ND (x) 6⊂ C, since L3 (C) = ∅.
Thus, there is y ∈ ND (x) \ C implying C \ {x} is not a vertex cover. Therefore, C is a
minimal vertex cover.
⇐) If x ∈ L3 (C), then by Proposition 2.4, ND (x) ⊆ C \ {x}. Hence, C \ {x} is a vertex
cover. A contradiction, since C is minimal. Therefore L3 (C) = ∅.
Definition 2.6. A vertex cover C of D is strong if for each x ∈ L3 (C) there is (y, x) ∈
E(D) such that y ∈ L2 (C) ∪ L3 (C) with y ∈ V + (i.e. w(y) 6= 1).
2
Remark 2.7. Let C be a vertex cover of D. Hence, by Proposition 2.4 and since
C = L1 (C) ∪ L2 (C) ∪ L3 (C), we have that C is strong if and only if for each x ∈ C such
that N (x) ⊂ C, there exist y ∈ N − (x) ∩ (C \ L1 (C)) with y ∈ V + .
Corollary 2.8. If C is a minimal vertex cover of D, then C is strong.
Proof. By Proposition 2.5, we have L3 (C) = ∅, since C. Hence, C is strong.
Remark 2.9. The vertex set V of D is a vertex cover. Also, if z ∈ V , then ND (z) ⊆
V \ z. Hence, by Proposition 2.4, L3 (V ) = V . Consequently, L1 (V ) = L2 (V ) = ∅. By
Proposition 2.5, V is not a minimal vertex cover of D. Furthermore since L3 (V ) = V ,
−
V is a strong vertex cover if and only if ND
(x) ∩ V + 6= ∅ for each x ∈ V .
Definition 2.10. If G is a cycle with E(D) = {(x1 , x2 ), . . . , (xn−1 , xn ), (xn , x1 )} and
V (D) = {x1 , . . . , xn }, then D is called oriented cycle.
Definition 2.11. D is called unicycle oriented graph if it satisfies the following conditions:
1) The underlying graph of D is connected and it has exactly one cycle C.
2) C is an oriented cycle in D. Furthermore for each y ∈ V (D) \ V (C), there is an
oriented path from C to y in D.
3) w(x) 6= 1 if degG (x) ≥ 1.
Lemma 2.12. If V (D) is a strong vertex cover of D and D1 is a maximal unicycle
oriented subgraph of D, then V (D′ ) is a strong vertex cover of D′ = D \ V (D1 ).
−
Proof. We take x ∈ V (D′ ). Thus, by Remark 2.9, there is y ∈ ND
(x) ∩ V + (D). If
y ∈ D1 , then we take D2 = D1 ∪ {(y, x)}. Hence, if C is the oriented cycle of D1 , then
C is the unique cycle of D2 , since degD2 (x) = 1. If y ∈ C, then (y, x) is an oriented
path from C to x. Now, if y ∈
/ C, then there is an oriented path L form C to y in D1 .
Consequently, L ∪ {(y, x)} is an oriented path form C to x. Furthermore, degD2 (x) = 1
and w(y) 6= 1, then D2 is an unicycle oriented graph. A contradiction, since D1 is
−
+
maximal. This implies y ∈ V (D′ ), so y ∈ ND
(D′ ). Therefore, by Remark 2.9,
′ (x) ∩ V
′
′
V (D ) is a strong vertex cover of D .
Lemma 2.13. If V (D) is a strong vertex cover of D, then there is an unicycle oriented
subgraph of D.
Proof. Let y1 be a vertex of D. Since V = V (D) is a strong vertex cover, there is y2 ∈ V
such that y2 ∈ N − (y1 ) ∩ V + . Similarly, there is y3 ∈ N − (y2 ) ∩ V + . Consequently,
(y3 , y2 , y1 ) is an oriented path. Continuing this process, we can assume there exist
y2 , y3 , . . . , yk ∈ V + where (yk , yk−1 , . . . , y2 , y1 ) is an oriented path and there is 1 ≤ j ≤
k − 2 such that (yj , yk ) ∈ E(D), since V is finite. Hence, C = (yk , yk−1 , . . . , yj , yk ) is
an oriented cycle and L = (yj , . . . , y1 ) is an oriented path form C to y1 . Furthermore, if
j = 1, then w(y1 ) 6= 1. Therefore, D1 = C ∪ L is an unicycle oriented subgraph of D.
Proposition 2.14. Let D = (V, E, w) be a weighted oriented graph, hence V is a strong
vertex cover of D if and only if there are D1 , . . . , Ds unicycle oriented subgraphs of D
such that V (D1 ), . . . , V (Ds ) is a partition of V = V (D).
3
Proof. ⇒) By Lemma 2.13, there is a maximal unicycle oriented subgraph D1 of D.
Hence, by Lemma 2.12, V (D′ ) is a strong vertex cover of D′ = D \ V (D1 ). So, by
Lemma 2.13, there is D2 a maximal unicycle oriented subgraph of D′ . Continuing this
process we obtain unicycle oriented subgraphs D1 , . . . , Ds such that V (D1 ), . . . , V (Ds )
is a partition of V (D).
⇐) We take x ∈ V (D). By hypothesis there is 1 ≤ j ≤ s such that x ∈ V (Dj ). We
assume C is the oriented cycle of Dj . If x ∈ V (C), then there is y ∈ V (C) such that
(y, x) ∈ E(Dj ) and w(y) 6= 1, since degDj (y) ≥ 2 and Dj is a unicycle oriented subgraph.
Now, we assume x ∈
/ V (C), then there is an oriented path L = (z1 , . . . , zr ) such that
z1 ∈ V (C) and zr = x. Thus, (zr−1 , x) ∈ E(D). Furthermore, w(zr−1 ) 6= 1, since
degDj (zr−1 ) ≥ 2. Therefore V is a strong vertex cover.
3
Edge ideals and their primary decomposition
As is usual if I is a monomial ideal of a polynomial ring R, we denote by G(I) the minimal
monomial set of generators of I. Furthermore, there exists a unique decomposition,
T
I = q1 ∩ · · · ∩ qr , where q1 , . . . , qr are irreducible monomial ideals such that I 6= i6=j qi
for each j = 1, . . . , r. This is called the irredundant irreducible decomposition of I.
Furthermore, qi is an irreducible monomial ideal if and only if qi = (xai11 , . . . , xaiss ) for
some variables xij . Irreducible ideals are primary, then a irreducible decomposition is a
primary decomposition. For more details of primary decomposition of monomial ideals
see [6, Chapter 6]. In this section, we define the edge ideal I(D) of a weighted oriented
graph D and we characterize its irredundant irreducible decomposition. In particular we
prove that this decomposition is an irreducible primary decomposition, i.e, the radicals
of the elements of the irredundant irreducible decomposition of I(D) are different.
Definition 3.1. Let D = (V, E, w) be a weighted oriented graph with V = {x1 , . . . , xn }.
The edge ideal of D, denote by I(D), is the ideal of R = K[x1 , . . . , xn ] generated by
w(x )
{xi xj j | (xi , xj ) ∈ E}.
+
Definition 3.2. A source of D is a vertex x, such that ND (x) = ND
(x). A sink of D is
−
a vertex y such that ND (y) = ND (y).
Remark 3.3. Let D = (V, E, w) be a weighted oriented graph. We take D′ = (V, E, w′ )
a weighted oriented graph such that w′ (x) = w(x) if x is not a source and w′ (x) = 1 if x
is a source. Hence, I(D) = I(D′ ). For this reason in this paper, we will always assume
that if x is a source, then w(xi ) = 1.
Definition 3.4. Let C be a vertex cover of D, the irreducible ideal associated to C is
the ideal
w(x )
IC = L1 (C) ∪ {xj j | xj ∈ L2 (C) ∪ L3 (C)} .
Lemma 3.5. I(D) ⊆ IC for each vertex cover C of D.
Proof. We take I = I(D) and m ∈ G(I), then m = xy w(y) , where (x, y) ∈ D. Since
C is a vertex cover, x ∈ C or y ∈ C. If y ∈ C, then y ∈ IC or y w(y) ∈ IC . Thus,
+
m = xy w(y) ∈ IC . Now, we assume y ∈
/ C, then x ∈ C. Hence, y ∈ ND
(x) ∩ C c , so
w(y)
x ∈ L1 (C). Consequently, x ∈ IC implying m = xy
∈ IC . Therefore I ⊆ IC .
4
Definition 3.6. Let I be a monomial ideal. An irreducible monomial ideal q that
contains I is called a minimal irreducible monomial ideal of I if for any irreducible
monomial ideal p such that I ⊆ p ⊆ q one has that p = q.
Lemma 3.7. Let D be a weighted oriented graph.
{xi1 , . . . , xis } is a vertex cover of D.
If I(D) ⊆ (xai11 , . . . , xaiss ), then
Proof. We take J = (xai11 , . . . , xaiss ). If (a, b) ∈ E(D), then abw(b) ∈ I(D) ⊆ J. Thus,
a
6 ∅. Therefore
xijj |abw(b) for some 1 ≤ j ≤ s. Hence, xij ∈ {a, b} and {a, b}∩{xi1 . . . xis } =
{xi1 , . . . , xis } is a vertex cover of D.
Lemma 3.8. Let J be a minimal irreducible monomial ideal of I(D) where G(J) =
{xai1s , . . . , xaiss }. If aj =
6 1 for some 1 ≤ j ≤ s, then there is (x, xij ) ∈ E(D) where
x∈
/ G(J).
Proof. By contradiction suppose there is aj 6= 1 such that if (x, xij ) ∈ E(D), then
a
x ∈ M = {xai11 , . . . , xaiss }. We take the ideal J ′ = (M \ {xijj }). If (a, b) ∈ E(D), then
abw(b) ∈ I(D) ⊆ J. Consequently, xaikk |abw(b) for some 1 ≤ k ≤ s. If k 6= j, then
a
abw(b) ∈ J ′ . Now, if k = j, then by hypothesis aj 6= 1. Hence, xijj |bw(b) implying
a
xij = b. Thus, (a, xij ) ∈ E(D), so by hypothesis a ∈ M \ {xijj }. This implies abw(b) ∈ J ′ .
Therefore I(D) ⊆ J ′ ( J. A contradiction, since J is minimal.
Lemma 3.9. Let J be a minimal irreducible monomial ideal of I(D) where G(J) =
{xai11 , . . . , xaiss }. If aj 6= 1 for some 1 ≤ j ≤ s, then aj = w(xij ).
/ M = {xai11 , . . . , xaiss }. Also,
Proof. By Lemma 3.8, there is (x, xij ) ∈ E(D) with x ∈
w(xi )
w(xi )
w(xi )
xxij j ∈ I(D) ⊆ J, so xaikk |xxij j for some 1 ≤ k ≤ s. Hence, xaikk |xij j , since
x∈
/ M . This implies, k = j and aj ≤ w(xij ). If aj < w(xij ), then we take J ′ = (M ′ )
a
w(xi )
where M ′ = {M \ {xijj }} ∪ {xij j }. So, J ′ ( J. Furthermore, if (a, b) ∈ E(D), then
m = abw(b) ∈ I(D) ⊆ J. Thus, xaikk |abw(b) for some 1 ≤ k ≤ s. If k 6= j, then xaikk ∈ M ′
a
implying abw(b) ∈ J ′ . Now, if k = j then xijj |bw(b) , since aj > 1. Consequently, xij = b
w(xi )
and xij j |m. Then m ∈ J ′ . Hence I(D) ⊆ J ′ ( J, a contradiction since J is minimal.
Therefore aj = w(xij ).
Theorem 3.10. The following conditions are equivalent:
1) J is a minimal irreducible monomial ideal of I(D).
2) There is a strong vertex cover C of D such that J = IC .
Proof. 2) ⇒ 1) By definition J = IC is a monomial irreducible ideal. By Lemma 3.5,
I(D) ⊆ J. Now, suppose I(D) ⊆ J ′ ⊆ J, where J ′ is a monomial irreducible ideal.
We can assume G(J ′ ) = {xbj11 , . . . , xbjss }. If x ∈ L1 (C), then there is (x, y) ∈ E(D) with
y ∈
/ C. Hence, xy w(y) ∈ I(D) and y r ∈
/ J for each r ∈ N. Consequently y r ∈
/ J ′ for
bi
w(y)
for some 1 ≤ i ≤ s, since
each r, implyinig y ∈
/ {xj1 , . . . , xjs }. Furthermore xji |xy
bi
w(y)
′
′
xy
∈ I(D) ⊆ J . This implies, x = xji ∈ J . Now, if x ∈ L2 (C), then there is
(y, x) ∈ E(D) with y ∈
/ C. Thus y ∈
/ J, so y ∈
/ {xbj11 , . . . , xbjss }. Also, xw(x) y ∈ I(D) ⊆ J ′ ,
5
then xbjii |xw(x) y for some 1 ≤ i ≤ s. Consequently, xbjii |xw(x) implies xw(x) ∈ J ′ . Finally
if x ∈ L3 (C), then there is (y, x) ∈ E(D) where y ∈ L2 (C) ∪ L3 (C) and w(y) 6= 1,
since C is a strong vertex cover. So, xw(x) y ∈ I(D) ⊆ J ′ implies xbjii |xw(x) y for some i.
Furthermore y ∈
/ J = IC , since y ∈ L2 (C) ∪ L3 (C) and w(y) 6= 1. This implies y ∈
/ J′
bi w(x)
w(x)
′
′
so, xji |x
then x
∈ J . Hence, J = IC ⊆ J . Therefore, J is a minimal monomial
irreducible of I(D).
1) ⇒ 2) Since J is irreducible, we can suppose G(J) = {xai11 , . . . , xaiss }. By Lemma 3.9, we
have aj = 1 or aj = w(xij ) for each 1 ≤ j ≤ s. Also, by Lemma 3.7, C = {xi1 , . . . , xis }
is a vertex cover of D. We can assume G(IC ) = {xbi11 , . . . , xbiss }, then bj ∈ {1, w(xij )} for
each 1 ≤ j ≤ s. Now, suppose bk = 1 and w(xik ) 6= 1 for some 1 ≤ k ≤ s. Consequently
xik ∈ L1 (C). Thus, there is (xik , y) ∈ E(D) where y ∈
/ C. So, xik y w(y) ∈ I(D) ⊆ J
ar
w(y)
and xir |xik y
for some 1 ≤ r ≤ s. Furthermore y ∈
/ C, then r = k and ak = ar = 1.
Hence, IC ∩ V (D) ⊆ J ∩ V (D). This implies, IC ⊆ J, since aj , bj ∈ {1, w(xij )} for each
1 ≤ j ≤ s. Therefore J = IC , since J is minimal. In particular ai = bi for each 1 ≤ i ≤ s.
Now, assume C is not strong, then there is x ∈ L3 (C) such that if (y, x) ∈ E(D), then
w(y) = 1 or y ∈ L1 (C). We can assume x = xi1 , and we take J ′ the monomial ideal with
a
w(z )
G(J ′ ) = {xai22 , . . . , xaiss }. We take (z1 , z2 ) ∈ E(D). If xijj |z1 z2 2 for some 2 ≤ j ≤ s,
w(z2 )
then z1 z2
w(z2 )
a
∈ J ′ . Now, assume xijj ∤ z1 z2
for each 2 ≤ j ≤ s. Consequently
w(z )
z1 z2 2
w(z )
∈ I(D) ⊆ J, then xai11 |z1 z2 2 .
z2 ∈
/ {xi2 . . . xis }, since aj ∈ {1, w(xij )}. Also
But xi1 ∈ L3 (C), so z1 , z2 ∈ NG [xi1 ] ⊆ C. If xi1 = z1 , then there is 2 ≤ r ≤ s
w(z )
such that z2 = xir . Thus xairr | z1 z2 2 . A contradiction, then xi1 = z2 , z1 ∈ C
and (z1 , xi1 ) ∈ E(D). Then, w(z1 ) = 1 or z1 ∈ L1 (C). In both cases z1 ∈ G(IC ).
w(z )
Furthermore z1 6= z2 since (z1 , z2 ) ∈ E(D). This implies z1 ∈ G(J ′ ). So, z1 z2 2 ∈ J ′ .
′
Hence, I(D) ⊆ J . This is a contradiction, since J is minimal. Therefore C is strong.
Theorem 3.11. If Cs is the set of strong vertex covers
of D, then the irredundant
T
irreducible decomposition of I(D) is given by I(D) = C∈Cs IC .
Proof. By
decomposition
T [3, Theorem 1.3.1], there is a unique′ irredundant irreducible
′
I(D) = m
i=1 Ii . If there isTan irreducible ideal Ij such that I(D) ⊆ Ij ⊆ Ij for some j ∈
{1, . . . , m}, then I(D) = ( i6=j Ii ) ∩ Ij′ is an irreducible decomposition. Furthermore this
decomposition is irredundant. Thus, Ij′ = Ij . Hence, I1 , . . . , Im are minimal irreducible
ideals of I(D). Now, if there is C ∈ Cs such that IC ∈
/ {I1 , . . . , Im }, thenTthere is
m
α1
αm
i
xα
∈
I
\
I
for
each
i
∈
{1,
.
.
.
,
m}.
Consequently,
m
=
lcm(x
i
C
ji
j1 , . . . , xjm ) ∈
i=1 Ii =
β1
βk
I(D) ⊆ IC . Furthermore, if C = {xi1 , . . . , xik }, then IC = (xi1 , . . . , xik ) where βj ∈
β
{1, w(xij )}. Hence, there is j ∈ {1, . . . , k} such that xijj |m. So, there is 1 ≤ u ≤ m
T
β
αu
u
such that xijj | xα
/ IC . Therefore I(D) = C∈Cs IC is the
ju . A contradiction, since xju ∈
irredundant irreducible decomposition of I(D).
Remark 3.12. If C1 , . . . , Cs are the strong vertex covers of D, then by Theorem 3.11,
IC1 ∩ · · · ∩ ICs is the irredundant irreducible decomposition of I(D). Furthermore, if
Pi = rad(ICi ), then Pi = (Ci ). So, Pi 6= Pj for 1 ≤ i < j ≤ s. Thus, IC1 ∩ · · · ∩ ICs
is an irredundant primary decomposition of I(D). In particular we have Ass(I(D)) =
{P1 , . . . , Ps }.
6
Example 3.13. Let D be the following oriented weighted graph
x1
3
x5
x2
4
2
x3
5
2
x4
whose edge ideal is I(D) = (x31 x2 , x42 x3 , x53 x4 , x3 x25 , x24 x5 ). From Theorem 3.10 and Theorem 3.11, the irreducible decomposition of I(D) is:
I(D) = (x31 , x3 , x24 ) ∩ (x31 , x3 , x5 ) ∩ (x2 , x3 , x24 ) ∩ (x2 , x53 , x5 ) ∩ (x2 , x4 , x25 ) ∩
(x31 , x42 , x53 , x5 ) ∩ (x31 , x42 , x4 , x25 ) ∩ (x2 , x53 , x24 , x25 ) ∩ (x31 , x42 , x53 , x24 , x25 ).
Example 3.14. Let D be the following oriented weighted graph
x1
x2
x3
x4
2
(x1 x22 , x2 x53 , x3 x74 ).
Hence, I(D) =
decomposition of I(D) is:
5
7
By Theorem 3.10 and Theorem 3.11, the irreducible
I(D) = (x1 , x3 ) ∩ (x22 , x3 ) ∩ (x2 , x74 ) ∩ (x1 , x53 , x74 ) ∩ (x22 , x53 , x74 ).
In Example 3.13 and Example 3.14, I(D) has embedding primes. Furthermore the
monomial ideal (V (D)) is an associated prime of I(D) in Example 3.13. Proposition 2.14
and Remark 3.12 give a combinatorial criterion for to decide when (V (D)) ∈ Ass(I(D)).
4
Unmixed weighted oriented graphs
Let D = (V, E, w) be a weighted oriented graph whose underlying graph is G = (V, E). In
this section we characterize the unmixed property of I(D) and we prove that this property
is closed under c-minors. In particular if G is a bipartite graph or G is a whisker or G is
a cycle, we give an effective (combinatorial) characterization of this property.
Definition 4.1. An ideal I is unmixed if each one of its associated primes has the same
height.
Theorem 4.2. The following conditions are equivalent:
1) I(D) is unmixed.
2) Each strong vertex cover of D has the same cardinality.
3) I(G) is unmixed and L3 (C) = ∅ for each strong vertex cover C of D.
7
Proof. Let C1 , . . . , Cℓ be the strong vertex covers of D. By Remark 3.12, the associated
primes of I(D) are P1 , . . . , Pℓ , where Pi = rad(ICi ) = (Ci ) for 1 ≤ i ≤ ℓ.
1) ⇒ 2) Since I(D) is unmixed, |Ci | = ht(Pi ) = ht(Pj ) = |Cj | for 1 ≤ i < j ≤ ℓ.
2) ⇒ 3) If C is a minimal vertex cover, then by Corollary 2.8, C ∈ {C1 , . . . , Cℓ }. By
hypothesis, |Ci | = |Cj | for each 1 ≤ i ≤ j ≤ ℓ, then Ci is a minimal vertex cover of D.
Thus, by Lemma 2.5, L3 (Ci ) = ∅. Furthermore I(G) is unmixed, since C1 , . . . , Cℓ are
the minimal vertex covers of G.
3) ⇒ 1) By Proposition 2.5, Ci is a minimal vertex cover, since L3 (Ci ) = ∅ for each
1 ≤ i ≤ ℓ. This implies, C1 , . . . , Cℓ are the minimal vertex covers of G. Since G is
unmixed, we have |Ci | = |Cj | for 1 ≤ i < j ≤ ℓ. Therefore I(D) is unmixed.
Definition 4.3. A weighted oriented graph D has the minimal-strong property if each
strong vertex cover is a minimal vertex cover.
Remark 4.4. Using Proposition 2.5, we have that D has the minimal-strong property
if and only if L3 (C) = ∅ for each strong vertex cover C of D.
Definition 4.5. D′ is a c-minor of D if there is a stable set S of D, such that D′ =
D \ NG [S].
Lemma 4.6. If D has the minimal-strong property, then D′ = D \ NG [x] has the
minimal-strong property, for each x ∈ V .
Proof. We take a strong vertex cover C ′ of D′ = D \ NG [x] where x ∈ V . Thus,
C = C ′ ∪ ND (x) is a vertex cover of D. If y ′ ∈ L3 (C ′ ), then by Proposition 2.4,
ND′ (y ′ ) ⊆ C ′ . Consequently, ND (y ′ ) ⊆ C ′ ∪ ND (x) = C implying y ′ ∈ L3 (C). Hence,
L3 (C ′ ) ⊆ L3 (C). Now, we take y ∈ L3 (C), then ND (y) ⊆ C. This implies y ∈
/ ND (x),
since x ∈
/ C. Then, y ∈ C ′ and ND′ (y) ∪ (ND (y) ∩ ND (x)) = ND (y) ⊆ C = C ′ ∪ ND (x).
So, ND′ (y) ⊆ C ′ implies y ∈ L3 (C ′ ). Therefore L3 (C) = L3 (C ′ ).
Now, if y ∈ L3 (C) = L3 (C ′ ), then there is z ∈ C ′ \ L1 (C ′ ) with w(z) 6= 1, such that
(z, y) ∈ E(D′ ). If z ∈ L1 (C), then there exist z ′ ∈
/ C such that (z, z ′ ) ∈ E(D). Since
′
′
′
′
z ∈
/ C, we have z ∈
/ C , then z ∈ L1 (C ). A contradiction, consequently z ∈
/ L1 (C).
Hence, C is strong. This implies L3 (C) = ∅, since D has the minimal-strong property.
Thus, L3 (C ′ ) = L3 (C) = ∅. Therefore D′ has the minimal-strong property.
Proposition 4.7. If D is unmixed and x ∈ V , then D′ = D \ NG [x] is unmixed.
Proof. By Theorem 4.2, G is unmixed and D has the minimal-strong property. Hence,
by [6], G′ = G \ NG [x] is unmixed. Also, by Lemma 4.6 we have that D′ has the
minimal-strong property. Therefore, by Theorem 4.2, D′ is unmixed.
Theorem 4.8. If D is unmixed, then a c-minor of D is unmixed.
Proof. If D′ is a c-minor of D, then there is a stable S = {a1 , . . . , as } such that D′ =
D \ NG [S]. Since S is stable, D′ = (· · · ((D \ NG [a1 ]) \ NG [a2 ]) \ · · · ) \ NG [as ]. Hence,
by induction and Proposition 4.7, D′ is unmixed.
Proposition 4.9. If V (D) is a strong vertex cover of D, then I(D) is mixed.
Proof. By Proposition 2.4 V (D) is not minimal, since L3 (V (D)) = V (D). Therefore, by
Theorem 4.2, I(D) is mixed.
8
Remark 4.10. If V = V + , then I(D) is mixed.
−
Proof. If xi ∈ V , then by Remark 3.3 ND
(xi ) 6= ∅, since V = V + . Thus, there is xj ∈ V
such that (xj , xi ) ∈ E(D). Also, w(xj ) 6= 1 and xj ∈ V = L3 (V ). So, V is a strong
vertex cover. Hence, by Proposition 4.9, I(D) is mixed.
In the following three results we assume that D1 , . . . , Dr are the connected components of D. Furthermore Gi is the underlying graph of Di .
Sr
Lemma 4.11. Let C be a vertex cover of D, then L1 (C) = i=1 L1 (Ci ) and L3 (C) =
S
r
i=1 L3 (Ci ), where Ci = C ∩ V (Di ).
Proof. We take x ∈ C, then x ∈ Cj for some 1 ≤ j ≤ r. Thus, ND (x) = NDj (x). In
Sr
+
+
+
+
particular ND
(x) = ND
(x), so C ∩ND
(x) = Cj ∩ND
(x). Hence, L1 (C) = i=1 L1 (Ci ).
j
j
On the other hand,
x ∈ L3 (C) ⇔ ND (x) ⊆ C ⇔ NDj (x) ⊆ Cj ⇔ x ∈ L3 (Cj ).
Sr
Therefore, L3 (C) = i=1 L3 (Ci ).
Lemma 4.12. Let C be a vertex cover of D, then C is strong if and only if each Ci =
C ∩ V (Di ) is strong with i ∈ {1, . . . , r}.
−
Proof. ⇒) We take x ∈ L3 (Cj ). By Lemma 4.11, x ∈ L3 (C) and there is z ∈ ND
(x)∩V +
−
with z ∈ C \ L1 (C), since C is strong. So, z ∈ NDj (x) and z ∈ V (Dj ), since x ∈ Dj .
Consequently, by Lemma 4.11, z ∈ Cj \ L1 (Cj ). Therefore Cj is strong.
⇐) We take x ∈ L3 (C), then x ∈ Ci for some 1 ≤ i ≤ r. Then, by Lemma 4.11,
−
(x) such that w(a) 6= 1 and a ∈ Ci \ L1 (Ci ), since Ci
x ∈ L3 (Ci ). Thus, there is a ∈ ND
i
is strong. Hence, by Lemma 4.11, a ∈ C \ L1 (C). Therefore C is strong.
Corollary 4.13. I(D) is unmixed if and only if I(Di ) is unmixed for each 1 ≤ i ≤ r.
Proof. ⇒) By Theorem 4.8, since Di is a c-minor of D.
⇐) By Theorem 4.2, Gi is unmixed thus G is unmixed. Now, if C is a strong vertex
cover, then by Lemma 4.11, Ci = C ∩ V (Di ) is a strong vertex cover.
S Consequently,
L3 (Ci ) = ∅, since I(Di ) is unmixed. Hence, by Lemma 4.11, L3 (C) = ri=1 L3 (Ci ) = ∅.
Therefore, by Theorem 4.2, I(D) is unmixed.
Definition 4.14. Let G be a simple graph whose vertex set is V (G) = {x1 , . . . , xn }
and edge set E(G). A whisker of G is a graph H whose vertex set is V (H) = V (G) ∪
{y1 , . . . , yn } and whose edge set is E(H) = E(G) ∪ {{x1 , y1 }, . . . , {xn , yn }}.
Definition 4.15. Let D and H be weighted oriented graphs. H is a weighted oriented
whisker of D if D ⊆ H and the underlying graph of H is a whisker of the underlying
graph of D.
Theorem 4.16. Let H a weighted oriented whisker of D, where V (D) = {x1 , . . . , xn }
and V (H) = V (D) ∪ {y1 , . . . , yn }, then the following conditions are equivalents:
1) I(H) is unmixed.
2) If (xi , yi ) ∈ E(H) for some 1 ≤ i ≤ n, then w(xi ) = 1.
9
Proof. 2) ⇒ 1) We take a strong vertex cover C of H. Suppose xj , yj ∈ C, then
yj ∈ L3 (C), since ND (yj ) = {xj } ⊆ C. Consequently, (xj , yj ) ∈ E(G) and w(xj ) 6= 1,
since C is strong. This is a contradiction by condition 2). This implies, |C ∩ {xi , yi }| = 1
for each 1 ≤ i ≤ n. So, |C| = n. Therefore, by Theorem 4.2, I(H) is unmixed.
1) ⇒ 2) By contradiction suppose (xi , yi ) ∈ E(H) and w(xi ) 6= 1 for some i. Since
w(xi ) 6= 1 and by Remark 3.3, we have that xi is not a source. Thus, there is xj ∈ V (D),
such that (xj , xi ) ∈ E(H). We take the vertex cover C = {V (D) \ xj } ∪ {yj , yi }, then
by Proposition 2.4, L3 (C) = {yi }. Furthermore ND (xi ) \ C = {xj } and (xj , xi ) ∈ E(H),
then xi ∈ L2 (C). Hence C is strong, since L3 (C) = {yi }, (xi , yi ) ∈ E(G) and w(xi ) 6= 1.
A contradiction by Theorem 4.2, since I(H) is unmixed.
Theorem 4.17. Let D be a bipartite weighted oriented graph, then I(D) is unmixed if
and only if
1) G has a perfect matching {{x11 , x21 }, . . . , {x1s , x2s }} where {x11 , . . . , x1s } and {x21 , . . . , x2s }
are stable sets. Furthermore if {x1j , x2i }, {x1i , x2k } ∈ E(G) then {x1j , x2k } ∈ E(G).
′
′
+ k
(xj ) = {xki1 , . . . , xkir } where {k, k ′ } = {1, 2}, then ND (xkiℓ ) ⊆
2) If w(xkj ) 6= 1 and ND
+ k
− k
ND (xj ) and ND (xiℓ ) ∩ V + = ∅ for each 1 ≤ ℓ ≤ r.
Proof. ⇐) By 1) and [1, Theorem 2.5.7], G is unmixed. We take a strong vertex cover
C of D. Suppose L3 (C) 6= ∅, thus there exist xki ∈ L3 (C). Since C is strong, there is
′
′
′
xkj ∈ V + such that (xkj , xki ) ∈ E(D), xkj ∈ C \ L1 (C) and {k, k ′ } = {1, 2}. Furthermore
′
′
+ k′
+ k′
ND
(xj ) ⊆ C, since xkj ∈
/ L1 (C). Consequently, by 3), ND (xki ) ⊆ ND
(xj ) ⊆ C and
′
′
− k
+
k
ND (xi )∩V = ∅. A contradiction, since xi ∈ L3 (C) and C is strong. Hence L3 (C) = ∅
and D has the strong-minimal property. Therefore I(D) is unmixed, by Theorem 4.2.
⇒) By Theorem 4.2, G is unmixed. Hence, by [1, Theorem 2.5.7], G satisfies 1).
+ k
+ k
If w(xkj ) 6= 1, then we take C = ND
(xj ) ∪ {xki | ND (xki ) 6⊆ ND
(xj )} and k ′ such that
′
′
+ k
{k, k ′ } = {1, 2}. If {xki , xki′ } ∈ E(G) and xki ∈
/ C, then xki′ ∈ ND (xki ) ⊆ ND
(xj ) ⊆ C.
k
This implies, C is a vertex cover of D. Now, if xi1 ∈ L3 (C), then ND (xki1 ) ⊆ C.
+ k
/ C. A contradiction, then L3 (C) ⊆
(xj ) implies xki1 ∈
Consequently ND (xki1 ) ⊆ ND
+ k
− k
− k
k
ND (xj ). Also, NG (xj ) 6= ∅, since w(xj ) 6= 1. Thus xkj ∈ L2 (C), since NG
(xj ) ∩ C = ∅.
+ k
+
k
Hence C is strong, since L3 (C) ⊆ ND (xj ) and xj ∈ V . Furthermore {x′1 , . . . , x′s }
is a minimal vertex cover, then by Theorem 4.2 |C| = s, since D is unmixed. We
′
′
+ k
/ C for each 1 ≤ ℓ ≤ r. So,
assume ND
(xj ) = {xki1 , . . . xkir }. Since C is minimal, xkiℓ ∈
′
− k
+ k
+
k
ND (xiℓ ) ⊆ ND (xj ). Now, suppose z ∈ ND (xiℓ ) ∩ V , then z = xkiℓ′ for some 1 ≤ ℓ′ ≤ r,
+ k
+ k
+ k′
since ND (xkiℓ ) ⊆ ND
(xj ). We take C ′ = ND
(xj ) ∪ {xki | i ∈
/ {i1 , . . . , ir }} ∪ ND
(xiℓ′ ).
+ k
′
k
Since ND (xiu ) ⊆ ND (xj ) for each 1 ≤ u ≤ r, we have that C is a vertex cover. If
′
′
′
+ k
(xj ) implies q ∈ {i1 , . . . , ir }.
{xkq , xkq } ∩ L3 (C) = ∅, then {xkq , xkq } ⊆ C ′ , so xkq ∈ ND
′
′
′
+ k
k
′
k
Consequently, xq ∈ ND (xiℓ′ ), since xq ∈ C . This implies, (xkj , xkq ), (xkiℓ′ , xkq ) ∈ E(D).
′
+ k′
+ k
Moreover, ND
(xiℓ′ ) ∪ ND
(xj ) ⊆ C ′ , then xkiℓ′ ∈
/ L1 (C ′ ) and xkj ∈
/ L1 (C ′ ). Thus, C ′ is
′
+ k
k
k′
+
′
(xj )
strong, since xj , xiℓ′ ∈ V . Furthermore, by Theorem 4.2, |C | = s. But xkiℓ ∈ ND
′
′
+ k
− k
k
k
k
′
+
and xiℓ ∈ ND (xiℓ′ ), hence xiℓ , xiℓ ∈ C . A contradiction, so ND (xiℓ )∩V = ∅. Therefore
D satisfies 2).
10
Lemma 4.18. If the vertices of V + are sinks, then D has the minimal-strong property.
Proof. We take a strong vertex cover C of D. Hence, if y ∈ L3 (C), then there is
(z, y) ∈ E(D) with z ∈ V + . Consequently, by hypothesis, z is a sink. A contradiction,
since (z, y) ∈ E(D). Therefore, L3 (C) = ∅ and C is a minimal vertex cover.
Lemma 4.19. Let D be a weighted oriented graph, where G ≃ Cn with n ≥ 6. Hence,
D has the minimal-strong property if and only if the vertices of V + are sinks.
Proof. ⇐) By Lemma 4.18.
⇒) By contradiction, suppose there is (z, y) ∈ E(D), with z ∈ V + . We can assume
G = (x1 , x2 , . . . , xn , x1 ) ≃ Cn , with x2 = y and x3 = z. We take a strong vertex
cover C in the following form: C = {x1 , x3 , . . . , xn−1 } ∪ {x2 } if n is even or C =
{x1 , x3 , . . . , xn−2 } ∪ {x2 , xn−1 } if n is odd. Consequently, if x ∈ C and ND (x) ⊆ C, then
x = x2 . Hence, L3 (C) = {x2 }. Furthermore (x3 , x2 ) ∈ E(D) with x3 ∈ V + . Thus, x3 is
not a source, so, (x4 , x3 ) ∈ E(D). Then, x3 ∈ L2 (C). This implies C is a strong vertex
cover. But L3 (C) 6= ∅. A contradiction, since D has the minimal-strong property.
D2
D1
1
x1
1
x2
x5
x3
w(x5 ) 6= 1
w(x2 ) 6= 1
x1
x2
1 x5
x3 1
w(x3 ) 6= 1
D4
D3
w(x1 ) 6= 1
1
x1
1
x2
x5
x3
w(x5 ) 6= 1
w(x3 ) 6= 1
w(x2 ) 6= 1
1
x1
x2
x5
x3
w(x5 ) 6= 1
w(x3 ) 6= 1
x4
x4
x4
x4
1
1
w(x4 ) 6= 1
1
Theorem 4.20. If G ≃ Cn , then I(D) is unmixed if and only if one of the following
conditions hold:
1) n = 3 and there is x ∈ V (D) such that w(x) = 1.
2) n ∈ {4, 5, 7} and the vertices of V + are sinks.
3) n = 5, there is (x, y) ∈ E(D) with w(x) = w(y) = 1 and D 6≃ D1 , D 6≃ D2 , D 6≃ D3 .
4) D ≃ D4 .
Proof. ⇒) By Theorem 4.2, D has the minimal- strong property and G is unmixed. Then,
by [1, Exercise 2.4.22], n ∈ {3, 4, 5, 7}. If n = 3, then by Remark 4.10, D satisfies 1). If
n = 7, then by Lemma 4.19, D satisfies 2). Now suppose n = 4 and D does not satisfies
2), then we can assume x1 ∈ V + and (x1 , x2 ) ∈ E(D). Consequently, (x4 , x1 ) ∈ E(G),
since w(x1 ) 6= 1. Furthermore, C = {x1 , x2 , x3 } is a vertex cover with L3 (C) = {x2 }.
Thus, x1 ∈ L2 (C) and (x1 , x2 ) ∈ E(D) so C is strong. A contradiction, since C is not
minimal. This implies D satisfies 2). Finally suppose n = 5. If D ≃ D1 , then C1 =
{x1 , x2 , x3 , x5 } is a vertex cover with L3 (C1 ) = {x1 , x2 }. Also (x5 , x1 ), (x3 , x2 ) ∈ E(D)
11
with x5 , x3 ∈ V + . Consequently, C1 is strong, since x5 , x3 ∈ L2 (C1 ). A contradiction,
since C1 is not minimal. If D ≃ D2 , then C2 = {x1 , x2 , x4 , x5 } is a vertex cover where
L3 (C2 ) = {x1 , x5 } and (x2 , x1 ), (x1 , x5 ) ∈ E(D) with x2 , x1 ∈ V + . Hence, C2 is strong,
since x2 , x1 ∈
/ L1 (C2 ). A contradiction, since C2 is not minimal. If D ≃ D3 , C3 =
{x2 , x3 , x4 , x5 } is a vertex cover where L3 (C3 ) = {x3 , x4 } and (x4 , x3 ), (x5 , x4 ) ∈ E(D)
with x4 , x5 ∈ V + . Thus, C3 is strong, since x4 , x5 ∈
/ L1 (C3 ). A contradiction, since C3 is
not minimal. Now, since n = 5 and by 3) we can assume (x2 , x3 ) ∈ E(D), x2 , x3 ∈ V +
and there are not two adjacent vertices with weight 1. Since x2 ∈ V + , (x1 , x2 ) ∈ E(D).
Suppose there are not 3 vertices z1 , z2 , z3 in V + such that (z1 , z2 , z3 ) is a path in G, then
w(x4 ) = w(x1 ) = 1. Furthermore, w(x5 ) 6= 1, since there are not adjacent vertices with
weight 1. So, C4 = {x2 , x3 , x4 , x5 } is a vertex cover of D, where L3 (C4 ) = {x3 , x4 }. Also
(x2 , x3 ) ∈ E(G) with w(x2 ) 6= 1. Hence, if (x3 , x4 ) ∈ E(D) or (x5 , x4 ) ∈ E(D), then C4
is strong, since x3 , x5 ∈ V + . But C4 is not minimal. Consequently, (x4 , x3 ), (x4 , x5 ) ∈
E(D) and D ≃ D4 . Now, we can assume there is a path (z1 , z2 , z3 ) in D such that
z1 , z2 , z3 ∈ V + . Since there are not adjacent vertices with weight 1, we can suppose there
is z4 ∈ V + such that L = (z1 , z2 , z3 , z4 ) is a path. We take {z5 } = V (D) \ V ((L)) and we
can assume (z2 , z3 ) ∈ E(D). This implies, (z1 , z2 ), (z5 , z1 ) ∈ E(D), since z1 , z2 ∈ V + .
Thus, C5 = {z1 , z2 , z3 , z4 } is a vertex cover with L3 (C5 ) = {z2 , z3 }. Then C5 is strong,
since (z1 , z2 ), (z2 , z3 ) ∈ E(D) with z2 ∈ L3 (C5 ) and z1 ∈ L2 (C5 ). A contradiction, since
C5 is not minimal.
⇐) If n ∈ {3, 4, 5, 7}, then by [1, Exercise 2.4.22] G is unmixed. By Theorem 4.2,
we will only prove that D has the minimal-strong property. If D satisfies 2), then by
Lemma 4.18, D has the minimal-strong property. If D satisfies 1) and C is a strong
vertex cover, then by Proposition 2.14, |C| ≤ 2. This implies C is minimal. Now,
suppose n = 5 and C ′ is a strong vertex cover of D with |C ′ | ≥ 4. If D ≃ D4 , then
−
−
x2 , x5 ∈
/ L3 (C ′ ), since (ND
(x2 ) ∪ ND
(x5 )) ∩ V + = ∅. So ND (x2 ) 6⊆ C ′ and ND (x5 ) 6⊆ C ′ .
−
′
′
Consequently, x1 ∈
/ C implies C = {x2 , x3 , x4 , x5 }. But x4 ∈ L3 (C ′ ) and ND
(x4 ) = ∅.
′
A contradiction, since C is strong. Now assume D satisfies 3). Suppose there is a
path L = (x1 , x2 , x3 ) in G such that w(x1 ) = w(x2 ) = w(x3 ) = 1. We can suppose
(x4 , x5 ) ∈ E(D) where V (D)\V (L) = {x4 , x5 }. Since w(x1 ) = w(x3 ) = 1, x2 ∈
/ L3 (C ′ ). If
−
′
′
′
x2 ∈
/ C , then C = {x1 , x3 , x4 , x5 } and x4 ∈ L3 (C ). But ND (x4 ) = {x3 } and w(x3 ) = 1.
A contradiction, hence x2 ∈ C ′ . We can assume x3 ∈
/ C ′ , since x2 ∈
/ L3 (C ′ ). This
′
′
implies C = {x1 , x2 , x4 , x5 } and L3 (C ) = {x1 , x5 }. Thus, (x5 , x1 ) ∈ E(D), x5 , x4 ∈ V + .
Consequently (x3 , x4 ) ∈ E(D), since x4 ∈ V + . A contradiction, since D 6≃ D2 . Hence,
there are not three consecutive vertices whose weights are 1. Consequently, since D
satisfies 3), we can assume w(x1 ) = w(x2 ) = 1, w(x3 ) 6= 1 and w(x5 ) 6= 1. If w(x4 ) =
1, then x3 , x5 ∈
/ L3 (C ′ ) since ND (x3 , x5 ) ∩ V + = ∅. This implies ND (x3 ) 6⊆ C ′ and
′
ND (x5 ) 6⊆ C . Then, x4 ∈
/ C ′ and C ′ = {x1 , x2 , x3 , x5 }. Thus, (x5 , x1 ), (x3 , x2 ) ∈ E(D),
′
since L3 (C ) = {x1 , x2 }. Consequently, (x4 , x5 ), (x4 , x3 ) ∈ E(D), since x5 , x3 ∈ V + . A
contradiction, since D 6≃ D1 . So, w(x4 ) 6= 1 and we can assume (x5 , x4 ) ∈ E(D), since
x4 ∈ V + . Furthermore (x1 , x5 ) ∈ E(D), since x5 ∈ V + . Hence, (x3 , x4 ) ∈ E(D), since
D 6≃ D3 . Then (x2 , x3 ) ∈ E(D), since x3 ∈ V + . This implies x1 , x2 , x3 , x5 ∈
/ L3 (C ′ ),
−
+
′
since ND (xi ) ∩ V = ∅ for i ∈ {1, 2, 3, 5}. A contradiction, since |C | ≥ 4. Therefore D
has the minimal-strong property.
12
5
Cohen-Macaulay weighted oriented graphs
In this section we study the Cohen-Macaulayness of I(D). In particular we give a combinatorial characterization of this property when D is a path or D is complete. Furthermore, we show the Cohen-Macaulay property depends of the characteristic of K.
Definition 5.1. The weighted oriented graph D is Cohen- Macaulay over the field K if
the ring R/I(D) is Cohen-Macaulay.
Remark 5.2. If G is the underlying graph of D, then rad(I(D)) = I(G).
Proposition 5.3. If I(D) is Cohen-Macaulay, then I(G) is Cohen-Macaulay and D has
the minimal-strong property.
Proof. By Remark 5.2, I(G) = rad(I(D)), then by [4, Theorem 2.6], I(G) is CohenMacaulay. Furthermore I(D) is unmixed, since I(D) is Cohen-Macaulay. Hence, by
Theorem 4.2, D has the minimal-strong property.
Example 5.4. In Example 3.13 and Example 3.14 I(D) is mixed. Hence, I(D) is not
Cohen-Macaulay, but I(G) is Cohen-Macaulay.
Conjecture 5.5. I(D) is Cohen- Macaulay if and only if I(G) is Cohen-Macaulay and
D has the minimal-strong property. Equivalently I(D) is Cohen-Macaulay if and only if
I(D) is unmixed and I(G) is Cohen-Macaulay.
Proposition 5.6. Let D be a weighted oriented graph such that V = {x1 , . . . , xk } and
whose underlying graph is a path G = (x1 , . . . , xk ). Then the following conditions are
equivalent.
1) R/I(D) is Cohen-Macaulay.
2) I(D) is unmixed.
3) k = 2 or k = 4. In the second case, if (x2 , x1 ) ∈ E(D) or (x3 , x4 ) ∈ E(D), then
w(x2 ) = 1 or w(x3 ) = 1 respectively.
Proof. 1) ⇒ 2) By [1, Corollary 1.5.14].
2) ⇒ 3) By Theorem 4.17, G has a perfect matching, since D is bipartite. Consequently
k is even and {x1 , x2 }, {x3 , x4 }, . . . , {xk−1 , xk } is a perfect matching. If k ≥ 6, then
by Theorem 4.17, we have {x2 , x5 } ∈ E(G), since {x2 , x3 } and {x4 , x5 } ∈ E(G). A
contradiction since {x2 , x5 } ∈
/ E(G). Therefore k ∈ {2, 4}. Furthermore by Theorem
4.17, w(x2 ) = 1 or w(x3 ) = 1 when (x2 , x1 ) ∈ E(D) or (x3 , x4 ) ∈ E(D), respectively.
3) ⇒ 1) We take I = I(D). If k = 2, then we can assume (x1 , x2 ) ∈ E(D). So,
w(x )
w(x )
I = (x1 x2 2 ) = (x1 ) ∩ (x2 2 ). Thus, by Remark 3.12, Ass(I) = {(x1 ), (x2 )}. This
implies, ht(I) = 1 and dim(R/I) = k − 1 = 1. Also, depth(R/I) ≥ 1, since (x1 , x2 ) ∈
/
Ass(I). Hence, R/I is Cohen-Macaulay. Now, if k = 4, then ht(I) = ht(rad(I)) =
ht(I(G)) = 2. Consequently, dim(R/I) = k − 2 = 2. Furthermore one of the following
w(x )
w(x )
w(x )
w(x )
w(x )
w(x )
sets {x2 − x1 1 , x3 − x4 4 }, {x2 − x1 1 , x4 − x3 3 }, {x1 − x2 2 , x4 − x3 3 } is a
regular sequence of R/I, then depth(R/I) ≥ 2. Therefore, I is Cohen-Macaulay.
Theorem 5.7. If G is a complete graph, then the following conditions are equivalent.
13
1) I(D) is unmixed.
2) I(D) is Cohen-Macaulay.
3) There are not D1 , . . . , Ds unicycles orientes subgraphs of D such that V (D1 ), . . . , V (Ds )
is a partition of V (D)
Proof. We take I = I(D). Since I(G) = rad(I) and G is complete, ht(I) = ht(I(G)) =
n − 1.
1) ⇒ 3) Since ht(I) = n − 1 and I is unmixed, (x1 , . . . , xn ) ∈
/ Ass(I). Thus, by Remark
3.12, V (D) is not a strong vertex cover of D. Therefore, by Proposition 2.14, D satisfies
3).
3) ⇒ 2) By Proposition 2.14, V (D) is not a strong vertex cover of D. Consequently,
by Remark 3.12, (x1 , . . . , xn ) ∈
/ Ass(I). This implies, depth(R/I) ≥ 1. Furthermore,
dim(R/I) = 1, since ht(I) = n − 1. Therefore I is Cohen-Macaulay.
2) ⇒ 1) By [1, Corollary 1.5.14].
If D is complete or D is a path, then the Cohen-Macaulay property does not depend
of the field K. It is not true in general, see the following example.
Example 5.8. Let D be the following weighted oriented graph:
1
x4
2 x1
1
x7
x10 1
1
x5
2 x2
2 x3
x11 1
x8 1
x6
1
x9
1
Hence,
I(D) = (x21 x4 , x21 x8 , x21 x5 , x21 x9 , x22 x10 , x22 x5 , x22 x11 , x22 x8 , x22 x6 , x23 x7 , x23 x10 , x23 x6 ,
x23 x9 , x4 x8 , x4 x7 , x4 x11 , x5 x10 , x5 x9 , x5 x11 , x6 x8 , x6 x9 , x6 x11 , x7 x10 , x7 x11 , x9 x11 ).
By [5, Example 2.3], I(G) is Cohen- Macaulay when the characteristic of the field K
is zero but it is not Cohen-Macaulay in characteristic 2. Consequently, I(D) is not
Cohen-Macaulay when the characteristic of K is 2. Also, I(G) is unmixed. Furthermore,
by Lemma 4.18, I(D) has the minimal-strong property, then I(D) is unmixed. Using
Macaulay2 [2] we show that I(D) is Cohen-Macaulay when the characteristic of K is
zero.
14
References
[1] I. Gitler, R. H. Villarreal, Graphs, Rings and Polyhedra, Aportaciones Mat. Textos,
35, Soc. Mat. Mexicana, México, 2011.
[2] D. Grayson and M. Stillman, Macaulay2, 1996. Available via anonymous ftp from
math.uiuc.edu.
[3] J. Herzog and T. Hibi, Monomial Ideals, Graduate Texts in Mathematics. 260,
Springer, 2011.
[4] J. Herzog, Y. Takayama, N. Terai, On the radical of a monomial ideal, Arch. Math.
85 (2005), 397–408.
[5] A. Madadi and R. Zaare-Nahandi, Cohen-Macaulay r-partite graphs with minimal
clique cover, Bull. Iranian Math. Soc. 40 (2014) No. 3, 609–617.
[6] R. H. Villarreal, Monomial Algebras, Second Edition, Monographs and Research
Notes in Mathematics, Chapman and Hall/CRC, 2015.
15
| 0math.AC
|
arXiv:0812.1986v1 [cs.SE] 10 Dec 2008
Control software analysis, Part II
Closed-loop properties
Eric Feron
∗
Fernando Alegre
†
March 7, 2018
Abstract: The analysis and proper documentation of the properties of
closed-loop control software presents many distinct aspects from the analysis of
the same software running open-loop. Issues of physical system representations
arise, and it is desired that such representations remain independent from the
representations of the control program. For that purpose, a concurrent program
representation of the plant and the control processes is proposed, although the
closed-loop system is sufficiently serialized to enable a sequential analysis. While
dealing with closed-loop system properties, it is also shown by means of examples
how special treatment of nonlinearities extends from the analysis of control
specifications to code analysis.
1
Closed-loop software system analysis
The first part of this report essentially dealt with the analysis of open-loop
control systems, that is, the analysis of control algorithms without regards for
their stabilizing effect on the plant: In [FA08], we were concerned with verifying
and quantifying the boundedness of all computed quantities in the controller,
regardless of the actual performance of the closed-loop system. In that regard,
the analysis is similar to that described in [CCF+ 05], although our work is much
more focused on control algorithms and, unlike the work in [CCF+ 05], it ignores
the many other software components present in a fully developed avionics system. In the present document, we focus on the analysis and verification of the
system’s closed-loop performance, beginning with closed-loop system stability.
One of the elements of this report is the modeling, by means of an example,
of closed-loop control systems as two concurrent processes, where one process
represents the real-time computer program and the other represents the system being controlled. The concurrent representation of these processes does
not make their analysis more complicated, and it provides the engineer with
greater flexibility by making the choice of controller representations and implementations independent from the representation of the plant being controlled.
∗ Dutton/Ducoffe Professor of Aerospace Software Engineering, Georgia Institute of Technology. [email protected]
† College of Computing, Georgia Tech. [email protected]
1
Again, the best way to perform the closed-loop system analysis is to begin with
specification-level closed-loop system analysis, and then to move on to incorporate more detailed controller implementations. This is the process illustrated in
this report.
1.1
Specification-level analyses: Stability
The analysis of the specifications of the lead-lag controller designed and presented in [FA08] begins with Figure 1. In that figure, the controlled system
is represented by a continuous differential equation, while the controller is a
discrete-time system that interacts with the continuous system by means of a
signal sampler at the input and a zero-order hold at the output. The dynamics
of the controller may be written
xc,k+1
uk
=
=
Ac xc,k + Bc SAT(yk )
Cc xc,k + Dc SAT(yk )
with
Ac =
0.4999
0.0100
−0.0500
1.0000
, Bc =
1
0
Cc = [564.48 0] , and Dc = −1280.
It is well-known from control theory [FP80] that the analysis of such a closedloop, sampled-data system can be performed by replacing the plant, the zeroorder hold and the sampler by the discrete-time system
xp,k+1
yk
with
Ap =
1.0000
−0.0100
= Ap xp,k + Bp uk
= Cp xp,k
0.0100
1.0000
, Bp =
0.00005
0.01
and Cp = [1 0]. Moreover, xp,k = xp (0.01k), where k ∈ N. Then the analysis
of the system shown in Fig. 1 is equivalent to the analysis of the discrete-time
system shown in Fig. 2. The stability analysis of this discrete-time system
is made more difficult than that of the controller alone [FA08] because of the
presence of a saturation nonlinearity aimed at limiting the range of sensor inputs
to the control system. This nonlinearity may be accounted for in the overall
system stability analysis by using a variety of computational techniques, such
as those described in [BEFB94]. In particular one of the available techniques
consists of computing a quadratic Lyapunov function for the system. It is, for
example, possible to show that the quadratic function
0.2205 0.0188 −0.0750 0.0177
T
0.0188 0.4736 0.0535
xc
xc
0.0015
V (x) =
P
, with P =
−0.0750 0.0535 0.1012 −0.0049
xp
xp
0.0177 0.0015 −0.0049 0.0015
2
u
∙
=
∙
uk xc,k+1
=
d
dt
x
ẋ
¸
¸∙
¸ ∙ ¸
0 1
x
0
+
u, x(0) = x0 , ẋ(0) = ẋ0
y
−1 0
ẋ
1
∙
¸
x
= [1 0]
.
ẋ
y
ZOH
uk
∙
−0.050
1.000
0.499
0.010
¸
xc,k +
∙
¸
1
0
SAT(yk )
yk
S
= [564.48 0] xc,k − 1280 SAT(yk )
Figure 1: Feedback system
xp,k+1
yk
uk
=
∙
1.0000 0.0100
−0.0100 1.0000
= [1 0] xp,k .
xc,k+1
uk
=
∙
0.499
0.010
¸
xp,k +
−0.050
1.000
¸
∙
0.00005
0.01
xc,k +
∙
1
0
¸
uk , x0 =
¸
SAT(yk )
= [564.48 0] xc,k − 1280 SAT(yk )
Figure 2: Equivalent discrete system
3
∙
x0
ẋ0
yk
¸
decays along all trajectories (xc,k , xp,k )k=1,2,... for initial conditions where the
controller is at rest (xc,0 = 0) and the initial state xp,0 of the mechanical system
lies inside the ellipse EQ with
0.1012 −0.0049
Q=
and EQ = {y ∈ R2 | y T Qy ≤ 1}
(1)
−0.0049 0.0015
The mechanisms used to reach that conclusion rely on standard stability considerations, most notably absolute stability theory: Consider the closed-loop
system dynamics
xk+1 = Axk + BSAT(Cxk ),
with x = [xTc xTp ]T ,
A=
Ac
Bp Cc
0.4990 −0.0500
0
0.0100 1.0000
0
0
=
0.0282
Ap
0
1.0000
5.6448
0
−0.0100
Bc
B=
, and C = [0 Cp ].
Bp Dc
0
0
,
0.0100
1.0000
Consider the set of x’s such that xT P x < 1. It is easy to show that (i) the set of
all x = [0 xTp ]T such that xp ∈ EQ is contained in EP , and (ii) that for all x ∈ EP ,
(SAT(Cx) − 0.2Cx)(SAT(Cx) − Cx) ≤ 0. Introducing yc,k = SAT(Cxk ), it is
therefore sufficient to show that
xTk+1 P xk+1 − xTk P xk ≤ 0
whenever (yc,k − 0.2Cxk )(yc,k − Cxk ) ≤ 0. An equivalent condition is
x
yc
T
AT
BT
P [A B] −
I
0
P [I 0]
x
yc
≤0
whenever (yc −0.2Cx)(yc −Cx) ≤ 0. Using the well-known S-procedure [AG64],
a sufficient condition is the existence of λ > 0 such that
T
A
I
0.2C T C −0.6C T
P
[A
B]
−
P
[I
0]
−
λ
≤ 0.
(2)
0
−0.6C
1
BT
The latter inequality can be easily shown to be true for λ = 6.76. It must
be kept in mind that all statements related to numerical objects and computations are made under the assumption that computations on reals are exact.
This assumption is obviously false and must be eventually addressed when implementing the operational software verification tools that may arise from this
paper, using for example techniques developped by Goubault [Gou01]. We are
now interested in showing how this proof of (local) closed-loop stability may
be translated at the code and system level. To be more precise, we will show
how invariance of the ellipsoid EP can be exploited to develop a proof of proper
4
% Controller Dynamics
1c: Ac = [0.4990, -0.0500; 0.0100,1.0000];
2c: Cc=[564.48, 0];
3c: Bc=[1;0]; D = -1280;
4c: xc = zeros(2,1);
5c: receive(y);
6c: while (1)
7c:
yc = max(min(y,1),-1);
8c:
u = Cc*xc + Dc*yc;
9c:
xc = Ac*xc + Bc*yc;
10c: send(u);
11c: receive(y);
12c: end
% Plant Dynamics
1p: Ap = [1.0000,0.0100;-0.0100,1.0000];
2p: Cp=[1,0];
3p: Bp = [0.00005; 0.01];
4p: while (1)
5p:
y = Cp*xp;
6p:
send(y);
7p:
receive(u);
8p:
xp = Ap*xp + Bp*u;
9p: end;
Table 1: Concurrent representation of plant and controller implementations
behavior, such as stability and performance, for the computer program that
implements the controller, as it interacts with the physical system. Unlike the
developments relative to controller open-loop properties, this proof necessarily
involves the presence of the controlled artifact. It would not make sense for
the system to be represented at various degrees of computer program implementations since these do not have any physical meaning. Thus we choose to
represent the plant and the computer program as two concurrent programs, as
shown in Table 1. The computer program representation of the mechanical system will remain invariant, written in a high-level language like MATLAB, while
the representation of the controller will be allowed to evolve to reflect the various
stages of its implementation, without changing the representation of the physical system. We assume the two programs communicate by means of send /
receive commands. It is assumed that the receive command is blocking, that
is, the program execution cannot proceed until the variable of interest has been
received. While the programs in Table 1 are purely event-driven, it is possible
to modify the algorithmic description of the plant to account for the presence of
time. The question of establishing proofs of stability of the closed-loop system at
the code level is necessarily tied to understanding the behavior of the controller
and that of the plant. The entire state-space then consists of the direct sum of
the controller’s state-space and that of the plant. We now propose to leverage
the approach developed in the first part of this report [FA08] to document the
corresponding system of two concurrent programs. This documentation can be
extended to other representations of the control program implementation: Table 2 shows one such implementation of the controller in pseudo-C. One of the
interesting aspects of the processes in this report is their concurrency, which
can make the structure of the state transitions rather complicated. However, a
close inspection of the programs reveals a relatively simple transition structure.
Our investigation thus begins with a forward analysis of both programs. Since
5
/* variable declarations */
double Ac[2][2];
double Cc[2];
double Bc[2];
double Dc;
double xc[2], yc,y,u;
int flags, flaga;
/* aux vars */
int i,j;
double xc_new[2];
/* code */
Ac[0][0]=0.4990; Ac[0][1]=-0.0500; Ac[1][0]=0.0100;Ac[1][1]=1.0000;
Cc[0]=564.48;Cc[1]=0.0;
Bc[0]=1.;Bc[1]=0.;
Dc=-1280.;
xc[0]=0.;xc[1]=0.;
receive(y);
while(1) {
yc = (y > 1. ? 1. : y) < -1. ? -1. : y;
u = 0.0;
for(i=0;i<2;i++) u += Cc[i]*xc[i];
u += Dc*yc;
for(i=0;i<2;i++) {
xc_new[i] = 0.0;
for(j=0;j<2;j++) xc_new[i] += Ac[i][j]*xc[j];
xc_new[i] += Bc[i]*yc;
}
for(i=0;i<2;i++) xc[i] = xc_new[i];
send(u);
receive{y}
}
Table 2: Implementation of controller in C
6
only local stability has been proven from the program specifications, we may
not expect any better result by inspecting and documenting the code. Declared
but non-initialized variables v will be assumed to belong to the empty set, that
is v ∈⊥. Likewise, the set of all possible values will be denoted >, and a variable v that may take any possible value will be characterized as v ∈ >. Since
both programs run concurrently, tracking state values and the sets they belong
to may be somewhat challenging. The only variable whose value is assumed
to be initialized prior to the plant execution is the initial plant state xp . For
the purpose of stability analysis, we assume that the state xp is initially within
the ellipsoid EQ , where Q is defined in (1). Thus, prior to the initiation of
the plant process, and using insight from the system specification analysis, we
form the condition {(xc , xp ) ∈ EP , (y, yc , u) ∈ >}. The first three lines of the
plant dynamics 1p, 2p, 3p do not change this condition, and so they may be
commented as
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
1p: Ap = [1.0000,0.0100;-0.0100,1.0000];
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
2p: Cp=[1,0];
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
3p: Bp = [0.00005; 0.01];
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
Line 4p: while (1), together with line 9p: end; defines the main loop of
the plant dynamics. We use the requirements analysis to postulate the following
set of pre- and post-conditions
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
4p: while (1)
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
..
.
{(xc , xp ) ∈ EP , (y, yc , u) ∈ >}
9p: end
{f alse}
The last assertion {f alse} simply indicates a never ending loop; this should
be expected from a dynamical system that never ceases to execute. In the following, and in order to simplify the notation, we will omit to mention the variables
whose value could be arbitrary (that is, the variables whose values belong to
>). With this convention, the previous set of commented lines becomes
7
{(xc , xp ) ∈ EP }
4p: while (1)
{(xc , xp ) ∈ EP }
..
.
{(xc , xp ) ∈ EP }
9p: end
{f alse}.
The following line, 5p: y = Cp*xp assigns a value to the variable y. The
corresponding pair of conditions is therefore
{(xc , xp ) ∈ EP }
5p: y = Cp*xp
{(xc , xp , y) ∈ GR },
with
GR =
x ∈ Rn
1
x
xT
R
≥0
.
The next line, 6p: send(y); does not affect the state variables of interest
and therefore becomes
{(xc , xp , y) ∈ GR }
6p: send(y)
{(xc , xp , y) ∈ GR }
At this point, the execution of the plant process is blocked by the receive
command in line 7p. We now turn our attention to the controller code. The
first four lines of the controller code 1c, 2c, 3c, 4c, are commented using the
statement {(xc , xp , y) ∈ GR , (yc , u) ∈ >} or, in short, {(xc , xp , y) ∈ GR }. Thus
we get
{(xc , xp , y) ∈ GR }
1c: Ac = [0.4990, -0.0500; 0.0100,1.0000];
{(xc , xp , y) ∈ GR }
2c: Cc=[564.48 0];
{(xc , xp , y) ∈ GR }
3c: Bc=[1; 0]; D = -1280;
{(xc , xp , y) ∈ GR },
4c: xc = zeros(2,1);
{(xc , xp , y) ∈ GR }.
At this point, the execution of the controller program is halted by the command 5c: receive(y) and can proceed only after the plant has executed line
6p. Valid pre- and post-conditions are
{(xc , xp , y) ∈ GR }
5c: receive(y);
{(xc , xp , y) ∈ GR }
8
We then encounter line 6c: while (1) which, together with line 12c:
end defines the main loop of the controller. We postulate the following pre- and
post-conditions for these instructions
{(xc , xp , y) ∈ GR }
6c: while (1)
{(xc , xp , y) ∈ GR }
.
.
.
{(xc , xp , y) ∈ GR }
12c: end
{f alse}.
The next line 7c: yc = max(min(y,1),-1), applies the nonlinear saturation operator to y. Common systems practice (as well as the methods used
to compute P in the first place) suggests that this nonlinear relation may be
accurately captured by the quadratic inequality
(yc − y)(yc − 0.2y) ≤ 0,
also named a sector bound on the saturation operator. This sector bound is not
true in general, but it holds for all variables in GR . This bound may also be
expressed in matrix form
T
0 0
xc
xp 0 0
y 0 0
0 0
yc
0
0
0.2
−0.6
xc
0
xp
0
−0.6 y
yc
1
T
xc
xc
xp
xp
=
y T y
yc
yc
≤ 0.
(3)
Thus one possible set of pre- and post- conditions for line 7c is
{(xc , xp , y) ∈ GR }
7c: yc = max(min(y,1),-1)
T
xc
xp
T
(xc , xp , y) ∈ GR ,
y
yc
xc
xp
≤0
y
yc
However, following the principles outlined in [FA08], we seek to capture all
variables within a single quadratic constraint, for the purpose of making proof
checking simple and mirroring the usage of the S-procedure to establish stability
proofs at the specification level. To do so, we rely on the following lemma:
Lemma: Assume the real vector variables z and w satisfy the constraints
1 zT
≥0
z U
for a given U = U T and
z
w
T
T
9
z
w
≤0
for T = T T . Then for any real µ, we have
z
w
∈ GV , V =
I
0
0
0
−µ
U
0
0
I
−1
U
T
0
0
I
whenever V , defined as such, exists and is positive definite.
Proof: The proof is simple and therefore omitted.
Applying this lemma with U = R, T defined as in (3) and µ = −λ = −6.76
yields the inequality
T T
1 [xc xp y yc ]
xc
≥ 0.
xp
V
y
yc
and the pre- and post-conditions
{(xc , xp , y) ∈ GR }
7c: yc = max(min(y,1),-1);
{(xc , xp , y, yc ) ∈ GV }
At this point, the variable y is not of use anymore and we may therefore
release it. Thus we replace the post-condition {(xc , xp , y, yc ) ∈ GV } of line 7c
by the post condition {(xc , xp , yc ) ∈ GW }, with
I
W = 0
0
The assignment 8c:
0
I
0
0
0
0
I
0
0 V 0
0
I
0
I
0
0
0
0
T
0
0
I
u = Cc*xc + Dc*yc is then easily commented as
{(xc , xp , yc ) ∈ GW }
8c: u = Cc*xc + Dc*yc
{(xc , xp , u, yc ) ∈ GX }
with
I
0
X=
Cc
0
0
I
0
0
0
I
0
0
W
Dc Cc
1
0
The next line of the controller loop 9c:
similarly to line 8c, to yield the triple
{(xc , xp , u, yc ) ∈ GX }
9c: xc = Ac*xc + Bc*yc
{(xc , xp , u) ∈ GY }
10
0
I
0
0
T
0
0
.
Dc
1
xc = Ac*xc + Bc*yc can be treated
with
Ac 0
Y = 0 I
0 0
0
0
1
Bc
Ac 0
0 X 0 I
0
0 0
0
0
1
T
Bc
0 .
0
Note that the variable yc is not used anymore and has therefore been released.
Line 10c does not influence the variables and therefore can be commented as
{(xc , xp , u) ∈ GY }
10c: send(u)
{(xc , xp , u) ∈ GY }
The only remaining line in the controller loop is then line 11c: receive(y),
which blocks the controller loop until a new value of y is available.
At that point, it becomes important to look again at the plant process,
which restarts at line 7p: receive(u). A necessary post-condition for this
line is {(xc , xp , u) ∈ GY }, but no pre-condition can be clearly extracted. Line
7p will therefore be only partially instrumented, replacing the pre-condition by
.
the symbol ... In addition, we indicate between brackets which line execution in
the controller code (line 10c) unlocks line 7p, yielding the “triple”
..
.
7p: receive(u) [10c]
{(xc , xp , u) ∈ GY }.
Consider then line 8p: xp = Ap*xp + Bp*u;. This assignment is properly
documented with pre- and post- conditions to yield
{(xc , xp , u) ∈ GY }
8p: xp = Ap*xp + Bp*u;
{(xc , xp ) ∈ GZ }
where
Z=
I
0
0
Ap
0
Bp
Y
I
0
0
Ap
0
Bp
T
,
and u has been released.
At that point, the plant dynamics meets line 9p: end. That line has already been instrumented with candidate pre-and post conditions, which therefore must be shown to be compatible with the post-condition of line 9p. More
precisely, we must show
(xc , xp ) ∈ GZ ⇒ (xc , xp ) ∈ EP .
(4)
Numerical computations show that this is indeed true (modulo floating point
operation arithmetic errors). While the entire loop describing the plant dynamics has been investigated, such is not the case of the controller dynamics,
whose lines 11c and 12c have not been executed yet. We therefore need to keep
11
tracking the execution of the two programs. While the control program is still
blocked at line 11c, the plant dynamics loop sends the program execution to line
4p, 5p, and 6p, where the pre- and post- conditions are already established and
coherent. Line 7p then blocks any further propagation of the plant dynamics,
while the controller program is now allowed to proceed past line 11c, for which
a valid post-condition is therefore {(xc , xp , y) ∈ GR } and we will instrument as
follows
..
.
11c: receive(y) [6p]
{(xc , xp , y) ∈ GR }
which is compatible with the following line, 12c, which has been already
documented.
The resulting commented programs then look like those shown in Table 3.
2
Sector-bounded nonlinearities and stability analysis
Following similar developments in [FA08], there are apparent differences between
the stability analysis performed in the requirements-level stability analysis, summarized by Eq. (2), and the forward analysis performed afterwards. Namely,
we want to show that the invariance condition (4) is satisfied if and only if the
condition (2) is satisfied. For that purpose, we begin with the forward analysis
of the system, beginning at line 7c: yc = max(min(y,1),-1). The assertion
T
xc
xc
xp
xp
(xc , xp , y) ∈ GR ,
T
≤
0
y
y
yc
yc
with
R=
P −1
CP −1
P −1 C T
CP −1 C T
implies the assertion
1
xc
xp
y
yc
0 0
0 0
, and T =
0 0
0 0
[xTc xTp y yc ]
V
with
V =
I
0
0
0
−µ
R
0
12
0
I
0
0
−0.6
1
0
0
0.2
−0.6
≥ 0,
−1
R
T
0
0
I
.
% Controller Dynamics
{(xc , xp , y) ∈ GR }
1c: Ac = [0.4990, -0.0500; 0.0100,1.0000];
{(xc , xp , y) ∈ GR }
2c: Cc=[564.48, 0];
{(xc , xp , y) ∈ GR }
3c: Bc=[1;0]; D = -1280;
{(xc , xp , y) ∈ GR }
4c: xc = zeros(2,1);
{(xc , xp , y) ∈ GR }
5c: receive(y)
{(xc , xp , y) ∈ GR }
6c: while (1)
{(xc , xp , y) ∈ GR }
7c: yc = max(min(y,1),-1)
{(xc , xp , y, yc ) ∈ GV }
skip
{(xc , xp , yc ) ∈ GW }
8c: u = Cc*xc + Dc*yc
{(xc , xp , u, yc ) ∈ GX }
9c: xc = Ac*xc + Bc*yc
{(xc , xp , u) ∈ GY }
10c: send(u)
{(xc , xp , u) ∈ GY }
.
.
.
11c: receive(y) [6p]
{(xc , xp , y) ∈ GR }
12c: end
{f alse}.
% Plant Dynamics
{(xc , xp ) ∈ EP }
1p: Ap = [1.0000,0.0100;-0.0100,1.0000];
{(xc , xp ) ∈ EP }
2p: Cp=[1,0];
{(xc , xp ) ∈ EP }
3p: Bp = [0.00005; 0.01];
{(xc , xp ) ∈ EP }
4p: while (1)
{(xc , xp ) ∈ EP }
5p: y = Cp*xp
{(xc , xp , y) ∈ GR }
6p: send(y);
{(xc , xp , y) ∈ GR }
.
.
.
7p: receive(u); [10c]
{(xc , xp , u) ∈ GY }
8p: xp = Ap*xp+ Bp*u;
{(xc , xp ) ∈ GZ }
skip
{(xc , xp ) ∈ EP }
9p: end
{f alse}
Table 3: Commented closed-loop control program
13
Long and somewhat tedious computations indicate that W , where W is
obtained by erasing the row and column of V corresponding to the the variable
y, satisfies
−1
P + 0.2µC T C −0.6µC T
W =
−0.6µC
−µ
−1
−1 T
P̃
0.6P̃ C
1
=
0.6C P̃ −1 − + 0.36µC P̃ −1 C T
µ
P̃ −1
0
I
0
I 0.6C T
1
,
=
0.6C 1
0
1
0
−
µ
with P̃ = P − 0.16µC T C. Following the forward analysis and substituting
matrix expressions for their values eventually yield
P̃ −1
0
T
I
0
I 0.6C T
1
A B
Z= A B
.
0.6C 1
0
1
0
−
µ
The invariance condition (4) therefore becomes
P̃ −1
0
I
0
I 0.6C T
1
A B
A
0.6C 1
0
1
0
−
µ
Using Schur complements a first time yields the inequality
I
−1
A B
−P
0.6C
T
T
I
0
−P̃ 0
A B
[30pt]
0.6C 1
0
µ
B
0
1
T
≤ P −1
≤ 0.
Using Schur complements a second time then yields the inequality
T
T
I
0
I
0
P̃
0
A B
P A B
−
≤ 0.
0.6C 1
0.6C 1
0 −µ
Multiplying this inequality to the left and right by
T
I
0
I
and
−0.6C 1
−0.6C
0
1
,
respectively, yields
A
B
T
P
A
B
−
P
0
0
0
which is the same as (2), with µ = −λ.
14
+µ
0.2C T C
−0.6C
−0.6C T
1
≤ 0,
3
Conclusions
This report describes an approach to documenting control programs, whereby
the control program code is annotated with logical expressions describing the
set of reachable program states. This approach constitutes the application of
the Floyd-Hoare paradigm to control programs. It is shown that the the extensive domain knowledge gathered by control theory about control system specifications is readily applicable to develop stability and performance proofs of the
corresponding control programs. Some system elements, such as sector-bounded
nonlinearities, can be handled via forward analysis, and this analysis was shown
to be equivalent to well known results in absolute stability theory.
The analyses discussed in this report may be used in several different contexts: First, they may be used in an autocoding environment, whereby diagrambased specification languages such as Simulink can be autocoded in a target
language such as C to automatically produce software with extensive, inlined
proofs of software stability and performance. Such a documented software may
then be easily verified independently by a proof checker.
Second, the analyses described in this paper may also be used to produce
stability and poerformance proofs of undocumented software, even if little or
no information is available about the software specifications. Producing such
proofs becomes an integral part of the software verification process. Although
time-consuming, this process is the only available option for much of the existing
body of real-time control software. A particular case of interest is concerned
with autocoded sofftware, where prior understanding of the autocoding process
may help speed up the verification task.
Acknowledgements
This report was prepared under partial support from the National Science Foundation under grant CSR/EHS 0615025, NASA under cooperative agreement
NNX08AE37A, and the Dutton-Ducoffe professorship at Georgia Tech.
References
[AG64]
M. A. Aizerman and F. R. Gantmacher. Absolute stability of regulator
systems. Information Systems. Holden-Day, San Francisco, 1964.
[BEFB94] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in System and Control Theory, volume 15 of SIAM
Studies in Applied Mathematics. SIAM, 1994.
[CCF+ 05] P. Cousot, R. Cousot, J. Feret, A. Miné, D. Monniaux,
L. Mauborgne, and X. Rival. The ASTRÉE analyzer. In In S. Sagiv
(Ed.), Programming Languages and Systems, 14th European Symposium on Programming, ESOP 2005, Held as Part of the Joint European Conferences on Theory and Practice of Software, ETAPS 2005,
15
Lecture Notes in Computer Science 3444, pages 21–30, Edinburgh,
UK, April 4–8, 2005, 2005. Springer.
[FA08]
E. Feron and F. Alegre. Control software analysis, part I: Open-loop
properties. Technical Report arXiv:0809.4812, arXiv.org, October
2008.
[FP80]
G. Franklin and J. D. Powell. Digital Control Of Dynamic Systems.
Addison-Wesley, 1980.
[Gou01]
E. Goubault. Static analyses of the precision of floating point operations. In P. Cousot, editor, Static Analysis: 8th International
Symposium, SAS 2001,Paris, France, July 16-18, 2001 : Proceedings, pages 234–257. Springer-Verlag, 2001.
16
| 6cs.PL
|
Random Forest Ensemble of Support Vector
Regression Models for Solar Power Forecasting
Mohamed Abuella, Student Member, IEEE
Badrul Chowdhury, Senior Member, IEEE
Energy Production and Infrastructure Center
Department of Electrical and Computer Engineering
University of North Carolina at Charlotte
Charlotte, USA
Email: [email protected]
Energy Production and Infrastructure Center
Department of Electrical and Computer Engineering
University of North Carolina at Charlotte
Charlotte, USA
Email: [email protected]
Abstract— To mitigate the uncertainty of variable renewable
resources, two off-the-shelf machine learning tools are deployed
to forecast the solar power output of a solar photovoltaic system.
The support vector machines generate the forecasts and the
random forest acts as an ensemble learning method to combine
the forecasts. The common ensemble technique in wind and solar
power forecasting is the blending of meteorological data from
several sources. In this study though, the present and the past
solar power forecasts from several models, as well as the
associated meteorological data, are incorporated into the random
forest to combine and improve the accuracy of the day-ahead
solar power forecasts. The performance of the combined model is
evaluated over the entire year and compared with other
combining techniques.
Keywords—Ensemble learning, post-processing, random forest,
solar power, support vector regression.
I.
INTRODUCTION
The wind and solar energy resources have created
operational challenges for the electric power grid due to the
uncertainty involved in their output in the short term. The
intermittency of these resources may adversely affect the
operation of the electric grid when the penetration levels of
these variable generations are high. Thus, wherever the
variable generation resources are used, it becomes highly
desirable to maintain higher than normal operating reserves
and efficient energy storage systems to manage the power
balance in the system. The operating reserves that use fossil
fuel generating units should be kept to a minimum in order to
get the maximum benefit from the deployment of the
renewable resources. Therefore, the forecast of these variable
generations becomes a vital tool in the operation of the power
systems and electricity markets [1].
As in wind power forecasting, the solar power also consists
of a variety of methods based on the time horizon being
forecasted, the data available to the forecaster and the
particular application of the forecast. The methods are broadly
categorized according to the time horizon in which they
generally show value. Methods that are common in solar power
forecasting include Numerical Weather Prediction (NWP) and
Model Output Statistics (MOS) to produce forecasts, as well as
hybrid techniques that combine ensemble forecasts and
Statistical Learning Methods [2]. Applying machine learning
techniques directly to historical time series of solar
photovoltaic (PV) production associated with NWP outcomes
have placed among the top models in the recent global
competition of energy forecasting, GEFCom2014 [3]. Just to
name a few of the machine learning tools, the artificial neural
networks (ANN) and support vector regression (SVR), gradient
boosting (GB), random forest (RF), etc. are believed to be the
most common.
Hybrid models of two or more statistical and physical
techniques are often combined to capture complex interactions
and provide useful insights and better forecasts. In ref. [4], the
authors implement a hybrid model that consists of ARMA and
ANN to forecast the solar irradiance by NWP data for 5
locations of a Mediterranean climate. They found the proposed
model outperforms the naïve persistence model and
improvement with respect to its core techniques as well. The
study reported in ref. [5] presents the benefits of combining the
data of solar irradiance that is derived from a satellite with
ground-measured data to improve the intraday forecasts in the
range up to six hours ahead. In ref [6], the authors combine
satellite images with ANN outcomes to forecast the solar
irradiance of leading time up to two hours for two sites in
California.
In ref. [7], several statistical combining methods are used to
combine multiple linear regression models for load forecasting,
and the authors conclude that the regression combining
technique is the best. While ref. [8] uses several statistical
models to forecast the hourly PV electricity production for the
next day at some power plants in France, the random forest
(RF) has shown a superior performance. Ref. [9] also found the
random forest has the best performance among others to
predict the daily solar irradiance variability of four sites with
different climatic conditions in Australia. The authors of [10]
used random forest with other models as well to forecast the
solar power in GEF2014 competition. They indicate that the
random forest and the gradient boosting technique are the most
accurate models. In the aforementioned studies, RF is not
implemented there as a combining method; it is a standalone
forecasting model that depends on its own trees.
The commonly used ensemble technique in wind and solar
power forecasting is to blend the weather data from several
sources. As in ref. [11], the authors compare several data-
Note: This is a pre-print of the full paper that published in Innovative Smart Grid Technologies, North America Conference, 2017, which can be referenced as below:
M. Abuella and B. Chowdhury, “Random Forest Ensemble of Support Vector Regression for Solar Power Forecasting,” in Proceedings of Innovative Smart Grid
Technologies, North America Conference, 2017.
driven models using input data from two NWPs and building
two artificial hybrid and stochastic ensemble models based on
ANN, the model that combines multiple models outperforms
the rest of the models. It points out that the ensemble is
enhanced by including forecasts with similar accuracy, but
generated from NWP data of higher variance and different
data-driven techniques. Ref. [12] uses Ensemble Prediction
System (EPS) to produce weather scenarios by running
multiple initials to quantify the uncertainty and then produce
the probabilistic solar power forecasts of sites in Italy. Ref.
[13] applies a physical post processing and ANN to improve
the solar irradiance forecasts of one and two days ahead.
of the bagged tree should be grown by a random number of
features and observation samples. Hence, this method is called
the random forest (RF).
The majority of the existing research literature on
combining forecasting methods of solar forecasts do not
include the previously generated forecasts to boost the model
performance. It can be useful to add these past models’
outcomes as well into the ensemble learning methods. The
research team from the National Renewable Energy Laboratory
(NREL) and IBM Thomas Watson Research Center [17]
deployed and tested several machine-learning techniques to
blend three NWPs outcomes. They conclude that the ensemble
approaches that take into account the diversity and the state
parameters of the models provide lower errors in the solar
irradiance forecasting. Although these studies forecast the solar
irradiance at different sites in the U.S., the time period is
limited since they do not investigate the performance of the
different seasons over the entire year.
Three parameters are required to be set in RF, the number
of trees B (forest size), m the number of predictors out of p
available variables (features) that are randomly chosen to be
used for each split, and the minimum number nmin of
observations per node (leaf size).
Using RF as a combining method of other models in solar
power is scarce. This paper adopts RF to combine the forecasts
for a PV system and investigates the performance throughout a
complete year. The rest of the paper is organized as follows:
Section II describes the ensemble method that is used to build
and combine the SVR models. Section III discusses the
methodology of combining the solar power forecasts. Section
IV presents the results and the evaluation of the model. Section
V provides the conclusions.
II.
ENSEMBLE LEARNING
The algorithms that use decision trees can be useful to
combine the different models’ outcomes efficiently. This
ensemble approach combines all the outputs from variant
models besides the features, such as the weather data that allow
the ensemble method to find the associative rules to determine
the best output. For instance, if the weather is sunny, then
model A performs best, and its output should be selected;
otherwise model B should be selected, and so on. This
ensemble learning approach has shown very promising results
in numerous machine learning benchmarks. For more details
on this topic, i.e., ensemble learning and its techniques as
bagging, boosting, stacking, and Bayesian averaging, the
interested reader may refer to [16], The general diagram of
combining the models is shown in Fig.1.
A. Random Forest
Since the classification and regression trees (CART) use
the bagging principle of the ensemble learning, and they are
built by the same data, these trees are sometimes correlated and
statistically dependent. Consequently, to make the trees more
variant and uncorrelated, Breiman [17] proposed that each split
Model A
Model B
Individual
forecasting
models
Method of
Combining The
Models
Model C
Combined Forecasts
Model N
Fig.1. General diagram of combining different models
The random forest building algorithm [16] has three major
steps.
• Create B sample datasets of size N from the training
data; these sample datasets can be replaced and
overlapped.
• For each sample dataset, grow a random forest tree Tb,
by repeating the following steps for each terminal node,
until the minimum node size nmin is reached:
I.
Select m predictors at random from the p
variables.
II.
Pick the best predictor among the m
selected predictors for the split-point.
III.
Split this point (node) into two daughter
nodes by setting certain decision rules.
• Finally, find the ensemble of the trees {Tb }1B , where B
is the number of trees in the random forest.
The prediction of a given point x of the response variable
can be obtained by averaging the individual tree’s outputs:
𝐵
1
(1)
𝑓̂𝑅𝐹 = ∑ 𝑇𝑏 (𝑥)
𝐵
𝑏=1
The ensemble learning algorithm repeatedly assembles the
input data to create regression trees that best fit the relationship
between the features and the output. This process of
decorrelation of the trees makes the random forest outcomes
less variable, and hence more reliable.
III.
MODELING
A. Data Description
The solar power system is in Australia and has a latitude
35°16'30"S, longitude 149°06'49"E, altitude 595m. The panel
type is Solarfun SF160-24-1M195, consisting of 8 panels, its
nominal power of 1560W, and panel orientation 38° clockwise
from the north, with panel tilt of 36°. The weather forecasts
data and the available measured solar power data starts from
April 2012 to May 2014, as shown in Fig.2.
Note: This is a pre-print of the full paper that published in Innovative Smart Grid Technologies, North America Conference, 2017, which can be referenced as below:
M. Abuella and B. Chowdhury, “Random Forest Ensemble of Support Vector Regression for Solar Power Forecasting,” in Proceedings of Innovative Smart Grid
Technologies, North America Conference, 2017.
2012
2012
2012
2012
2012
2012
2012
2012
2012
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2014
2014
2014
2014
2014
(a)
The forecast methodology is shown in Fig.4.a. The
forecasting day should be excluded from the data, while the
rest of the available data is used to train the SVR models. The
super learner is the random forest where the available weather
and the previous forecasts are blended together to find the
associative rules to achieve better solar power forecasts for the
next day, as shown in Fig.4.b, and they should be as close as
possible to the observed solar power that minimize the forecast
errors of that day.
Day Month
1
:
:
:
:
:
:
30
(b)
31
Fig.2. (a) The weather data and their numeral codes,
(b) available data size
[x − min(X)]
∗ {b − a}
[max(X) − min(X)]
:
:
:
:
:
:
:
:
:
:
:
July
:
:
:
April
May
00:00
:
23:00
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
Day Month
(2)
Where x is a sample from data variable X, {a, b} is the desired
range of the normalized data, such as {0, 1}, and X (min,
max)=the minimum and maximum values of the observed data.
There is also a standardization technique, especially when the
variance of the data is high, which is making the data to have a
zero mean and a unit standard deviation, as follows:
Forecasts
àat 00:00
0.068
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
Combined
Forecasts
Forecasts
0.066
0.069
0.0655
0.068
0.065
0.067
RMSE
(3)
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
(b)
0.066
0.065
0.064
0.066
0.0645
0.065
0.064
0.063
0.064
B. Model Building
The different models are achieved by using different
parameterization within the same model, i.e., support vector
regression models (SVRs). Refer to [18] for more details about
this model. The models are built as in Fig.3, where the
available data is divided into two sets, one dataset consisting of
all 26 months and another dataset consisting of the most recent
12 months only, as shown in Fig.2.a. Then, each of these
datasets is used to build 12 SVR models based on two
normalization techniques as in Equations (2) and (3), after that
two different SVR’s hyper-parameters of C and Gamma are
also chosen, and different combination of weather variables are
used to produce 12 models from each two datasets to get the
total of 24 SVR models.
Original 14
Inputs
Original 14
Inputs
C=16
New Inputs
C=16
New Inputs
Gamma=1
Heat index & wind
Gamma=1
Heat index & wind
Without Cloud
Cover Inputs
Without Cloud
Cover Inputs
Normalized (B)
Original 14
Inputs
Original 14
Inputs
C=10
New Inputs
C=10
New Inputs
Gamma=0.8
Heat index & wind
Gamma=0.8
Heat index & wind
Without Cloud
Cover Inputs
:
:
:
:
:
:
:
:
:
:
:
July
:
:
:
April
May
00:00
:
23:00
RF does not need cross-validation to estimate the
parameters because it has a built-in estimation of accuracy. The
parameter selection is carried out by the wrapping strategy or a
greedy search for the best evaluation results among the
available training data, the parameters search of RF are shown
in Fig.5, [number of trees B=100, samples m=6 (i.e.,18/3), leaf
size nmin =5]. It is worth mentioning that a change with a
reasonable range of these parameters does not affect the RF
performance significantly. Thus, the robustness and flexibility
of RF are the main advantages of this ensemble method.
0.062
0
Normalized (A)
Weather Data Models’ Outcomes PV Power
Fig.4. Demonstrates (a) the 24 hours ahead forecasting and (b)
the combining methodology scheme for May 31st.
RMSE
Xstandardized
June
31
(Model’s
Outcomes)
0.067
[x − mean(X)]
=
std(𝑋)
1
:
:
:
:
:
:
30
(a)
Normalizing the data is of paramount importance since the
scale used for the values for each variable might be different.
The best practice is to normalize the data and transform all the
values to a common scale, as shown in equation (2).
XScaled = a +
Weather Data PV Power
June
(Past Observations)
Year
April
May
June
July
August
September
October
November
December
January
February
March
April
May
June
July
August
September
October
November
December
January
February
March
April
May
RMSE
Month
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
(Past Observations)
No.
Without Cloud
Cover Inputs
Fig.3. Construction of 12 SVR models from a dataset. Another 12 SVRs from
the other dataset. The total number of SVR models is 24
200
400
600
800
1000
Number of Trees
0.063
0
0.0635
5
10
15
20
Number of Features
(a)
0.063
0
20
40
60
80
100
Minimum Number of Samples at Each Node
(b)
(c)
Fig.5.The search for random forest parameters. (a) The forest size, (b) the
features number at each node, (c) the leaf size (minimum number of
samples at each node).
IV.
RESULTS AND EVALUATION
The following metrics are used to evaluate the accuracy of
the forecasts and the model performance: graphs, Root Mean
Square Error (RMSE) as calculated by (4), and a comparison
with other combining methods. For comparison purposes, the
simple average combining method (Avg.), and the best model
(Model 4) out of the 24 SVRs are adopted to evaluate the
performance of the combined forecasts. Also, the improvement
or the skill factor is used to compare the performance of the
combined forecasts with respect to the other methods as in (5).
1
𝑛
𝑅𝑀𝑆𝐸 = √ ∑(𝑌𝑖 − 𝑌̂ 𝑖 )
𝑛
2
𝑖=1
(4)
where 𝑌̂ is the forecast of the solar power and Y is the observed
value of the solar power. 𝑌̂ and Y are normalized to the
nominal installed capacity of the solar power system, 𝑛 is the
number of hours - it can be day hours or month hours. The
objective is to minimize the RMSE for all forecasting hours to
Note: This is a pre-print of the full paper that published in Innovative Smart Grid Technologies, North America Conference, 2017, which can be referenced as below:
M. Abuella and B. Chowdhury, “Random Forest Ensemble of Support Vector Regression for Solar Power Forecasting,” in Proceedings of Innovative Smart Grid
Technologies, North America Conference, 2017.
yield more accurate forecasts. If the training and testing of the
model are carried out for just the daylight hours while filtering
out the night hours (which have zero solar power generation),
the RMSE should also be determined for these daylight hours
only without including the night hours.
In general, the aggregated mean (i.e., the statistical
average) of the monthly RMSEs indicates that the combined
forecasts from the ensemble method have the most accurate
forecasts (RMSEmean=0.0725), while, the total improvement of
the ensemble method over the average and the best model is
9% and 5 % respectively.
𝐼𝑚𝑝𝑟𝑜𝑣𝑒𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 (%) =
Notice that the monthly RMSE is calculated for all hours of
the month where n in (4) is not the day hour, but rather the
month hours. Thus, it could be 744, or any other number of
hours depending on the month.
(𝑂𝑡ℎ𝑒𝑟 𝑚𝑒𝑡ℎ𝑜𝑑 − 𝐸𝑛𝑠𝑒𝑚𝑏𝑙𝑒 𝑀𝑒𝑡ℎ𝑜𝑑)
∗ 100
𝑂𝑡ℎ𝑒𝑟 𝑀𝑒𝑡ℎ𝑜𝑑
(5)
TABLE I
RESULT OF DIFFERENT MODELS AND THE COMBINED
FORECASTS
The best model (model-4) is made of normalization
technique (A), SVR’s hyper-parameters C=10 gamma =8, and
the original 14 weather inputs by using a dataset of all
available months.
Best Model (4) =
All-months dataset + Normalize (A)
+ Parm10_08 + Orig14ins
(6)
In the graphical illustration, the month of October is chosen
for comparison of all SVRs models with other combined
models as shown in Fig.6. For the aggregation of daily RMSEs
over the whole month, it is obvious the ensemble forecasts
(black line) has a lower RMSE than SVR models and the
simple average combining method (dark blue line). However,
in a few days, the best SVR model (model-4) or the average
method could be more accurate, as in 6th, 7th, 16th, 28th, and
29th days, when they have a lower RMSE.
The fluctuation in the daily RMSE can be seen among the
24 SVR models, and it’s a normal trend in the forecasting
models. However, the ensemble method produces more
accurate combined forecasts and the performance of this
combining model becomes more stable than in the individual
SVR models.
The RMSE of the Models
0.22
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
M13
M14
M15
M16
M17
M18
M19
M20
M21
M22
M23
M24
Combined Forecasts
Average Forecasts
0.2
0.18
0.16
RMSE
0.14
0.12
0.1
0.08
0.06
0.04
0.02
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Days
Improvement of
RMSE
Month
Best
Model (4)
June
July
August
September
October
November
December
January
February
March
April
May
Aggregated Mean
0.0734
0.0878
0.0859
0.0843
0.0851
0.0826
0.0747
0.0627
0.0772
0.0800
0.0645
0.0560
0.0762
Ensemble Over:
Simple
Ensemble
Best Model (4) Simple Avg.
Average
0.0784
0.0877
0.0877
0.0921
0.0974
0.0923
0.0798
0.0670
0.0811
0.0822
0.0653
0.0559
0.0806
0.0764
0.0849
0.0833
0.0815
0.0701
0.0737
0.0643
0.0580
0.0730
0.0846
0.0640
0.0561
0.0725
-4%
3%
3%
3%
18%
11%
14%
7%
5%
-6%
1%
0%
5%
3%
3%
5%
12%
28%
20%
19%
13%
10%
-3%
2%
0%
9%
The Random Forest is a nonlinear model and a black-box
and difficult to analyze. In order to get an idea about the
performance of the model, a statistical analysis is conducted.
Firstly, the importance estimation of the RF inputs (features) is
found. This would be available since one of the algorithm steps
is to estimate the best features to split the nodes, as explained
in Section II.A. October and March are chosen for this
analysis, because in these months, the largest change in the
ensemble method performance occurs. The features are the
weather variables and the 24 SVRs’ outcomes. As shown in
Fig.7 and Fig.8, the features do not have the similar pattern of
importance; thereby, in the data training by RF, they would
give different results and different performance.
Secondly, a statistical analysis by finding the standard
deviation and the correlation between the SVR models’
outcomes is conducted over the full year. Fig.9 presents the
histograms of this statistical metrics.
Fig.6. Daily RMSE of different models and combined forecasts, October
Importance of Features, October
Importance of Features, March
1.6
1.4
1.4
1.2
Importance Score
1.2
Importance Score
To get a broader evaluation of the combined forecasts
performance, the comparison is conducted with the best model
(model 4) and the simple average method over a complete year,
as shown in TABLE I. It is clear that the ensemble method has
lower monthly RMSEs. The improvement rate of the ensemble
method over the other methods is calculated as in (5). In some
months such as October, the ensemble method has an
improvement rate of 18% and 28% over the best model and the
average method respectively. For two months where the
improvement rate is negative, such as in March the simple
average and the best model are better than the ensemble, and
also in June, the best model has the lowest monthly RMSE.
1
0.8
0.6
0.4
0.6
0.4
0.2
0.2
0
1
0.8
1
2
3
4
5
6
7
8
Weather Features
(a)
9
10
11
12
13
14
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Weather Features
(b)
Fig.7. The estimation of weather features importance by Random Forest,
(a) October, (b) March
Note: This is a pre-print of the full paper that published in Innovative Smart Grid Technologies, North America Conference, 2017, which can be referenced as below:
M. Abuella and B. Chowdhury, “Random Forest Ensemble of Support Vector Regression for Solar Power Forecasting,” in Proceedings of Innovative Smart Grid
Technologies, North America Conference, 2017.
Importance of outcomes, March
0.7
0.6
0.6
0.5
0.5
Importance Score
Importance Score
Importance of outcomes, October
0.7
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
[5]
0.4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Models’ Outocmes
15
16
17
18
19
20
21
22
23
24
25
0
0
1
2
3
4
5
6
7
(a)
8
9
10
11
12
13
14
Models’ Outocmes
15
16
17
18
19
20
21
22
23
24
25
[6]
(b)
Fig.8. The estimation of models’ outcomes importance by Random
Forest, (a) October, (b) March
[7]
[8]
For October, the standard deviation is high, and hence the
correlation of the SVRs’ outcomes, and thus the ensemble
method, has the best performance. However, in the last
months - March, April, and May, the standard deviation is low
and the correlation is high, and the performance of the
ensemble method in these months is not as good as the other
months, as indicated in TABLE I.
1.000
0.995
0.990
0.985
0.980
0.975
0.970
0.965
0.960
0.955
0.950
0.016
Std. Dev.
0.014
0.012
0.010
0.008
0.006
0.004
0.002
0.000
[10]
Correlation
[11]
Correlation
Std.Dev.
Std. Dev. and Correlation of Models
0.018
[9]
[12]
[13]
Fig.9. The standard deviation and the correlation of different models
V.
CONCLUSION
Combining the forecasts by the random forest leads to more
accurate forecasts throughout the year. These combined
forecasts are produced from an intelligent weighting approach
that takes into account the weather situations, the past forecast
of the forecasting models (24 SVRs), and their different
temporal horizons - these all are used as associative rules in the
ensemble method to yield accurate forecasts and a stable
performance. The simple average as a combining method is not
enough to get better forecasts since it does not capture or
represent all the varieties in the weather data, and hence the
solar power forecast. The forecasting errors are inherited
mostly from the NWP model errors, and some of the errors
resulting from the technical degrading of the physical systems
(PV modules efficiency, orientation, etc.). Adding the past
generated forecasts increases the accuracy of the combined
forecasts.
REFERENCES
[1]
[2]
[3]
[4]
[14]
J. M. Morales, A. J. Conejo, H. Madsen, P. Pinson, and M. Zugno,
“Integrating renewables in electricity markets - Operational problems,”
Springer, vol. 205, p. 429, 2014.
A. Tuohy, J. Zack, S. E. Haupt, J. Sharp, M. Ahlstrom, S. Dise, E. Grimit,
C. Mohrlen, M. Lange, M. G. Casado, J. Black, M. Marquis, and C.
Collier, “Solar Forecasting: Methods, Challenges, and Performance,”
IEEE Power Energy Mag., vol. 13, no. 6, pp. 50–59, 2015.
T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and R. J.
Hyndman, “Probabilistic energy forecasting: Global Energy Forecasting
Competition 2014 and beyond,” Int. J. Forecast., 2016.
C. Voyant, M. Muselli, C. Paoli, and M.-L. Nivet, “Numerical weather
prediction (NWP) and hybrid ARMA/ANN model to predict global
radiation,” Energy, vol. 39, no. 1, pp. 341–355, 2012.
[15]
[16]
[17]
[18]
L. M. Aguiar, B. Pereira, P. Lauret, F. D’\iaz, and M. David, “Combining
solar irradiance measurements, satellite-derived data and a numerical
weather prediction model to improve intra-day solar forecasting,” Renew.
Energy, vol. 97, pp. 599–610, 2016.
R. Marquez, H. T. C. Pedro, and C. F. M. Coimbra, “Hybrid solar
forecasting method uses satellite imaging and ground telemetry as inputs
to ANNs,” Sol. Energy, vol. 92, pp. 176–188, 2013.
J. Liu, “Combining sister load forecasts,” MSc Thesis, University of
North Carolina at Charlotte, 2015.
M. Zamo, O. Mestre, P. Arbogast, and O. Pannekoucke, “A benchmark of
statistical regression methods for short-term forecasting of photovoltaic
electricity production, part I: Deterministic forecast of hourly
production,” Sol. Energy, vol. 105, pp. 792–803, 2014.
J. Huang, A. Troccoli, and P. Coppin, “An analytical comparison of four
approaches to modelling the daily variability of solar irradiance using
meteorological records,” Renew. Energy, vol. 72, no. October, pp. 195–
202, 2014.
A. A. Mohammed, W. Yaqub, and Z. Aung, “Probabilistic Forecasting of
Solar Power: An Ensemble Learning Approach,” in Intelligent Decision
Technologies, Springer, 2015, pp. 449–458.
M. Pierro, F. Bucci, M. De Felice, E. Maggioni, D. Moser, A. Perotto, F.
Spada, and C. Cornaro, “Multi-Model Ensemble for day ahead prediction
of photovoltaic power generation,” Sol. Energy, vol. 134, pp. 132–146,
2016.
S. Sperati, S. Alessandrini, and L. Delle Monache, “An application of the
ECMWF Ensemble Prediction System for short-term solar power
forecasting,” Sol. Energy, vol. 133, pp. 437–450, 2016.
M. Pierro, F. Bucci, C. Cornaro, E. Maggioni, A. Perotto, M. Pravettoni,
and F. Spada, “Model output statistics cascade to improve day ahead solar
irradiance forecast,” Sol. Energy, vol. 117, pp. 99–113, 2015.
W. Cheung, J. Zhang, A. Florita, B.-M. Hodge, S. Lu, H. F. Hamann, Q.
Sun, and B. Lehman, “Ensemble Solar Forecasting Statistical
Quantification and Sensitivity Analysis,” 2015.
S. Lu, Y. Hwang, I. Khabibrakhmanov, F. J. Marianno, X. Shao, J.
Zhang, B.-M. Hodge, and H. F. Hamann, “Machine learning based multiphysical-model blending for enhancing renewable energy forecastimprovement via situation dependent error correction,” in Control
Conference (ECC), 2015 European, 2015, pp. 283–290.
T. Hastie, R. Tibshirani, J. Friedman, and others, The elements of
statistical learning, 2 Edition. Springer-Verlag New York, 2009.
L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32,
2001.
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support
Vector Regression,” in Proceedings of the American Society for
Engineering Management 2016 International Annual Conference, 2016.
Mohamed Abuella (S’14) received his Bachelor of Technology degree from
College of industrial Technology, Misurata, Libya in 2008, and M.S. degree
from Southern Illinois University Carbondale in 2012. He is currently a PhD
student in the Department of Electrical and Computer Engineering at
University of North Carolina at Charlotte. His research interest is in planning
and operations of electrical power systems.
Badrul Chowdhury (S’83–M’87–SM’93) received the B.S. degree from the
Bangladesh University of Engineering and Technology, Dhaka, Bangladesh,
in 1981, and the M.S. and Ph.D. degrees from the Virginia Polytechnic
Institute and State University, Blacksburg, VA, USA, in 1983 and 1987,
respectively, all in electrical engineering. He is currently a Professor with the
Department of Electrical and Computer Engineering with joint appointment
with the Department of Systems Engineering and Engineering Management,
University of North Carolina at Charlotte, Charlotte, NC, USA. His current
research interests include power system modeling, analysis and control, and
renewable and distributed energy resource modeling and integration in smart
grids. Dr. Chowdhury is a Member of Tau Beta Pi and Phi Kappa Phi.
Note: This is a pre-print of the full paper that published in Innovative Smart Grid Technologies, North America Conference, 2017, which can be referenced as below:
M. Abuella and B. Chowdhury, “Random Forest Ensemble of Support Vector Regression for Solar Power Forecasting,” in Proceedings of Innovative Smart Grid
Technologies, North America Conference, 2017.
| 5cs.CE
|
Erdős-Ginzburg-Ziv theorem for finite commutative
arXiv:1309.5588v2 [math.CO] 20 Oct 2013
semigroups
Sukumar Das Adhikaria∗
a Harish-Chandra
b Center
Weidong Gaob†
Guoqing Wangc‡
Research Institute, Chhatnag Road, Jhusi, Allahabad 211 019, India
for Combinatorics, LPMC-TJKLC, Nankai University, Tianjin 300071, P. R. China
c Department
of Mathematics, Tianjin Polytechnic University, Tianjin, 300387, P. R. China
Abstract
Let S be a finite additively written commutative semigroup, and let exp(S) be its exponent which is defined as the least common multiple of all periods of the elements in
S. For every sequence T of elements in S (repetition allowed), let σ(T ) ∈ S denote the
sum of all terms of T . Define the Davenport constant D(S) of S to be the least positive integer d such that every sequence T over S of length at least d contains a proper
subsequence T ′ with σ(T ′ ) = σ(T ), and define E(S) to be the least positive integer ℓ
such that every sequence T over S of length at least ℓ contains a subsequence T ′ with
l |S| m
|T | − |T ′ | = exp(S)
exp(S) and σ(T ′ ) = σ(T ). When S is a finite abelian group, it is well
l |S| m
exp(S) = |S| and E(S) = D(S) + |S| − 1. In this paper we investigate
known that exp(S)
l |S| m
whether E(S) ≤ D(S) + exp(S)
exp(S) − 1 holds true for all finite commutative semigroups S. We provide a positive answer to the question above for some classes of finite
commutative semigroups, including group-free semigroups, elementary semigroups, and
archimedean semigroups with certain constraints.
∗
Email : [email protected]
Email : [email protected]
‡
Corresponding author’s email: [email protected]
†
1
Key Words: Erdős-Ginzburg-Ziv Theorem; Zero-sum; Finite commutative semigroups;
Elementary semigroups; Archimedean semigroups
1 Introduction
Zero-Sum Theory is a rapidly growing subfield of Combinatorial and Additive Number Theory.
The main objects of study are sequences of terms from an abelian group (see [8] for a survey,
or [16] for a recent monograph). Pushed forward by a variety of applications the last years
have seen (at least) three substantially new directions:
(a) The study of weighted zero-sum problems in abelian groups.
(b) The study of zero-sum problems in (not necessarily cancellative) commutative semigroups.
(c) The study of zero-sum problems (product-one problems) in non-commutative groups
(see [4, 9, 10, 15]).
In this paper we shall focus on direction (b). Let G be an additive finite abelian group.
A sequence T of elements in G is called a zero-sum sequence if the sum of all terms of T
equals to zero, the identify element of G. Investigations on zero-sum problems were initiated
by pioneering research on two themes, one of which is the following result obtained in 1961
by P. Erdős, A. Ginzburg and A. Ziv.
Throrem A. [5] (Erdős-Ginzburg-Ziv) Every sequence of 2n − 1 elements in an additive finite
abelian group of order n contains a zero-sum subsequence of length n.
Another starting point is the study on Davenport constant D(G) (named after H. Davenport)
of a finite abelian group G, which is defined as the smallest integer d such that, every sequence
T of d elements in G contains a nonempty zero-sum subsequence. Though attributed to H.
Davenport who proposed the study of this constant in 1965, K. Rogers [20] had first studied it
in 1962 and this reference was somehow missed out by most of the authors in this area.
2
Let G be a finite abelian group. For every integer ℓ with exp(G) | ℓ, let sℓ (G) denote the least
integer d such that every sequence T over G of length |T | ≥ d contains a zero-sum subsequence
of length ℓ. For ℓ = exp(G), we abbreviate s(G) = sexp(G) (G) which is called EGZ constant,
and for ℓ = |G| we abbreviate E(G) = s|G| (G), which is sometimes called Gao-constant (see
[15, page 193]). In 1996, the second author of this paper established a connection between
Erdős-Ginzburg-Ziv Theorem and Davenport constant.
Throrem B. (Gao, [6, 7]) If G is a finite abelian group, then sℓ (G) = D(G) + ℓ − 1 holds for
all positive integer ℓ providing that ℓ ≥ |G| and exp(G) | ℓ.
From Theorem B we know that
E(G) = D(G) + |G| − 1.
(∗)
This formula has stimulated a lot of further research (see [1–3, 11, 14, 16, 18, 25] for example).
Among others, the full Chapter 16 in the recent monograph [16] is devoted to this result.
Indeed, the final Corollary in this chapter, (Corollary 16.1, page 260), provides a far-reaching
generalization of the initial formula (*), which is called the Ψ-weighted Gao Theorem. The
formula E(G) = D(G) + |G| − 1 has also been generalized to some finite groups (not necessarily
commutative, for example see [4] and [9]).
In this paper, we aim to generalize E(G) = D(G) + |G| − 1 to some abstract finite commutative semigroups. To proceed, we shall need some preliminaries. The notations on zero-sum
theory used in this paper are consistent with [8] and notations on commutative semigroups are
consistent with [13]. For sake of completeness, we introduce some necessary ones.
• Throughout this paper, we shall always denote by S a finite commutative semigroup and by
G a finite commutative group. We abbreviate “finite commutative semigroup” into “f.c.s.”.
Let F (S) be the (multiplicatively written) free commutative monoid with basis S. Then
any A ∈ F (S), say A = a1 a2 · . . . · an , is a sequence of elements in the semigroup S. The
identify element of F (S) is denoted by 1 ∈ F (S) (traditionally, the identity element is also
called the empty sequence). For any subset S 0 ⊂ S, let A(S 0 ) denote the subsequence of A
consisting of all the terms from S 0 .
3
The operation of the semigroup S is denoted by “+”. The identity element of S, denoted 0S
(if exists), is the unique element e of S such that e + a = a for every a ∈ S. The zero element
of S, denoted ∞S (if exists), is the unique element z of S such that z + a = z for every a ∈ S.
Let
σ(A) = a1 + a2 + · · · + an
be the sum of all terms in the sequence A. If S has an identity element 0S , we allow A = 1, the
empty sequence and adopt the convention that σ(1) = 0S . Denote
if S has an identity element;
S,
0
S =
S ∪ {0}, if S does not have an identity element.
For any element a ∈ S, the period of a is the least positive integer t such that ra = (r + t)a
for some integer r > 0. We define exp(S) to be the period of S, which is the least common
multiple of all the periods of the elements of S. Let A, B ∈ F (S) be sequences on S. We call
B a proper subsequence of A if B | A and B , A. We say that A is reducible if σ(B) = σ(A) for
some proper subsequence B of A (note that, B is probably the empty sequence 1 if S has the
identity element 0S and σ(A) = 0S ). Otherwise, we call S irreducible.
Definition 1.1. Let S be an additively written commutative semigroup.
1. Let d(S) denote the smallest ℓ ∈ N0 ∪ {∞} with the property:
For any m ∈ N and c1 , . . . , cm ∈ S there exists a subset J ⊂ [1, m] such that |J| ≤ ℓ and
m
X
cj =
X
c j.
j∈J
j=1
2. Let D(S) denote the smallest ℓ ∈ N ∪ {∞} such that every sequence A ∈ F (S) of length
|A| ≥ ℓ is reducible.
3. We call d(S) the small Davenport constant of S, and D(S) the (large) Davenport constant
of S.
The small Davenport constant was introduced in [11, Definition 2.8.12], and the large Davenport constant was first studied in [24]. For convenience of the reader, we state the following
(well known) conclusion.
4
Proposition 1.2. Let S be a finite commutative semigroup. Then,
1. d(S) < ∞.
2. D(S) = d(S) + 1.
Proof. 1. See [11, Proposition 2.8.13].
2. Take an arbitrary sequence A ∈ F (S) of length at least d(S) + 1. By the definition of
d(S), there exists a subsequence A′ of A such that |A′ | ≤ d(S) < |A| and σ(A′ ) = σ(A). This
proves
D(S) ≤ d(S) + 1.
Now it remains to prove d(S) ≤ D(S) − 1. Take an arbitrary sequence T ∈ F (S). Let T ′ be the
minimal (in length) subsequence of T such that σ(T ′ ) = σ(T ). It follows that T ′ is irreducible
and |T ′ | ≤ D(S) − 1. By the arbitrariness of T , we have
d(S) ≤ D(S) − 1.
Definition 1.3. Define E(S) of any f.c.s. S as the smallest positive integer ℓ such that, every
sequence A ∈ F (S) of length ℓ contains a subsequence B with σ(B) = σ(A) and |A|−|B| = κ(S),
where
&
'
|S|
κ(S) =
exp(S).
exp(S)
(1)
Note that if S = G is a finite abelian group, the invariants E(S) and D(S) are consistent with
the classical invariants D(G) and E(G), respectively. We suggest the following generalization
of E(G) = D(G) + |G| − 1.
Conjecture 1.4. For any f.c.s. S,
E(S) ≤ D(S) + κ(S) − 1.
5
We remark that E(S) ≥ D(S) + κ(S) − 1, the converse of the above inequality, holds trivially
when S contains an identity element 0S , i.e., when S is a finite commutative monoid. The
extremal sequence can be obtained by any irreducible sequence T of length D(S) − 1 adjoined
with a sequence of length κ(S) − 1 of all terms equaling 0S . So, Conjecture 1.4, if true, would
imply the following.
Conjecture 1.5. For any finite commutative monoid S,
E(S) = D(S) + κ(S) − 1.
Nevertheless, to make the study more general, the existence of the identity element is not
necessary. We shall verify that Conjecture 1.4 holds for some important f.c.s., including groupfree f.c.s, elementary f.c.s, and archimedean f.c.s with certain constraints.
2 On group-free semigroups
We begin this section with some definitions.
On a commutative semigroup S the Green’s preorder, denoted ≦H , is defined by
a ≦H b ⇔ a = b + t
for some t ∈ S0 . Green’s congruence, denoted H, is a basic relation introduced by Green for
semigroups which is defined by:
a H b ⇔ a ≦H b and b ≦H a.
For an element a of S, let Ha be the congruence class by H containing a. We call a f.c.s. S
group-free, provided that all its subgroups are trivial, equivalently, exp(S) = 1. The group-free
f.c.s is fundamental for Semigroup Theory due to the following property.
Property C. (see [13], Proposition 2.4 of Chapter V) For any f.c.s. S, the quotient semigroup
S H of S by H is group-free.
6
We first show that Conjecture 1.4 holds true for any group-free f.c.s., for which the following lemma will be necessary.
Lemma 2.1. (See [13], Proposition 2.3 of Chapter V) For any group-free f.c.s. S, the Green’s
congruence H is the equality on S.
Now we are ready to give the following.
Theorem 2.2. For any group-free f.c.s. S,
E(S) ≤ D(S) + κ(S) − 1.
Proof. Take an arbitrary sequence T ∈ F (S ) of length D(S) + κ(S) − 1. By the definition of
D(S), there exists a subsequence T ′ of T with |T ′ | ≤ D(S) − 1 and σ(T ′ ) = σ(T ). Let T ′′ be a
subsequence of T containing T ′ , i.e.,
T ′ | T ′′ ,
with length
|T ′′ | = D(S) − 1.
(2)
Note that T ′′ is perhaps equal to T ′ for example when |T ′ | = D(S) − 1. We see that
σ(T ′ ) = σ(T ) ≦H σ(T ′′ ) ≦H σ(T ′ ),
and thus σ(T ′′ ) H σ(T ). By Lemma 2.1, we have
σ(T ′′ ) = σ(T ).
Combined with (2), we have the theorem proved.
Definition 2.3. A commutative nilsemigroup S is a commutative semigroup with a zero element
∞S in which every element x is nilpotent, i.e., nx = ∞S for some n > 0.
Since any finite commutative nilsemigroup is group-free, we have the following immediate
corollary of Theorem 2.2.
Corollary 2.4. Let S be a finite commutative nilsemigroup. Then E(S) ≤ D(S) + κ(S) − 1.
7
In the rest of this paper, we need only to consider the case of exp(S) > 1. Two classes
of important f.c.s., finite elementary semigroups and finite archimedean semigroups, will be
our emphasis as both these semigroups are basic components in two kinds of decompositions
of semigroups, namely, subdirect decompositions and semilattice decompositions, correspondingly. Both decompositions have been the mainstay of Commutative Semigroup Theory for
many years (see [13]).
3 On elementary semigroups
With respect to the subdirect decompositions, Birkhoff in 1944 proved the following.
Theorem D. ([13], Theorem 1.4 of Chapter IV) Every commutative semigroup is a subdirect
product of subdirectly irreducible commutative semigroups.
Hence, we shall give the following result with respect to the subdirect decomposition.
Theorem 3.1. For any subdirectly irreducible f.c.s. S,
E(S) ≤ D(S) + κ(S) − 1.
To prove Theorem 3.1, several preliminaries will be necessary.
Lemma 3.2. [12] Any subdirectly irreducible f.c.s. is either a nilsemigroup, or an abelian
group, or an elementary semigroup. In particular, any f.c.s. is a subdirect product of a commutative nilsemigroup, an abelian group, and several elementary semigroups.
Definition 3.3. A commutative semigroup S is elementary in case it is the disjoint union S =
G ∪ N of a group G and a nilsemigroup N, in which N is an ideal of S, the zero element ∞N of
N is the zero element ∞S of S and the identity element of G is the identity element of S .
Lemma 3.4. ([13], Proposition 3.2 of Chapter IV) On any commutative nilsemigroup N, the
relation PN on N given by a PN b ⇔ ∞ : a = ∞ : b is a congruence on N with {∞N } being a
PN class, where ∞ : a = {x ∈ N 0 : x + a = ∞N }.
8
Lemma 3.5. ([13], Proposition 5.1 of Chapter IV) In an elementary semigroup S = G ∪ N,
the action of every g ∈ G on N permutes every PN -class.
Lemma 3.6. Let N be a finite commutative nilsemigroup, and let a, b be two elements in N. If
a + b = a then a = ∞N .
Proof. It is easy to see that a = a + b = a + 2b = · · · = a + nb = ∞N for some n ∈ N, done.
Lemma 3.7. Let N be a finite commutative nilsemigroup, and let a, b be two elements of N
with a <H b. Then ∞ : b $ ∞ : a.
Proof. The conclusion ∞ : b ⊆ ∞ : a is clear. Hence, we need only to show that ∞ : b , ∞ : a.
If a = ∞N , then ∞ : a = N 0 , N ⊇ ∞ : b, done. Hence, we assume ∞N <H a. It follows that
there exists some minimal element c of N such that
∞N <H c ≦H a.
Then there exists some x ∈ N 0 such that
a + x = c.
Since a <H b, then there exists some y ∈ N such that
b + y = a.
Since c , ∞N , by Lemma 3.6 we have c + y <H c and hence combined with the minimality of
c, we have c + y = ∞N . Therefore, it follows that a + (x + y) = (a + x) + y = c + y = ∞N and
b + (x + y) = (b + y) + x = a + x = c, and thus, x + y ∈ ∞ : a and x + y < ∞ : b, and we are
done.
Proof of Theorem 3.1. By Lemma 3.2, Theorem B and Corollary 2.4, it suffices to consider
the case of S = G ∪ N is an elementary semigroup.
Take an arbitrary sequence T ∈ F (S) of length D(S) + κ(S) − 1. Let
T 1 = T (G)
9
be the subsequence of T consisting of all the terms from G, and let
T 2 = T (N)
be the subsequence of T consisting of all the terms from N. Then T 1 and T 2 are disjoint and
T1 · T2 = T .
If T 2 = 1, the empty sequence, then T = T 1 is a sequence of elements in the subsemigroup
G (group). Since D(S) ≥ D(G) and κ(S) ≥ |S| ≥ |G| is a multiple of exp(G) = exp(S), it
follows from Theorem B that there exists a subsequence T ′ of T with
|T ′ | = κ(S)
and
σ(T ′ ) = 0G = 0S .
Let
T ′′ = T · T ′−1 .
We see that σ(T ′′ ) = σ(T ′′ ) + 0S = σ(T ′′ ) + σ(T ′ ) = σ(T ), and we are done. Hence, we need
only to consider the case that
T 2 , 1.
(3)
Assume σ(T ) = ∞S . Then there exists a subsequence U of T with
σ(U) = σ(T )
and |U| ≤ D(S) − 1. Take a subsequence U ′ of T with
U | U′
and
|U ′ | = D(S) − 1.
We check that σ(U ′ ) = σ(U) + σ(U ′ U −1 ) = ∞S + σ(U ′ U −1 ) = ∞S = σ(T ), and we are done.
Hence, we need only to consider the case that
σ(T ) , ∞S .
10
(4)
Define the subgroup
K = {g ∈ G : g + σ(T 2 ) = σ(T 2 )}
of G, i.e., K is the stabilizer of σ(T 2 ) in G when considering the action of G on N. We claim
that
D(S) ≥ |T 2 | + D(G K).
(5)
Take a sequence W ∈ F (G) ⊂ F (S) such that ϕG K (W) is zero-sum free in the quotient group
G K with
|W| = D(G K) − 1,
where ϕG K denotes the canonical epimorphism of G onto G K. To prove (5), it suffices to
verify that W · T 2 is irreducible in S. Suppose to the contrary that W · T 2 contains a proper
subsequence V with
σ(V) = σ(W · T 2 ).
(6)
Let
V = V1 · V2
with
V1 | W
and
V2 | T 2 .
(7)
By (3), we have σ(W · T 2 ) ∈ N, which implies
V2 , 1.
By Lemma 3.5, we have that σ(V2 ) PN σ(V) and σ(T 2 ) PN σ(W · T 2 ). Combined with (6), we
have that
σ(V2 ) PN σ(T 2 ).
(8)
σ(T 2 ) , ∞N .
(9)
By (4), we have
By (7), we have
σ(T 2 ) ≦H σ(V2 ),
11
where ≦H denotes the Green’s preorder in the nilsemigroup N. Combined with (8) and Lemma
3.7, we derive that σ(T 2 ) H σ(V2 ). It follows from Lemma 2.1 that
σ(T 2 ) = σ(V2 ).
Combined with (7), (9) and Lemma 3.6, we conclude that
V2 = T 2 .
Recalling that V is a proper subsequence of W·T 2 , we have V1 , W. Since ϕG K (W) is zero-sum
free in the group G K, we derive that σ(V1 ) − σ(W) < K, and thus, σ(V) = σ(V1 ) + σ(V2 ) =
σ(V1 ) + σ(T 2 ) , σ(W) + σ(T 2 ) = σ(W · T 2 ), a contradiction with (6). This proves (5).
By (5), we have that
|T 1 | = |T | − |T 2 |
≥ D(S) + κ(S) − 1 − (D(S) − D(G K))
= D(G K) + κ(S) − 1.
Applying Theorem B, we derive that T 1 contains a subsequence T 1′ with
|T 1′ | = κ(S)
such that ϕG K (σ(T 1′ )) = 0G K , i.e.,
σ(T 1′ ) ∈ K.
(10)
Let T 2′ = T · T 1′−1 . Observe T 2 | T 2′ . By (10), we check that
σ(T ) = σ(T 1′ ) + σ(T 2′ )
= σ(T 1′ ) + (σ(T 2 ) + σ(T 2′ · T 2−1 ))
= (σ(T 1′ ) + σ(T 2 )) + σ(T 2′ · T 2−1 )
= σ(T 2 ) + σ(T 2′ · T 2−1 )
= σ(T 2′ ).
This completes the proof of the theorem.
We remark that since the elementary semigroup S = G∪N has an identity element 0S = 0G ,
the equality in the above theorem holds as noted in the introductory section, i.e., E(S) =
D(S) + κ(S ) − 1.
12
4 On archimedean semigroups
In this section, we shall deal with the class of semigroups associated to semilattice decomposition of semigroups. The semilattice decompositions were obtained for f.c.s by Schwarz
[21] and Thierrin [23], and then were extended to all commutative semigroups by Tamura and
Kimura [22], in which they proved the following.
Theorem E. Every commutative semigroup is a semilattice of commutative archimedean semi-
groups.
Definition 4.1. A commutative semigroup S is called archimedean provided that for any two
elements a, b ∈ S, there exist m, n > 0 and x, y ∈ S with ma = b + x and nb = a + y.
To be precise, for any commutative semigroup S there exists a semilattice Y and a partition
S
S = a∈Y Sa into subsemigroups Sa (one for every a ∈ Y) with Sa + Sb ⊆ Sa∧b for all a, b ∈ Y,
and moreover, each component Sa is archimedean. Hence, we shall consider Conjecture 1.4
on archimedean semigroups in what follows. To proceed with it, several preliminaries will be
necessary.
Definition 4.2. We call a commutative semigroup S nilpotent if | |
S + {z
···+S
} | = 1 for some
t
t > 0. For any commutative nilpotent semigroup S, the least such positive integer t is called
the nilpotency index and is denoted by L(S).
Note that, when the commutative semigroup S is finite, S is nilpotent if and only if S is
a nilsemigroup. With respect to finite semigroups, the famous Kleitman-Rothschild-Spencer
conjecture (see [17]) states that, on a statistical basis, almost all finite semigroups are nilpotent
of index at most three, for which there is considerable evidence, but gaps in the original proof
have remained unfilled. For the commutative version of this conjecture, there is also some
evidence.
We need to give some important notions, namely the Rees congruence and the Rees quotient. Let I be an ideal of a commutative semigroup S. The relation J defined by
a J b ⇔ a = b or a, b ∈ I
13
is a congruence on S, the Rees congruence of the ideal I. Let S I denote the quotient semigroup S J, which is called the Rees quotient semigroup of S by I. The Rees congruence
and the resulting Rees quotient semigroup introduced by Rees [19] in 1940 have been among
the basic notions in Semigroup Theory. In some sense, the Rees quotient semigroup is obtained
by squeezing I to a zero element (if I , ∅) and leaving S \ I untouched. Hence, it is not hard
to obtain the following lemma.
Lemma 4.3. For any ideal I of a f.c.s. S,
D(S) ≥ D(S I).
Lemma 4.4. ([13], Chapter III, Proposition 3.1) A commutative semigroup S which contains
an idempotent e (for instance a f.c.s.) is archimedean if and only if it is an ideal extension of
an abelian group G by a commutative nilsemigroup N; then S has a kernel K = He = e + S
and S K is a commutative nilsemigroup.
Lemma 4.5. For any finite commutative nilsemigroup N,
L(N) ≤ D(N) ≤ L(N) + 1.
Proof. By the definition of L(N), there exists a sequence T ∈ F (N) with |T | = L(N) − 1 and
σ(T ) , ∞N . By Lemma 3.6, we have that T is irreducible, which implies D(N) ≥ |T | + 1 =
L(N). On the other hand, since any sequence in F (N) of length L(N) has a sum ∞N , we have
D(N) ≤ L(N) + 1, and we are through.
Now we are in a position to put out our result on archimedean semigroup as follows.
Theorem 4.6. For any finite archimedean semigroup S, E(S) ≤ D(S) + κ(S). Moreover, if the
nilsemigroup S K has a nilpotency index at most three, then E(S) ≤ D(S) + κ(S) − 1, where
K denotes the kernel of S.
Proof. Let e be the unique idempotent of S. By Lemma 4.4, we have that the kernel K = He =
e + S and the Rees quotient semigroup
N = S K
14
(11)
is a nilsemigroup. By Theorem B and Corollary 2.4, we need only to consider the case that
both K and N are nontrivial. We claim that
D(S) ≥ max(D(N), D(K) + 1).
(12)
D(S) ≥ D(N) follows from (11) and Lemma 4.3. Now take a minimal zero-sum sequence U
of elements in the group K (the kernel of S) with length |U| = D(K). Since N is nontrivial,
the semigroup S has no identity element, which implies that U is irreducible in S, and thus,
D(S) ≥ |U| + 1 = D(K) + 1. This proves (12).
Now we take a sequence T ∈ F (S) with |T | = D(S) + κ(S) − ǫ, where ǫ = 0 or ǫ = 1
according to the conclusions to prove in what follows. Since |T | = D(S) + κ(S) − ǫ ≥ D(N) +
κ(G) − ǫ ≥ D(N), it follows from Lemma 4.5 that σ(T ) belongs to the kernel of S, i.e.,
σ(T ) ∈ K.
(13)
Let ψK : S → K be the canonical retraction of S onto K, i.e.,
ψK (a) = e + a
for every a ∈ S. Notice that ψK (T ) is a sequence of elements in the kernel K with length
|ψK (T )| = |T | = D(S) + κ(S) − ǫ ≥ D(K) + κ(S).
Since |κ(S)| ≥ |K| and exp(K) = exp(S) | κ(S), it follows from Theorem B that there exists a
subsequence T ′ of T with
|T ′ | = κ(S)
such that ψK (T ′ ) is a zero-sum sequence in the kernel K, i.e.,
ψK (σ(T ′ )) = σ(ψK (T ′ )) = e.
Now we assert the following.
Claim. If D(S) − ǫ ≥ L(N) then E(S) ≤ D(S) + κ(S) − ǫ.
Since |T T ′−1 | = D(S) − ǫ ≥ L(N), it follows that
σ(T T ′−1 ) ∈ K.
15
(14)
Combined with (13) and (14), we have that
σ(T T ′−1 ) = σ(T T ′−1 ) + e
= σ(T T ′−1 ) + ψK (σ(T ′ ))
= σ(T T ′−1 ) + (e + σ(T ′ ))
= (σ(T T ′−1 ) + σ(T ′ )) + e
= σ(T ) + e
= σ(T ).
Recall |T ′ | = κ(S). This proves the claim.
By (12), Lemma 4.5, and applying the above claim with ǫ = 0, we conclude that
E(S) ≤ D(S) + κ(S).
It remains to show E(S) ≤ D(S) + κ(S) − 1 when S K has a nilpotency index at most three.
Take ǫ = 1. By the above claim, we may assume without loss of generality that D(S) ≤ L(N).
Since K is nontrivial, we have D(K) ≥ 2. Combined with L(N) ≤ 3 and (12) and Lemma 4.5,
we conclude that
D(S) = L(N) = 3,
and D(K) = 2 which implies
K = C2 ,
the group of two elements. Take a subsequence T ′′ of T with length
|T ′′ | ≤ D(S) − 1
and
σ(T ′′ ) = σ(T ).
If |T ′′ | = D(S) − 1, we are done. Now assume |T ′′ | < D(S) − 1, equivalently,
|T T ′′−1 | > κ(S).
By (13), we have
σ(T ′′ ) ∈ K,
16
(15)
and thus, σ(ψK (T T ′′−1 )) = ψK (σ(T T ′′−1 )) = e + σ(T T ′′−1 ) = e. Since exp(S) = exp(K) = 2,
we can find a subsequence U of T T ′′−1 of length exactly κ(S) such that
σ(ψK (U)) = e.
Since T ′′ | T U −1 , it follows from (15) that σ(T U −1 ) ∈ K, and thus,
σ(T U −1 ) = σ(T U −1 ) + e
= σ(T U −1 ) + σ(ψK (U))
= σ(T U −1 ) + ψK (σ(U))
= σ(T U −1 ) + e + σ(U)
= σ(T ) + e
= σ(T ).
This completes the proof of the theorem.
Acknowledgements
This work is supported by NSFC (11301381, 11271207, 11001035), Science and Technology
Development Fund of Tianjin Higher Institutions (20121003). This work was initiated during
the first author visited the Center for Combinatorics of Nankai University in 2010, he would
like to thank the host’s hospitality. In addition, the authors would like to thank the referee for
many valuable comments and suggestions.
References
[1] S.D. Adhikari and Y.G. Chen, Davenport constant with weights and some related questions - II, J. Combinatorial Theory, Ser. A, 115 (2008) 178–184.
[2] S.D. Adhikari, Y.G. Chen, J.B. Friedlander, S.V. Konyagin and F. Pappalardi, Contributions to zero-sum problems, Discrete Math., 306 (2006) 1–10.
[3] S.D. Adhikari and P. Rath, Davenport constant with weights and some related questions,
Integers: Electronic journal of combinatorial number theory, 6 (2006) A30.
17
[4] J. Bass, Improving the Erdős-Ginzburg-Ziv theorem for some non-abelain groups, J.
Number Theory, 126 (2007) 217–236.
[5] P. Erdős, A. Ginzburg and A. Ziv, Theorem in additive number theory, Bull. Res. Council
Israel 10F (1961) 41–43.
[6] W.D. Gao, A combinatorial problem of finite Abelian group, J. Number Theory, 58 (1996)
100–103.
[7] W.D. Gao, On zero-sum subsequences of restricted size-II, Discrete Math., 271 (2003)
51–59.
[8] W.D. Gao and A. Geroldinger, Zero-sum problems in finite abelian groups: a survey,
Expo. Math., 24 (2006) 337–369.
[9] W.D. Gao and Y.L. Li, The Erdős-Ginzburg-Ziv theorem for finite solvable groups, J. Pure
Appl. Algebra, 214 (2010) 898–909.
[10] A. Geroldinger and D.J. Grynkiewicz, The large Davenport constant I: Groups with a
cyclic index 2 subgroup, J. Pure Appl. Algebra, 217 (2013) 863–885.
[11] A. Geroldinger and F. Halter-Koch, Non-Unique Factorizations. Algebraic, Combinatorial and Analytic Theory, Pure and Applied Mathematics, vol. 278, Chapman &
Hall/CRC, 2006.
[12] P.A. Grillet, On subdirectly irreducible commutative semigroups, Pacific J. Math., 69
(1977) 55–71.
[13] P.A. Grillet, Commutative semigroup, Kluwer Academic Publishers, (2001).
[14] D.J. Grynkiewicz, L.E. Marchan and O. Ordaz, A weighted generalization of two theorems of Gao, Ramanujan J., 28 no. 3, (2012) 323–340.
[15] D.J. Grynkiewicz, The large Davenport constant II: General upper bounds, J. Pure Appl.
Algebra, 217 (2013) 2221–2246.
18
[16] D.J. Grynkiewicz, Structural Additive Theory, Developments in Mathematics, Springer,
2013.
[17] D.J. Kleitman, B.R. Rothshild and J.H. Spencer, The number of semigroups of order n,
Proc. Amer. Math. Soc., 55 (1976) 227–232.
[18] F. Luca, A generalization of a classical zero-sum problem, Discrete Math., 307 (2007)
1672–1678.
[19] D. Rees, On semi-groups, Proc. Cambridge Phil. Soc., 36 (1940) 387–400.
[20] K. Rogers, A Combinatorial problem in Abelian groups, Proc. Cambridge Phil. Soc., 59
(1963) 559–562.
[21] Š . Schwarz, Contribution to the theory of torsion semigroups[Russian], Czechoslovak
Math. J., 3 (1953) 7–21.
[22] T. Tamura and N. Kimura, On decompositions of a commutative semigroup, Kodai Math.
Sem. Rep., 6 (1954) 109–112.
[23] G. Thierrin, Sur quelques properiétés de certaines classes de demi-groupes, C.R. Acad.
Sci. Paris, 239 (1954) 1335–1337.
[24] G.Q. Wang and W.D. Gao, Davenport constant for semigroups, Semigroup Forum, 76
(2008) 234–238.
[25] P.Z. Yuan and X.N. Zeng, Davenport constant with weights, European J. Combin., 31
(2010) 677–680.
19
| 0math.AC
|
94
COOPER ET AL.
UAI 2004
Bayesian Biosurveillance of Disease Outbreaks
Gregory F. Cooper
RODS Laboratory
Center for Biomedical Informatics
University of Pittsburgh
[email protected]
Denver H. Dash
Intel Research
Santa Clara
[email protected]
Abstract
Early, reliable detection of disease outbreaks is a
critical problem today. This paper reports an
investigation of the use of causal Bayesian
networks to model spatio-temporal patterns of a
non-contagious disease (respiratory anthrax
infection) in a population of people. The number
of parameters in such a network can become
enormous, if not carefully managed. Also,
inference needs to be performed in real time as
population data stream in. We describe techniques
we have applied to address both the modeling and
inference challenges. A key contribution of this
paper is the explication of assumptions and
techniques that are sufficient to allow the scaling
of Bayesian network modeling and inference to
millions of nodes for real-time surveillance
applications. The results reported here provide a
proof-of-concept that Bayesian networks can
serve as the foundation of a system that
effectively performs Bayesian biosurveillance of
disease outbreaks.
1 INTRODUCTION
Early, reliable detection of outbreaks of disease, whether
natural (e.g., West Nile virus and SARS) or bioterroristinduced (e.g., anthrax and smallpox), is a critically
important problem today. We need to detect outbreaks as
early as possible in order to provide the best response and
treatment, as well as improve the chances of identifying the
source, whether natural or bioterroristic. An analysis of one
bioagent release scenario estimated that as many as 30,000
people per day could die. The induced long-term economic
costs were estimated to be as high as 250 million dollars
per hour of the outbreak (Kaufmann 1997, Wagner 2001).
Early detection could dramatically reduce these losses.
Outbreaks often present signals that are weak and noisy
early in the event. If we hope to achieve rapid and reliable
detection, it likely will be necessary to integrate multiple
weak signals that together provide a relatively stronger
signal of an outbreak. Combining spatial and temporal data
is an important instance of such integration. For example,
John D. Levander, Weng-Keen Wong,
William R. Hogan, Michael M. Wagner
RODS Laboratory
Center for Biomedical Informatics
University of Pittsburgh
even though the number of patients in a given city with
fever, who were seen in emergency departments in the past
24 hours, may not be noticeably higher than average,
nonetheless, for the past 12 hours it may be significantly
higher for a given neighborhood of the city.
Because of the noise in signals early in the event, early
detection is almost always detection under uncertainty. In
the research reported here, we use probability as a measure
of uncertainty. A well-organized probabilistic approach
allows for the rational combination of multiple, small
indicators into a big-picture. Since the modeling of risk
factors, diseases, and symptoms often is causal, we use
causal Bayesian networks as our probabilistic modeling
method. Bayesian networks comprise an established,
unifying framework that is already recognized in the field
of epidemiology (Greenland 2000) as a promising
approach to epidemiological modeling, although to our
knowledge there are no reports in the literature of Bayesian
networks that have been applied to perform Bayesian
biosurveillance.
This paper describes an approach in which a causal
Bayesian network is used to model an entire population of
people. We concentrate on modeling non-contagious
outbreak diseases, such as airborne anthrax or West Nile
encephalitis that is transmitted by mosquitoes. Modeling an
entire population of people in just one city-wide area leads
to a Bayesian network model with millions of nodes. For
example, the model reported here contains approximately
20 million nodes. Each individual in the population is
represented by a 14-node subnetwork, which captures
important syndromic information that is commonly
available for health surveillance (such as emergency
department chief complaints), while avoiding any
information that could personally identify the individual
(e.g., name, social security number, and home street
address).
Given current data about individuals in the population, we
use a Bayesian network to infer the posterior probabilities
of outbreak diseases in the population. To provide timely
detection, inference needs to be performed in real time,
such that the biosurveillance system “keeps up” with the
data streaming in. Once the probability of an outbreak
exceeds a particular threshold, an alert is generated by the
UAI 2004
COOPER ET AL.
Bayesian-network-based biosurveillance system; this alert
can serve to warn public health officials
Using such a large Bayesian network presents both
modeling and inference challenges. To help make
modeling more tractable in terms of computational space,
we use the following approach: if some groups of people
are indistinguishable, according to the data being captured,
we model them with a single subpopulation subnetwork.
To speed up inference, we use a method that need only
update the network state based on new information about
an individual in the population (such as newly available
clinical information, based on the person visiting an
emergency department in seeking care).
A key contribution of this paper is the explication of
assumptions and techniques that are sufficient to allow the
scaling of Bayesian network modeling and inference to
millions of nodes for real-time surveillance applications,
thus providing a proof-of-concept that Bayesian networks
can serve as the foundation of a system that effectively
performs Bayesian biosurveillance of disease outbreaks.
With this foundation in place, many extensions are
possible, and we outline several of them in the final section
of the paper.
In remainder of this paper, we first outline our general
approach for using causal Bayesian networks to represent
non-contagious diseases that can cause outbreaks of
disease. Next, we introduce the specific network we have
constructed to monitor for an outbreak caused by the
outdoor release of anthrax spores. We then describe an
experiment that involves injecting simulated cases of
patients with anthrax (which were generated from a
separate model) onto background data of real cases of
patients who visited emergency departments during a
period when there were no known outbreaks of disease
occurring. We measure how long it takes the Bayesian
network system to detect such simulated outbreaks.
Finally, we discuss these results and suggest directions for
future research.
2 METHODOLOGY
In this section, we present a methodology that is sufficient
to allow explicit modeling of a large population of
individuals in a real-time setting. In Section 2.1 we detail
the modeling assumptions that we use, and in Section 2.2
we show how those assumptions can be exploited to
perform fast real-time inference.
2.1 MODELING
Our methodology uses Bayesian networks (BNs) to
explicitly model an entire population of individuals. Since
in this paper we are specifically interested in disease
outbreak detection from syndromic information, we will
refer to models of these individuals as person models,
although obviously the same ideas could be applied to
model other entities that might provide information about
95
disease outbreaks, such as biosensors and livestock. From
an object-oriented perspective, each person model can be
viewed as a class; using a class to represent a particular
person creates an object (Koller 1997).
We explicitly model each person in the population, and
thus in our BN there will exist (at least conceptually) an
object Pi for each person. An example of a complete model
for four people is shown in Figure 1, where each person in
the population is represented with a simple six-node
network structure. In this particular example, there is only
one person model (class), but our methodology can allow
for more.
In this paper, we restrict our methodology to model noncontagious diseases. We partition all the nodes X in the
network into three parts:
1.
2.
3.
A set of global nodes G,
A set of interface nodes, I, and
A set of person subnetworks P = {P1,P2,…,Pn}.
The set G, defined as G = X\{IP}, contains nodes that
represent global features common to all people. For the
example in Figure 1, G consists of two nodes: Terror Alert
Level (having states Green, White, Yellow, Orange, and
Red), and Anthrax Release (having states Yes and No). Set
I contains factors that directly influence the status of the
outbreak of disease in people in the population. Each Pi
subnetwork (object) represents a person in the population.
Structurally, we make the following two assumptions.
Assumption 1: The interface nodes, I, d-separate the
person subnetworks from each other, and any arc between
a node I in I and a node X in some person subnetwork Pi is
oriented from I to X.
Thus, we do not allow arcs between the person models.
Assumption 2: The interface nodes, I, d-separate the nodes
in G from the nodes in P, and any arc between a node G in
G and a node I in I is oriented from G to I.
Figure 2 presents the above two assumptions in
diagramatic form.
For non-contagious diseases that may cause outbreaks,
Assumptions 1 and 2 are reasonable when I contains all of
the factors that significantly influence the status of an
outbreak disease in individuals in the population. In the
case of bioterrorist-released bioagents, for example, such
information includes the time and place of release of the
agent. Key characteristics of nodes in I are that they have
arcs to the nodes in one or more person models, and they
induce the conditional independence relationships
described in Assumptions 1 and 2. Often the variables in I
will be unmeasured. It is legitimate, however, to have
measured variables in I. For example, the regional smog
level (not shown in Figure 1) might be a measured variable
that influences the disease status of people in the
population, and thus it would be located in I.
96
COOPER ET AL.
UAI 2004
G
P4
I
P1
P2
P3
Figure 1: A simplified four-person model for detecting an outbreak of anthrax. Each person Pi in the population is
represented explicitly by a six-node subnetwork. Observed variables are marked with a ground symbol.
P1
G
I
Pn
Figure 2: The closed regions represent Bayesian
subnetworks. The circles on the edges of the
subnetworks denote nodes that are connected by arcs
that bridge the subnetworks. Only two such “I/O”
nodes are shown per subnetwork, but in general there
could be any number. The arrows between
subnetworks show the direction in which the
Bayesian-network arcs are oriented between the
subnetworks. The braces show which nodes can
(possibly) be connected by arcs. In subnetwork I, the
I/O nodes on the left and those on the right are not
necessarily distinct.
Let T be a variable in G that represents a disease outbreak.
In Figure 1, T is the node Anthrax Release. The goal of our
biosurveillance method is to continually derive an updated
posterior probability over the states of T as data about the
population streams in.
We consider spatio-temporal data in deriving the posterior
probability of T. For example, we consider information
about when patient cases appear at the emergency
department (ED), as well as the home location (at the level
of zip codes) of those patients. In our current
implementation, spatio-temporal information is explicitly
represented by nodes in the network, such as Location of
Release, Time of Release, and patient Home Zip. We note
that in Figure 1, the Disease Status nodes contain values
that indicate when the disease started (if ever) and ended.
This temporal representation has the advantage (over, for
example, dynamic Bayesian networks (DBNs)) of being
relatively compact. The method allows us to create a
network with fewer parameters than the corresponding
DBN, and simplifies our method for performing real-time
inference.
2.2 INFERENCE
When performing inference for biosurveillance, our goal is
to continuously monitor target variable T by deriving its
updated posterior probability as new data arrives. At any
given time, there are two general sources of evidence we
consider:
1.
General evidence at
g = { G = g: G G}, and
the
global
level:
2.
The collective set of evidence e that we observe
from the population of people: e = {X = e: X Pi ,
Pi P}.
In our application, g might consist of the observation that
the Terror Alert Level = Orange, and e might include
information about the patients who have visited EDs in the
region in recent days, as well as demographic information
(e.g., age, gender, and home zip code) for the people in the
region who have not recently visited the ED.
Given e and g, our goal is to calculate the following:
P (T | e, g ) k P (e | T , g ) P (T | g ) ,
(1)
UAI 2004
COOPER ET AL.
where the proportionality constant is
k 1
P (e | T , g ) P (T | g ) .
T
Since T and g are in G, it follows from Assumptions 1 and
2 that the term P(T | g) in Equation 1 can be calculated
using Bayesian network inference on just the portion of the
model that includes G. Performing BN inference over just
the nodes in G is much preferable to inference over all the
nodes in X, because in the model we evaluated the number
of nodes in X is approximately 107.
The term P(e | T, g) in Equation 1 can be derived as
follows:
P (e | T , g )
P(e | I i ) P(I i | T , g),
probability table for all configurations of I and for all
possible states of evidence ei. Such a table would be
infeasibly large. As described in the next two sections, we
use two techniques to deal with the large size of the
inference problem: Equivalence Classes and Incremental
Updating. Using Equivalence Classes saves both space and
reduces inference time. Using Incremental Updating also
reduces inference time, often dramatically so.
2.2.1 Equivalence Classes
If some person subnetworks are identical in structure and
parameters, and they are instantiated to the same evidence,
then fewer calls to the inference engine are needed.
Equation 2 can be written as:
i
because by Assumption 2 the set I renders the nodes in P
(including e) independent from the nodes in G (including T
and g). The above summation can be very demanding
computationally, because e usually contains many nodes;
therefore, we next discuss its computation in greater detail.
We first show an example from Figure 1. Here we are
modeling exactly four people in the population. The two on
the left have identical attributes, as do the two on the right.
We want to calculate the probability of this configuration
of evidence, given the interface nodes. For this example,
we have two distinct sets of evidence, e1 = {Home
Zip=15213,
Age=20-30,
Gender=M,
Date
Admitted=never1, Respiratory symptoms=unknown} and e2
= {Home Zip=15260, Age=20-30, Gender=F, Date
Admitted=today, Respiratory symptoms=yes}. We need to
calculate:
P (e | I ) P ( P1 e1 , P2 e1 , P3 e 2 , P4 e 2 | I ) .
By Assumption 1, I d-separates each person model from
each other, so this equation can be factored as follows:
P (e | I )
( 2)
P( P1 e1 | I ) P ( P2 e1 | I ) P( P3 e 2 | I )
P ( P4 e 2 | I ).
It follows from Assumptions 1 and 2 that we can derive
each quantity P( Pi e j | I i ) via BN inference using just
the model fragment defined over the set of nodes in Pi I.
However, this quantity must be calculated for all
configurations I = i of the interface nodes. Performing this
calculation for each of millions of person models would be
infeasible within the time limits required for real-time
biosurveillance. We could cache these conditional
probability tables so that at run-time they amount to a
constant-time table lookup. This technique is problematic,
however, because it requires caching of a conditional
1
“Never” means the person is in the population at large and has not
recently been admitted to the ED.
97
P (e | I ) P ( P1 e1 | I ) 2 P ( P3 e 2 | I ) 2 .
We define an equivalence class
Qi
j
as a pair
j
Qi Pi , e j , where Pi is a person model and ej is a
(possibly incomplete) set of evidence over the variables in
Pi.2 A given evidence state e for the entire population
corresponds to a unique set of equivalence classes and
the instance count of each class. Using this set, the general
expression for the quantity P(e | I) is as follows:
P( e | I )
P( E j e j | I )
N ij
,
(3)
j
Qi
j
where Nij is the instance count of equivalence class Qi ,
that is, it is the number of people for whom we model with
person model Pi and for which the evidence is Ej = ej.
If the person model is relatively simple, then there could be
many fewer equivalence classes than there are members in
the population. For our example, since all person models
are identical, the number of equivalence classes is equal to
the number of possible ways to instantiate the variables of
the person model (including not observing the state). In
this case, we would have at most (101 zip codes) (2
genders) (10 age ranges) (4 relative dates of admission)
(3 respiratory states) = 24,240 states. In practice, the
actual number of equivalence classes present at any one
time would probably be smaller, since rarer equivalence
classes may not appear.
In previous work, object-oriented Bayesian networks
(Koller 1997) and related work (Srinivas 1994; Xiang
1999) have been used to improve the efficiency of BN
inference. The method we have described in this section
takes advantage of those computational savings, as well as
the savings that accrue from performing inference only
once for objects of a given class that share the same
evidence.
2
We abuse notation somewhat by using Pi here to denote a class, whereas
previously it has been used to denote an object.
98
COOPER ET AL.
2.2.2 Incremental Updating
When we apply this technique to a population of millions
of people, calculating P(e | I) will be a time-consuming
task, even with the savings that results from using
equivalence classes. Since we would like to perform this
calculation very frequently (e.g., every hour as ED patients
are coming into EDs throughout the modeled region), it is
important to avoid re-calculating P(e | I) for the entire
population every hour. To do so, we use the fact that as a
single person moves from one equivalence class Qi j to
another Qik , P(e | I) can be updated to P(e | I) as follows:
P (e' | I ) P (e | I )
P (Qik | I )
(4)
j
P(Qi | I )
When biosurveillance monitoring is begun, the set of
evidence e represents background information about the
population. Currently, as background information, we use
U.S. Census information to provide the age, gender, and
home zip code for the people in the region being monitored
for disease outbreaks.
UAI 2004
for use in detecting disease outbreaks that would result
from outside, airborne release of anthrax spores. The
model plus the inference algorithms constitute a Bayesian
biosurveillance system that we call PANDA (Populationwide ANomaly Detection and Assessment). We conclude
this section with a description of the method we used to
evaluate PANDA, as well as the results of that evaluation.
3.1 MODEL FOR OUTBREAK DETECTION
In our empirical tests we use a model similar to the
example model shown in Figure 1, with two primary
differences: (1) we do not use the Terror Alert Level node,
and (2) we use a more complex person model. Figure 3
shows the person model we use. The meanings of the
nodes are as follows: 3
Time of Release: This is the day that anthrax was
released, if ever. It has the states never, today,
yesterday, and day before yesterday.
Location of Release: This is the location at which the
anthrax was released, if released anywhere. It has the
states: nowhere, and one state for each of about 100
zip codes being covered by the model. In the current
model, we assume only a single point of release.
Home Zip: This node represents the location of the
person’s home zip code; it can take on one of about
100 zip codes in Allegheny county, Pennsylvania,
which is the region being modeled. There is currently
a “catch-all” zip code called other that represents
patients who do not live in Allegheny county, but who
are seen in EDs there.
Age Decile: This node represents the individual’s age,
which can take one of 9 values: 0, 1…8 corresponding
to (0-10 years), (10-20 years), …, (>80 years),
respectively.
Gender=F, Date Admitted=never, Respiratory symptoms
=unknown}.
Applying the incremental updating rule allows us to reduce
the number of updates that need to be processed each hour
to dozens (= rate of patient visits to all the EDs in the
region) rather than the millions (= the number of people in
the regional population).
Gender: This represents the gender of the individual,
taking values female and male.
Anthrax Infection: This node represents whether or not
the individual has been infected with a respiratory
anthrax infection within the past 72 hours. This node
takes states: AAA (indicating that anthrax was absent
for the past 3 days), AAI (indicating that within the
past 24 hours the patient was infected with anthrax),
AII (indicating that the patient was infected with
anthrax between 24 and 48 hours ago and is still
infected today), and finally, III (indicating that the
patient was infected between 48 and 72 hours ago and
continues to be infected today). There are in principle
4 other states that this node could have (IAA, IIA, IAI,
After P(e | I) is calculated once for the entire population,
we can apply Equation 4 and update this quantity
incrementally as we observe people enter the ED in realtime. As we observe a person from equivalence class
j
Qik enter the ED, we find the class Qi that this person
must have originated from in the background population.
For example, if we observe a patient in the ED with the
following attributes: Qik ={Home Zip=15260, Age=20-30,
Gender=F, Date Admitted=today, Respiratory symptoms
=yes}, then we know that she originated from the
j
background class Qi = {Home Zip=15260, Age=20-30,
By caching equivalence-classes and applying incremental
updating, we can process an hour’s worth of ED patient
cases (about 26 cases) from a region of 1.4 million people
in only 11 seconds using a standard Pentium III PC and the
Hugin BN inference engine v6.2 (Hugin 2004). Thus, there
is enough computing reserve to “keep ahead” of the real
time data, even when in the future we extend our model to
be considerably richer in detail, and we widen the
geographic region being monitored for a disease outbreak.
3 EMPIRICAL EVALUATION
This section describes the detection model that we
evaluated. The model represents a preliminary prototype
3
For each variable that is underlined, its conditional probability table was
estimated from a training set consisting of one year’s worth of ED patient
data from the year 2000. The variables in bold were estimated from U.S.
Census data about the region. The remaining variables had their
respective probabilities (whether prior or conditional) assessed
subjectively; these assessments were informed by the literature and by
general knowledge about infectious diseases.
UAI 2004
COOPER ET AL.
99
ED Admission Due to Other
ED Admission Due to Anthrax
Figure 3: The person model used in the evaluation. We used Hugin software (Hugin 2004)
to implement and display this model.
and AIA), however, we make the assumption that once
a person gets anthrax, he or she maintains the disease
for at least 3 days, so these other states have
probability 0. In a future work, we plan to extend the
Anthrax Infection variable (as well as other temporal
variables described here) to model over more than
three days.
Other ED Disease: This variable is conceptually
similar to Anthrax Infection, but it denotes instead
some other disease or disorder, which by definition is
sufficient to cause the individual to go into the ED, but
is not anthrax. This node has the same type of states as
Anthrax Infection.
Respiratory from Anthrax: Indicates that the individual
is showing respiratory symptoms (e.g., cough) due to
anthrax. It has states similar to those of Anthrax
Infection.
Respiratory from Other: Respiratory symptoms from
ED disease other than anthrax.
Respiratory Symptoms: Node indicating whether or
not the patient exhibits respiratory symptoms. It is a
“logical OR” function of Respiratory from Anthrax
and Respiratory from Other.
Respiratory When Admitted: This node represents
whether the person has respiratory symptoms at the
present time. If the person has been admitted to the ED
today, then we typically know the answer, otherwise
we do not. This node has states True, False, and
Unknown.
ED Admission Due to Anthrax: Indicates that the
person was admitted to the ED due an anthrax
infection.
ED Admission Due to Other: Indicates that the person
was admitted to the ED due to a disease other than
anthrax.
ED Admission: Indicates the day (if any) that the
person was admitted to the ED within the past 72
hours. It is a “logical OR” function of ED Admission
Due to Anthrax and ED Admission Due to Other. We
currently do not model the possibility that a person
could be admitted more than once. To do so, the deidentified data that we receive on each patient could be
extended to include a unique integer for the patient
that does not reveal the patient’s personal identity.
We emphasize that the current model is an initial
prototype, which we intend to refine further.
3.2 SIMULATION MODEL
We evaluated the performance of PANDA on data sets
produced by injecting simulated ED cases into a
background of actual ED cases obtained from several
hospitals in Allegheny county. In accordance with HIPAA
regulations, all personal identifying information was
removed from these actual ED cases. The simulated cases
of anthrax were produced by a simulator (Hogan 2004) that
models the effects of an airborne anthrax release using an
independently developed Gaussian plume model of
atmospheric dispersion of anthrax spores (Hanna 1982).
Given weather conditions and parameters for the location,
height, and amount of the airborne anthrax release, the
Gaussian plume model derives the concentration of anthrax
spores that are estimated to exist in each zip code, which in
turn determines the severity of the outbreak for that zip
code. The output from the simulator consists of a list of
anthrax cases, where each case consists of a date-time field
and a zip code. The full details of the model are in (Hogan
2004). For our experiments, we selected historical
meteorological conditions (e.g., wind direction and speed)
for Allegheny county from a random date as the
meteorological input to the simulator. The height of the
COOPER ET AL.
simulated release is sampled from a prior distribution,
created using expert judgment (Hogan 2004). This
distribution is skewed towards heights less than 1500 feet.
Finally, the release locations are sampled from a prior
distribution which favors release locations that would
affect large numbers of people given the current
meteorological conditions (Hogan 2004).
The output of the simulator cannot be used directly by
PANDA because a full evidence vector for a case includes
information about the patient’s age and gender. As a result,
we took the partially complete patient cases produced by
the simulator and probabilistically assigned the age and
gender fields using the person-model Bayesian network.
The age of the patient is sampled from the conditional
distribution of age given the home zip of the patient and
given the fact that the patient had respiratory symptoms
when admitted. We use a similar procedure for
determining the gender.
The anthrax-release simulator that we used generally
generates multiple down-wind cases of anthrax that span
several zip codes. The simulator also includes a minimum
incubation period of 24 hours after the release during
which no cases of anthrax are generated. Beyond that
minimum period, the incubation period varies, with greater
airborne concentrations of anthrax leading to a shorter
incubation period, in general, than lesser concentrations.
In order to evaluate the detection capability of PANDA, we
generated data sets corresponding to simulated releases of
anthrax of the following amounts: 1.0, 0.5, 0.25 and
0.125.4 For each release amount, we create 96 data sets,
each with a unique release location. For each month in
2002, we choose 8 random release dates and times to use
with the simulator, thus producing a total of 96 different
anthrax-release data sets. We used only 91 data sets for the
0.125 concentration because five of the data sets generated
had no reported anthrax cases. PANDA was applied to
monitor the data from each data set, starting on midnight of
January 4, 2001 and extending through to six days after the
simulated anthrax release occurred.
We measured the performance of PANDA using an
AMOC curve (Fawcett 1999), which plots time to
detection as a function of false alarms per week. The points
on the AMOC curve are generated by determining the false
positive rate (0 to 1) and detection time of the algorithm
over a range of alarm thresholds, where an alarm is
considered to be raised if the posterior probability of an
Anthrax Release exceeds the given alarm threshold. Since
no known releases of anthrax have ever occurred in the
region under study, the false positive rate was measured by
determining the fraction of monitored hours that the release
probability exceeded the alarm threshold under
consideration for the period starting on January 4, 2001
and continuing until 24 hours after the simulated release
4
The units of concentration are not reported here in order to avoid
providing results that could pose some degree of security risk.
UAI 2004
date for the particular data set. In order to measure the
timeliness of detection, we counted the number of hours
that passed between the time of the simulated anthrax
release and the time that the first anthrax-outbreak
posterior probability (produced by PANDA) exceeded the
alarm threshold. If no alarms are raised after the simulated
release point, the detection time is set to be 144 hours.
3.3 RESULTS
Figure 4 illustrates the AMOC curve for PANDA over the
four anthrax concentrations. Since the incubation period of
the simulation is set at a minimum of 24 hours, the earliest
possible detection time is shown with a dotted line at the
24 hour mark. As expected, the detection time decreases as
the simulated release amount increases, since a larger
release is more easily detected. In particular, at zero false
positives, the detection time is approximately 132, 84, 58,
and 46 hours for respective simulated release
concentrations of 0.125, 0.25, 0.5 and 1.0. The maximum
width of the 95% confidence intervals for the detection
times at concentrations of 0.125, 0.25, 0.5 and 1.0 are +/5.21, 4.00, 2.67 and 1.68 hours, respectively.
AMOC Curve
140
Detection Time (Hours)
100
120
Dosage 1.0
100
Dosage 0.5
80
Dosage 0.25
60
Dosage 0.125
40
Optimal
20
0
0
0.2
0.4
0.6
0.8
False Positive Rate
Figure 4: An AMOC curve showing the detection
capabilities of PANDA over different anthrax
concentrations.
The majority of false positives with release probabilities
over 50% occurred during a 10-hour period from January
20, 2002, 11:00 pm to January 21, 2002, 9:00 am and also
during a 17-hour period from midnight August 18, 2002 to
August 19, 2002, 5:00 pm.
We note that the model parameters were based in part on
data from the year 2000, whereas the evaluation was based
on using test data from 2002. So, some false positives may
have been due to a lack of synchronization between the
model and the test data. When tested on data from the year
2001, there were no false alarms above the 50% level.
3.4 INCORPORATING THE SPATIAL
DISTRIBUTION OF CASES INTO THE MODEL
In this section, we describe a change in the PANDA model
to account for the situation in which an anthrax release
UAI 2004
COOPER ET AL.
101
infects people in more than one zip code. In particular, we
added a new interface node, called angle of release, which
describes the orientation of the airborne anthrax release
and takes on the eight possible values of N, NE, E, SE, S,
SW, W, or NW, as shown in Figure 5. For computational
reasons, we also decomposed the previous anthrax
infection node into an exposed to anthrax node and a new
anthrax infection node. Figure 6 depicts the modified
person model. We will refer to this modified version of
PANDA as the spatial model and the previous version will
be referred to as the non-spatial model.
consider them to be potentially exposed to anthrax if their
home zip code is within a rectangular region that originates
at the centroid of the hypothesized release zip code and is
rotated according to the angle of the release variable. As an
example, suppose the release occurs in 15213. The two
dots in Figure 5 represent zip code centroids. There are 8
rectangular regions centered at the centroid of zip code
15213. If a person has a home zip in 15132, and the angle
of release is SE, then we would consider that person to be
potentially exposed to anthrax. The actual probability of
being exposed to anthrax is computed by decaying the
value 1.0 by a half for every 3 miles of distance between
the release zip code’s centroid and the person’s home zip
code centroid. The distance of 3 miles was obtained by
tuning the model over data sets produced by the simulator;
these datasets were distinct from the ones we used to
evaluate PANDA. The width of the rectangle is set to be
approximately 3 miles, which was chosen by calculating
the average area per zip code in Allegheny county,
determining the diameter of a circle with this average area,
and then assigning that diameter as the width. The length
of the rectangle is assumed to extend to infinity, as shown
by the arrows in Figure 5.
Figure 5: The rotating rectangular regions centered at (for
example) the centroid of zip code 15213 that are used to
determine if a person is exposed to anthrax.
We evaluated the spatial model over the 96 simulated data
sets for the 1.0 concentration that were previously used.
The false positive rate was measured over the period of
January 1, 2002 until 24 hours after the start of the
simulated release for the particular data set. The results are
shown in Figure 7, and they indicate that the spatial model
improves the detection time significantly. The largest
difference in detection time is approximately 9.6 hours.
The maximum widths of the 95% confidence intervals for
the spatial and non-spatial model results are +/- 1.68 and
1.64 hours, respectively.
The exposed to anthrax node represents the probability that
the person is exposed to anthrax during a release, given the
zip code of the home location of the person, the location of
the release, and the angle of the release. The spatial model
assigns a probability of 1.0 for exposed to anthrax to
anyone who has the same home zip code as the zip code of
the hypothesized anthrax release, regardless of the angle of
release. For people outside of the release zip code, we
Figure 6: The person model modified to incorporate spatial information
102
COOPER ET AL.
Detection Time (Hours)
AMOC Curve for Spatial versus Non-Spatial Model
50
45
40
35
30
25
20
15
10
5
0
Spatial Model
Non-spatial Model
0
0.2
0.4
0.6
0.8
False Positive Rate
Figure 7: An AMOC curve comparing the spatial model
results with those from the non-spatial model
4 RELATED RESEARCH
This section provides a representative sample of the
spectrum of biosurveillance approaches that have been
reported in the literature. For a more comprehensive and
detailed survey, see (Moore 2003) and (Wong, 2004). The
most studied and applied techniques are univariate
detection algorithms based on time-series models or
regression. These algorithms include the Serfling method
(Serfling 1963, Tsui 2001), ARIMA model (Hamilton
1994, Reis 2003), univariate hidden Markov (HMM)
models (Rabiner 1989), the Kalman filter (Hamilton
1994), and change-point statistics (Carlstein 1988).
Another large group of detection algorithms are based on
the field of statistical quality control, including techniques
such as CuSum (Hutwagner, 2003) and EWMA
(Williamson 1999). All of these algorithms monitor a
single variable, such as the rate of patient visits to
emergency departments, looking for values of the variable
that are significantly abnormal. The time-series algorithms
differ from each other in how they define and detect what
is abnormal.
Methods that monitor the spatial dimension are rarer. The
most prominent method is the Spatial Scan algorithm
(Kulldorf 1997), which searches over a region, looking for
subregions that appear abnormal along some single
dimension (e.g., disease counts), relative to the remaining
regions. Recent work has improved the speed of the Spatial
Scan method using a multi-resolution algorithm (Neill
2003) as well as generalized it to include a time dimension
(Kulldorff 2001).
WSARE (Wong 2003) and BCD (Buckeridge 2004) are
two multivariate methods that take as input both spatial
data (e.g., patient zip codes) and temporal data (e.g., the
time at which patients visit the emergency department), as
well as patient features, such as age, gender, and
symptoms. WSARE uses rules to represent anomalies, and
it searches over the rule space in an efficient and
statistically sound manner. BCD monitors in a frequentist
manner whether a Bayesian network learned from past data
(during a "safe" training period) appears to have a
UAI 2004
distribution that differs from the distribution of more recent
data. If so, an anomaly may have occurred.
All of the approaches mentioned above use frequentist
statistics -- none are Bayesian. The current paper is novel
in introducing, implementing, and evaluating a spatiotemporal,
multivariate
Bayesian
approach
to
biosurveillance.
5 SUMMARY AND FUTURE WORK
This paper introduced a biosurveillance method that uses
causal Bayesian networks to model non-contagious
diseases in a population. By making two independence
assumptions in the model, both of which appear plausible,
and by performing inference using equivalence classes and
incremental updating, it is possible to achieve tractable
Bayesian biosurveillance in a region with 1.4 million
people. We implemented and evaluated an outbreak
detection system called PANDA. Overall, the run time
results and the detection performance of this initial
evaluation are encouraging, although additional studies are
needed and are in process.
There are several straightforward extensions to PANDA
that we plan to implement in the near term, including (1)
increasing the number of days being modeled, (2)
modeling on an hourly basis, rather than a daily one, and
(3) adding nodes for prevailing wind direction and wind
speed to the model. We also plan to incorporate into the
model a set of variables that represent the amount of over
the counter (OTC) medication sales of a particular type
(e.g., cough medication sales) per subregion (e.g., zip
code) per day. In future work, we will extend the BN
model to represent additional non-contagious outbreak
diseases, as well as non-outbreak diseases that might be
easily confused with outbreak diseases. We also intend to
investigate models of contagious diseases, which will be
more complex, because there is much less independence in
these models. Developing inference algorithms that are fast
enough to permit real-time biosurveillance of contagious
diseases will be challenging. Thus, we expect eventually to
need to use approximate inference algorithms. Finally,
throughout this work, we plan to perform extensive testing
of the run time and detection performance of Bayesian
detection algorithms and then compare those results to the
detection performance of other methods.
Acknowledgments
We thank Andrew Moore and the other members of the
Pitt-CMU Detection Group for helpful discussions. We
also thank the UAI reviewers for helpful comments. This
research was supported in part by grants from the National
Science Foundation (IIS-0325581), the Defense Advanced
Research Projects Agency (F30602-01-2-0550), and the
Pennsylvania Department of Health (ME-01-737).
UAI 2004
COOPER ET AL.
References
(Buckeridge 2004) D.L. Buckeridge, H. Burkom, M.
Campbell, A.W. Moore. Outbreak detection algorithms:
A synthesis of research from the DARPA BioALIRT
Project (in preparation).
(Carlstein 1988) E. Carlstein. Nonparametric change-point
estimation. Annals of Statistics, 1988;16:188-197.
(Fawcett 1999) T. Fawcett, F. Provost. Activity
monitoring: Noticing interesting changes in behavior. In:
Proceedings of the Fifth International Conference on
Knowledge Discovery and Data Mining; 1999.
(Greenland 2000)S. Greenland. Causal analysis in the
health sciences. Journal of the American Statistical
Association, 2000;95:286-289.
(Hamilton 1994) J. Hamilton. Time Series Analysis.
Princeton, NJ: Princeton University Press; 1994.
(Hanna 1982) S. Hanna, G. Briggs, J. Hosker. Handbook of
Atmospheric Diffusion. Document DOE/TIC-11223,
Washington, D.C.: Department of Energy; 1982.
(Hogan 2004) W.R. Hogan, G.F. Cooper, M.M. Wagner. A
Bayesian anthrax aerosol release detector. RODS
Technical Report; 2004.
(Hugin 2004) Hugin software. http://www.hugin.com/
(Hutwagner 2003) L. Hutwagner, W. Thompson, G.M.
Seeman. The bioterrorism preparedness and response
early aberration reporting system (EARS). Journal of
Urban Health, 2003;80:i89-i96.
(Kaufmann 1997) A. Kaufmann, M. Meltzer, G. Schmid.
The economic impact of a bioterrorist attack: Are
prevention and postattack intervention programs
justifiable? Emerging Infectious Diseases, 1997;3:83-94.
(Koller, 1997) D. Koller and A. Pfeffer. Object-oriented
Bayesian networks. In: Proceedings of the Conference
on Uncertainty in Artificial Intelligence, 1997; 302-313.
(Kulldorf 1997) M. Kulldorf. A spatial scan statistic.
Communications in Statistics--Theory and Methods,
1997;26:1481-1496.
(Kulldorff 2001) M. Kulldorff. Prospective time periodic
geographical disease surveillance using a scan statistic.
Journal of the Royal Statistical Society, Series A,
2001;164:61-72.
(Moore 2003) A. Moore, G. Cooper, R. Tsui, M. Wagner.
Summary of biosurveillance-relevant technologies.
Available at: http://www-2.cs.cmu.edu/~awm/biosurvmethods.pdf, 2003.
(Neill 2003) D.B. Neill, A.W. Moore. A fast multiresolution method for detection of significant spatial
disease clusters. In: Advances in Neural Information
Processing Systems (NIPS), 2003.
103
(Rabiner 1989) L.R. Rabiner. A tutorial on hidden Markov
models and selected applications in speech recognition.
In: Proceedings of IEEE, 1989;77:257-285.
(Reis 2003) B. Reis, K.D. Mandl. Time series modeling for
syndromic surveillance. BMC Medical Informatics and
Decision Making; 2003; v3.
(Serfling 1963) R.E. Serfling. Methods for current
statistical analysis of excess pneumonia-influenza
deaths. Public Health Reports, 1963;78:494-506.
(Srinivas 1994) S. Srinivas. A probabilistic approach to
hierarchical model-based diagnosis. In: Proceedings of
the Conference on Uncertainty in Artificial Intelligence,
1994; 538-545.
(Tsui 2001) F.C. Tsui, M.M. Wagner, V. Dato, C. Chang.
Value of ICD-9-coded chief complaints for detection of
epidemics. In: Proceedings of the Fall Symposium of the
American Medical Informatics Association; 2001;711715.
(Wagner 2001) M.M. Wagner, F.C. Tsui, J.U. Espino,
V.M. Dato, D.F. Sittig, R.A. Caruana, L.F. McGinnis,
D.W. Deerfield, M.J. Druzdzel, D.B. Fridsma. The
emerging science of very early detection of disease
outbreaks. Journal of Public Health Management and
Practice, 2001;7:51-59.
(Wein 2003) L.M. Wein, D.L. Craft, E.H. Kaplan.
Emergency response to an anthrax attach. Proceedings
of the National Academcy of Sciences, 2003;100:43464351.
(Williamson 1999) G.D. Williamson, G.W. Hudson. A
monitoring system for detecting aberrations in public
health surveillance reports. Statistics in Medicine,
1999;18:3283-3298.
(Wong 2003) W.K. Wong, A. Moore, G. Cooper, M.
Wagner. Bayesian network anomaly pattern detection
for disease outbreaks. In: Proceedings of the
International Conference on Machine Learning; 2003;
808-815.
(Wong 2004) Data Mining for Early Disease Outbreak
Detection. Doctoral Dissertation, School of Computer
Science, Carnegie Mellon University, 2004.
(Xiang 1999) Y. Xiang and F.V. Jensen. Inference in
multiply sectioned Bayesian networks with extended
Shafer-Shenoy and lazy propagation. In: Proceedings of
the Conference on Uncertainty in Artificial Intelligence,
1999; 680-687.
| 5cs.CE
|
1
Scaling Exponent and Moderate Deviations
Asymptotics of Polar Codes for the AWGN Channel
arXiv:1706.02458v1 [cs.IT] 8 Jun 2017
Silas L. Fong, Member, IEEE and Vincent Y. F. Tan, Senior Member, IEEE
Abstract
This paper investigates polar codes for the additive white Gaussian noise (AWGN) channel. The scaling exponent µ of polar
codes for a memoryless channel qY |X with capacity I(qY |X ) characterizes the closest gap between the capacity and non-asymptotic
achievable rates in the following way: For a fixed ε ∈ (0, 1), the gap between the capacity I(qY |X ) and the maximum non-asymptotic
rate Rn∗ achieved by a length-n polar code with average error probability ε scales as n−1/µ , i.e., I(qY |X ) − Rn∗ = Θ(n−1/µ ). It is
well known that the scaling exponent µ for any binary-input memoryless channel (BMC) with I(qY |X ) ∈ (0, 1) is bounded above
by 4.714, which was shown by an explicit construction of polar codes. Our main result shows that 4.714 remains to be a valid
upper bound on the scaling exponent for the AWGN channel. Our proof technique involves the following two ideas: (i) The capacity
√
of the AWGN channel can be achieved within a gap of O(n−1/µ log n) by using an input alphabet consisting of n constellations
and restricting the input distribution to be uniform; (ii) The capacity of a multiple access channel (MAC) with an input alphabet
consisting of n constellations can be achieved within a gap of O(n−1/µ log n) by using a superposition of log n binary-input polar
codes.
In addition, we investigate the performance of polar codes in the moderate deviations regime where both the gap to capacity
and the error probability vanish as n grows. An explicit construction of polar codes is proposed to obey a certain tradeoff between
the gap to capacity and the decay rate of the error probability for the AWGN channel.
Index Terms
AWGN channel, multiple access channel, moderate deviations, polar codes, scaling exponent
I. I NTRODUCTION
A. The Additive White Gaussian Noise Channel
This paper investigates low-complexity codes over the classical additive white Gaussian noise (AWGN) channel [1, Ch. 9],
where a source wants to transmit information to a destination and each received symbol is the sum of the transmitted symbol
and an independent Gaussian random variable. More specifically, if Xk denotes the symbol transmitted by the source in the k th
time slot, then the corresponding symbol received by the destination is
Yk = Xk + Zk
(1)
S. L. Fong and V. Y. F. Tan were supported by National University of Singapore (NUS) Young Investigator Award under Grant R-263-000-B37-133.
S. L. Fong is with the Department of Electrical and Computer Engineering, NUS, Singapore 117583 (e-mail: [email protected]).
V. Y. F. Tan is with the Department of Electrical and Computer Engineering, NUS, Singapore 117583, and also with the Department of Mathematics, NUS,
Singapore 119076 (e-mail: [email protected]).
February 28, 2018
DRAFT
2
where Zk is the standard normal random variable. When the transmission lasts for n time slots, i.e., each transmitted codeword
consists of n symbols, it is assumed that Z1 , Z2 , . . . , Zn are independent and each transmitted codeword xn , (x1 , x2 , . . . , xn )
must satisfy the peak power constraint
n
1X 2
xk ≤ P
n
(2)
k=1
where P > 0 is a constant which denotes the permissible power. If we would like to transmit a uniformly distributed message
W ∈ {1, 2, . . . , ⌈2nR ⌉} across this channel, it was shown by Shannon [2] that the limit of the maximum coding rate R as n
approaches infinity (i.e., capacity) is
C(P ) ,
1
log(1 + P ).
2
(3)
B. Polar Codes
Although the capacity of a memoryless channel was proved by Shannon [2] in 1948, low-complexity channel codes that
achieve the capacity have not been found until Arıkan [3] proposed to use polar codes with encoding and decoding complexities
being O(n log n) for achieving the capacity of a binary-input memoryless symmetric channel (BMSC). This paper investigates
the scaling exponent of polar codes [4] for the AWGN channel, a ubiquitous channel model in wireless communications.
The scaling exponent µ of polar codes for a memoryless channel qY |X with capacity
I(qY |X ) , max I(X; Y )
pX
(4)
characterizes the closest gap between the channel capacity and non-asymptotic achievable rates in the following way: For a fixed
ε ∈ (0, 1), the gap between the capacity I(qY |X ) and the maximum non-asymptotic rate Rn∗ achieved by a length-n polar code
with average error probability ε scales as n−1/µ , i.e., I(qY |X ) − Rn∗ = Θ(n−1/µ ). It has been shown in [4]–[6] that the scaling
exponent µ for any BMSC with I(qY |X ) ∈ (0, 1) lies between 3.579 and 4.714, where the upper bound 4.714 was shown by
an explicit construction of polar codes. Indeed, the upper bound 4.714 remains valid for any general binary-input memoryless
channel (BMC) [7, Lemma 4]. It is well known that polar codes are capacity-achieving for BMCs [8]–[11], and appropriately
chosen ones are also capacity-achieving for the AWGN channel [12]. In particular, for any R < C(P ) and any β < 1/2,
polar codes operated at rate R can be constructed for the AWGN channel such that the decay rate of the error probability is
β
O(2−n ) [12] and the encoding and decoding complexities are O(n log n). However, the scaling exponent of polar codes for
the AWGN channel has not been investigated yet.
In this paper, we construct polar codes for the AWGN channel and show that 4.714 remains to be a valid upper bound on the
scaling exponent. Our construction of polar codes involves the following two ideas: (i) By using an input alphabet consisting
of n constellations and restricting the input distribution to be uniform as suggested in [12], we can achieve the capacity of the
√
AWGN channel within a gap of O(n−1/µ log n); (ii) By using a superposition of log n binary-input polar codes1 as suggested
in [13], we can achieve the capacity of the corresponding multiple access channel (MAC) within a gap of O(n−1/µ log n) where
the input alphabet of the MAC has n constellations (i.e., the size of the Cartesian product of the input alphabets corresponding
to the log n input terminals is n). The encoding and decoding complexities of our constructed polar codes are O(n log2 n). On
1 In
this paper, n is always a power of 2
February 28, 2018
DRAFT
3
the other hand, the lower bound 3.579 holds trivially for the constructed polar codes because the polar codes are constructed by
superposing log n binary-input polar codes whose scaling exponents are bounded below by 3.579 [5].
In addition, Mondelli et al. [4, Sec. IV] provided an explicit construction of polar codes for any BMSC which obey a certain
tradeoff between the gap to capacity and the decay rate of the error probability. More specifically, if the gap to capacity is set
1−γ
1
to vanish at a rate of Θ n− µ
for some γ ∈ 1+µ
, 1 , then a length-n polar code can be constructed such that the error
−1 γµ+γ−1
γh
−n 2 ( γµ )
where h2 : [0, 1/2] → [0, 1] denotes the binary entropy function. This tradeoff was
probability is O n · 2
developed under the moderate deviations regime [14] where both the gap to capacity and the error probability vanish as n grows.
For the AWGN channel, we develop a similar tradeoff under the moderate deviations regime by using our constructed polar
codes described above.
C. Paper Outline
This paper is organized as follows. The notation used in this paper is described in the next subsection. Section II presents the
background of this work, which includes existing polarization results for the BMC which are used in this work. Sections III-A
to III-C state the formulation of the binary-input MAC and present new polarization results for the binary-input MAC. Sections IV-A to IV-B state the formulation of the AWGN channel and present new polarization results for the AWGN channel.
Section V establishes the definition of the scaling exponent for the AWGN channel and establishes the main result — 4.714 is
an upper bound on the scaling exponent of polar codes for the AWGN channel. Section VI presents an explicit construction of
polar codes for the AWGN channel which obey a certain tradeoff between the gap to capacity and the decay rate of the error
probability under the moderate deviations regime. Concluding remarks are provided in Section VII.
D. Notation
The set of natural numbers, real numbers and non-negative real numbers are denoted by N, R and R+ respectively. For any
sets A and B and any mapping f : A → B, we let f −1 (D) denote the set {a ∈ A |f (a) ∈ D} for any D ⊆ B. We let 1{E}
be the indicator function of the set E. An arbitrary (discrete or continuous) random variable is denoted by an upper-case letter
(e.g., X), and the realization and the alphabet of the random variable are denoted by the corresponding lower-case letter (e.g.,
x) and calligraphic letter (e.g., X ) respectively. We use X n to denote the random tuple (X1 , X2 , . . . , Xn ) where each Xk has
the same alphabet X . We will take all logarithms to base 2 throughout this paper.
The following notations are used for any arbitrary random variables X and Y and any real-valued function g with domain X .
We let pY |X and pX,Y = pX pY |X denote the conditional probability distribution of Y given X and the probability distribution of
(X, Y ) respectively. We let pX,Y (x, y) and pY |X (y|x) be the evaluations of pX,Y and pY |X respectively at (X, Y ) = (x, y). To
R
make the dependence on the distribution explicit, we let PpX {g(X) ∈ A} denote X pX (x)1{g(x) ∈ A} dx for any set A ⊆ R.
The expectation of g(X) is denoted as EpX [g(X)]. For any (X, Y, Z) distributed according to some pX,Y,Z , the entropy of X
and the conditional mutual information between X and Y given Z are denoted by HpX (X) and IpX,Y,Z (X; Y |Z) respectively.
For simplicity, we sometimes omit the subscript of a notation if it causes no confusion. The relative entropy between pX and
qX is denoted by
D(pX kqX ) ,
February 28, 2018
Z
X
pX (x) log
pX (x)
qX (x)
dx.
(5)
DRAFT
4
The 2-Wasserstein distance between pX and pY is denoted by
sZ Z
sX,Y (x, y)(x − y)2 dydx .
W2 (pX , pY ) , sinf :
X,Y
sX =pX ,
sY =pY
X
(6)
Y
We let N ( · ; µ, σ 2 ) : R → [0, ∞) denote the probability density function of a Gaussian random variable whose mean and
variance are µ and σ 2 respectively, i.e.,
(z−µ)2
1
e− 2σ2 .
N (z; µ, σ 2 ) , √
2
2πσ
II. BACKGROUND : P OINT- TO -P OINT C HANNELS
AND
(7)
E XISTING P OLARIZATION R ESULTS
In this section, we will review important polarization results related to the scaling exponent of polar codes for binary-input
memoryless channels (BMCs).
A. Point-to-Point Memoryless Channels
Consider a point-to-point channel which consists of one source and one destination, denoted by s and d respectively. Suppose
node s transmits information to node d in n time slots. Before any transmission begins, node s chooses message W destined for
node d, where W is uniformly distributed over the alphabet
W , {1, 2, . . . , M }
(8)
which consists of M elements. For each k ∈ {1, 2, . . . , n}, node s transmits Xk ∈ X based on W and node d receives Yk ∈ Y
in time slot k where X and Y denote respectively the input and output alphabets of the channel. After n time slots, node d
declares Ŵ to be the transmitted W based on Y n . Formally, we define a length-n code as follows.
Definition 1: An (n, M )-code consists of the following:
1) A message set W as defined in (8). Message W is uniform on W.
2) An encoding function fk : W → X for each k ∈ {1, 2, . . . , n}, where fk is used by node s for encoding Xk such that
Xk = fk (W ).
3) A decoding function ϕ : Y n → W used by node d for producing the message estimate Ŵ = ϕ(Y n ).
Definition 2: The point-to-point memoryless channel is characterized by an input alphabet X , an output alphabet Y and a
conditional distribution qY |X such that the following holds for any (n, M )-code: For each k ∈ {1, 2, . . . , n}, pW,X n ,Y n =
Qn
pW,X n k=1 pYk |Xk where pYk |Xk (yk |xk ) = qY |X (yk |xk ) for all xk ∈ X and yk ∈ Y.
For any (n, M )-code defined on the point-to-point memoryless channel, let pW,X n ,Y n ,Ŵ be the joint distribution induced by
the code. By Definitions 1 and 2, we can factorize pW,X n ,Y n ,Ŵ as
pW,X n ,Y n ,Ŵ = pW pX n |W
Y
n
k=1
pYk |Xk pŴ |Y n .
(9)
B. Polarization for Binary-Input Memoryless Channels
Definition 3: A point-to-point memoryless channel characterized by qY |X is called a binary-input memoryless channel (BMC)
if X = {0, 1}.
February 28, 2018
DRAFT
5
We follow the formulation of polar coding in [10]. Consider any BMC characterized by qY |X . Let pX be the probability
distribution of a Bernoulli random variable X, and let pX n be the distribution of n independent copies of X ∼ pX , i.e.,
Qn
pX n (xn ) = k=1 pX (xk ) for all xn ∈ X n . For each n = 2m where m ∈ N, the polarization mapping of a length-n polar code
is given by
Gn ,
1 0
1 1
⊗m
= G−1
n
(10)
where ⊗ denotes the Kronecker power. Define pU n |X n such that
[U1 U2 . . . Un ] = [X1 X2 . . . Xn ]Gn
(11)
where the addition and product operations are performed over GF(2), define
pYk |Xk (yk |xk ) , qY |X (yk |xk )
for each k ∈ {1, 2, . . . , n} and each (xk , yk ) ∈ X × Y where qY |X characterizes the BMC (cf. (2)), and define
pU n ,X n ,Y n , pX n pU n |X n
n
Y
k=1
pYk |Xk .
(12)
In addition, for each k ∈ {1, 2, . . . , n}, define the Bhattacharyya parameter associated with time k as
Z [pX ;qY |X ] (Uk |U k−1 , Y n )
Z
q
X
,2
pU k−1 ,Y n (uk−1 , y n ) pUk |U k−1 ,Y n (0|uk−1 , y n )pUk |U k−1 ,Y n (1|uk−1 , y n )dy n
uk−1 ∈U k−1
X
=2
uk−1 ∈U k−1
(13)
Yn
Z
Yn
q
pUk ,U k−1 ,Y n (0, uk−1 , y n )pUk ,U k−1 ,Y n (1, uk−1 , y n )dy n ,
(14)
where the distributions in (13) and (14) are marginal distributions of pU n ,X n ,Y n defined in (12). The following result is based
on [4, Sec. III] and has been used in [7] to show that 4.714 is an upper bound on the scaling exponent for any BMC. To simplify
notation, let
β , 4.714
(15)
in the rest of this paper.
Lemma 1 ( [4, Sec. III], [7, Lemma 2]): There exists a universal constant t > 0 such that the following holds. Fix any BMC
characterized by qY |X and any pX . Then for any m ∈ N and n , 2m , we have2
1
k ∈ {1, 2, . . . , n} Z [pX ;qY |X ] (Uk |U k−1 , Y n ) ≤
n
III. P ROBLEM F ORMULATION
OF
1
n4
B INARY-I NPUT MAC S
AND
≥ IpX qY |X (X; Y ) −
t
n1/β
.
(16)
N EW P OLARIZATION R ESULTS
Polar codes have been proposed and investigated for achieving any rate tuple inside the capacity region of a binary-input
multiple access channel (MAC) [13,15]. The goal of this section is to use the polar codes proposed in [13] to achieve the
2 This
lemma remains to hold if the quantities
this lemma are replaced by n1ν for any ν > 2.
February 28, 2018
1
n4
are replaced by
1
nν
for any ν > 0. The main result of this paper continues to hold if the quantities
1
n4
in
DRAFT
6
symmetric sum-capacity of a binary-input MAC.
A. Binary-Input Multiple Access Channels
Consider a MAC [1, Sec. 15.3] which consists of N sources and one destination. Let I , {1, 2, . . . , N } be the index set of
the N sources and let d denote the destination. Suppose the sources transmit information to node d in n time slots. Before any
transmission begins, node i chooses message Wi destined for node d for each i ∈ I, where Wi is uniformly distributed over
Wi , {1, 2, . . . , Mi }
(17)
which consists of Mi elements. For each k ∈ {1, 2, . . . , n}, node i transmits Xi,k ∈ Xi based on Wi for each i ∈ I and node d
receives Yk ∈ Y in time slot k where Xi denotes the input alphabet for node i and Y denotes the output alphabet. After n time
slots, node d declares Ŵi to be the transmitted Wi based on Y n for each i ∈ I.
To simplify notation, we use the following convention for any T ⊆ I. For any random tuple (X1 , X2 , . . . , XN ), we let
XT , (Xi : i ∈ T ) be the corresponding subtuple, whose realization and alphabet are denoted by xT and XT respectively.
Similarly, for each k ∈ {1, 2, . . . , n} and each random tuple (X1,k , X2,k , . . . , XN,k ) ∈ XI , we let XT,k , (Xi,k : i ∈ T ) denote
the corresponding random subtuple, and let xT,k and XT,k denote respectively the realization and the alphabet of XT,k . Formally,
we define a length-n code for the binary-input MAC as follows.
Definition 4: An (n, MI )-code, where MI , (M1 , M2 , . . . , MN ), consists of the following:
1) A message set Wi for each i ∈ I as defined in (17), where message Wi is uniform on Wi .
MAC
MAC
2) An encoding function fi,k
: Wi → Xi for each i ∈ I and each k ∈ {1, 2, . . . , n}, where fi,k
is used by node i for
MAC
encoding Xi,k such that Xi,k = fi,k
(Wi ).
3) A decoding function ϕMAC : Y n → WI used by node d for producing the message estimate ŴI = ϕMAC (Y n ).
Definition 5: The multiple access channel (MAC) is characterized by N input alphabets specified by XI , an output alphabet specified by Y and a conditional distribution qY |XI such that the following holds for any (n, MI )-code: For each
k ∈ {1, 2, . . . , n},
p
WI ,XIn ,Y n
=
Y
i∈I
p
Wi ,Xin
Y
n
k=1
pYk |XI,k
(18)
where pYk |XI,k (yk |xI,k ) = qY |XI (yk |xI,k ) for all xI,k ∈ XI and yk ∈ Y.
B. Polarization for Binary-Input MACs
Definition 6: A MAC characterized by qY |XI is called a binary-input MAC if XI = {0, 1}N .
Consider any binary-input MAC characterized by qY |XI . For each i ∈ I, let pXi be the probability distribution of a Bernoulli
Qn
random variable Xi , and let pXin be the distribution of n independent copies of Xi ∼ pXi , i.e., pXin (xni ) = k=1 pXi (xi,k ) for
all xni ∈ Xin . Recall the polarization mapping Gn defined in (10). For each i ∈ I, define pUin |Xin such that
[Ui,1 Ui,2 . . . Ui,n ] = [Xi,1 Xi,2 . . . Xi,n ]Gn
February 28, 2018
(19)
DRAFT
7
where the addition and product operations are performed over GF(2), and define
Y
Y
n
pUIn ,XIn ,Y n ,
pXin pUin |Xin
pYk |XI,k .
i∈I
(20)
k=1
In addition, for each i ∈ I and each k ∈ {1, 2, . . . , n}, define [i − 1] , {1, 2, . . . , i − 1} and define the Bhattacharyya parameter
associated with node i and time k as
n
Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
, Y n)
Z r
X
X
pUi,k ,U k−1 ,X n
,2
∈{0,1}(i−1)n
uik−1 ∈Uik−1 xn
[i−1]
Yn
i
[i−1]
k−1
, xn[i−1] , y n )pUi,k ,U k−1 ,X n ,Y n (1, uik−1 , xn[i−1] , y n )dy n ,
,Y n (0, ui
i
[i−1]
(21)
where the distributions in (21) are marginal distributions of pUIn ,XIn ,Y n defined in (20). The following lemma is a direct
consequence of Lemma 1.
Lemma 2: There exists a universal constant t > 0 such that the following holds. Fix any binary-input MAC characterized by
qY |XI and any pXI . Then for any m ∈ N and n , 2m , we have3
1 n
n
, Y n) ≤
k ∈ {1, 2, . . . , n} Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
n
1
n4
o
≥ IpXI qY |XI (Xi ; X[i−1] , Y ) −
t
n1/β
(22)
for each i ∈ I.
Proof: Fix any i ∈ I. Construct pX[i−1] ,Y |Xi by marginalizing pXI qY |XI and view pX[i−1] ,Y |Xi as the conditional distribution
that characterizes a BMC. The lemma then follows directly from Lemma 1.
C. Polar Codes That Achieve the Symmetric Sum-Capacity of a Binary-Input MAC
Throughout this paper, let p∗Xi denote the uniform distribution on {0, 1} for each i ∈ I and define p∗XI ,
p∗XI (xI ) =
1
2N
Q
i∈I
p∗Xi , i.e.,
(23)
for any xI ∈ {0, 1}N .
Definition 7: For a binary-input MAC characterized by qY |XI , the symmetric sum-capacity is defined to be
Csum , Ip∗X
I
qY |XI (XI ; Y
).
The following definition summarizes the polar codes for the binary-input MAC proposed in [13, Sec. IV].
Definition 8 ( [13, Sec. IV]): Fix an n = 2m where m ∈ N. For each i ∈ I, let Ji be a subset of {1, 2, . . . , n}, define
Jic , {1, 2, . . . , n} \ Ji , and let bi,Jic , (bi,k ∈ {0, 1} : k ∈ Jic ) be a binary tuple. An (n, JI , bI,JIc )-polar code, where
JI , (Ji : i ∈ I) and bI,JIc , (bi,Jic : i ∈ I), consists of the following:
1) An index set for information bits transmitted by node i denoted by Ji for each i ∈ I. The set Jic is referred to as the
index set for frozen bits transmitted by node i.
3 This
lemma remains to hold if the quantities
this lemma are replaced by n1ν for any ν > 2.
February 28, 2018
1
n4
are replaced by
1
nν
for any ν > 0. The main result of this paper continues to hold if the quantities
1
n4
in
DRAFT
8
2) A message set Wi , {1, 2, . . . , 2|Ji | } for each i ∈ I, where Wi is uniform on Wi .
3) An encoding bijection fiMAC : Wi → Ui,Ji for encoding Wi into |Ji | information bits denoted by Ui,Ji for each i ∈ I
such that
Ui,Ji = fiMAC(Wi ),
where Ui,Ji and Ui,Ji are defined as Ui,Ji ,
uniform on
Wi , fiMAC (Wi )
Q
k∈Ji
Uk and Ui,Ji , (Ui,k : k ∈ Ji ) respectively. Since message Wi is
is a sequence of i.i.d. uniform bits such that
P{Ui,Ji = ui,Ji } =
1
2|Ji |
(24)
for all ui,Ji ∈ {0, 1}|Ji| , where the bits are transmitted through the polarized channels indexed by Ji . For each i ∈ I and
each k ∈ Jic , let
Ui,k = bi,k
(25)
be the frozen bit to be transmitted by node i in time slot k. After Uin has been determined, node i transmits Xin where
[Xi,1 Xi,2 . . . Xi,n ] , [Ui,1 Ui,2 . . . Ui,n ]G−1
n .
(26)
k−1
4) A sequence of successive cancellation decoding functions ϕMAC
× {0, 1}(i−1)n × Y n → {0, 1} for each i ∈ I
i,k : {0, 1}
and each k ∈ {1, 2, . . . , n} such that the recursively generated (Û1,1 , . . . , Û1,n ), (Û2,1 , . . . , Û2,n ), . . . , (ÛN,1 , . . . , ÛN,n )
and (X̂1,1 , . . . , X̂1,n ), (X̂2,1 , . . . , X̂2,n ), . . . , (X̂N,1 , . . . , X̂N,n ) are produced as follows. For each i ∈ I and each k =
n
n
were constructed before the construction of Ûi,k , node d constructs the
1, 2, . . . , n, given that Ûik−1 , Û[i−1]
and X̂[i−1]
estimate of Ui,k through computing
k−1
n
, Y n)
Ûi,k , ϕMAC
, X̂[i−1]
i,k (Ûi
where
k−1
ûi,k , ϕMAC
, x̂n[i−1] , y n )
i,k (ûi
0
if k ∈ Ji and pUi,k |U k−1 ,X n ,Y n (0|ûik−1 , x̂n[i−1] , y n ) ≥ pUi,k |U k−1 ,X n ,Y n (1|ûik−1 , x̂n[i−1] , y n ),
i
i
[i−1]
[i−1]
= 1
if k ∈ Ji and pUi,k |U k−1 ,X n ,Y n (0|ûik−1 , x̂n[i−1] , y n ) < pUi,k |U k−1 ,X n ,Y n (1|ûik−1 , x̂n[i−1] , y n ),
i
i
[i−1]
[i−1]
bi,k if k ∈ J c .
i
(27)
After obtaining Ûin , node d constructs the estimate of Xin through computing
[X̂i,1 X̂i,2 . . . X̂i,n ] , [Ûi,1 Ûi,2 . . . Ûi,n ]G−1
n
(28)
and declares that
Ŵi , fiMAC
is the transmitted Wi where fiMAC
−1
−1
(Ûi,Ji )
denote the inverse function of fiMAC .
Remark 1: By inspecting Definition 4 and Definition 8, we see that every (n, JI , bI,JIc )-polar code is also an
February 28, 2018
DRAFT
9
n, (2|J1 | , 2|J2 | , . . . , 2|JN | ) -code.
Definition 9: The uniform-input (n, JI )-polar code is defined as an (n, JI , BI,JIc )-polar code where BI,JIc consists of i.i.d.
uniform bits that are independent of the message WI .
Definition 10: For the uniform-input (n, JI )-polar code defined for the MAC, the probability of decoding error is defined as
[
P{ŴI 6= WI } = P
{Ui,Ji 6= Ûi,Ji }
i∈I
where the error is averaged over the random messages and the frozen bits. The code is also called a uniform-input (n, JI , ε)-polar
code if the probability of decoding error is no larger than ε.
The following proposition bounds the error probability in terms of Bhattacharyya parameters, and it is a generalization of the
well-known result for the special case N = 1 (e.g., see [3, Proposition 2]). The proof of Proposition 3 can be deduced from [13,
Sec. IV], and is contained in Appendix A for completeness.
Proposition 3: For the uniform-input (n, JI )-polar code defined for the MAC qY |XI , we have
P{ŴI 6= WI } ≤
N X
X
i=1 k∈Ji
n
, Y n ).
Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
(29)
The following proposition follows from combining Lemma 2, Definition 7 and Proposition 3.
Proposition 4: There exists a universal constant t > 0 such that the following holds. Fix any N -source binary-input MAC
characterized by qY |XI . Fix any m ∈ N, let n = 2m and define
n
o
∗
n
JiSE , k ∈ {1, 2, . . . , n} Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
, Y n ) ≤ n14
(30)
for each i ∈ I where p∗XI is the uniform distribution as defined in (23) and the superscript “SE” stands for “scaling exponent”.
Then, the corresponding uniform-input (n, JISE )-polar code satisfies
PN
i=1
n
JiSE
≥ Csum −
tN
n1/β
(31)
and
P{ŴI 6= WI } ≤
N
.
n3
(32)
Proof: Let t > 0 be the universal constant specified in Lemma 2 and fix an n. For each i ∈ I, it follows from Lemma 2
and Proposition 3 that
JiSE
t
≥ Ip∗X qY |XI (Xi ; X[i−1] , Y ) − 1/β
I
n
n
(33)
and
P{ŴI 6= WI } ≤
February 28, 2018
N
n3
(34)
DRAFT
10
for the uniform-input (n, JISE )-polar code. Since p∗XI =
Ip∗X
I
QN
i=1
qY |XI (Xi ; X[i−1] , Y
p∗Xi , it follows that
) = Ip∗X
I
qY |XI (Xi ; Y
|X[i−1] )
(35)
holds for each i ∈ I, which implies that
N
X
i=1
Ip∗X
I
qY |XI (Xi ; X[i−1] , Y
) = Ip∗X
I
qY |XI (XI ; Y
).
(36)
Consequently, (31) follows from (33), (36) and Definition 7, and (32) follows from (34).
Remark 2: Proposition 4 shows that the sum-capacity of a binary-input MAC with N sources can be achieved within a gap
of O(N n−1/β ) by using a superposition of N binary-input polar codes.
IV. P ROBLEM F ORMULATION
OF THE
AWGN C HANNEL AND N EW P OLARIZATION R ESULTS
A. The AWGN Channel
It is well known that appropriately designed polar codes are capacity-achieving for the AWGN channel [12]. The main
contribution of this paper is proving an upper bound on the scaling exponent of polar codes for the AWGN channel by using
uniform-input polar codes for binary-input MACs described in Definition 8. The following two definitions formally define the
AWGN channel and length-n codes for the channel.
Definition 11: An (n, M, P )-code is an (n, M )-code described in Definition 1 subject to the additional assumptions that
X = R and the peak power constraint
P
(
n
1X 2
Xk ≤ P
n
k=1
)
=1
(37)
is satisfied.
Definition 12: The AWGN channel is a point-to-point memoryless channel described in Definition 2 subject to the additional
assumption that Y = R and qY |X (y|x) = N (y; x, 1) for all x ∈ R and y ∈ R.
Definition 13: For an (n, M, P )-code defined on the AWGN channel, we can calculate according to (9) the average probability
of error defined as P Ŵ 6= W . We call an (n, M, P )-code with average probability of error no larger than ε an (n, M, P, ε)-
code.
B. Uniform-Input Polar Codes for the AWGN Channel
Recall that we would like to use uniform-input polar codes for binary-input MACs described in Definition 8 to achieve the
capacity of the AWGN channel, i.e., C(P ) in (3). The following definition describes the basic structure of such uniform-input
polar codes.
Definition 14: Fix an n = 2m where m ∈ N. An (n, JI , P, A)avg -polar code with average power P and input alphabet A
consists of the following:
1) An input alphabet A ⊂ R ∪ {0− } with |A| = n such that
1
n
February 28, 2018
X
a∈A\{0− }
a2 ≤ P,
(38)
DRAFT
11
where R ∪ {0− } can be viewed as a line with 2 origins.4 We index each element of A by a unique length-m binary tuple
xMAC
, (xMAC
, xMAC
, . . . , xMAC
I
1
2
m ),
(39)
and let ρ : {0, 1}m → A be the bijection that maps the indices to the elements of A such that ρ(xMAC
) denotes the element
I
in A indexed by xMAC
.
I
2) A binary-input MAC qY |XIMAC induced by A as defined through Definitions 5 and 6 with the identifications N = m and
I = {1, 2, . . . , m}.
3) A message set Wi , {1, 2, . . . , 2|Ji | } for each i ∈ {1, 2, . . . , m}, where WI is the message alphabet of the uniform-input
(n, JI )-polar code for the binary-input MAC qY |XIMAC as defined through Definitions 8 and 9 such that
|WI | =
m
Y
i=1
|Wi | = 2
Pm
i=1
|Ji |
.
(40)
In addition, WI is uniform on WI . We view the uniform-input (n, JI )-polar code as an n, (2|J1 | , 2|J2 | , . . . , 2|JN | ) -code
MAC
(cf. Remark 1) and let {fi,k
| i ∈ I, k ∈ {1, 2, . . . , n}} and ϕMAC denote the corresponding set of encoding functions
and the decoding function respectively (cf. Definition 4).
4) An encoding function fk : WI → A defined as
MAC
MAC
MAC
fk (WI ) , ρ(f1,k
(W1 ), f2,k
(W2 ), . . . , fm,k
(Wm ))
for each k ∈ {1, 2, . . . , n}, where fk is used for encoding WI into Xk such that
fk (WI ) if fk (WI ) 6= 0− ,
Xk =
0
if fk (WI ) = 0− .
(41)
(42)
Note that both the encoded symbols 0 and 0− in A result in the same transmitted symbol 0 ∈ R according to (42). By
construction, f1 (WI ), f2 (WI ), . . ., fn (WI ) are i.i.d. random variables that are uniformly distributed on A and hence X1 ,
X2 , . . ., Xn are i.i.d. real-valued random variables (but not necessarily uniform).
5) A decoding function ϕ : Rn → WI defined as
ϕ , ϕMAC
(43)
ŴI = ϕ(Y n ).
(44)
such that
Remark 3: For an (n, JI , P, A)avg -polar code, the flexibility of allowing A to contain 2 origins is crucial to proving the main
result of this paper. This is because the input distribution which we will use to establish scaling results for the AWGN channel
in Theorem 1 can be viewed as the uniform distribution over some set that contains 2 origins, although the input distribution in
the real domain as specified in (48) to follow is not uniform.
4 Introducing
the symbol 0− allows us to create a set of cardinality n which consists of n − 2 non-zero real numbers and 2 origins 0 and 0−
February 28, 2018
DRAFT
12
Proposition 5: There exists a universal constant t > 0 such that the following holds. Suppose we are given an (n, JISE , P, A)avg -
polar code defined for the AWGN channel qY |X with a 2-origin A (i.e., A ⊇ {0, 0− }). Define X , A \ {0− } ⊂ R where X
contains 1 origin and n − 2 non-zero real numbers. Then, the (n, JISE , P, A)avg -polar code is an (n, M )-code (cf. Definition 1)
which satisfies
t log n
1
log M ≥ Ip′X qY |X (X; Y ) − 1/β ,
n
n
P{ŴI 6= WI } ≤
(45)
log n
,
n3
(46)
and
P {X n = xn } =
n
Y
k=1
P {Xk = xk } =
for all xn ∈ X n where p′X is the distribution on X defined as
1
n
′
pX (a) =
2
n
n
Y
p′X (xk )
(47)
k=1
if a 6= 0,
(48)
if a = 0.
Proof: The proposition follows from inspecting Proposition 4 and Definition 14 with the identifications N = log n and
Pm
log M = i=1 |Ji |.
The following lemma, a strengthened version of [12, Th. 6], provides a construction of a good A which leads to a controllable
gap between C(P ) and Ip′X qY |X (X; Y ) for the corresponding (n, JISE , P, A)avg -polar code. Although the following lemma is
intuitive, the proof is technical and hence relegated to Appendix B.
Lemma 6: Let qY |X be the conditional distribution that characterizes the AWGN channel, and fix any γ ∈ [0, 1). For each
n = 2m where m ∈ N, define
Dn ,
ℓ
ℓ∈
n
1, 2, . . . ,
n
,...n − 1
2
,
(49)
define
sX (x) , N x; 0, P 1 −
1
n(1−γ)/β
(50)
for all x ∈ R, define ΦX to be the cdf of sX , and define
X , Φ−1
X (Dn ).
(51)
Note that X contains 1 origin and n − 2 non-zero real numbers, and we let p′X be the distribution on X as defined in (48). In
Qn
addition, define the distribution p′X n (xn ) , k=1 p′X (xk ). Then, there exists a constant t′ > 0 that depends on P and γ but
not n such that the following statements hold for each n ∈ N:
Ep′X [X 2 ] ≤ P 1 −
February 28, 2018
1
n(1−γ)/β
,
(52)
DRAFT
13
Pp′X n
(
n
1X 2
Xk > P
n
k=1
)
≤
e3
,
(53)
√
t′ log n
.
n(1−γ)/β
(54)
e
1 − 1−γ
β
n2
and
C(P ) − Ip′X qY |X (X; Y ) ≤
A shortcoming of Proposition 5 is that the (n, JISE , P, A)avg -polar code may not satisfy the peak power constraint (37) and hence
it may not qualify as an (n, 2
PN
i=1
|JiSE |
, P )-code (cf. Definition 11). Therefore, we describe in the following definition a slight
modification of an (n, JISE , P, A)avg -polar code so that the modified polar code always satisfies the peak power constraint (37).
Definition 15: The 0-power-outage version of an (n, JI , P, A)avg -polar code is an (n, 2
PN
i=1
|Ji |
)-code which follows identical
encoding and decoding operations of the (n, JI , P, A)avg -polar code except that the source will modify the input symbol in a
time slot k if the following scenario occurs: Let X n be the desired codeword generated by the source according to the encoding
operation of the (n, JI , P, A)avg -polar code, where the randomness of X n originates from the information bits (Ui,Ji : i ∈ I) and
Pk
the frozen bits BI,JIc . If transmitting the desired symbol Xk at time k results in violating the power constraint n1 ℓ=1 Xℓ2 > P ,
the source will transmit the symbol 0 at time k instead. An (n, 2
PN
i=1
|Ji |
, ε)-code is called an (n, JI , P, A, ε)peak -polar code if
it is the 0-power-outage version of some (n, JI , P, A)avg -polar code.
By Definition 15, every (n, JI , P, A, ε)peak -polar code satisfies the peak power constraint (37) and hence achieves zero power
Pn
outage, i.e., P{ n1 k=1 Xk2 > P } = 0. Using Definition 15, we obtain the following corollary which states the obvious fact
that the probability of power outage of an (n, JI , P, A)avg -polar code can be viewed as part of the probability of error of the
0-power-outage version of the code.
Corollary 7: Given an (n, JI , P, A)avg -polar code, define
ε1 , P{ŴI 6= WI }
and
ε2 , P
(
n
1X 2
Xk > P
n
k=1
(55)
)
.
(56)
Then, the 0-power-outage version of the (n, JI , P, A)avg -polar code is an (n, JI , P, A, ε1 + ε2 )peak -polar code that satisfies the
peak power constraint (37).
V. S CALING E XPONENTS
AND
M AIN R ESULT
A. Scaling Exponent of Uniform-Input Polar Codes for MACs
We define scaling exponent of uniform-input polar codes for the binary-input MAC as follows.
Definition 16: Fix an ε ∈ (0, 1) and an N -source binary-input MAC qY |XI with symmetric sum-capacity Csum (cf. Definition 7). The scaling exponent of uniform-input polar codes for the MAC is defined as
− log n
m
P
µPC-MAC
,
lim
inf
inf
.
n
=
2
,
there
exists
a
uniform-input
(n,
J
,
ε)-polar
code
on
q
I
Y |XI
N
ε
|J |
m→∞ JI
log Csum − i=1n i
February 28, 2018
DRAFT
14
1
Definition 16 formalizes the notion that we are seeking the smallest µ ≥ 0 such that |Csum − RnMAC | = O n− µ holds where
RnMAC ,
|J |
n
denotes the rate of an (n, JI , ε)-polar code. It has been shown in [5, Sec. IV-C] and [4, Th. 2] that
3.579 ≤ µPC-BMC
≤ β = 4.714
ε
∀ε ∈ (0, 1)
(57)
for the special case N = 1 where the binary-input MAC reduces a BMC. We note from [16, Th. 48] (also [17] and [18]) that
the optimal scaling exponent (optimized over all codes) for any non-degenerate DMC (as well as BMC) is equal to 2 for all
ε ∈ (0, 1/2).
Using Proposition 3 and Definition 16, we obtain the following corollary, which shows that 4.714, the upper bound on µPC-BMC
ε
in (57) for BMCs, remains to be a valid upper bound on the scaling exponent for binary-input MACs.
Corollary 8: Fix any ε ∈ (0, 1) and any binary-input MAC qY |XI . Then,
µPC-MAC
≤ β = 4.714.
ε
B. Scaling Exponent of Uniform-Input Polar Codes for the AWGN channel
Definition 17: Fix a P > 0 and an ε ∈ (0, 1). The scaling exponent of uniform-input polar codes for the AWGN channel is
defined as
µPC-AWGN
, lim inf inf
P,ε
m→∞ JI ,A
− log n
log C(P ) −
PN
i=1
n
n = 2m , there exists a uniform-input (n, JI , P, A, ε)peak -polar code .
|Ji |
1
Definition 17 formalizes the notion that we are seeking the smallest µ ≥ 0 such that C(P ) − RnAWGN = O n− µ
where
RnAWGN
,
PN
i=1
n
|Ji |
holds
denotes the rate of an (n, JI , P, ε)peak -polar code. We note from [16, Th. 54] and [18, Th. 5] that the
optimal scaling exponent of the optimal code for the AWGN channel is equal to 2 for any ε ∈ (0, 1/2). The following theorem
is the main result of this paper, which shows that 4.714 is a valid upper bound on the scaling exponent of polar codes for the
AWGN channel.
Theorem 1: Fix any P > 0 and any ε ∈ (0, 1). There exists a constant t∗ > 0 that does not depend on n such that the
following holds. For any n = 2m where m ∈ N, there exists an A such that the corresponding (n, JISE , P, A)peak -polar code
defined for the AWGN channel qY |X satisfies
N
1 X SE
t∗ log n
|Ji | ≥ C(P ) − 1/β ,
n i=1
n
(58)
and
P{ŴI 6= WI } ≤
e3
log n
+
1− 1 .
n3
en 2 β
(59)
In particular, we have
µPC-AWGN
≤ β = 4.714.
P,ε
(60)
Proof: Fix a P > 0, an ε ∈ (0, 1) and an n = 2m where m ∈ N. Combining Proposition 5 and Lemma 6, we conclude
that there exist a constant t∗ > 0 that does not depend on n and an A such that the corresponding (n, JISE , P, A)avg -polar code
February 28, 2018
DRAFT
15
defined for the AWGN channel qY |X satisfies (58)
P{ŴI 6= WI } ≤
log n
,
n3
(61)
and
P
(
n
1X 2
Xk > P
n
k=1
)
≤
e3
1− 1
β
.
(62)
en 2
Using (61), (62) and Corollary 7, we conclude that the (n, JISE , P, A)avg -polar code is an (n, JISE , P, A)peak -polar code that
satisfies (58) and (59). Since
log n
e3
+
1− 1 ≤ ε
n3
en 2 β
(63)
for all sufficiently large n, it follows from (58), (59) and Definition 17 that (60) holds.
VI. M ODERATE D EVIATIONS R EGIME
A. Polar Codes That Achieve the Symmetric Capacity of a BMC
The following result is based on [4, Sec. IV], which developed a tradeoff between the gap to capacity and the decay rate of the
error probability for a BMC under the moderate deviations regime [14] where both the gap to capacity and the error probability
vanish as n grows.
Lemma 9 ( [4, Sec. IV]): There exists a universal constant tMD > 0 such that the following holds. Fix any γ ∈
1
1+β , 1
and
any BMC characterized by qY |X . Recall that p∗X denotes the uniform distribution on {0, 1}. Then for any n = 2m where m ∈ N,
we have
o
−1 γβ+γ−1
1 n
tMD
k ∈ {1, 2, . . . , n} Z [p∗X ;qY |X ] (Uk |U k−1 , Y n ) ≤ 2−nγh2 ( γβ )
≥ Ip∗X qY |X (X; Y ) − (1−γ)/β
n
n
(64)
where h2 : [0, 1/2] → [0, 1] denotes the binary entropy function.
B. Polar Codes that Achieve the Symmetric Sum-Capacity of a Binary-Input MAC
The following lemma, whose proof is omitted because it is analogous to the proof of Lemma 2, is a direct consequence of
Lemma 9.
1
, 1 and any binaryLemma 10: There exists a universal constant tMD > 0 such that the following holds. Fix any γ ∈ 1+β
Q
input MAC characterized by qY |XI . Recall that p∗XI = i∈I p∗Xi . Then for any n = 2m where m ∈ N, we have
−1 γβ+γ−1
γh2 (
∗
)
1
γβ
n
, Y n ) ≤ 2−n
k ∈ {1, 2, . . . , n} Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
n
tMD
≥ Ip∗X qY |XI (Xi ; X[i−1] , Y ) − (1−γ)/β
(65)
I
n
for each i ∈ I.
Combining Lemma 10, Definition 7 and Proposition 3, we obtain the following proposition, whose proof is analogous to the
proof of Proposition 4 and hence omitted.
February 28, 2018
DRAFT
16
Proposition 11: There exists a universal constant tMD > 0 such that the following holds. Fix any γ ∈
N -source binary-input MAC characterized by qY |XI . In addition, fix any m ∈ N, let n = 2m and define
−1 γβ+γ−1
γh
)
k−1
n
−n 2 ( γβ
MD
[pXI ;qY |XI ]
n
Ji , k ∈ {1, 2, . . . , n} Z
(Ui,k |Ui , X[i−1] , Y ) ≤ 2
1
1+β , 1
and any
(66)
for each i ∈ I where the superscript “MD” stands for “moderate deviations”. Then, the corresponding uniform-input (n, JIMD )polar code described in Definition 9 satisfies
PN
JiMD
tMD N
≥ Csum − (1−γ)/β
n
n
i=1
(67)
and
−1
γh
2
P{ŴI 6= WI } ≤ N n 2−n
( γβ+γ−1
)
γβ
.
(68)
C. Uniform-Input Polar Codes for the AWGN Channel
Proposition 12: There exists a universal constant tMD > 0 such that the following holds. Fix any γ ∈
1
1+β , 1
. Suppose
we are given an (n, JIMD , P, A)avg -polar code (cf. Definition 14) defined for the AWGN channel qY |X with a 2-origin A
(i.e., A ⊇ {0, 0−}). Define X , A \ {0− } ⊂ R where X contains 1 origin and n − 2 non-zero real numbers. Then, the
(n, JIMD , P, A)avg -polar code is an (n, M )-code (cf. Definition 1) which satisfies
tMD log n
1
log M ≥ Ip′X qY |X (X; Y ) − (1−γ)/β ,
n
n
−1
γh
2
P{ŴI 6= WI } ≤ (log n)n 2−n
( γβ+γ−1
)
γβ
(69)
,
(70)
and
n
n
P {X = x } =
n
Y
k=1
P {Xk = xk } =
n
Y
p′X (xk )
(71)
k=1
for all xn ∈ X n where p′X is the distribution on X as defined in (48).
Proof: The proposition follows from inspecting Proposition 11 and Definition 14 with the identifications N = log n and
P
MD
log M = m
|.
i=1 |Ji
The following theorem develops the tradeoff between the gap to capacity and the decay rate of the error probability for
(n, JIMD , P, A)peak -polar codes defined for the AWGN channel.
Theorem 2: Fix a γ ∈
1
1+β , 1
. There exists a constant t∗MD > 0 that depends on P and γ but not n such that the following
holds for any n = 2m where m ∈ N. There exists an (n, JIMD , P, A, ε)peak -polar code defined for the AWGN channel qY |X that
satisfies
N
t∗ log n
1 X MD
|Ji | ≥ C(P ) − MD
,
n i=1
n(1−γ)/β
February 28, 2018
(72)
DRAFT
17
and
−1
γh
2
P{ŴI 6= WI } ≤ (n log n + e3 )2−n
)
( γβ+γ−1
γβ
.
(73)
Proof: By Proposition 12 and Lemma 6, there exists a constant t∗MD > 0 that depends on P and γ but not n such that for
any n = 2m where m ∈ N, there exist an A and the corresponding (n, JIMD , P, A)avg -polar code that satisfies (72),
−1
γh2
P{ŴI 6= WI } ≤ (log n)n 2−n
)
( γβ+γ−1
γβ
,
(74)
and
P
(
n
1X 2
Xk > P
n
k=1
)
≤
e3
e
1 − 1−γ
β
n2
.
(75)
It remains to show (73). Using (74), (75) and Corollary 7, we conclude that the (n, JIMD , P, A)avg -polar code is an (n, JIMD , P, A, ε)peak polar code that satisfies
−1
γh2
ε , (n log n)2−n
≤ (n log n + e3 )2
( γβ+γ−1
)
γβ
+
e3
1 − 1−γ
β
n2
e
−1 γβ+γ−1
γh
)
−n 2 ( γβ
(76)
(77)
where the inequality follows from the fact that h2 (x) ≥ 2x for all x ∈ [0, 1/2]. This concludes the proof.
Remark 4: A candidate of A in Theorem 2 can been explicitly constructed according to Lemma 6 by the identification
A ≡ X ∪ {0− }.
VII. C ONCLUDING R EMARKS
In this paper, we provided an upper bound on the scaling exponent of polar codes for the AWGN channel (Theorem 1). In
addition, we have shown in Theorem 2 a moderate deviations result — namely, the existence of polar codes which obey a certain
tradeoff between the gap to capacity and the decay rate of the error probability for the AWGN channel.
Since the encoding and decoding complexities of the binary-input polar code for a BMC are O(n log n) as long as we allow
pseudorandom numbers to be shared between the encoder and the decoder for encoding and decoding the randomized frozen
bits (e.g., see [3, Sec. IX]), the encoding and decoding complexities of the polar codes for the AWGN channel defined in
Definition 14 and Definition 15 are O(n log n) × log n = O n log2 n . By a standard probabilistic argument, there must exist
a deterministic encoder for the frozen bits such that the decoding error of the polar code for the AWGN channel with the
deterministic encoder is no worse than the polar code with randomized frozen bits. In the future, it may be fruitful to develop
low-complexity algorithms for finding a good deterministic encoder for encoding the frozen bits. Another interesting direction for
future research is to compare the empirical performance between our polar codes in Definitions 14 and 15 and the state-of-the-art
polar codes. One may also explore various techniques (e.g., list decoding, cyclic redundancy check (CRC), etc.) to improve the
empirical performance of the polar codes constructed herein.
February 28, 2018
DRAFT
18
A PPENDIX A
P ROOF
OF
P ROPOSITION 3
Unless specified otherwise, all the probabilities in this proof are evaluated according to the distribution induced by the uniforminput (n, JI )-polar code. Consider
N
X
P {Ŵi 6= Wi } ∩ {Ŵ[i−1] = W[i−1] }
P{ŴI 6= WI } =
(78)
i=1
N
X
n
n
}
= X[i−1]
P {Ûin 6= Uin } ∩ {X̂[i−1]
=
(79)
i=1
N X
n
X
n
n
= X[i−1]
}
P {Ûi,k 6= Ui,k } ∩ {Ûik−1 = Uik−1 } ∩ {X̂[i−1]
=
(80)
i=1 k=1
n
n
where (79) is due to the fact by Definition 8 that {Ŵi = Wi } = {Ûin = Uin } and {Ŵ[i−1] = W[i−1] } = {X̂[i−1]
= X[i−1]
}. For
each i ∈ I and each k ∈ Ji , we have
n
n
}
= X[i−1]
P {Ûi,k 6= Ui,k } ∩ {Ûik−1 = Uik−1 } ∩ {X̂[i−1]
Z
X
X
X
≤
pUik ,X n ,Y n (uki , xn[i−1] , y n )
n
∈X[i−1]
ui,k ∈{0,1} uk−1 ∈U k−1 xn
[i−1]
i
i
n
× 1 pUi,k |U k−1 ,X n
i
≤2
X
n
[i−1]
X
n
uk−1
∈Uik−1 x[i−1] ∈X[i−1]
i
r
× pUi,k |U k−1 ,X n
i
[i−1]
,Y
Z
n
[i−1]
Yn
(ui,k |uik−1 , xn[i−1] , y n ) ≤ pUi,k |U k−1 ,X n
Yn
i
pU k−1 ,X n
i
[i−1]
[i−1]
k−1
, xn[i−1] , y n )
,Y n (ui
,Y
n
o
(1 − ui,k |uik−1 , xn[i−1] , y n ) dy n
k−1
, xn[i−1] , y n )pUi,k |U k−1 ,X n ,Y n (1|uik−1 , xn[i−1] , y n )dy n
,Y n (0|ui
i
[i−1]
n
= Z [pXI ;qY |XI ] (Ui,k |Uik−1 , X[i−1]
, Y n)
(81)
(82)
(83)
where (81) follows from (27). In addition, it follows from (25) and (27) that
n
n
P {Ûi,k 6= Ui,k } ∩ {Ûik−1 = Uik−1 } ∩ {X̂[i−1]
= X[i−1]
} =0
(84)
for each i ∈ I and each k ∈ Jic . Combining (80), (83) and (84), we obtain (29).
A PPENDIX B
P ROOF
OF
L EMMA 6
Let qY |X be the conditional distribution that characterizes the AWGN channel and fix a γ ∈ [0, 1). Recall the definitions of
p′X and sX in (48) and (50) respectively and recall that ΦX is the cdf of sX . Fix a sufficiently large n ≥ 36 that satisfies
1+P
1
∈ 0,
(85)
2
n(1−γ)/β
and
Φ−1
X
February 28, 2018
1
n1−(1−γ)/β
≤ −1.
(86)
DRAFT
19
In addition, recall the definition of X in (51) and let g : R → X be a quantization function such that
ℓ
g(t) , Φ−1
X
n
where ℓ ∈ {1, 2, . . . , n2 , . . . n − 1} is the unique integer that satisfies
Φ−1 ℓ ≤ t < Φ−1 ℓ+1
if t ≥ 0,
X
X
n
n
Φ−1 ℓ−1 < t ≤ Φ−1 ℓ if t < 0.
X
X
n
n
(87)
(88)
In words, g quantizes every a ∈ R to its nearest point in X whose magnitude is smaller than |a|. Let
X̂ , g(X)
(89)
n
o
PsX |X̂| ≤ |X| = 1,
(90)
be the quantized version of X. By construction,
PsX
1
ΦX (X̂) − ΦX (X) ≤
n
=1
(91)
and
PsX
1
ℓ+1
ℓ
−1
,
Φ
=
X ∈ Φ−1
X
X
n
n
n
(92)
for all ℓ ∈ {0, 1, . . . , n − 1}. It follows from (92) and the definition of p′X in (48) that
Ep′X [X 2 ] = EsX [X̂ 2 ]
(93)
and
Pp′X n
where sX n (xn ) ,
Qn
k=1 sX (xk ).
(
n
1X 2
Xk > P
n
k=1
)
= PsX n
(
n
1X 2
X̂k > P
n
k=1
)
Consequently, in order to show (52) and (53), it suffices to show
1
EsX [X̂ 2 ] ≤ P 1 − (1−γ)/β
n
(94)
(95)
and
PsX n
(
n
1X 2
X̂k > P
n
k=1
)
≤
e3
e
1 − 1−γ
β
n2
(96)
respectively. Using (90) and the definition of sX in (50), we obtain (95). In order to show (96), we consider the following chain
of inequalities:
)
n
1X 2
X̂k > P
PsX n
n
k=1
( n
)
1X 2
≤ PsX n
Xk > P
n
(
(97)
k=1
February 28, 2018
DRAFT
20
= PsX n
(
(
1−γ
1
Pn
2
√
n2− β
k=1 Xk
>
n
+
√
1
1
nP 1 − n(1−γ)/β
1 − n(1−γ)/β
≤ PsX n e
√
nP
Pn
2
k=1 Xk
1−
1
n(1−γ)/β
>e
√
1 − 1−γ
β
n+n 2
)
)
1 − 1−γ
−n/2 √
β
n+n 2
−
1
≤ 1− √
e
n/2
1 − 1−γ
n/2 √
β
n+n 2
−
1
= 1+ √
e
n/2 − 1
≤e
=e
√n
n−2
√
2 n
√
n−2
·e
−
√
· e−n
≤ e3 · e−n
1 − 1−γ
β
n+n 2
(98)
(99)
(100)
(101)
(102)
1 − 1−γ
2
β
(103)
1 − 1−γ
2
β
(104)
where
•
(97) is due to (90).
•
(100) is due to Markov’s inequality.
•
•
(102) is due to the fact that (1 + 1t )t ≤ e for all t > 0.
√
(104) is due to the assumption that n ≥ 6.
It remains to show (54). To this end, we let qZ denote the distribution of the standard normal random variable (cf. (1)) and
consider
1
C P − (1−γ)/β − Ip′X qY |X (X; Y )
n
1
= C P − (1−γ)/β − Ip′X qZ (X; X + Z)
n
= IsX qZ (X; X + Z) − Ip′X qZ (X; X + Z)
(105)
(106)
where (106) is due to the definition of sX in (50). In order to simplify the RHS of (106), we invoke [19, Corollary 4] and obtain
√
√
IsX qZ (X; X + Z) − Ip′X qZ (X; X + Z) ≤ (log e)(3 1 + P + 4 P )W2 (sY , p′Y ).
(107)
After some tedious calculations which will be elaborated after this proof, it can be shown that the Wasserstein distance in (107)
satisfies
W2 (sY , p′Y
)≤
s
κ log n
n
2(1−γ)
β
where
4P
κ , P + 4P +
log e
2
1 + log
,
r
(108)
P
2π
!
.
(109)
On the other hand, since
log(1 + P − ξ)
log(1 + P )
ξ
2ξ 2
≥
−
−
log e
log e
1+P
(1 + P )2
February 28, 2018
(110)
DRAFT
21
by Taylor’s theorem and
for each ξ ∈ 0, 1+P
2
C P−
1
n
1−γ
β
1
n(1−γ)/β
∈ 0, 1+P
by (85), we have
2
log e
≥ C(P ) −
2
1
n
1−γ
β
(1 + P )
+
n
2
2(1−γ)
β
(1 + P )2
!
.
(111)
Using (106), (107), (108) and (111), we obtain
r
√
√
log e
κ log n
C (P ) − Ip′X qY |X (X; Y ) ≤ (log e)(3 1 + P + 4 P )
+
2(1−γ)/β
2
n
1
n
1−γ
β
(1 + P )
2
+
n
2(1−γ)
β
(1 + P )2
!
.
(112)
Consequently, (54) holds for some constant t′ > 0 that does not depend on n.
D ERIVATION
OF
(108)
Consider the distribution (coupling) rX,X̂,Y,Ŷ defined as
rX,X̂,Y,Ŷ (x, x̂, y, ŷ) = sX (x)qZ (y − x)1{x̂ = g(x)}1{ŷ = g(x) + y − x}
and simplify the Wasserstein distance in (107) as follows:
Z Z
2
′
rY,Ŷ (y, ŷ)(y − ŷ)2 dŷdy
(W2 (sY , pY )) ≤
R R
Z Z
rX,X̂ (x, x̂)(x − x̂)2 dx̂dx
=
ZR R
sX (x)(x − g(x))2 dx
=
(113)
(114)
(115)
(116)
R
where
•
(114) follows from the definition of W2 in (6) and the facts due to (113) that rY = sY and rŶ = p′Y .
•
(115) follows from the fact due to (113) that PrX,X̂,Y,Ŷ {Y − Ŷ = X − X̂} = 1.
•
(116) is due to (113).
Following (116), we define ξn to be the positive number that satisfies
Z ∞
1
sX (x)dx = 1−(1−γ)/β
n
ξn
(117)
and consider
Z
R
sX (x)(x − g(x))2 dx
=
Z
−ξn
−∞
≤
Z
=2
−ξn
sX (x)(x − g(x))2 dx +
sX (x)x2 dx +
∞
ξn
ξn
−ξn
−∞
Z
Z
2
sX (x)x dx +
Z
ξn
−ξn
Z
ξn
−ξn
sX (x)(x − g(x))2 dx +
sX (x)(x − g(x))2 dx +
Z
∞
Z
∞
ξn
sX (x)(x − g(x))2 dx
sX (x)x2 dx
(118)
(119)
ξn
sX (x)(x − g(x))2 dx
(120)
where (119) follows from the fact due to (87) that x − g(x) ≥ 0 for all x ∈ R. In order to bound the first term in (120), we let
Pn , P 1 −
February 28, 2018
1
n(1−γ)/β
(121)
DRAFT
22
and consider
Z
∞
sX (x)x2 dx = Pn
Z
ξn
ξn
<
Z
∞
ξn
<
where
•
(122) follows from integration by parts.
•
(123) is due to the simple fact that
Z
∞
ξn
•
∞
sX (x)dx + ξn sX (ξn )
sX (x)dx 2Pn + ξn2
1
2P + ξn2
n1−(1−γ)/β
Z ∞
ξn2
Pn
sX (x)dx >
1 + 2 sX (x)dx
Pn + ξn2 ξn
x
Pn
ξn sX (ξn ).
=
Pn + ξn2
(122)
(123)
(124)
(125)
(126)
(124) involves using (117).
In order to bound the term in (124), we note that
ξn ≥ 1
(127)
by (86) and (117) and would like to obtain an upper bound on ξn through the following chain of inequalities:
Z ∞
1
sX (x)dx
=
n1−(1−γ)/β
ξn
Z ∞
xsX (x)dx
≤
(128)
(129)
ξn
= Pn sX (ξn )
r
2
ξn
Pn − 2P
n
=
e
2π
r
P − ξn2
e 2P
<
2π
(130)
(131)
(132)
where
•
(128) is due to (117).
•
(129) is due to (127).
Since
ξn2
2P
≤
log e
r !
1−γ
P
1−
log n + log
β
2π
(133)
by (132) and β = 4.714 > 3, it follows from (124) that
Z
∞
1
2
sX (x)x dx <
ξn
<
n1−
log n
n
February 28, 2018
1−γ
β
2(1−γ)
β
2P
2P +
log e
2P +
2P
log e
r !!
P
1−γ
1−
log n + log
β
2π
r !!
P
1 + log
.
2π
(134)
(135)
DRAFT
23
In order to bound the second term in (120), we consider
Z ξn
sX (x)(x − g(x))2 dx
−ξn
=
Z
ξn
−ξn
2
dx
sX (x) Φ−1 Φ(x) − Φ−1 Φ(g(x))
2
1
1
sX (x)
≤
dx
×
n sX (ξn )
−ξn
Z ξn
Pn2
sX (x)
≤
dx
2(1−γ)
−ξn
n β
P2
< 2(1−γ)
n β
Z
ξn
(136)
(137)
(138)
(139)
where
•
(137) is due to (91), the mean value theorem and the fact that the derivative of Φ is always positive and uniformly bounded
below by sX (ξn ) on the interval [−ξn , ξn ].
•
(138) is due to (130).
Combining (116), (120), (135) and (139) and recalling the definition of κ in (109), we obtain (108).
R EFERENCES
[1] T. M. Cover and J. A. Thomas, Elements of Information Theory.
Hoboken, NJ: John Wiley and Sons Inc., 2006.
[2] C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, pp. 379–423, 1948.
[3] E. Arıkan, “Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels,” IEEE Trans. Inf.
Theory, vol. 55, no. 7, pp. 1–23, Jul 2009.
[4] M. Mondelli, S. H. Hassani, and R. Urbanke, “Unified scaling of polar codes: Error exponent, scaling exponent, moderate deviations, and error floors,”
IEEE Trans. Inf. Theory, vol. 62, no. 12, pp. 6698 – 6712, 2016.
[5] S. H. Hassani, K. Alishahi, and R. Urbanke, “Finite-length scaling for polar codes,” IEEE Trans. Inf. Theory, vol. 60, no. 10, pp. 5875–5898, 2014.
[6] D. Goldin and D. Burshtein, “Improved bounds on the finite length scaling of polar codes,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 6966 – 6978,
2014.
[7] S. L. Fong and V. Y. F. Tan, “On the scaling exponent of polar codes for binary-input energy-harvesting channels,” IEEE J. Sel. Areas Commun., vol. 34,
no. 12, pp. 3540 – 3551, 2016.
[8] E. Şaşoğlu, I. Telatar, and E. Arıkan, “Polarization of arbitrary discrete memoryless channels,” in Proc. IEEE Inf. Theory Workshop, Seoul, Korea, October
2009, pp. 114–118.
[9] D. Sutter, J. M. Renes, F. Dupuis, and R. Renner, “Achieving the capacity of any DMC using only polar codes,” in Proc. IEEE Inf. Theory Workshop,
Lausanne, Switzerland, September 2012, pp. 114–118.
[10] J. Honda and H. Yamamoto, “Polar coding without alphabet extension for asymmetric models,” IEEE Trans. Inf. Theory, vol. 59, no. 12, pp. 7829–7838,
2013.
[11] M. Mondelli, R. Urbanke, and S. H. Hassani, “How to Achieve the Capacity of Asymmetric Channels,” in Proc. Allerton Conference on Communication,
Control and Computing, October 2014, pp. 789–796.
[12] E. Abbe and A. Barron, “Polar coding schemes for the AWGN channel,” in Proc. IEEE Int. Symp. Inf. Theory, St Petersburg, Russia, July 2011, pp.
194–198.
[13] E. Abbe and E. Telatar, “Polar codes for the m-user multiple access channel,” IEEE Trans. Inf. Theory, vol. 58, no. 8, pp. 5437 – 5448, 2012.
[14] Y. Altuğ and A. B. Wagner, “Moderate deviations in channel coding,” IEEE Trans. Inf. Theory, vol. 20, no. 8, pp. 4417 – 4426, 2014.
[15] H. Mahdavifar, M. El-Khamy, J. Lee, and I. Kang, “Achieving the uniform rate region of general multiple access channels by polar coding,” IEEE Trans.
Commun., vol. 64, no. 2, pp. 467 – 478, 2016.
[16] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56, no. 5, pp. 2307–2359,
2010.
February 28, 2018
DRAFT
24
[17] V. Strassen, “Asymptotische abschätzungen in Shannons informationstheorie,” Trans. Third Prague Conf. Inf. Theory, pp. 689–723, 1962,
http://www.math.cornell.edu/~pmlut/strassen.pdf.
[18] M. Hayashi, “Information spectrum approach to second-order coding rate in channel coding,” IEEE Trans. Inf. Theory, vol. 55, no. 11, pp. 4947–4966,
2009.
[19] Y. Polyanskiy and Y. Wu, “Wasserstein continuity of entropy and outer bounds for interference channels,” IEEE Trans. Inf. Theory, vol. 62, no. 7, pp.
3992 – 4002, 2016.
February 28, 2018
DRAFT
| 7cs.IT
|
THIS VERSION: 28 NOVEMBER 2017
1
Performance Measures in Electric Power Networks
under Line Contingencies
arXiv:1711.10348v1 [cs.SY] 28 Nov 2017
Tommaso Coletta, and Philippe Jacquod, Member, IEEE
Abstract—Classes of performance measures expressed in terms
of H2 -norms have been recently introduced to quantify the
response of coupled dynamical systems to external perturbations.
For the specific case of electric power networks, these measures
quantify for instance the primary effort control to restore
synchrony, the amount of additional power that is ohmically
dissipated during the transient following the perturbation or,
more conceptually, the coherence of the synchronous state. So far,
investigations of these performance measures have been restricted
to nodal perturbations. Here, we go beyond earlier works and
consider the equally important, but so far neglected case of line
perturbations. We consider a network-reduced power system,
where a Kron reduction has eliminated passive buses. Identifying
the effect that a line fault in the physical network has on the
Kron-reduced network, we find that performance metrics depend
on whether the faulted line connects two passive, two active buses
or one active to one passive bus. In all cases, performance metrics
depend quadratically on the original load on the faulted line times
a topology dependent factor. Our theoretical formalism being
restricted to Dirac-δ perturbations, we investigate numerically
the validity of our results for finite-time line faults. We find good
agreement with theoretical predictions for longer fault durations
in systems with more inertia.
Index Terms—Power generation control, power system dynamics performance measures, line contingency.
I. I NTRODUCTION
In normal operation, electric power grids are synchronized.
Their operating state corresponds to equal frequencies and
voltage angle differences ensuring power conservation at all
buses. Such synchronous states are not specific to power grids.
They occur in many different coupled dynamical systems,
depending on the balance between the internal dynamics of the
individual systems and the coupling between them [1], [2]. For
the specific case of electric power grids, the individual systems
are either generators or loads, and they are coupled to one
another by power lines. Individual units have effective internal
dynamics determined by their nature – they may be rotating
machines, inertialess new renewable energy sources, load
impedances and so forth – and by the amount of power they
generate or consume [3]. Rapid changes are currently affecting
the structure of power grids which will no doubt impact
their operating states [4]. With higher penetration levels of
renewable energy sources, generations become decentralized,
they have less inertia and they fluctuate more strongly [5].
Future power grids will be subjected more often to stronger
external perturbations to which they may react more strongly.
T. Coletta and Ph. Jacquod are with the School of Engineering of
the University of Applied Sciences of Western Switzerland CH-1951 Sion,
Switzerland. Emails: (tommaso.coletta, philippe.jacquod)@hevs.ch.
There is thus a clear need to better assess power grid vulnerability. Investigations have been initiated on the robustness
of the synchronous operating state of electric power grids to
external perturbations. An approach has been advocated in
consensus and synchronization studies [6]–[9], which starts
from a stable operating state, perturbs it and quantifies the
magnitude of the induced transient excursion through various
performance measures. Focusing on Dirac-δ, nodal perturbations – instantaneous changes in generation or consumption –
quadratic performance measures have been proposed, which
can be formulated as L2 and squared H2 norms of linear
systems. The approach is mathematically elegant because these
norms can be conveniently expressed in terms of observability
Gramians [10]. Exported to electric power grids, performance
measures allow to evaluate additional transmission losses incurred during the transient as synchronous machines oscillate
relative to one another, [11], [12], or the primary control effort
necessary to restore synchrony [13]. Under the assumptions
that synchronous machines have uniform inertia and damping
coefficients, and that transmission lines are homogeneous –
that they have constant resistance over reactance ratios –
Ref. [11] obtained that additional transmission losses depend
only on the number of buses in the network, and not on
its topology. Relaxing the homogeneous line assumption,
Ref. [12] related the same performance measure to a graph theoretical distance metric known as the resistance distance [14],
[15]. To the best of our knowledge, investigations of performance measures of synchronized states have considered
nodal perturbations only. In this manuscript, we extend these
investigations to line contingencies which are at least as
important for evaluating the robustness of electric power grids.
The main contribution of our work is to extend the observability Gramian formalism to assess performance measures
for transients caused by line contingencies. In this case, the
perturbation acts on the network Laplacian matrix, it is thus
a multiplicative perturbation, a priori fundamentally different
and harder to treat than the additive nodal perturbations
considered so far. A second difficulty we overcome is that
analytical results can be obtained only for uniform inertia to
damping ratio, forcing us to consider Kron-reduced networks
without inertialess passive nodes. To consider relevant line
contingencies, we therefore need to map single-line faults in
the physical network onto the Kron-reduced network. Our
results differentiate between faults on power lines connecting
two passive nodes, two active nodes, or one passive to one
active node in the physical network. Because we consider
Dirac-δ perturbations, we compare our theory to numerical
simulations with finite-time line faults. We confirm our ana-
THIS VERSION: 28 NOVEMBER 2017
lytical results for finite-time fault lasting typically up to few
AC cycles. We find that the agreement between theoretical
Dirac-δ and numerical perturbations is better for longer fault
durations in systems with larger inertias.
The works mentioned above focus on performance measures that are relevant to AC electric power networks and
which can be computed analytically as H2 norms for a statespace system. Generally, the observability Gramian required
to evaluate H2 norms is defined implicitly by a Lyapunov
equation which is typically solved numerically. A secondary
contribution of our work is to derive an analytic solution of the
Lypaunov equation valid for generic performance measures,
under the assumption that synchronous machines have uniform
damping over inertia ratios. By revisiting some of the results
of Refs. [11], [12] and [13], we show how specific assumptions
lead to performance measures that no longer depend on the
network topology, and clarify the mathematical mechanism
by which this occurs. Our results allow to compute other
performance measures and to interpret them in terms of the
average resistance distance [14], [15] a quantity also known
as the inverse closeness centrality [16].
This paper is organized as follows. Section II introduces
the mathematical notations and defines the effective resistance
distance. The high voltage AC electric network model, and the
observability Gramian formalism are outlined in Sec. III. Section IV derives a closed-form expression for the observability
Gramian in general cases. The broad range of applicability
of our results for the observability Gramian is illustrated
by evaluating a variety of network performance measures in
Sec. V. The new application of the Gramian formalism to
line contingencies is discussed in Sec. VI and is supported
by the numerical simulations presented in Sec. VII. A brief
conclusion is given in Sec. VIII.
II. M ATHEMATICAL NOTATION
Given the vector v ∈ RN and the matrix M ∈ RN ×N we
denote their transpose by v > and M > . For any two vectors
u, v ∈ RN , uv > is the matrix in RN ×N having as i, j th
component the scalar ui vj and diag({vi }) denotes the diagonal
matrix having v1 , . . . , vN as diagonal entries. Let êl ∈ RN
with l ∈ {1, . . . , N } denote the unit vector with components
(êl )i = δil . We define e(l,q) ∈ RN as e(l,q) = êl − êq .
We denote undirected weighted graphs by G = (N , E, W)
where N is the set of its N vertices, E is the set of
edges, and W = {wij } is the set of edge weights, with
wij = 0 whenever i and j are not connected by an edge, and
wij = wji > 0 otherwise. The graph Laplacian
L ∈ RN ×N
P
is the symmetric matrix given by L = i<j wij e(i,j) e>
(i,j) .
We denote by {λ1 , . . . λN } and {u(1) , . . . , u(N ) } the eigenvalues and orthonormalized eigenvectors of L. The zero row
and column sum property
√ of L implies that λ1 = 0 and
that u(1) = (1, . . . , 1)/ N . All remaining eigenvalues of
L are strictly positive in connected graphs λi > 0 for
i = 2, . . . , N . The orthogonal matrix T ∈ RN ×N having
u(i) as ith column diagonalizes L, i.e. T > LT = Λ where
Λ = diag({λi }). The Moore-Penrose pseudoinverse of L is
-1
>
given by L† = T diag({0, λ-1
and is such that
2 , . . . , λN })T
2
>
LL† = L† L = I − u(1) u(1) with I ∈ RN ×N denoting the
identity matrix.
The effective resistance distance between any two nodes
†
i and j of the network is defined as Ωij = e>
(i,j) L e(i,j)
[14], [15]. This quantity is a graph theoretical distance metric
satisfying the properties: i) Ωii = 0 ∀i ∈ N , ii) Ωij ≥ 0 ∀i 6=
j ∈ N , and iii) Ωij ≤ Ωik + Ωkj ∀i, j, k ∈ N . It is known
as the resistance distance because if one replaces the edges of
G by resistors with a conductance 1/Rij = wij , then Ωij is
equal to the equivalent network resistance when a current is
injected at node i and extracted at node j with no injection
anywhere else. The expression of the resistance distance in
terms of the eigenvalues and eigenvectors of L is given by
P
(l)
(l) 2
Ωij = l≥2 λ-1
l (ui − uj ) [17]–[19].
III. P OWER N ETWORK M ODEL AND QUADRATIC
PERFORMANCE MEASURES
We consider high voltage transmission power networks
in the DC approximation (i.e. constant voltage magnitudes,
purely susceptive transmission lines and small voltage phase
differences). The steady state power flow equations relating
the active power injections P to the voltage phases θ at
every node read P = Lb θ. Here, Lb is the Laplacian matrix
of the graph modeling the electric network and whose edge
weights are given by the susceptances of the transmission lines
wij = bij ≥ 0. We assume that each node of the network has
a synchronous machine (generator or consumer) of rotational
inertia mi > 0 and damping coefficient di > 0. The network
dynamics is governed by the swing equations [3]. In the frame
rotating at the nominal frequency of the network they read
M θ̈ = −D θ̇ + P − Lb θ,
(1)
with M = diag({mi }) and D = diag({di }). Subject to a
power injection perturbation δp(t), the system deviates from
the nominal operating point (θ ? , ω) := (L†b P , 0) according to
θ(t) = θ ? + δθ(t), and ω(t) = δ θ̇(t). The swing dynamics
is determined by
0
0
I
δθ
δ θ̇ =
+
. (2)
M -1 δp
−M -1 Lb −M -1 D
ω
ω̇
Using δθ = M 1/2 δθ and ω = M 1/2 ω we rewrite Eq. (2) as
˙
0
0
I
δθ +
δθ
=
,
M -1/2 δp
−M -1/2 Lb M -1/2 −M -1 D
ω
ω̇
|
{z
}
A
(3)
which symmetrizes the four blocks of the stability matrix A.
Eqs. (2) and (3) capture the transient dynamics resulting from
the perturbation δp(t). For asymptotically stable systems and
perturbations that are short and weak enough that they leave
the dynamics inside the basin of attraction of θ ? , the operating
point will eventually return to (δθ, ω) = (0, 0).
We want to characterize the transient by evaluating quadratic
performance measures of the type
Z ∞
(1,1)
δθ
Q
0
>
>
P=
δθ ω Q ω dt , Q =
0
Q(2,2)
0
(4)
THIS VERSION: 28 NOVEMBER 2017
3
where we assumed that δp(t) is such that δp(t < 0) = 0 and
the symmetric matrix Q ∈ R2N ×2N depends on the specific
performance measure to investigate. For Dirac-δ perturbations
δp(t) = δ(t)p0 and initial conditions (δθ(0), ω(0)) = (0, 0),
Eq. (3) is explicitly solved yielding
0
δθ(t)
= eAt
.
(5)
M -1/2 p0
ω(t)
{z
}
|
B
The performance measure P, Eq. (4), can be expressed as
P = B > XB ,
(6)
R ∞ A> t M At
with the observability Gramian X = 0 e
Q e dt, and
-1/2 (1,1)
Q
M -1/2
0
QM = M
. (7)
0
M -1/2 Q(2,2) M -1/2
For asymptotically stable systems (i.e. when all eigenvalues
of A have negative real part), the observability Gramian X
satisfies the Lyapunov equation
A> X + XA = −QM .
(8)
P
P
For Laplacian systems, i (Lb )ij = j (Lb )ij = 0, therefore
A has a marginally stable mode A[M 1/2 u(1) , 0]> = 0.
The standard approach to deal with this marginally stable
mode is to consider performance measures Q such that
u(1) ∈ ker(Q(1,1) ), in which case the observability Gramian
is well defined by Eq. (8) with the additional constraint
X[M 1/2 u(1) , 0]> = 0 [13]. In this work, we propose a new
approach to treat the marginally stable mode by introducing a
regularizing parameter in the Laplacian making it nonsingular.
Proposition 1. The Laplacian Lb under the transformation
Lb → Lb + I with > 0 is non singular and its inverse is
-1
-1
-1
>
L-1
b = T diag({ , (λ2 + ) , . . . , (λN + ) })T ,
(9)
where λi ’s and T are the eigenvalues and the orthogonal
matrix diagonalizing Lb for = 0.
p
G = (A, B, QM ). The squared H2 norm kGk2H2 measures
the sum of the system’s
responses to Dirac-δ impulses at every
P
node, kGk2H2 = i P (i) , where P (i) is the L2 norm of the
system’s output for a Dirac-δ impulse at node i (i.e. δp(i) (t) =
δ(t)êi for i = 1, . . . , N ). This quantity is easy to evaluate in
terms of the observability Gramian and is given by the trace
kGk2H2 = Tr[B > XB] .
In the formalism we just outlined, the difficulty of evaluating
performance measures resides in solving the Lyapunov equation for the observability Gramian X. Despite some specific
choices of Q for which analytical solutions have been found –
see in particular Refs. [11]–[13] for Q corresponding to resistive losses and primary control effort – this task is generally
performed numerically. In the case of generic performance
measures little is known about the solutions of the Lyapunov
equation. The next section fills this gap by explicitly deriving
an analytical expression for the observability Gramian in terms
of the eigenvectors of M -1/2 Lb M -1/2 . We then show how our
results allow to relate performance measures of the type P of
Eq. (6) to the network topology.
IV. C LOSED FORM EXPRESSION FOR THE O BSERVABILITY
G RAMIAN
Proposition 3. Let A be a non symmetric, diagonalizable
matrix with eigenvalues µi 6= 0. Let TR (TL ) denote the
matrix whose columns (rows) are the right (left) eigenvectors
of A. The observability Gramian X, solution of the Lyapunov
Eq. (8) is given by
X −1
(TL )li (TL )qj TR> QM TR lq .
(11)
Xij =
µl + µq
l,q
Proof. By definition, TL and TR fulfill the bi-orthogonality
condition TL TR = I, thus TL ATR = µ with µ = diag({µi }).
Using this transformation in Eq. (8) one has
µX + Xµ = −TR> QM TR ,
X = TR> XTR ,
Proof. The effect of Lb → Lb + I is to shift all eigenvalues
of Lb by (i.e. λi → λi + ) but to leave all the eigenvectors
u(l) unchanged. For > 0 Lb is non singular and one readily
obtains Eq. (9) for its inverse.
which yields
Proposition 2. Under the transformation Lb → Lb + I with
> 0, the system defined by Eq. (3) is asymptotically stable
and has no marginally stable mode.
Finally, using X = TL> XTL one obtains Eq. (11).
Proof. For > 0, Lb + I is positive definite. Under this
condition, Ref. [20] showed that all eigenvalues of A have a
negative real part.
Under the transformation of Proposition 2, Eq. (8) is sufficient to define the observability Gramian with no additional
constraint. In this approach we take the limit → 0 at the
end of the calculation of a performance measure to recover
the physically relevant quantities.
The quadratic performance
measure P can be expressed as
R∞
the L2 norm Pp= 0 y > (t)y(t) dt ≡ kykL2 of the system’s
output y(t) = QM [δθ(t), ω(t)]> . Eqs. (3) and (5) together
with y(t) define an input/output system which we denote by
(10)
X lq =
−1
(T > QM TR )lq .
µl + µq R
(12)
(13)
Since Proposition 3 holds for µi 6= 0 ∀i, it is satisfied
by the matrix A defined in Eq. (3) under the transformation
of Proposition 2. We note that Eqs. (11) and (13) are ill
defined if one takes the limit → 0. In what follows we
assume 6= 0 and take the limit → 0 only after computing
performance measures. Next we relate the explicit expression
of the observability Gramian, Eq. (11), to the eigenvectors of
M -1/2 Lb M -1/2 .
Assumption 1. All synchronous machines have uniform damping over inertia ratios di /mi = γ ∀i.
Proposition 4. Consider the power system model defined in
Eq. (3) under Assumption 1. The left and right transformation
THIS VERSION: 28 NOVEMBER 2017
4
matrices TL and TR diagonalizing A are related to the eigenvectors of M -1/2 Lb M -1/2 through the linear transformations
given in Eqs. (19) and (20).
Proof. We assume di /mi = γ ∀i, while still allowing inertias
to be different from one synchronous machine to the other.
Under this assumption one has that M -1 D = γI. Thus
M -1 D and M -1/2 Lb M -1/2 commute and have a common
basis of eigenvectors. Since M -1/2 Lb M -1/2 is symmetric, it
has a real spectrum with eigenvalues denoted by λM
i , and it is
diagonalized by an orthogonal matrix TM
(14)
TM> M -1/2 Lb M -1/2 TM = ΛM := diag({λM
i }) .
The transformation
>
0
I
TM 0
TM 0
A
=
−ΛM −γI
0 TM
0 TM>
(15)
leads after index reordering to a matrix with a block diagonal
structure composed of 2 × 2 blocks of the form
0
1
.
(16)
−λM
i −γ
Diagonalizing the 2 × 2 blocks, Eq. (16), one obtains the
eigenvalues of A,
q
1
(−γ
±
Γ
)
,
Γ
=
γ 2 − 4λM
(17)
µ±
=
i
i
i ,
i
2
for i = 1, . . . , N . From the right and left eigenvectors of
Eq. (16) one readily obtains the full transformation which
diagonalizes A,
diag({µ+
})
0
i
TL ATR =
,
(18)
0
diag({µ−
i })
with TL TR = I,
p
p
diag({i/ p
Γj })
diag({1/ pΓj })
TM 0
,
TR =
−
0 TM diag({µ+
j / Γj }) diag({iµj / Γj })
(19)
and
p
p
>
diag({−µ−
TM 0
j / pΓj }) diag({1/p Γj })
TL =
.
0 TM>
diag({−iµ+
j / Γj }) diag({i/ Γj })
(20)
Eqs. (19) and (20) relate the eigenvectors of A to those
of M -1/2 Lb M -1/2 . This allows to express the observability Gramian of Eq. (11) in terms of the eigenvectors of
M -1/2 Lb M -1/2 .
The linearity of the Lyapunov Eq. (8) with respect to both
X and QM implies that for performance measures involving
both frequency and voltage phase degrees of freedom [i.e. both
Q(1,1) 6= 0 and Q(2,2) 6= 0 in Eq. (4)], the observability
Gramian is given by a linear combination of two observability Gramians, each obtained solving a separate Lyapunov
equation. Thus, without loss of generality, we address separately two classes of performance measures: those involving
frequency degrees of freedom only and those involving phase
degrees of freedom only. We note that from Eqs. (5), (6) and
(10) it is clear that only the X (2,2) block of the observability
Gramian is relevant to evaluate L2 and squared H2 norms.
Proposition 5 (Observability Gramian for frequency based
performance measures). Consider the power system model
defined in Eq. (3) and satisfying Proposition 2. Under Assumption 1, the X (2,2) block of the observability Gramian
associated with the quadratic performance measure defined in
Eq. (4) with Q(1,1) = 0, and Q(2,2) 6= 0 is given by
(2,2)
Xij
=
N
X
(TM )il (TM> )qj (TM> M -1/2 Q(2,2) M -1/2 TM )lq
l,q=1
#
M
γ(λM
l + λq )
,
×
M 2
M
M
2γ 2 (λM
l + λq ) + (λq − λl )
"
(21)
where λM
l and TM are the eigenvalues and the orthogonal
matrix diagonalizing M -1/2 Lb M -1/2 respectively.
Proof. Inserting Eqs. (19) and (20) into Eq. (11), under the
assumption that Q(1,1) = 0 and taking the indices i , j ∈ {N +
1, . . . , 2N } to access X (2,2) yields
(2,2)
Xij
N
(2,2)
X
(TM> )li (TM> )qj (TM )nl (TM )pq Qnp
√
mn mp Γl Γq
=
l,q=1
n,p=1
×
"
−
µ+
q µl
+
µ−
l + µq
+
+
µ−
q µl
−
µ+
l + µq
−
+
µ+
q µl
+
µ+
l + µq
−
−
µ−
q µl
−
µ−
l + µq
#
,
which simplifies to Eq. (21) using Eq. (17).
Proposition 6 (Observability Gramian for phase based
performance measures). Consider the power system model
defined in Eq. (3) and satisfying Proposition 2. Under Assumption 1, the X (2,2) block of the observability Gramian
associated with the quadratic performance measure defined in
Eq. (4) with Q(1,1) 6= 0, and Q(2,2) = 0 is given by
(2,2)
Xij
=
N
X
(TM )il (TM> )qj (TM> M -1/2 Q(1,1) M -1/2 TM )lq
l,q=1
"
#
2γ
×
,
M 2
M
M
2γ 2 (λM
l + λq ) + (λq − λl )
(22)
where λM
l and TM are the eigenvalues and the orthogonal
matrix diagonalizing M -1/2 Lb M -1/2 respectively.
Proof. Inserting Eqs. (19) and (20) into Eq. (11), under the
assumption that Q(2,2) = 0 and taking the indices i , j ∈ {N +
1, . . . , 2N } to access X (2,2) yields
(2,2)
Xij
=
N
(1,1)
X
(TM> )li (TM> )qj (TM )nl (TM )pq Qnp
√
mn mp Γl Γq
l,q=1
n,p=1
×
1
1
1
1
+
−
−
+
−
+
− ,
µ−
µ+
µ+
µ−
l + µq
l + µq
l + µq
l + µq
which simplifies to Eq. (22) using Eq. (17).
Eqs. (21) and (22) are explicit expressions for the observability Gramian valid for generic performance measures.
Similar expressions were recently derived in Ref. [21] working
in the Laplace frequency domain. For performance measures
involving both frequency and phase degrees of freedom, the
THIS VERSION: 28 NOVEMBER 2017
5
observability Gramian is given by a linear combination of
Eqs. (21) and (22). We note that the results of Propositions 4,
5, and 6 obtained for the model defined in Eq. (1) generalize
to a different class of 2nd P
order coupled oscillator models
involving relative damping j −dij (θ̇i − θ̇j ) as opposed to
absolute damping −di θ̇i . Such models have been used to
describe vehicle platoon formations with relative position and
relative velocity feedback [6], [12], [22]. The corresponding
observability Gramians are given in Appendix IX.
Computing Tr(X (2,2) ) from Eqs. (22) and (25) yields
X
X
>
Tr[X (2,2) ]=
gαβ
(T )il (T > )qi (T > e(α,β) e(α,β) T )lq
hαβi
hαβi
V. P ERFORMANCE MEASURES
We next illustrate how our general expressions for the
observability Gramian, Eqs. (21) and (22), can be applied to a
variety of performance measures. To do that we first reproduce
the results of Refs. [11], [12] and [13]. As a side product our
results clarify the mathematical mechanism by which specific
parameter choices lead to performance measures P or squared
H2 norms that no longer depend on the network topology, a
somehow surprising observation in Refs. [11] and [13].
1) Primary control effort: Ref.
the priR ∞[13]
P defines
2
mary control effort as P = 0
d
ω
dt.
Injecting
i
i
i
Q(1,1) = 0, Q(2,2) = D into Eq. (21), and recalling that
M -1/2 DM -1/2 = γI, gives
#
"
N
M
X
(TM )il (TM> )qj (TM> TM )lq γ 2 (λM
l + λq )
. (23)
M 2
M
M
2γ 2 (λM
l + λq ) + (λq − λl )
l,q=1
Using TM> TM = TM TM> = I, Eq. (23) simplifies to
X (2,2) =
1
I,
2
and leads to the squared H2 norm kGk2H2 = 1/2
agreement with Ref. [13].
X
hαβi
where Ωαβ is the effective resistance distance between nodes
α and β computed with respect to Lb . In the limit of homogeneous lines (i.e. gαβ /bαβ ≡ r/x for all lines), Eq. (28) further
simplifies to kGk2H2 = (1/2d)(r/x)(N −1) in agreement with
Ref. [11]. Remarkably, kGk2H2 depends on the total number
N of nodes but not on the network topology in that case.
coherence: The performance measure P =
R ∞3) Phase
>
δθ
δθ
dt
quantifies the integrated global voltage phase
0
deviation from the nominal operating point during a transient.
We evaluate it under the assumption of uniform inertias (i.e.
TM ≡ T and λM
i = λi /m) and for transients induced by
a power injection pulse. In this case, we have Q(1,1) = I,
Q(2,2) = 0, δp(t) = δ(t)ês and B = [0, m-1/2 ês ]> for a
power injection impulse at node s. The observability Gramian,
Eq. (22) becomes
(2,2)
P
i
gαβ e(α,β) e>
(α,β) .
(25)
In Eq. (25), hα, βi denotes all pairs of nodes connected
by a resistive line of conductance gαβ . For uniform inertias, M -1/2 Lb M -1/2 and Lb have the same eigenvectors,
TM ≡ T , with eigenvalues differing by a division with m,
λM
i ≡ λi /m ∀i. The uniform inertia assumption also simplifies
the computation of the squared H2 norm to
kGk2H2 = Tr[B > XB] = m−1 Tr[X (2,2) ] .
Xij
=
where
L-1
b
m−1
i , in
hα,βi
(26)
(27)
l≥2
√
(1)
where we used T T > = I and ui = 1/ N ∀i to drop
the l = 1 term in the sum before taking the limit → 0.
Combining Eqs. (26) and (27) we finally reproduce the result
of Ref. [12]
1 X
gαβ Ωαβ ,
(28)
kGk2H2 =
2d
(24)
2) Transmission losses: Under the assumption of uniform
inertias (i.e. mi ≡ m, and di = mi γ ≡ d ∀i), Ref. [12]
evaluates the transmission losses incurred during a transient
by computing the squared H2 norm of the system, for the
performance measure Q(1,1) = Lg , Q(2,2) = 0 with
Lg =
l=1
N
X
1 X
(l) 2
(l)
gαβ
λ−1
=
l (uα − uβ ) ,
2γ
hαβi
(2,2)
Xij =
i,l,q
2γ
×
2γ 2 (λl + λq ) + (λq − λl )2 /m
N
X
1 X
2
gαβ
=
λ−1
l [(T )αl − (T )βl ] ,
2γ
N
1 X −1
1 -1
λl (T )il (T > )lj ⇒ X (2,2) =
L , (29)
2γ
2γ b
l=1
is defined in Proposition 1. We obtain
N
1 -1 (1) 2 X
-1 (l) 2
(s)
P =
us +
(λl + ) us
.
2d
(30)
l≥2
The superscript in P (s) specifies that we are computing the
response to a perturbation occurring at node s and u(l) is the
lth eigenvector of Lb . The marginally stable mode of A implies
that this quantity diverges as the regularizing parameter → 0.
We instead compute the difference P (s) − P (t) for which the
two divergences cancel
X
X
1
(s)
(t)
P −P =
Ωsi −
Ωti ,
(31)
2dN
i
i
√
P
(1)
where we have used that ui = 1/ N ∀i, and that i Ωsi =
P
P
P
(l) 2
−1
N l≥2 λ−1
+ l≥2 λ−1
i Ωsi
l us
l . The quantity N
entering Eq. (31) is the average resistance distance separating
node s from the rest of the network. Its inverse is known as the
closeness centrality [16]. From Eq. (31) one concludes that the
THIS VERSION: 28 NOVEMBER 2017
6
difference P (s) − P (t) is positive (negative) if the centrality of
node t is greater (smaller) than that of node s. In other terms,
power fluctuations occurring at nodes which are more central
have a smaller impact on the voltage phase fluctuations during
the transient.
The examples discussed above illustrate how for specific parameter choices the expressions for the performance measures
simplify because of the orthogonality of the eigenvectors of
Lb . This sometimes leads to results that no longer depend on
the topology of the network. For instance, this occurs for i)
the transmission losses measure, Eq. (25), when ii) inertias are
assumed uniform, for iii) Dirac-δ perturbations averaged over
all nodes of the network as captured by the squared H2 norm,
and iv) for lines having constant susceptance over conductance
ratio. If any one of these four assumptions is relaxed, the
performance measure depends nontrivially on the topology of
the network.
VI. P ERFORMANCE MEASURES UNDER LINE
CONTINGENCIES : F ORMALISM
The formalism outlined above can be applied to more
general perturbations than the power injection fluctuations
considered so far. We consider nonsingular single line faults
that do not split the network into two disconnected parts, and
introduce a time dependent network Laplacian
Lb (t) = Lb − δ(t)bαβ e(α,β) e>
(α,β) ,
(32)
which describes an infinitesimally short fault of the α − β line
at t = 0.
The results presented so far rely on the assumption that all
nodes of the network are governed by the swing dynamics of
Eq. (1), i.e. all nodes have synchronous machines with inertia.
In real power networks however, some so-called passive nodes
are inertialess. Kron reduction allows to eliminate passive
nodes and to formulate the dynamics of the network in terms of
a swing equation, similar to Eq. (1), involving only the voltage
phases of the synchronous machines, all with a finite inertia,
on an effective network. For a review of Kron reduction, we
refer the reader to Refs. [23] and [24].
To characterize the transient behavior resulting from a line
contingency, we first need to formulate how a line fault in
the physical network, Eq. (32), impacts the swing dynamics
of the Kron reduced network of synchronous machines. If we
denote by Ng = {1, . . . , g} and Nc = {g + 1, . . . , N } the
node subsets representing synchronous machines and passive
nodes respectively, we rewrite the DC power flow equation in
the physical network as
(g,g) (g,c) ?
θg
Pg
Lb
Lb
=
,
(33)
(c,g)
(c,c)
θc?
Pc
L
L
b
(g,g)
(g,c)
(c,g)
M θ̈g = −D θ̇g + Pred − Lred θg ,
(35)
with M = diag({mi }) and D = diag({di }) for i ∈ Ng .
Because Kron reduction has eliminated all passive nodes, all
remaining nodes have inertia.
When discussing transients resulting from line contingencies, it is natural to question how a line fault in the physical
network, Eq. (32), affects the swing dynamics of the Kron
reduced network Eq. (35). This analysis requires to distinguish
three cases:
1) The faulted line connects two synchronous machines:
In this case α, β ∈ Ng and the fault, Eq. (32), only affects the
(g,g)
Lb
block of the Laplacian Lb . In terms of Lred and Pred
the fault is described by
Lred → Lred − δ(t)bαβ e(α,β) e>
(α,β) , (36)
Pred → Pred ,
where e(α,β) ∈ R|Ng | . The swing equation (35), relative to the
nominal operating point θg (t) = θg? + δθ g (t), becomes
?
M δ θ̈ g = −Dδ θ̇ g −Lred δθ g +δ(t)bαβ e(α,β) e>
(α,β) (θg +δθg ) .
(37)
With the initial condition (δθg (0), ωg (0)) = (0, 0), the solution to Eq. (37) is
0
δθ g (t)
= eAt M -1/2 b e
>
? ,
αβ (α,β) e(α,β) θg
ω g (t)
(38)
|
{z
}
B
where δθ g = M 1/2 δθ g , ω g = M 1/2 ωg , and A is the matrix
defined in Eq. (3) with Lb appropriately replaced by Lred .
2) The faulted line connects two passive nodes: In this case
(c,c)
α, β ∈ Nc and the fault only affects the Lb
block of the
(g,g)
(g,c)
(c,g)
Laplacian Lb while the blocks Lb , Lb
and Lb
remain
unchanged. This impacts both Lred and Pred which become
(g,c)
Pred − δ(t)bαβ
Lred −
Lb
(c,c) -1
[Lb
(c,c) -1
] e(α,β) e>
(α,β) [Lb
] Pc
(c,c)
Lred
(34)
,
(c,c) -1
1 − bαβ e>
] e(α,β)
(α,β) [Lb
(g,c)
(c,c)
(c,c) -1 (c,g)
Lb [Lb ]-1 e(α,β) e>
] Lb
(α,β) [Lb
δ(t)bαβ
(c,c)
1 − bαβ e>
]-1 e(α,β)
(α,β) [Lb
,
(39)
b
where Lb , Lb , Lb
and Lb are blocks of the Laplacian Lb . Applying Kron reduction to Eq. (33) to eliminate θc?
yields
h
i
(g,g)
(g,c) (c,c) -1 (c,g)
(g,c) (c,c) -1
− Lb Lb
Lb
θg? .
Pg − Lb Lb
Pc = Lb
{z
} |
|
{z
}
Pred
Starting from the physical injections Pg , Pc and the physical
network with Laplacian Lb , Eq. (34) defines an effective
vector of power injections Pred and an effective Laplacian Lred
on a reduced graph [23]. The swing dynamics on the reduced
graph reads
where e(α,β) ∈ R|Nc | , and where we used the ShermanMorrison formula [25]
(c,c)
[Lb
(c,c) -1
-1
−bαβ e(α,β) e>
(α,β) ] = [Lb
]
(c,c)
(c,c) -1
[Lb ]-1 e(α,β) e>
]
(α,β) [Lb
+bαβ
(c,c) -1
>
1 − bαβ e(α,β) [Lb ] e(α,β)
,
(40)
THIS VERSION: 28 NOVEMBER 2017
7
(c,c)
to express the inverse of the rank-1 perturbation of Lb .
Injecting Eqs. (39) in Eq. (35), and solving the swing equation
with initial conditions (δθg (0), ωg (0)) = (0, 0) yields
0
(g,c)
(c,c) -1
-1/2
>
?
δθ g (t)
= eAt −bαβ M Lb [Lb ] e(α,β) e(α,β) θc ,
ω g (t)
(c,c)
[Lb ]-1 e(α,β)
1 − bαβ e>
(α,β)
{z
}
|
B
(41)
where δθ g = M 1/2 δθ g , ω g = M 1/2 ωg , and A is the matrix
defined in Eq. (3) with Lb replaced by Lred .
3) The faulted line connects a synchronous machine and a
passive node: In this case α ∈ Ng and β ∈ Nc , and the four
blocks of Lb change according to
(g,g)
Lb
(c,c)
Lb
(g,g)
→Lb
(c,c)
→Lb
(c,g)
(c,g)
Lb →Lb
(g,c)
(g,c)
Lb →Lb
(g,c)
[êα + Lb
1−
(g,c)
Lred − δ(t)bαβ
[êα + Lb
+
+
bαβ êβ ê>
α ,
bαβ êα ê>
β ,
(42)
(c,c) -1
] êβ ]ê>
β [Lb
(c,c) -1
] + [Lb
] Pc
(c,c) -1
,
(c,c) -1
>
] êβ ][ê>
α + êβ [Lb
(c,g)
] Lb
(c,c) -1
]ββ
1 − bαβ [Lb
]
,
(43)
(c,g)
] Lb
(g,c)
(g,c)
[L-1
= −L-1
b]
red Lb
>
(c,c)
bαβ [Lb ]-1
ββ
[Lb
(c,c) -1
(c,c)
(L-1
= [Lb
b)
(c,g)
(g,c)
,
[L-1
= [L-1
b]
b]
− bαβ êβ ê>
β ,
(c,c) -1
[Lb
Proof. The observability Gramian associated to the phase
coherence measure is given in Eq. (29) and is the inverse of
the reduced Laplacian Lred . To compute P for the three types
of line contingencies, we use B defined in Eqs. (38), (41) and
(44) respectively and the matrix block inversion property [26]
− bαβ êα ê>
α ,
where êα ∈ R|Ng | and êβ ∈ R|Nc | . Pred and Lred become
Pred − δ(t)bαβ
if the faulted line connects a synchronous machine α and
a passive node β. In Eqs. (45)–(47), Ωαβ is the resistance
distance computed with respect to the physical network Lb ,
prior to Kron reduction.
(g,c)
L-1
red Lb
(c,c) -1
[Lb
]
(c,c) -1
[Lb
] ,
(g,g)
[L-1
= L-1
b]
red .
(48)
Eq. (48) holds for Lred regularized according to Proposition
2. Eqs. (45)–(47) are obtained after some algebra, and taking
the limit → 0 at the end of the calculation.
The results of Proposition 7 show that for the three types
of line contingencies, the voltage phase deviation from the
nominal operating point is proportional to the square of the
power flowing on the line prior to the fault, times a topological
factor. The latter is equal to the resistance distance when the
faulted line connects two synchronous machines, Eq. (45).
For the other two types of line faults, Eqs. (46) and (47),
the resistance distance factor of Eq. (45) is complemented
by terms which account for the topology of the network of
passive nodes, with no straightforward interpretation, except
(c,c) -1
for the term e>
] e(α,β) in Eq. (46), which is equal
(α,β) [Lb
to the resistance distance between α and β in the network of
passive nodes augmented by a ground node (see Th. 3.9 of
Ref. [23]).
where, again, we used the Sherman-Morrison formula [25] to
(c,c)
compute the inverse of Lb − bαβ êβ ê>
β . Finally, injecting
Eqs. (43) in Eq. (35), and solving the swing equation with
initial conditions (δθg (0), ωg (0)) = (0, 0) yields
"
#
0
(g,c)
(c,c) -1
-1/2
?
?
δθ g (t)
= eAt bαβ M (êα + Lb [Lb ] êβ )(θg,α − θc,β ) , Proposition 8 (Primary control effort under line con(c,c) -1
ω g (t)
1 − bαβ [Lb
]ββ
{z
} tingency). Consider the Kron reduced power system model
|
B
of Eq. (35) satisfying Proposition
R ∞ P2 and2 Assumption 1. The
(44)
primary control effort P = 0
1/2
1/2
i di ωi dt required during
where δθ g = M δθ g , ω g = M ωg , and A is the matrix
the transient caused by a line contingency modeled by Eq. (32)
defined in Eq. (3) with Lb replaced by Lred .
is given by
Having solved the swing equation for the three types of line
?
?
[bαβ (θg,α
− θg,β
)]2 −1
faults we are now ready to present our main results.
P=
(mα + m−1
(49)
β ),
2
Proposition 7 (Phase coherence under line contingency).
Consider the Kron reduced power system model of Eq. (35) if the faulted line connects two synchronous machines, by
satisfying Proposition 2, AssumptionR1, and mi = m ∀i ∈ Ng .
?
?
[bαβ (θc,α
− θc,β
)]2
∞
P=
(50)
The phase coherence measure P = 0 δθ >
g δθ g dt evaluated
2
for a line contingency modeled by Eq. (32) is given by
(c,c)
(c,g)
(g,c)
(c,c)
e>
]-1 Lb M -1 Lb [Lb ]-1 e(α,β)
(α,β) [Lb
?
?
[bαβ (θg,α
− θg,β
)]2
×
,
(c,c) -1
Ωαβ ,
(45)
P=
[1 − bαβ e>
] e(α,β) ]2
(α,β) [Lb
2d
if the faulted line connects two synchronous machines, by
if the faulted line connects two passive nodes, and by
(c,c)
?
?
]-1 e(α,β)
[bαβ (θc,α
− θc,β
)]2 Ωαβ − e>
(α,β) [Lb
,
(c,c) -1
2d
[1 − bαβ e>
] e(α,β) ]2
(α,β) [Lb
(46)
if the faulted line connects two passive nodes, and by
P=
(c,c)
P=
?
?
[bαβ (θg,α
− θc,β
)]2 Ωαβ − [Lb ]-1
ββ
,
(c,c) -1 2
2d
[1 − bαβ [L
] ]
b
ββ
(47)
?
?
[bαβ (θg,α
− θc,β
)]2
(51)
2
(c,c) -1 (c,g)
(g,c)
(c,c)
>
[ê>
] Lb ]M -1 [êα + Lb [Lb ]-1 êβ ]
α + êβ [Lb
×
,
(c,c)
2
[1 − bαβ [Lb ]-1
ββ ]
P=
if the faulted line connects a synchronous machine α and a
passive node β.
THIS VERSION: 28 NOVEMBER 2017
8
Proof. The observability Gramian associated to the primary
control effort is given in Eq. (24). To compute P = B > XB
for the above three types of lines contingencies, we use B
defined in Eqs. (38), (41) and (44) respectively.
Eq. (49) shows that the effort of primary control which
results from the outage of the α − β line is proportional to
the square of the power flowing on the line prior to the fault
−1
times the prefactor (m−1
α + mβ ). The latter indicates that
the primary control effort is large if the rotational inertias of
the synchronous machines at both ends of the faulted line
are small. For the other types of line contingencies, Eqs. (50)
and (51) predict a more involved dependence of P with the
inertias of the synchronous machines. Quite interestingly, for
both Eqs. (50) and (51), only the inertias of the synchronous
machines directly connected to P
the passive nodes matter. This
>
|Ng |
is easily seen writing M -1 = i m−1
i êi êi for êi ∈ R
(c,g)
th
and noticing that Lb êi = 0 if the i synchronous machine
is not connected to any of the passive nodes.
VII. P ERFORMANCE MEASURES UNDER LINE
CONTINGENCIES : N UMERICAL A NALYSIS
To illustrate our results we perform numerical investigations taking as physical network the IEEE-118 testcase [28].
Assuming that PQ buses are passive, we simulate the swing
dynamics for the reduced model, Eq. (35), where all PQ
buses have been eliminated by Kron reduction. To model
temporary line disconnections, we consider time-dependent
network Laplacians,
Lb (t) = Lb + Θ(t)Θ(τ − t)bαβ e(α,β) e>
(α,β) ,
(52)
where Θ(t) is the Heavyside step function, τ is the clearing
time, and α − β is the faulted line. We perform numerical
simulations for all possible line contingencies in the network,
each
the performance measures
R ∞ Ptime 2evaluating Rnumerically
∞P
2
δθ
(t)
dt
and
d
ω
(t)
dt. We compare the nui
i
i i i
0
0
merical results to our theoretical predictions. Since the latter
hold for Dirac-δ perturbations [see Eq. (32)], our theoretical
results are expected to be accurate for small clearing times as
δ(t) = limτ →0 Θ(t)Θ(τ − t)/τ .
Fig. 1, shows the behavior of the phase coherence measure
rescaled by the square of the power flowing on the line prior
to the fault for all lines connecting two synchronous machines in the physical network. As predicted for homogeneous
inertias and sufficiently short clearing times, Eq. (45), this
quantity is linear in the resistance distance. The validity of
the observability Gramian prediction for Dirac-δ perturbations
extends to longer clearing times for larger inertia of the
synchronous machines (red crosses). This is so because voltage
phase oscillations are more easily absorbed locally by larger
inertia. These results show that for longer clearing times, and
more generally for perturbations that are extended in time,
alternative approaches to the observability Gramian are needed
to accurately evaluate performance measures [29].
Fig. 2, shows the behavior of the primary control effort
rescaled by the square of the power flowing on the line
prior to the fault for all lines connecting two synchronous
machines in the physical network. For heterogeneous inertias
and sufficiently short clearing times we expect, according to
−1
Eq. (49), that this quantity scales linearly with (m−1
α + mβ ).
This prediction is confirmed for sufficiently large inertias. For
lower values of inertia, the linear tendency holds for short
clearing times, but breaks down for longer faults. The red
crosses in Figs. 1 and 2 correspond to somehow exaggerated
values of inertia. They are here to illustrate that our theoretical
predictions have larger domain of applicability in the presence
of larger inertias.
Finally, Fig. 3 shows the phase coherence and the primary
control effort measures for all possible line contingencies (170
lines in total, of which 46 connect two synchronous machines,
35 connect two passive nodes and 89 connect a passive node
to a synchronous machine) and τ = 20 ms. The transmission
lines are sorted according to the square of the power flowing
on the line in normal operation. Numerical results confirm that
our theoretical predictions are accurate indicators of transient
vulnerability under line contingency and that the transient
excursion is given by the square of the power flowing on the
transmission line prior to the fault times a topological factor
for the phase coherence measure [Eqs. (45)–(47)] or an inertia
dependent factor for the primary control effort [Eqs. (49)–
(51)].
Remarkably, our results show that the transient performance
under line fault is not a monotonic function of the power
load of the faulted line, the square of which is indicated by
the black line in Fig. 3. Both the network topology and the
distribution of inertia in the grid strongly impact the transients.
We observe that the most critical lines are not always the most
highly loaded ones. For the phase coherence measure (left
panel in Fig. 3), the line carrying the 6th largest power leads
to the largest integrated transient excursions even though it
carries 30% less power than the line carrying the most power.
We saw (but do not show here) this non monotonic behavior
also when lines are sorted according to their relative load (the
load relative to their capacity). A similar non monotonicity is
observed for the primary control effort (right panel in Fig. 3).
The fault of the line carrying the 3rd largest power causes the
largest primary control effort, and the line carrying 44% of
the largest transmitted power (line ranked 14th with respect to
absolute load) is the 4th most critical one.
VIII. C ONCLUSION
The standard formalism used until now to evaluate performance measures of electric power grids was restricted
to nodal perturbations [11]–[13] and we extended it to line
perturbations. We showed numerically that, despite its restriction to Dirac-δ perturbation (instantaneous in time), the
formalism correctly evaluates performance measures even in
the physically relevant case of perturbations with finite, but
not too long duration. One would naively guess that the most
critical lines are those that are the most heavily loaded, either
relatively to their thermal limit or in absolute value. Quite
surprisingly, we found that faults on lines transmitting less
than half of the heaviest line load in the network require more
primary effort control or perturb the network’s coherence more
than lines with higher loads. This is so, because performance
THIS VERSION: 28 NOVEMBER 2017
9
0.16
h
i
P/ τ 2b2αβ (θα? − θβ? )2/2d
τ = 0.02 [s]
τ = 0.04 [s]
τ = 0.06 [s]
τ = 0.08 [s]
0.12
0.08
0.04
mi = m? = 20/2πf
mi = 10m?
0.00
0.00
0.04
0.08
0.12
mi = m? = 20/2πf
mi = 10m?
0.16 0.00
0.04
0.08
Ωαβ
0.12
mi = m? = 20/2πf
mi = 10m?
0.16 0.00
0.04
0.08
Ωαβ
0.12
mi = m? = 20/2πf
mi = 10m?
0.16 0.00
0.04
0.08
Ωαβ
0.12
0.16
Ωαβ
Fig. 1. Voltage phase variance (from the nominal operating point) incurred during a transient resulting from a line contingency as a function of the resistance
distance separating the nodes of the faulted line. Each data point corresponds to the fault of a line connecting two generators in the physical network.
Simulation parameters: IEEE-118 testcase with uniform inertia at all nodes, f = 50 Hz, di /mi = 0.5 [s-1 ], and mi = m? = 2H/2πf , H = 10 [s] (typical
values from [27], blue circles) and mi = 10m? (red crosses). From left to right: fault clearing times τ corresponding to 1, 2, 3, and 4 AC cycles. The straight
line gives our theoretical prediction Eq. (45).
120
12
hmi i = 200/2πf
h
i
P/ τ 2b2αβ (θα? − θβ? )2/2
τ = 0.02 [s]
100
12
hmi i = 200/2πf
τ = 0.04 [s]
10
12
hmi i = 200/2πf
τ = 0.06 [s]
10
12
hmi i = 200/2πf
τ = 0.08 [s]
10
10
80
8
8
8
8
60
6
6
6
6
40
4
4
4
4
20
2
2
2
2
hmi i = 20/2πf
0
1
2
(m−1
α
+
3
4
m−1
β )
hmi i = 20/2πf
0
5
1
2
(m−1
α
· hmii
+
3
m−1
β )
4
5
0
hmi i = 20/2πf
1
2
(m−1
α
· hmii
3
+
m−1
β )
4
5
hmi i = 20/2πf
0
1
2
(m−1
α
· hmii
+
3
4
m−1
β )
· hmii
5
0
Fig. 2. Primary control effort required during a transient resulting from a line contingency as a function of the sum of the inverse inertias of the synchronous
machines at both ends of the faulted line. Each data point corresponds to the fault of a line connecting two generators in the physical network. Simulation
parameters: IEEE-118 testcase with inertias uniformly distributed in the interval [0.2hmi, 1.8hmi] with hmi = 2H/2πf , H = 10 [s] (typical values from
[27], blue circles) and hmi = 200/2πf (red crosses), f = 50 Hz, and di /mi = 0.5 [s-1 ]. From left to right: fault clearing times τ corresponding to 1, 2, 3,
and 4 AC cycles. The straight line gives our theoretical prediction Eq. (49).
1.0
1.0
τ = 0.02 [s]
1.0
0.8
τ = 0.02 [s]
1.0
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.4
0.4
0.2
0.4
0.2
0.0
145
150
155
160
165
0.4
170
Numerics (normalized)
0.2
0.2
Theory (normalized)
Squared power flow (normalized)
0.0
0
20
40
60
0.0
145
150
155
160
165
170
Numerics (normalized)
80
Theory (normalized)
Squared power flow (normalized)
100
120
140
160
Line number
0.0
0
20
40
60
80
100
120
140
160
Line number
Fig. 3. Plot of the normalized performance measure (red), theoretical prediction (green), and square of the power flowing on the line prior to the fault (black)
for line contingencies with clearing times τ = 0.02 [s]. Transmission lines are ordered according to the power flowing on the line prior to the fault. Left
panel: phase coherence measure, uniform inertias mi = m? = 2H/2πf , H = 10 [s], f = 50 Hz, and di /mi = 0.5 [s-1 ]. Right panel: primary control
effort, inertias uniformly distributed in the interval [0.2hmi, 1.8hmi] with hmi = 2H/2πf , H = 10 [s], f = 50 Hz, and di /mi = 0.5 [s-1 ].
measures, Eqs. (45)–(47) and Eqs. (49)–(51), are given by the
square of the original load on the faulted line times a term
depending on the topology of the network. The performance
measures we calculated could therefore be used in N − 1
contingeny analysis to quickly identify the most critical lines,
based on topological characteristics of the network together
with the load they carry.
Future works should investigate nodal N −1 faults, where a
bus with all its connected lines is removed from the network.
At present this seems to be hard to calculate as such a fault is
difficult to map from the physical to the Kron reduced network.
ACKNOWLEDGMENT
We thank B. Bamieh for useful discussions. This work was
supported by the Swiss National Science Foundation under an
AP Energy Grant.
THIS VERSION: 28 NOVEMBER 2017
10
IX. A PPENDIX
Proposition 9 (Oscillators with relative damping). Consider
the 2nd order, coupled oscillator model with relative damping
M ẍ = −Ld ẋ + P − Lb x,
(53)
where M = diag({mi }) and both Lb and Ld are Laplacian matrices associated to the network of coupling and
damping interactions. If the matrices M -1/2 Lb M -1/2 and
M -1/2 Ld M -1/2 commute, the observability Gramians of
Propositions 5 and 6 generalize to
(2,2)
Xij
=
N
X
(TM )il (TM> )qj (TM> M -1/2 Q(2,2) M -1/2 TM )lq
l,q=1
#
M M
(ηqM λM
l + ηq λ q )
(54)
×
M M
M 2
M
(ηqM + ηlM )(ηqM λM
l + ηl λq ) + (λq − λl )
"
and
(2,2)
Xij
=
N
X
(TM )il (TM> )qj (TM> M -1/2 Q(1,1) M -1/2 TM )lq
l,q=1
#
(ηqM + ηlM )
(55)
×
M M
M 2
M
(ηqM + ηlM )(ηqM λM
l + ηl λq ) + (λq − λl )
"
where λM
l and TM are the eigenvalues and the orthogonal
matrix diagonalizing M -1/2 Lb M -1/2 , and ηlM are the eigenvalues of M -1/2 Ld M -1/2 .
Proof. Since M -1/2 Lb M -1/2 and M -1/2 Ld M -1/2 commute
they have common eigenvectors TM . The results of Proposition
4 generalize simply replacing γ by ηiM in Eqs. (17), (19)
and (20). The X (2,2) block of the observability Gramian is
calculated analogously as in Propositions 5 and 6 to obtain
Eqs. (54) and (55).
[12] T. W. Grunberg and D. F. Gayme, “Performance measures for linear
oscillator networks over arbitrary graphs,” IEEE Transactions on Control
of Network Systems, vol. PP, no. 99, pp. 1–1, 2016.
[13] B. K. Poolla, S. Bolognani, and F. Dörfler, “Optimal placement of
virtual inertia in power grids,” IEEE Transactions on Automatic Control,
vol. PP, no. 99, pp. 1–1, 2017.
[14] D. J. Klein and M. Randić, “Resistance distance,” Journal of Mathematical Chemistry, vol. 12, no. 1, pp. 81–95, 1993.
[15] K. Stephenson and M. Zelen, “Rethinking centrality: Methods and
examples,” Social Networks, vol. 11, no. 1, pp. 1 – 37, 1989.
[16] E. Bozzo and M. Franceschet, “Resistance distance, closeness, and
betweenness,” Social Networks, vol. 35, no. 3, pp. 460 – 469, 2013.
[17] D. J. Klein, “Graph geometry, graph metrics and Wiener,” Commun.
Math. Comput. Chem., no. 35, pp. 7–27, 1997.
[18] W. Xiao and I. Gutman, “Resistance distance and laplacian spectrum,”
Theoretical Chemistry Accounts, vol. 110, no. 4, pp. 284–289, 2003.
[19] T. Coletta and P. Jacquod, “Resistance distance criterion for optimal
slack bus selection,” arXiv preprint arXiv:1707.02845, 2017.
[20] ——, “Linear stability and the braess paradox in coupled-oscillator
networks and electric power grids,” Phys. Rev. E, vol. 93, p. 032222,
Mar 2016.
[21] F. Paganini and E. Mallada, “Global performance metrics for synchronization of heterogeneously rated power systems: The role of machine
models and inertia,” arXiv preprint arXiv:1710.07195, 2017.
[22] F. Lin, M. Fardad, and M. R. Jovanovic, “Optimal control of vehicular
formations with nearest neighbor interactions,” IEEE Transactions on
Automatic Control, vol. 57, no. 9, pp. 2203–2218, 2012.
[23] F. Dörfler and F. Bullo, “Kron reduction of graphs with applications
to electrical networks,” IEEE Transactions on Circuits and Systems I,
vol. 60, no. 1, pp. 150–163, 2013.
[24] ——, “Spectral analysis of synchronization in a lossless structurepreserving power network model,” in First IEEE International Conference on Smart Grid Communications, 2010, pp. 179–184.
[25] J. Sherman and W. J. Morrison, “Adjustment of an inverse matrix
corresponding to a change in one element of a given matrix,” Ann.
Math. Statist., vol. 21, no. 1, pp. 124–127, 1950.
[26] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. New York,
NY, USA: Cambridge University Press, 2012.
[27] P. Kundur, N. J. Balu, and M. G. Lauby, Power system stability and
control. McGraw-hill New York, 1994, vol. 7.
[28] U. of Washington, “Power systems testcase archive,” 1993. [Online].
Available: https://www2.ee.washington.edu/research/pstca
[29] M. Tyloo, T. Coletta, and P. Jacquod, “Robustness of synchrony in
complex networks and generalized Kirchhoff indices,” arXiv preprint
arXiv:1710.07536, 2017.
R EFERENCES
[1] Y. Kuramoto, Lecture Notes in Physics, vol. 39, pp. 420–422, 1975.
[2] A. Pikovsky, M. Rosemblum, and K. J, Synchronization: A Universal
Concept in Nonlinear Sciences. Cambridge University Press, 2001.
[3] J. Machowski, J. W. Bialek, and J. R. Bumby, Power system dynamics:
stability and control. John Wiley, 2008.
[4] “Power systems of the future: The case for energy storage, distributed
generation, and microgrids,” IEEE Smart Grid, Tech. Rep. Nov. 2012.
[5] S. Backhaus and M. Chertkov, “Getting a grip on the electrical grid,”
Phys. Today, vol. 66, no. 5, pp. 42–48, 2013.
[6] B. Bamieh, M. R. Jovanovic, P. Mitra, and S. Patterson, “Coherence
in large-scale networks: Dimension-dependent limitations of local feedback,” IEEE Transactions on Automatic Control, vol. 57, no. 9, pp.
2235–2249, 2012.
[7] T. Summers, I. Shames, J. Lygeros, and F. Dörfler, “Topology design for
optimal network coherence,” in European Control Conference. IEEE,
2015, pp. 575–580.
[8] M. Siami and N. Motee, “Systemic measures for performance and
robustness of large-scale interconnected dynamical networks,” in 53rd
Annual Conference on Decision and Control. IEEE, 2014, pp. 5119–
5124.
[9] M. Fardad, F. Lin, and M. R. Jovanovic, “Design of optimal sparse
interconnection graphs for synchronization of oscillator networks,” IEEE
Transactions on Automatic Control, vol. 59, no. 9, pp. 2457–2462, 2014.
[10] K. Zhou, J. Doyle, and K. Glover, “Robust and optimal control,” 1996.
[11] E. Tegling, B. Bamieh, and D. F. Gayme, “The price of synchrony:
Evaluating the resistive losses in synchronizing power networks,” IEEE
Transactions on Control of Network Systems, vol. 2, no. 3, pp. 254–266,
2015.
Tommaso Coletta Received the M.Sc. degree in physics and the Ph.D. degree
in theoretical physics from the Ecole Polytechnique Fédérale de Lausanne
(EPFL), Lausanne, Switzerland in 2009 and 2013 respectively. He has been
a Postdoctoral researcher at the Chair of Condensed Matter Theory at the
Institute of Theoretical Physics of EPFL. Since 2014 he is a Postdoctoral
researcher at the engineering department of the University of Applied Sciences
of Western Switzerland, Sion, Switzerland working on complex networks and
power systems.
Philippe Jacquod Philippe Jacquod received the Diplom degree in theoretical
physics from the ETHZ, Zürich, Switzerland, in 1992, and the PhD degree in
natural sciences from the University of Neuchâtel, Neuchâtel, Switzerland, in
1997. He is a professor with the engineering department, University of Applied
Sciences of Western Switzerland, Sion, Switzerland. From 2003 to 2005 he
was an assistant professor with the theoretical physics department, University
of Geneva, Geneva, Switzerland and from 2005 to 2013 he was associate, then
full professor with the physics department, University of Arizona, Tucson,
USA. Currently, his main research topics is in power systems and how they
evolve as the energy transition unfolds. He is co-organizing an international
conference series on related topics. He has published about 100 papers in
international journals, books and conference proceedings.
| 3cs.SY
|
arXiv:1712.03280v1 [cs.AI] 8 Dec 2017
Nintendo Super Smash Bros. Melee:
An "Untouchable" Agent
Ben Parr (bparr), Deepak Dilipkumar (ddilipku), Yuan Liu (yuanl4)
Deep Reinforcement Learning and Control (10-703)
Carnegie Mellon University
Pittsburgh, PA 15213
Abstract
Nintendo’s Super Smash Bros. Melee fighting game can be emulated on modern
hardware allowing us to inspect internal memory states, such as character positions.
We created an AI that avoids being hit by training using these internal memory
states and outputting controller button presses. After training on a month’s worth
of Melee matches, our best agent learned to avoid the toughest AI built into the
game for a full minute 74.6% of the time.
1
Introduction
Super Smash Bros. Melee is a video game created by Nintendo in 2001 that is still played competitively today. Based on “current viewership, sponsorship, player base and, most importantly, future
growth potential,” ESPN listed Melee as number six on their 2016 Top 10 Esports Draft (Erzberger
(2016)), making Melee the longest-played game in the Esports Draft. Competitive play is a one player
versus another player fighting game where each player chooses one character to play as from the list
of 25 playable characters. The goal of the game is to knock the opponent’s character off the stage and
into the abyss four times before the opponent knocks your character off four times. Each character
has a wide array of offensive and defensive abilities. For example, the character Marth has a sword
which allows him to attack from a further distance, but at a slower speed. The game was originally
created for the Nintendo GameCube console, and so runs at 60 frames per second (fps).
Figure 1: A screenshot of our agent (Fox, left) using a special attack on the sword-wielding opponent
(Marth, right) on the entirely flat Final Destination stage.
The GameCube is no longer technologically advanced. In fact, Melee can be played on most modern
computers using Dolphin software that emulates the GameCube hardware. This allows us to inspect
specific memory addresses, such as the memory address containing the position of each character.
Fortunately, Vlad Firoiu has written this infrastructure-like code already (Firoiu et al. (2017)). Firoiu
also trained agents for the game using reinforcement learning. For the class project, we decided to
train our own independent agents, not based on his reinforcement learning code. Specifically, we
trained a Deep Q-Network (DQN) agent, a Double Deep Q-Network (Double DQN) agent, and a
Dueling Deep Q-Network (Dueling DQN) agent, as well as a preliminary Asynchronous Advantage
Actor-Critic (A3C) agent. Each agent plays as the Fox character and is given the values inside
the chosen GameCube memory addresses, and decides which evasive maneuver to take to remain
untouched by the opposing Marth character.
Melee has AI built into the game itself, thus allowing a single player option where you play against
the built-in computer. The Melee agent has nine levels of difficulty. Unlike our agent, the built-in
Melee agent is fixed, and so does not learn from experience. We use the built-in Melee agent to
train our own agent. We generated training data in parallel using Google Cloud virtual machines.
After training, we evaluated our agents by the number of frames it remained untouched while playing
against the toughest AI built into the game.
2
Related Work
DeepMind Technologies has successfully learned to play Atari 2600 games using an emulator Mnih
et al. (2015). DeepMind used a convolutional neural network combined with a Q-learning variant to
learn to play games using just the games’ pixel output. We plan to build on this success by focusing
on a relatively newer game with more complex interactions. Instead of pixels, we look at the game’s
internal memory, including features such as characters’ positions and velocities.
The work most relevant to this project is that of Firoiu et al. (2017), which applies Reinforcement
Learning to create an agent for Melee. Utilizing an actor-critic model with a deep architecture
and self-play, they were able to achieve excellent results with the Captain Falcon character on the
Battlefield stage, going on to even beat a number of professional players. Their agent is however
known to have difficulties dealing with unusual opponent behavior that has not been seen in training,
such as the opponent simply crouching.
There have been more recent papers that build on the DeepMind’s earlier success using variations
of the regular DQN architecture. One of these is the Double DQN introduced by van Hasselt et al.
(2015). This architecture uses two networks playing alternating roles in order to reduce the tendency
of the regular DQN to over-estimate action values. Another recent success was the Dueling DQN of
Wang et al. (2015). This uses two parallel networks that estimate the value function and advantage
function, and finally combine these to estimate the Q-value functions.
Another class of algorithms that has gained popularity in the era of Deep Reinforcement Learning
is the policy gradient method, such as the actor-critic algorithm. In particular, the Asynchronous
Advantage Actor Critic method (A3C) introduced in Mnih et al. (2016) has become one of the
standard algorithms for distributed learning in many game-playing applications of reinforcement
learning.
3
Methods
We implemented four different reinforcement learning algorithms and applied them to Melee. The
four algorithms are DQN, Double DQN, Dueling DQN and A3C. Each algorithm interacted with a
custom environment, which for example, rewarded taking no damage. Finally, each algorithm trained
on Melee gameplay samples generated in parallel using Google Cloud virtual machines.
3.1
Environment Setup
We use the Dolphin GameCube emulator to run Melee. The game includes 9 AIs with increasing
difficulty, all of which are used during training. Even though our agent predominately trained against
the toughest AI, training against all 9 AIs provided our agent a wider amount of combat scenarios.
All of our agents play as the Fox character versus the Marth character (controlled by an in-game
2
AI) on Final Destination. This Marth opponent, along with a sporadic frame-skipping issue (see
Discussion), make the environment stochastic.
By looking into the internal game memory, we are able to access information about the current match.
These are included in the state that we pass into the algorithm. Specifically, the state includes both
the agent’s and opposing Marth’s current position, velocity, action state, duration of current action
state, direction facing, whether charging an attack, whether airborn, shield size, jumps used, hitlag
counter, and damage.
The agent has access to five actions to choose from: do nothing, left dodge, right dodge, standing
dodge, and Fox’s shine (essentially a shield that pushes opponents away). These actions were chosen
since our primary goal was to create an agent that could not be hit, for which these defensive actions
would be sufficient. The shine attack in particular is tied for the fastest move in the game, and thus
should be able to deal with situations that come up which the other moves can’t handle.
As for the reward scheme, the agent gets a reward of 1/60 for every frame that it is not moving, and
has zero damage. The number 1/60 was chosen to keep the rewards small. We noticed that keeping
rewards small in this way helped to stabilize our mean-max Q values. The agent gets a reward of 0
for choosing to dodge or shine, in order to discourage it from spamming these actions.
The most important aspect of the setup is the terminal flag. As soon as the agent is hit, it receives a
terminal flag and the match is reset. This way, it gets a strong negative signal for getting hit, as it can
only receive rewards as long as its damage is at 0. This way, the agent only sees frames from the
beginning of the game during the start of training. Once it learns to get past these states, it starts to
see states from later on. This scheme can be described as a form of curriculum learning, in the sense
that the agent has to learn how to fend off attacks later in the game only after it learns to get through
the initial frames without being hit.
3.2
Training Algorithms and Specifications
We implement multiple state-of-the-art Deep Reinforcement Learning techniques to train our agent.
The first of these is the DQN introduced in Mnih et al. (2015). We approximate the value function
Qw using a deep neural network parametrized by weights w and update the weights as:
0 0
w = w + α r + γ max
Q
(s
,
a
)
−
Q
(s,
a)
∇w Qw (s, a)
w
w
0
a ∈A
We use target fixing and a replay memory as suggested in the paper.
Next, we use the Double DQN as described in van Hasselt et al. (2015). This uses the same update
rule as for the DQN, but the action selection and value estimation for the target during training are
done by two different networks. The online network is used to do action selection, and the target
network is used for value estimation. So the update (with Qw1 as on the online network and Qw2 as
the target network) can be written as :
w1 = w1 + α r + γQw2 (s0 , arg max Qw1 (s0 , a0 )) − Qw1 (s, a) ∇w1 Qw1 (s, a)
a0 ∈A
This is known to reduce the tendency of the DQN to overestimate its Q-values.
We also use the Dueling DQN as described in Wang et al. (2015). The most important aspect of a
Dueling DQN is the architecture. It consists of two streams of fully connected layers, one which
estimates the state value function, and one which estimates the advantage function. These two streams
are combined to get an estimate of the Q-value for a particular (state, action) pair:
Q(s, a) = V (s) + A(s, a) −
1 X
A(s, a0 )
|A| 0
a
After training using the three DQN algorithms, we also experimented to a limited extent with the
A3C algorithm introduced in Mnih et al. (2016). This again involves two “streams”, both of which
3
are neural networks. One - the actor - estimates the policy directly, as a probability distribution over
the actions that can be taken. The other - the critic - is used to estimate state value functions to guide
the actor to the correct policy in a principled manner. We have implemented a variant of actor-critic,
as described in Lillicrap et al. (2015), which uses target fixing and a replay memory like in the other
methods. This also fits very well with our Google Cloud Parallelization method, which is described
in the next section.
Our DQN and Double DQN use two hidden layers, having 128 and 256 units. Dueling and A3C
each have two streams. The streams share one hidden layer with 128 units, and then branch out to
separate fully connected layers having 512 units. In Dueling, the streams are eventually combined as
described above, whereas in A3C, they are used to predict two separate quantities (value function and
policy).
All of our experiments use γ = 0.99, RMSProp for optimization with a learning rate of 0.00025,
batch size of 32, and a linear decay greedy policy for exploration, with decaying from 1 to 0.1
in the first quarter of training. The replay memory size was 1, 000, 000, and the target network
was updated every 10, 000 training iterations. During training, the agent faces an AI whose level is
decided at random. Specifically, we trained against the toughest in-game AI 70% of the time, with
the remaining 30% distributed across the other AI levels. This was done in order to avoid “overfitting”
to a particular opponent. See the Discussion section for more information on this.
We use three metrics to compare these methods throughout training. The first is the reward gained
by the agent based on our reward scheme. The second is the average game length that the agent
survives for without being hit, in terms of number of frames. The final one is the mean max Q-value,
calculated on 1000 states randomly sampled before training. The third metric gives us an idea of how
well the network believes it is performing. The second metric tells us how well the agent is doing
in terms of our actual goal - dodging the opponent without getting hit. Finally, links between the
first and second metrics will give us a good idea of how well the reward scheme is able to guide our
agent to our actual goal. This way, we are able to comprehensively evaluate and compare each of our
algorithms.
3.3
Episode Parallelization
Figure 2: Infrastructure outline where a single Pittsburgh Supercomputing Center (PSC) job trains on
gameplay generated on N Google Cloud workers
A single GPU-enabled manager trains on gameplay data that is generated by 50 workers playing
Melee in parallel, allowing each model to train on a full month of gameplay data generated in only
8 hours. First, the manager creates an initial model, and uploads the model to the workers. The
workers then enter a loop: generate a few minutes of gameplay samples from the current fixed model,
upload the samples to the manager, download the latest model, repeat. Each Melee worker uses a
fixed model Qw to generate replay memory tuples (s, a, r, s0 ). This allows the manager to train on
a stream of incoming gameplay data generated faster than the manager could generate on its own.
4
Finally, the manager completes the loop by periodically saving an updated model every 15 times the
manager receives gameplay samples from its workers.
Fortunately, one of the cheapest available Google Cloud virtual machine, the g1-small machine, is
able to run Dolphin emulated Melee. We chose Google Cloud since they provide a free $300 credit to
new users. We were able to run all training for free by only using this $300 credit.
4
Experimental Results
The results of our experiments for DQN, Double DQN and Dueling DQN are shown in Figure 3 and
Figure 4. Figure 3 shows rewards and game lengths for all 3 models. First, we see that the results are
very erratic, and fluctuate through training. However, all of the models show a clear improvement in
performance about 20% of the way through training.
(a) DQN Rewards
(b) DQN Game Lengths
(c) Double Rewards
(d) Double Game Lengths
(e) Dueling Rewards
(f) Dueling Game Lengths
Figure 3: Rewards and Game Lengths for DQN, Double DQN and Dueling DQN
The regular DQN does not appear to perform as well as the other two in terms of game lengths.
The Double DQN had the most consistently good results, with game lengths remaining high for a
significant portion of training. The Dueling DQN however, had the highest peaks, meaning that it
produced the best overall agent. Also for context, it is useful to note that at 60 fps, 2250 frames is
5
37.5 seconds, meaning that each agent is often able to last above 30 seconds without being hit at all
for the Double and Dueling networks.
Another useful point to note in these results is the link between the reward and game length. The
patterns between these two metrics for each model are quite similar, indicating that our reward scheme
is a good proxy to lead the agent towards our goal of dodging the opponent.
Figure 4 shows the Mean Max Q values for these 3 models. The DQN seems to be improving steadily,
although for our reward scheme, the value of roughly 2.5 is heavily overestimated. The Double DQN
is meant to remedy the DQN’s tendency to overestimate its own performance, and we see that it does
indeed do that - the mean max Q values for the Double DQN are a lot more reasonable. The graph
for Dueling is surprisingly erratic, and almost cyclic in its fluctuations.
Figure 4: Mean Max Q over training.
We also include results from our preliminary A3C runs in Figure 5. Note that the third graph shows
the estimated value function on the held out states, as opposed to the max Q-value, as it is the value
function that the critic evaluates. This is still comparable to the mean max Q-value that we computed
for the other methods.
All three graphs indicate the same issue - the agent appears to get stuck early on and is unable to
improve any further during training. Watching it playing, we see that it simply chooses to do nothing,
and collect the rewards that it can gain in the couple of seconds before it is hit. This is reflected in
the game length graph as well, which is stuck at around 120. We believe that this is likely due to
training being stuck in a local minimum, since standing still is actually the only way for the agent
to increase its reward in the short-term. One possible explanation for this is the need for a more
sophisticated exploration strategy. An epsilon greedy policy did not seem to work, and an entropy
based exploration scheme (which was used to generate these graphs) also does poorly. Using the
right weight for the entropy term in the loss function is also a potential issue. Additionally, using a
different optimizer (such as ADAM) might help our agent to break out of this local minimum.
As our experiments with the DQN variants were successful, we did not spend a significant amount
time ironing out the issues with the A3C. We realize that it has tremendous potential, especially with
the parallelization scheme that we already use, and so we hope to solve these problems and get A3C
working well in the future.
Videos of our agent playing can be found at https://goo.gl/x67ioE. These are all minute-long
matches between our agent (Fox) and the highest level AI in the game (Marth), in which our agent
completely avoids being hit. Our best agent, trained using the Dueling DQN, is able to dodge a level
9 AI for a minute or longer, 74.6% of the time.
6
(a) Rewards
(b) Game Lengths
(c) Value Function
Figure 5: A3C results
5
Discussion
The major limitation of our project was the opposing player. Of course, we would have preferred
training against human players who play Melee professionally. Based on our resources, we instead
trained against the AI players built into the game itself. So the agent “overfits” the toughest in-game
AI’s strategy, which we were able to alleviate to some extent by randomizing the CPU level during
training.
One solved issue was that our initial DQN models learned to do nothing, and so, would be hit by the
opponent Marth almost immediately. We solved this issue by not using a ReLU activation on the
final layer of our deep networks. Interestingly, a ReLU on the final layer seemed to completely block
learning in Melee, where initially zero Q-values never change during training. One open issue is the
erratic Dueling DQN Q-values, which spike above 8.0 after 80% training. Reducing the target update
frequency could help make training more stable and produce a smoother mean max Q.
Finally, we observed sporadic frame skipping issues when emulating Melee, where our agent would
hit a button but Melee would ignore it. This makes Melee harder for our agent. Future work could
look into technology used by Tool-Assisted Speedrun (TAS) to remove the frame skipping issue, and
thus make the opponent player the only stochastic part of the environment.
6
Conclusion
DQN performs worse than Double DQN and Dueling DQN. Double DQN showed consistent performance during training, in terms of game lengths and rewards, and also has reasonable Q-values.
However, the Dueling DQN had the highest peaks on the game lengths graph, meaning that it produced the best agents. After training on one month of gameplay data with the Dueling DQN, our best
agent is able to dodge the highest level AI in the game for at least a full minute 74.6% of the time.
Our code can be found at https://github.com/bparr/melee-ai and videos of our agent playing
can be found at https://goo.gl/x67ioE.
7
7
Acknowledgements
Thank you to Vlad Firoiu, without whom this class project would not be possible (Firoiu et al. (2017)).
All the best at Google DeepMind.
References
Erzberger, T. (2016). The 2016 top 10 esports draft. ESPN Esports.
Firoiu, V., Whitney, W. F., and Tenenbaum, J. B. (2017). Beating the world’s best at super smash
bros. with deep reinforcement learning. arXiv preprint arXiv:1702.06230.
Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D.
(2015). Continuous control with deep reinforcement learning. CoRR, abs/1509.02971.
Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T. P., Harley, T., Silver, D., and Kavukcuoglu,
K. (2016). Asynchronous methods for deep reinforcement learning. CoRR, abs/1602.01783.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A.,
Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou,
I., King, H., Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. (2015). Human-level control
through deep reinforcement learning. Nature, 518(7540):529–533.
Sutton, R. S. and Barto, A. G. (1998). Introduction to Reinforcement Learning. MIT Press, Cambridge,
MA, USA, 1st edition.
van Hasselt, H., Guez, A., and Silver, D. (2015). Deep reinforcement learning with double q-learning.
CoRR, abs/1509.06461.
Wang, Z., de Freitas, N., and Lanctot, M. (2015). Dueling network architectures for deep reinforcement learning. CoRR, abs/1511.06581.
8
| 2cs.AI
|
Query-Free Attacks on Industry-Grade Face Recognition Systems under
Resource Constraints
Di Tang
arXiv:1802.09900v1 [cs.LG] 13 Feb 2018
Chinese University of Hong Kong
XiaoFeng Wang
Indiana University
Kehuan Zhang
Chinese University of Hong Kong
Abstract
To attack a deep neural network (DNN) based Face
Recognition (FR) system, one needs to build substitute
models to simulate the target, so the adversarial examples
discovered could also mislead the target. Such transferability is achieved in recent studies through querying the
target to obtain data for training the substitutes. A realworld target, likes the FR system of law enforcement,
however, is less accessible to the adversary. To attack
such a system, a substitute with similar quality as the target is needed to identify their common defects. This is
hard since the adversary often does not have the enough
resources to train such a model (hundreds of millions of
images for training a commercial FR system).
We found in our research, however, that a resourceconstrained adversary could still effectively approximate
the target’s capability to recognize specific individuals,
by training biased substitutes on additional images of
those who want to evade recognition (the subject) or the
victims to be impersonated (called Point of Interest, or
PoI). This is made possible by a new property we discovered, called Nearly Local Linearity (NLL), which models the observation that an ideal DNN model produces
the image representations whose distances among themselves truthfully describe the differences in the input images seen by human. By simulating this property around
the PoIs using the additional subject or victim data, we
significantly improve the transferability of black-box impersonation attacks by nearly 50%. Particularly, we successfully attacked a commercial system trained over 20
million images, using 4 million images and 1/5 of the
training time but achieving 60% transferability in an impersonation attack and 89% in a dodging attack.
1
Introduction
With its commercial success, deep learning (DL) based
face recognition (FR) is haunted by the security risks
posed by the adversary adaptive to new AI inventions.
Prior research shows that adversarial examples can be
found to mislead even the state-of-the-art recognition algorithms [4, 7, 27, 29], causing them to misclassify these
examples. More specifically, such adversarial examples
are images derived from those correctly recognized by a
DL model, for the purpose of inducing classification errors while maintaining the level of changes low so they
can appear less distinguishable from the original images
by humans. Indeed, a recently approach [2] alters merely
16 pixels to ensure misclassification on 32×32 images.
Attacking a strawman. On the other hand, such adversarial learning risks need to be put into perspective.
Still we are less clear how realistic the discovered threats
could be, given that most of them are reliant on a whitebox assumption about the target (the FR system they aim
at), that is, the availability of full information about the
target’s parameters. In practice, however, an industrygrade system’s parameters are often commercial secret
and cannot be easily acquired by unauthorized parties.
A more realistic way to understand a DL system’s
security properties is the black-box approach, in which
the adversary queries the target, utilizes the features inferred through the queries to learn a substitute model and
then searches the substitute for the adversarial examples
that also work on the target. Such an approach is based
upon transferability of adversarial examples across different models [13]: some examples mislabeled by one
DL model are also found to be misclassified by another.
To ensure a high transferability, existing approaches aggressively query the target to obtain adequate inputoutput samples for accurately simulating the target. As
a prominent example, a recent black-box attack needs to
interact with the target for at least 1,000 times [17].
With all the progresses being made, still a big gap
exists between hypothetic attacks proposed and credible threats with practical impacts. Particularly, querying
security-critical FR systems is often expensive or even
infeasible in practice: e.g., an FR ATM can immediately
alert a card holder to a potential fraud once an impersonation attempt fails, making further probes less likely to
continue. Another problem of prior transferability studies is the simple dataset used to train their models, e.g.,
the MNIST database [12] includes only tens of thousands of images for recognizing handwritten digits. A
real-world FR system, however, is typically trained over
tens or even hundreds of millions of images for identifying millions of people. Less clear is whether what is
learnt from such small-scale studies over toy examples is
indeed applicable to real FR systems.
pose of finding the right makeup to cheat an FR ATM
into authenticating him as the victim. This attack, which
we call asymmetric cross-class image transfer or EXCIT,
is found to be completely feasible in our research, due to
a new property called nearly local linearity (NLL) discovered in our study.
More specifically, under a well-trained DL model,
the difference between a pair of images’ representations
(e.g., the Cosine distance between the vectors produced
by the DL model) should be nearly linear to their similarity as seen by human eye: in other words, when these images become increasingly dissimilar, the difference between their representations grows (nearly) proportionally. This NLL property, as discovered in our research,
can be approximated at a given PoI (e.g., the fugitive) or
a pair of PoIs (e.g., the victim and the impersonator) during the training of a substitute model: using the victim
or the attacker’s additional images, we can minimize the
distances between the scores they receive from the model
and what are expected according to NLL. We found that
such a model can effectively simulate a better-trained target’s behaviors around the PoIs.
In our research, we implemented EXCIT and first evaluated it under the settings of our transferability study.
We observed that the new technique vastly enhanced
the effectiveness of the attacks, particularly for impersonations, from 20% to 50%, even when the adversary
only used 10% of the training data and half of the layers
(thus saving the training time by 5 orders of magnitude).
Further we ran this approach against industry-grade systems including ColorReco, Facevis, Face++ and SenseTime (the SenseTime system trained over tens of millions of photos). Using 4 million images collected from
the web (the largest scale for this type of research), EXCIT was found to significantly elevate the chance of successful cross-model attacks, from 11% to 60%, without
any communication with the target before the attack.
Cross-class transferability. To better understand the
security guarantee of real-world FR systems, we revisited transferability in our research, assuming that the
adversary cannot get any feedback from the target and
has limited resources. In the study, we trained multiple
common deep neural networks (DNN), including VGG,
GoogLeNet and ResNet, and evaluated the transferability of adversarial examples across these models, under
various settings (shadower substitute networks, different structures and fewer data) to simulate a resourceconstrained attacker. This research sheds new light on
transferability: e.g., for ResNet, the transferability from
a 50-layer substitute to a 101-layer target is about 16.8%,
compared with 24.6% between the 101-layer substitute
and the same target, in an impersonation attack. More
interesting is the significant impact of training data
sizes: the transferability has dropped from 24.6% in an
impersonation attack to 14.5% when the substitute was
trained on a dataset one order of magnitude smaller than
that of the target, and further to 7.1% for a training set
two orders of magnitude smaller. Intuitively, substitutes
learned with fewer data or a shallower model would have
a looser boundary and therefore is less likely to ensure
a misclassification on the target (which is better trained
with more data). Overall, however, we only witnessed a
very limited success on transferability, particularly when
it comes to the impersonation attack, only around 20%
under different settings.
To attack an industry-grade FR system without querying it, a set of high-quality substitutes need to be built
to find common defects of the systems similar to or even
better trained than the target. However, constructing such
substitutes is hard, particularly with limited resource. In
our research, we studied what the adversary could do
to narrow this gap and enhance his odds of success. A
unique observation we have is that even though the target generally has a more precise decision boundary, the
substitute could still partially approach this boundary,
at points of interest (PoI): for example, a criminal may
leverage a large number of his own and his victim’s photos to boost the substitute’s accuracy with regard to the
identification of (just) these two individuals, for the pur-
Contributions. The contributions of the paper are outlined as follows:
• The NLL property and understanding of transferability. Our large-scale study reveals the impacts the training
data size can have on the successful transferring of an
adversarial instance from one model to another. More
importantly, we discovered the nearly linear relation between input images and their representations (in terms
of their differences) under an ideal model, which enables our query-free attack and might lead to better understanding of the fundamental defects in DL models.
• New techniques for query-free attacks. Based upon the
new discovery, we designed a new attack technique that
finds adversarial instances against a well-trained target
model without querying the target and using limited re-
2
sources. At the center of the technique is leverage of
additional victim and attacker images and the NLL property to train a substitute capable of simulating the target
model around PoIs, even when the adversary only possess a small amount of training data and much less computing resources. This makes an important step toward
understanding the realistic threat of adversarial learning.
More generically in the image processing area, three
DNN models have been extensively used. VGG16 [24] running 16 cascaded convolution layers was
reported to achieve state-of-the-art recognition results in the ImageNet Large-Scale Visual Recognition
Challenge 2014 [20] (ILSVRC-2014), together with
GoogLeNet [26], which involves 22 layers and an Inception architecture invented by Google for combining information from multi-views. Empowered by the
pervasiveness of GPU and Batch Normalization technologies [8], ResNet-152 [5] winning the ILSVRC-2015
classification task is armed with 152 layers and capable
of transferring shadow features to deep layers.
• Implementation and evaluation. We implemented the
technique and evaluated it over industry FR systems.
2
2.1
Background
Deep Learning and Face Recognition
Deep Neural Network. Deep Neural network (DNN) is
a function that projects the input domain onto an output
domain for classification and other purposes. Following
prior research [15], a DNN for image processing can be
formalized below:
2.2
Adversarial Learning
The potential of deploying DNN to real-world systems
(e.g., self-driving cars) faces the security challenges of
adversarial learning, an attack that manipulates the inputs to a DNN to cause misclassification. This attack was
first discussed by Szegedy et al. [27], who point out the
existence of adversarial examples, i.e., perturbed input
x0 similar to the original input x but misclassified by the
DNN into a different category. Such attacks can be targeted or not. In the latter case, the attacker seeks adversarial examples misclassified into any categories except
the one they belong to. For instance, in FR, the adversary wants to dodge a face detection system by slightly
changing his appearance from x to x0 , causing the classification result argmaxi F(x0 )i 6= argmaxi F(x)i . Here,
the DNN outputs a vector that describes the probabilities for the input belonging to different individuals. During a targeted attack, the adversary intends to impersonate a given individual t, by seeking a makeup x0 causing
argmaxi F(x0 )i = t.
F(x) = so f tmax(Z(x)) = y,
where x is the image serving as the input to the DNN,
y is its output, typically a vector of probabilities for the
image to be in different classes, and Z(x) is the “logits”,
a function describing all DNN layers except the “softmax” layer, whose outputs are unscaled log probabilities
serving as the inputs to “softmax”.
During its operations, a DNN first converts its inputs
into a representation, a high number of parameters that
capture the features of the inputs, and then hands it over
to the last a few layers of the network to generate the
output. For image classification and FR in specific, a
DNN can be further described as follows:
Z(x) = C ◦ R(x),
where R(X) outputs the representation of the input and
C(·) is the classification function that produces the “logits” based upon the representation. A well-trained DNN
is characterized by its capability to generate similar representations for similar inputs. This avoids the pitfall
when two similar inputs actually are mapped to very different representations and as a result, are assigned into
different classes. Note that in our research, the similarity
between two representations is measured by the cosine
distance between them.
Attack methods. To find adversarial examples, people
need to define the similarity between two images (the inputs), x and x0 , based upon a distance metric. Prior research on adversarial learning uses the L p distance, with
p being 0, 2 or ∞:
n
1
kx − x0 k p = ( ∑ |xi − xi0 | p ) p .
i=1
Here xi − xi0 is the subtraction between the i-th pixel
of two input images. Minimizing the L0 distance, we
can get x0 with the smallest number of pixels differing
from those on the original input x. The Jacobian-based
Saliency Map (JSMA) [18] is an attack optimized under
the L0 distance. It iteratively picks pixels that have the
most impact on the results and modifies them, until either a given threshold (an upper bound for the number of
pixels) is reached or an adversarial example is found.
Minimizing the L2 distance, we can obtain x0 that has
the least modification, in terms of Euclidean distance,
across all pixels on x and x0 . The first attempt using
Face recognition systems. Since the introduction of
deep convolutional neural networks (CNN) [11], FR
technologies have been evolving rapidly. As a prominent example, DeepFace [28] close the gap between the
recognition capabilities of human beings and machines.
Further, DeepID3 [25] attained a 99.53% accuracy on the
LFW dataset [6] that exceeds the human performance,
99.2%. More recently, FaceNet [22] exploited a deep
architecture to achieve a 99.63% accuracy on the same
dataset.
3
2.3
this distance is L-BFGS [27] that minimizes the L2 distance under the box-constraint, i.e., x0 ∈ [0, 1]n , where n
is the number of pixels. It exploited the classical gradient descend method to find the optimal solution with a
pre-defined learning rate lr:
x0 = x + lr · 5x F(x),
Threat Model
We consider an adversary that intends to perform a dodging attack or a impersonation attack on a target FR model
that he cannot query. The adversary does not have access
to the internal parameters of the target but has limited
information about its architecture (e.g., ResNet, VGG
or GoogLeNet) and its depth (e.g., about 100 layers for
ResNet, though the precise number of layers is still unknown to him). All such information about a commercial system is often made available through various public sources, such as research papers (e.g., the design of
Face++ was described in the paper [3]), technical reports
and other online documents.
The target model studied in our research is assumed to
be trained over a large amount of data, tens or even hundreds of millions of images, as those commercial FR systems are. On the other hand, the adversary does not have
that level of resources, though we do assume that he can
still acquire millions of images publicly available online,
as we did in the study. Further, the adversary can obtain
thousands of images of himself and the victim he want to
impersonate, and also sufficient resources from the cloud
to train the substitute model over the data. We believe
that these assumptions are all realistic, as demonstrated
in our research: particularly, all the computing power required for training our attack model can be purchased
from Amazon at an approximate cost of 10,000 dollars.
Specially, if the adversary can not obtain sufficient images of the victim from the internet, they can follow the
victim and record videos to get enough images that are
taken in various scenarios and from different angles.
(1)
Minimizing L∞ distance, we can find x0 with the smallest maximum-changes to the pixels. Under this distance,
the optimization algorithm seeks a region of pixels with
similar intensities to modify. An example of the prior attack is Fast Gradient Sign Method (FGSM) [4], which
iteratively updates x0 to produce an adversarial example
by stepping away a small stride along with the direction
of 5x F(x0 ).
In our research, we use L2 distance to measure the
changes to an image for finding its adversarial examples.
Transferability. As mentioned earlier, transferability
is the key to practical adversarial learning, when the
adversary cannot directly access the internal parameters of the target model. Prior research [13] demonstrates that around 20% adversarial examples discovered
from one of the three models (ResNet-152, VGG-16 and
GoogLeNet) are also misclassified by other two models
under a dodging attack. A more recent study [16] further
shows that transferability can happen even across different machine learning techniques: DNN, Logistic Regression (LR), Support Vector Machine (SVM) and Nearest
Neighbors (kNN). Particularly, more than 60% of adversarial examples discovered in LR or SVM were found to
be still effective on the other model. When it comes to
the impersonation attack, also 20% adversarial examples
were reported to work across different DNN models [13].
These examples were found using an ensemble-based approach that descends along the summation of the gradients of several models.
A primary limitation of these prior studies is that
they are all based upon relatively “small” datasets, such
as ILSVRC-2014 including 1000 categories. Compared with the industry-grade FR systems such as Facevisa, ColorReco, which are trained to classify tens or
even hundreds of thousands of identities, what has been
learned from these studies can be less conclusive. Also
importantly, the prior research either considers that the
substitute is built upon similar or even identical datasets
as the target , or at the very least, assumes that the adversary is capable of continuously querying the target to
collect data (query results) for training the substitute. As
discussed earlier, in many cases, these assumptions are
still a far cry from reality. Our research instead looked
into the transferability over a large dataset, when the adversary cannot query the target and does not have enough
data to train his substitute model.
3
Understanding Transferability across
Asymmetric Models
To understand whether a target model the adversary cannot query is still subject to attacks, we need to find out
the challenges in simulating the target’s operations, under the limited resources and information. For this purpose, we conducted the largest study on transferability,
using a dataset with 4 million images. Our research reveals the importance of training data size to a successful
cross-model attack.
3.1
Settings
Our study utilized MegaFace Challenge 2 [14], a dataset
including 672K identities and their above 4 million photos, and Caffe [9], an open-source deep learning framework, to train FR DNN models in our experiments. All
such experiments were conducted on a 8-GPUs server
with each GPU armed with 12GB memory.
In our studies, we assume our target model is F ∗ (·)
and it outputs a vector, F ∗ (x), for a input image x.
Our study covers both dodging attacks and the imper4
sonation attacks. Here, we say that an adversarial example x0 , that is similar to x, causes a dodging attack
if argmaxi F ∗ (x0 )i 6= argmaxi F ∗ (x)i , where F ∗ (·)i represents the i-th element of the output vector. In the case of
an impersonation attack, we have F ∗ (x0 )t > 0.5, where t
is the victim to be impersonated and we ensure t 6= o, the
true owner or x.
To find adversarial examples, in our study, we chose
L2 distance as our metric to optimize the following objective function [2]:
minimize
subjects and victims involved in the attacks. Over these
models, we analyzed the transferability of the dodging
and impersonation attacks, using an ensemble learning
that integrates the common adversary examples found in
4 substitutes trained independently to find those causing
the target to misclassify. More specifically, in the experiments, we studied both the dodging attacks, using 635
photos from 100 individuals randomly selected from our
dataset, and also the impersonation attacks, based upon
randomly chosen 600 photo pairs (each pair with the images of two different individuals). These images were
inside the training sets of both the target and substitutes.
The results are presented in Table 1. As we can see
from the table, for dodging, we observed a transferability
about 95%, and for impersonation, it became 20%. This
finding is pretty much in line with what is reported in the
prior research, indicating that transferring adversarial examples across different models are feasible, though less
effective in the case of impersonation.
1
ktanh(w) − xk22 + c · f (tanh(w)).
2
For the dodging attack, f is defined as
f (x0 ) = max(Z(x0 )o − max{Z(x0 )i : i 6= o}, −κ).
For the impersonation attack, f becomes
f (x0 ) = max(max{Z(x0 )i : i 6= t} − Z(x0 )t , −κ).
Depths. Further we looked into the impacts of depths
on transferability. For this purpose, we utilized ResNet,
since the depth of its structure can be easily adjusted.
More specifically, we built 4 ResNets with 50, 65, 80
and 101 cascaded convolutional layers respectively. The
compositions of their structures are also presented in Table 1. Using these structures, we also trained models on
different 100K identities’ photos (about 600K photos for
each model).
In this study, we ran those models as substitutes to attack the target (ResNet-101), both dodging and impersonation. As expected, the complexity of the network
(its depth) indeed affects transferability: more layers
make the DNN more capable and enhance transferability. Again, transferability tends to be low for the impersonation attack, around 16% when using ResNet-50 to
attack ResNet-101.
x0
Here, the adversarial example we found is
=
tanh(w∗ ), where w∗ is the optimal solution of above function. In that function, c is a parameter that balances the
importance of two components, the first component minimizing the L2 distance between the adversarial example x0 and the original image x, the second component
modeling the goal of this attack, either dodging or impersonation. Also, κ is a threshold indicating when the
attack goal is achieved. We set c = 20, κ = 20 for both
the dodging attack and the impersonation attack in our
experiments.
And we further improve the performance of above
function by exploiting the standard ensemble-based approach. Specifically, we will use K = 4 substitutes and
assemble them to solve the following function:
K
1
ktanh(w) − xk22 + c · ∑ f k (tanh(w)).
2
k=1
(2)
where f k (·) is the k-th substitutes.
minimize
3.2
3.3
Impacts of Data Size
Training size and transferability. An important observation is that a real-world adversary typically cannot get
as many photos as a large organization uses to train its
industry-grade FR system. An important question we
were asking is what impacts a relatively smaller dataset
could have on the chance of a successful cross-model attack. For this purpose, we trained VGG, GoogLeNet and
ResNet-101 models on three datasets, with 10K identities (60K photos), 100K identities (600K photos) and
300K identities (1.5M photos) respectively. Again, all
these individuals were randomly drawn from our dataset
and we made sure that there was no overlap across substitutes’ training set and target model’s training set. In
the experiment, we utilized the substitutes of the same
structure to attack (dodging and impersonating) the one
Impacts of Structural Features
Structures. As mentioned earlier, to understand the impacts of DNN structures on transferability, we looked
into the 3 most prominent structures: VGG1 , GoogLeNet
2 and ResNet3 . In our study, we trained a target model
and 4 substitutes for each structure using 600K photos
from 100K identities from the MegaFace Challenge 2
dataset. Such a training set was selected to ensure that it
did not overlap with those of other models except for the
1 https://github.com/davidgengenbach/vgg- caffe
2 https://github.com/BVLC/caffe/tree/master/models/bvlc- googlenet
3 https://github.com/KaimingHe/deep- residual- networks
5
Table 1: Transferability among Different Structures: Cell (i, j) corresponds to the transferability of adversarial examples that are generated from the ensemble of four model i for dodging/impersonation attack on model j.
ResNet-50
ResNet-65
ResNet-80
ResNet-101
GoogLeNet
VGG-16
ResNet-101
Dodging Impersonation
95.1%
16.8%
95.6%
18.0%
96.3%
23.2%
98%
24.6%
95%
16.5%
93.4%
17%
GoogLeNet
Dodging Impersonation
96.7%
20.2%
97.8%
23.1%
94.4%
18.1%
Table 2: Structures of Different Depths: All ResNet layers can be divided into 5 stages, starting with a convolutional layer, followed by four stages, each including a
different number of “bottleneck” blocks. Different stage
has different output size.
Stage 1
Stage 2
Stage 3
Stage 4
Stage 5
ResNet-50
1 Conv
3 Blocks
4 Blocks
5 Blocks
3 Blocks
ResNet-65
1 Conv
3 Blocks
4 Blocks
10 Blocks
3 Blocks
ResNet-80
1 Conv
3 Blocks
4 Blocks
15 Blocks
3 Blocks
VGG-16
Impersonation
96.4%
20.3%
94.6%
18.5%
97.2%
22.3%
Dodging
cles, public paper and other sources. On the other hand,
a deeper network with more data certainly need more
computing resources to train. For example, on our system with 8 GPUs, training a ResNet-101 model took 9
hours for a data set of 60K images, while only half of
the time was needed for a 50-layer model over the same
images. Most importantly, collecting a large number of
high-quality images is often a challenge for the adversary: for example, SenseTime Ltd’s model is reported to
be built from above 20M images and the dataset of this
scale could not be found on the Internet, up to our knowledge. Therefore, we believe that whether transferability
could be enhanced in the presence of a relatively small
set of training data is critical question for assessing the
practical impacts of adversarial learning on FR systems.
Also to attack a real-world system without querying
it, the adversary needs to estimate his chance of success
based upon the features of a given adversarial example,
for example, the percentage of the pixels modified. This
can not be easily done since the adversary does not have
access to the target system and therefore cannot figure
out the probability of success by testing his adversarial
examples on the target. However, our study described
above shows that the transferability between the substitute and the target can actully be gagued using the transferability between a model learnt from a smaller dataset
and the substitute. This is because the probability of success in the latter case is expected to be higher than that in
the former. A more formal analysis of this observation is
presented in Section 4.2.
ResNet-101
1 Conv
3 Blocks
4 Blocks
22 Blocks
3 Blocks
built upon 300K identities’ photos, The results are presented in Table 3. As we can see here, training data size
turns out to significantly affect transferability and such
an influence is also consistent across different structures.
Compared with the structural impacts, transferability became lower when we reduced the data size from 300K to
100K (1.5M to 600K images). Compared with the impact of depth, the attack was less likely to succeed when
we downsized the data (300K to 100K) than when we
removed layers (from 101 to 50).
Further we ran the models with 10K identities to exploit those with 100K, as illustrated in the right column
of Table 3. An interesting observation is that the transferability from 10K models to 100K models actually is
lower (e.g., 14.5%) than that for 100K to 300K (e.g.,
17.5%), even though the difference in the training data
is even larger in the latter case (200K) than the former
(90K). Intuitively, with the increase in training data, a
substitute becomes closer to a perfect model and adding
more data then can be less effective in improving the
model’s precision than the time when the model only
learns from a small set of data and therefore much less
accurate. Further analysis of the observation leads to the
conclusion that the enhancement of transferability will
slow down when the data size goes up (see Section 4.2).
4
Query-Free Asymmetric Attack
To enhance transferability, ideally we need to make the
substitute very similar to or even more accurate than the
target model. Although this sounds like a mission impossible for most real-world adversaries, given their limited
resources, particularly a much smaller training set they
are able to acquire, still something can be done to narrow the gap between the two models. A key observation
is that a unique resource the adversary often has is abundant photos of the subject (often himself) in a dodging
attack and also those of the victim in an impersonation
Discussion. Our study shows that although both structural features and training data size affect transferability, apparently the impact of the latter is more prominent.
In practice, the structural information of many commercial FR systems can often be found, from research arti6
Table 3: Transferability among Different Dataset Sizes: Cell (i, j) corresponds to the transferability of adversarial
examples that are generated from the ensemble of four model i for dodging/impersonation attack on model j.
ResNet-101
GoogLeNet
VGG-16
300K
100K
10K
300K
100K
10K
300K
100K
10K
Dodging
98.8%
81.3%
34.5%
97.3%
79%
32%
97.5%
77%
32.3%
300K
Impersonation
25.2%
17.5%
7.1%
24.1%
16.2%
5.1%
23.9%
16.3%
6.3%
attack. Leveraging such images, we could train a model
biased toward the subject or the subject and victim pair.
Even though such a model can be overfit and therefore
its overall accuracy could be way below that of the target
model, all we care about here is just the target’s behavior around the subject and/or the victim, which we could
potentially simulate in the substitute using the resource
(extra photos).
However, effective use of such resource turns out to
be challenging. Table 4 shows the experimental results
when we directly duplicate those photos of subjects and
the victims to 600K photos and add them to the original
600K (photos) training data for augmenting the transferability under the VGG, GoogLeNet and ResNet models
in attacking the targets trained over 1.5M photos of 300K
identities. From the table, we do not see a significant improvement in the effectiveness of the attack, compared
with those without such data augmentation.
Dodging
83%(81.3%)
80.2%(79%)
77.3%(77%)
71.2%
69.4%
66.7%
100K
Impersonation
14.5%
13.3%
12.9%
low 30% (for impersonation) to above 60% on commercial systems (Section 5.3).
4.1
Nearly Local Linearity
Essentially, an adversarial example exploits an imperfectly trained model, inducing its nonlinear behavior,
that is, a small perception change causing a big representation change.
Observation. A key observation we have is that the representations produced by an ideal DNN FR should accurately model the human perceptions: when two images
look very different to the human, the distance between
the representations should be large, and when the images
appear to be similar, the distance should become small.
So our idea to enhance the substitute model, with regard
to a given subject or a subject-victim pair, is to approximate such nearly linear relations in the substitute, between the subject and the victim or the subject and other
identities, for simulating the behaviors of a better trained
model around these data points. To this end, right metrics
need to be chosen to measure the human perception and
the distance between the representations. In our research,
we found that L2 (as used in the prior research [2]) and
Cosine distances can serve these purposes.
In our research, we trained three ResNet-101 models on three datasets with 10K identities (60K photos),
100K identities (600K photos) and 600K identities (4M
photos) respectively. From each dataset, we selected,
uniformly at random, 10K photo pairs, with the photos
from two different identities in one pair. Then between
each pair (a, b), we synthesized a series of 99 images by
equidistant interpolation. Formally, the k-th image can
k
(b − a). Specially, x0 = a
be represented as xk = a + 100
100
and x = b.
Then, we ran all three models on these interpolated
images to get their representations. Altogether, 106 representations were produced from the 10K image pairs.
Further we calculated the mean for the Cosine distances
between the representation of each image xk , denoted
by R(xk ), and that of b, R(b), across all their peers.
Table 4: Results for Naively Augmented Data: the original transferability is in the bracket.
ResNet-101
GoogLeNet
VGG-16
Dodging
Impersonation
18.2%(17.5%)
15.8%(16.2%)
16.8%(16.3%)
Intuitively, the subject and victim’s photos alone are
insufficient for simulating the relations (with regard to
facial features) established by a better trained model between them and between the subject and other identities
in the dataset. Such relations need to be built upon other
images between the pairs (e.g., those more similar to the
subject than the victim) and the way a well-trained DNN
scores them according to their features. Following we
show how such relations can be modeled using a new
property discovered in our research, called Nearly Local Linearity (NLL), the key technique behind our EXCIT attack, which helps augment the substitute model
for simulating the target’s behaviors around the subject
and/or the victim, boosting the transferability from be7
4.2
Fig 1 illustrates the relations between the L2 distance (divided by kb − ak2 ) between xk and a and their representation’s mean Cosine distance, together with ρ, which
is used to measure the mean slope of each model: ρ =
10−6 ∑k,(a,b) k−1 cos(R(xk ), R(b)).
0.5
0
0.5
0
0
0.5
1
1
0
0.5
0
0.5
1
1
1
0
0.5
0.5
normalized L2
rho=0.76836
Cosine
Cosine
Cosine
0.5
normalized L2
rho=1.0795
1
0
0.5
0
0
normalized L2
rho=1.0098
1
The discovery of the Nearly Local Linearity property in
the ideal DNN FR model enables us to train a substitute
to not only leverage the extra resource at the adversary’s
disposal but also integrate such resource into the model
by approximating the relations between these additional
photos and the existing images in the dataset. For this
purpose, we need to synthesize a set of “transitional” images as those interpolated photos mentioned earlier, and
redefine the optimization goals when training the substitute to connect these photos to others in an expected way.
These are the key steps for our EXCIT attack, as elaborated below.
1
Cosine
1
Cosine
Cosine
1
0.5
Subject-oriented data augmentation. To synthesize
“transitional” images and enrich the training dataset, we
designed a subject-oriented data augment algorithm (see
Algorithm 1). Our approach takes the original dataset
D, a subject-victim (identity) pair (o,t) and the number of images to synthesize n as its input and outputs
the augmented dataset Daug . For a dodging attack, o
and t will be set to the same input so the algorithm will
look for transitional images between all o’s images A
and randomly-selected other images B in the dataset.
Given the constraint of the total number of synthetic images n and the requirement that for each image pair, at
least 10 synthetic photos need to be injected, we have
n
/|A |. For each photo pair (a, b) from A × B,
|B| ≤ 10
our approach randomly (λ ∼ U(0, 1)) interpolate |A n||B|
transitional images.
0
0
normalized L2
0.5
1
0
normalized L2
0.5
1
normalized L2
Figure 1: NLL on ResNet-101: There are 3 columns,
from left to right, representing 600K, 100K, 10K model
respectively. The first row demonstrates the average results, and the second row demonstrates the results of a
randomly selected image pair.
As we can see from the figures, the relation between
the L2 distance and the Cosine distance approaches linear
with the increase of the training data size. Particularly, it
becomes almost linear for the ResNet model trained over
600K individuals (4 million images), with ρ = 1.01. Under the identical experimental settings, we observed the
same L2 and Cosine distance relation in the VGG and
GoogLeNet models, as illustrated in Fig 2. This indicates that the relations between the subject and interpolated images, and between the subject and the other victim can be captured by this NLL property.
0.5
0
0.5
0
0
0.5
1
1
0
0.5
0
1
1
1
0.5
0
0.5
0.5
normalized L2
rho=0.8097
Cosine
Cosine
Cosine
0.5
normalized L2
rho=0.94739
1
0
0.5
0
0
normalized L2
rho=0.99966
1
Training NLL-augmented models. With enriched data,
we can train a substitute to approximate the NLL property for given PoIs (the subject or the subject-victim
pair). In this way, the substitute is expected to be closer
to or even surpass a better-trained model at the PoIs,
which will elevate the transferability of the adversary
example our discovered. To build such substitutes, we
first train the model on the original dataset D using the
standard softmax loss function. After convergence, we
fine-tune the substitute on our augmented dataset with
n synthetic images, using the following loss function to
minimize the distance between the output of the image x,
F(x) and its expected value, which given a pair of identities u and v, has 1 − λ as the value for the uth element on
the output vector and λ as the vth element and all other
values set to 0.
1
Cosine
1
Cosine
Cosine
1
0.5
0
0
normalized L2
rho=0.99682
0.5
1
0
normalized L2
rho=0.96625
0.5
1
normalized L2
rho=0.79764
Figure 2: NLL property: There are 3 columns, from left
to right, representing 600K, 100K, 10K model respectively. The first row demonstrates results of VGG-16
models, and the second row demonstrates the results of
GoogLeNet models.
Concept. Formally, we define the NLL property as follows:
|R(b)·R(xλ )|
kR(b)k2 kR(xλ )k2
λ
k2
≈ λ = 1 − kb−x
kb−ak
2
The EXCIT Attack
− ∑x [(1 − λ )logF(x)u + λ logF(x)v ]
Here, the loss function uses the Kullback–Leibler divergence to compare the distribution of F(x) (with F(x)u
and F(x)v being its uth and vth elements respectively)
against the output vector (with its corresponding elements being 1 − λ and λ ).
(3)
where
xλ = a + λ (b − a), λ ∈ [0, 1]
8
where different substitutes cannot agree on the direction.
This weighted search algorithm is illustrated in Algorithm 2. In each step, it modifies the intermediate result
x0 to move along a certain direction (20-th to 21-th line),
as indicated by the sum of weighted average derivative
calculated from the substitutes (9-th line). Also, it drops
the average on the dimension where the substitutes do
not agree on the direction. More specifically, the algorithm sets the dimensions whose average derivative is
small while it deviation is large (15-th to 19 line). For
this purpose, we normalize the deviation of every dimension by the max value in this dimension (13-th line) and
the average by the max average value (16-th line). This
allows us to choose one uniform threshold for every dimension, which is set to 0.3. The algorithm for the dodging attack is similar.
Algorithm 1: Subject-oriented data augmentation algorithm.
Input: D, (o,t), n
Output: Daug
1 A = {a : a ∈ D, o = argmaxi R(a)i };
2 if o = t then
n
3
m = 10
/|A |;
4
Select a subset B from D, s.t., B ∩ A = 0/ and
|B| = m;
5 end
6 else
7
B = {b : b ∈ D,t = argmaxi R(b)i };
8 end
9 Daug = D;
10 for a in A do
11
for b in B do
12
for k = 1 to |A n||B| do
13
Sample λ from U(0, 1);
14
c = (1 − λ )a + λ b;
15
Daug = Daug ∪ {c};
16
end
17
end
18 end
Algorithm 2: Modified ensemble-based algorithm
for impersonation attacks.
1
2
3
4
5
Finding adversarial examples. After generating multiple substitutes enhanced by NLL, we need to effectively
assemble them to find common adversarial examples.
This can be done directly, using the standard ensemblebased approach (Eq 2). However, a question is how to effectively use our substitutes trained on NLL augmented
dataset. The standard approach is to average the gradients discovered from individual substitute model. However, we found a way to outperform the standard approach by computing a weighted average and optimize
the following function:
6
7
8
9
10
11
12
13
14
15
16
minimize
1
ktanh(w) − xk22
2
17
18
K
+c · ∑ f k (tanh(w))(1 − cos(Rk (tanh(w)), x)). 19
k=1
20
(4)
Specifically, since each substitute we use already approximates the NLL property around the victim, we can
determine the “quality” of the gradient it provides based
upon the Cosine distance between its representation of
the updated image (the intermediate result for finding
the adversarial example) and that of the victim’s image:
the smaller it is, the more similar these images would be
based upon their L2 distance. Therefore, we set higher
weights in favor of the gradients from the substitutes producing smaller representation differences. In the meantime, we discard the gradient values in the dimension
21
22
Input: {Rk (·)}, { f k (·)}, a, b, c, K, lr
Output: x0
d = Dim(x0 );
x0 = a;
w = tanh−1 (a);
while Not converge do
for k = 1 to K do
αk = 1 − cos(Rk (x0 ), Rk (b));
sk = αk 5w f k (x0 );
end
g = ∑i si / ∑ j α j ;
for j = 1 to d do
h j = {s1 j , s2 j , ...};
p j = mean(h j );
q j = std(h j )/max(h j );
end
for j = 1 to d do
p
if q jj /max({p1 , p2 , ...}) ≤ 0.3 then
g j = 0;
end
end
w = w − lr(c · g + 5w 12 kx0 − xk22 );
x0 = tanh(w);
end
4.3
Analysis
To find out how EXCIT enhances the transferability in
a query-free, black-box attack, we analyzed our implementation using a ResNet-101 model trained on 1.5M
images of 300K identities as the target, and a set of
ResNet-101 models trained on 600K images of 100K
identities (no overlap with the target’s training set ex9
Percentage
1
0.5
5
10
L2 distance
15
20
0.5
10
15
5
20
10
15
20
L2 distance
L2 distance
Figure 4: Impersonation performance: The left figure
shows the distribution of modifications made by approaches with and without NLL enhance. The right figure shows the transferabilities of them.
Table 5: Impersonation Transferability of NLLenhanced Approach on Different Structures.The number in the bracket is the transferability using 300Ksubstitutes to attack the target 300K-model.
with NLL
without NLL
ResNet-101
49% (71.3%)
17.5% (22.7%)
GoogLeNet
45.8%
16.1%
VGG-16
43.1%
16.4%
size study (Section 3.3).
Distance from the subject. Without querying the target,
naturally the adversarial examples discovered by EXCIT,
through simulating a “better” model, tend to be farther
away from the subject. What we want to know, however, is for a given distance from the subject, whether
the examples found by our approach still have a higher
probability of success, compared with the attack without
the NLL enhancement. For this purpose, we compare
both methods’ average transferability under various L2
distances. The results are given in Table 6. A more detailed study, by restricting the searches within a certain
distances is given in the Appendix B. From these results,
we observe that NLL improves the transferability in every distance.
Table 6: Transferability under Different Distance.
0.5
No NLL
NLL
0
0
No NLL
NLL
0
5
1
Transferability
Percentage
No NLL
NLL
0.5
0
Transferability. As we can see from Fig 3 and Fig 4, in
both attacks, EXCIT improved the transferability, which
were evident for dodging (from 81.2% to 89.6%) and
dramatic for impersonation (from 17.5% to above 49%).
Interesting here is that for the dodging attack, our approach even close to the attack using the substitutes
trained on the same data size of the target, indicating that
our EXCIT model was actually effective in recognizing
the subject. This is further supported by the findings for
the impersonation attack, in which none of the substitute
models without the NLL enhancement could come even
close to our performance, even for those as well-trained
as the target. Actually, even for a ResNet-101 trained on
1.5M images of 300K identities, we found that our substitutes got a transferability of 49%.
1
1
No NLL
NLL
Transferability
cept attack subjects and victims) as substitutes. The latter were further augmented with the NLL property under different attack settings, using the similar subject
and victim data for the data-size study reported in Section 3.2. Specifically, in the dodging attacks, the same
set of 100 identities and their 635 images were utilized.
For each identity, 4 substitutes were trained over about
300K transitional images automatically synthesized. In
the impersonation attacks, we selected 600 photo pairs
from 10 identity pairs (subject-victim). For all identities
involved in impersonation attacks, we ensure that each of
them have at least 100 photos in our data set. And for a
certain identity pair, each photo pair (not only those selected photo pairs) between them were augmented with
30 transitional photos for training the 4 substitutes. Under both attack settings, the target was attacked with the
common adversarial examples for all 4 substitutes. The
results were compared with our findings in the data-size
study (Section 3.3).
5
10
15
Dodging with NLL
Dodging without NLL
Impersonation with NLL
Impersonation without NLL
20
L2 distance
Figure 3: Dodging performance: The left figure shows
the distribution of modifications made by approaches
with and without NLL enhance. The right figure shows
the transferabilities of them.
Further our study shows that EXCIT also works on
other DNN structures: we trained a VGG model and
a GoogLeNet over the 600K images to perform an impersonation attack on their corresponding targets with
the same structures but trained on the 1.5M images, and
found that (Table 5) both attacks achieved around 45%
transferability, way above the 16% reported in our data-
< 10
87.7%
79.5%
50.5%
14.5%
< 15
88.3%
81.2%
47.2%
17.5%
< 20
89.6%
81.2%
49%
17.5%
Training cost. To understand how EXCIT helps a
resource-limited adversary, we evaluated the cost for
training substitutes over our 8-GPU server (with the
12GB memory for each GPU). As illustrated in Table 9,
it took about 200 hours to train a ResNet-101 over 1.5M
images and more than 500 hours to build the model over
4M images, while constructing a substitute on 600K images used 75 hours. Note that we could fully parallelize
the training of 4 substitutes, but could not do this for
a target model over the same amount of computing re10
5.1
sources, due to the communication overheads.
To further analyze the cost EXCIT could reduce,
we trained 4 EXCITs of ResNet-50 and 4 EXCITs of
ResNet-101 over 60K photos to perform impersonation
attacks against a ResNet-101 target trained over 600K
photos and 4M images. As we can see in Tabel 7, with
the small amount of training data, our approach elevated
the transferability of these attacks to 30-40% for the
600K target and around 20% for the 4M target, while
the training time stayed at 5 to 9 hours per substitute.
Table 7: Impersonation Transferability with NLLenhanced Approach of 60K Photos’ Models. Cell (i, j)
is the result of using model i to attack model j.
ResNet-50 60K
ResNet-101 60K
600K
38.5%
44.2%
4M
16.8%
23.1%
Also we compared the efficiency of our approach
against the attacks without the NLL enhancement. In the
latter case, the only way to improve the transferability
is to train more substitutes to find their common adversary examples. In our study, we conducted three experiments using 4, 8 and 16 substitutes respectively, each
model trained over 60K images. The results of ensemble
adversarial learning over these models are presented in
Table 8. As we can see here, when attacking the 4M target, even with 16 substitutes, the attack could not achieve
the same level of transferability as the 4 NLL-enhanced
substitutes, even for the those with only 50 layers. In
this case, the cost of EXCIT, in terms of training time, is
no more than 16.8% of the direct attack (with 16 substitutes).
Table 8: Impersonation Transferability of Ensemblebased Models (no NLL) Trained over 60K Photos.
5.2
5
Images
60K
60K
600K
4M
Time
5h
9h
75h
>500h
Attack on Online APIs
The three online APIs attacked in our research are ColorReco5 , FaceVisa6 and Face++7 . These models were
trained with a large amount of data. For example, Face++
was built upon 5M photos of 20K identities [30] and
FaceVisa was upon 2M photos. Also they all demonstrated a high recognition accuracy over the Labeled
Faces in the Wild (LFW) dataset [6]: 99.4% for ColorReco, 99.5% for FaceVisa and 99.5% for Face++. In our
experiments, we ran a python script to automatically upload our test photo pairs to ColorReco and FaceVisa. For
Face++, we had to do it manually due to the requirement
of CAPTCHA solving.
The success rates of our attacks are presented in Table 10. Note that in the experiments, the thresholds
4 models 8 models 16 models
7.2%
12.7%
16.5%
Table 9: Cost for Training Different Models.
Depth
50
101
101
101
Experimental Settings
Unlike the models built in our analysis, which were
trained over the subject and victim’s images and output
a vector to specify whether an input image belongs to
these identities, a real world FR system takes two photos as inputs and calculates a score about the similarity
of the individuals in these two photos. Here is how we
determined whether an adversarial example worked on
these systems:
In a successful dodging attack, we expect that the target system outputs a low score (< T hdod ) for two images: one is the subject’s original photo and the other
is the adversarial example generated by our approach
from the original photo. In a successful impersonation
attack, the target system is supposed to output a high
score (> T himp ) for two images: the victim’s photo and
the adversarial example generated by our approach from
the subject’s image, indicating that they all belong to the
same individual.
In our experiments, we first trained 4 ResNet-101
models on randomly sampled 3M photos from 600K
identities in the MegaFace Challenge 2 dataset4 . From
all identities, we selected 10 individuals as the subjects
in our dodging attack. For the impersonation, we exploited 10 subject-victim pairs. Every identity involved
has at least 100 images in the dataset. In the experiments,
we randomly chose 10 of each individual’s images for
the dodging attack and 10 photo pairs for each subjectvictim pair to execute the impersonation attack. For each
of these subjects or subject-victim pairs, we further enhanced the 4 models using the NLL augmentation and
then ran our algorithm (Algorithm 2) to find adversarial
examples to attack them.
Memory
8x4.5G
8x6.5G
8x7.5G
8x8G
Evaluation on Real-World Systems
We evaluated our approach, NLL-enhanced attacks, on
four real world systems. Three of them are online, with
APIs available for the public, and the last one is a commercial system without open access, one of the products
from SenseTime Ltd. We performed both dodging and
impersonation attacks against them. The details of our
experiments and our findings are elaborated below.
4 We did not use all 4M for each substitute in an attempt to make
these substitutes diverse.
5 http://www.colorreco.com/faceCompare
6 http://www.facevisa.com/web/index/demo
7 https://www.faceplusplus.com/face-comparing/#demo
11
15
15
10
10
10
5
0
# of pairs
15
# of pairs
# of pairs
for different APIs are different and defined by the APIs
themselves. Besides, we considered that an attack failed
if the target FR system could not detect face from the adversarial example submitted, even for the dodging attack.
As we can see from the table, our approach achieved a
higher accuracy in the dodging attack, compared with the
attacks without the NLL enhancement (Table 5). A much
bigger boost, however, is observed for the impersonation
attack, in which EXCIT raised the success rates for all
three systems from around 20% to 67-82%. Fig 5 further
illustrates the distributions for the scores of our submitted photo pairs.
5
0
0
0.5
1
Score
60%. Fig 6 further shows examples for the successful attacks. We have reported our findings to SenseTime and
are helping them improve their system.
5
0
0
0.5
1
0
Score
0.5
1
Score
Figure 6: Successful impersonation attacks on SenseTime. The three columns are the subject photos, our
generated adversarial examples and the target photos respectively. The modification of the first case is 13.98 and
the second case is 7.32.
Figure 5: Distributions of scores for impersonation attacks: from left to right, they are the results of ColorReco, Facevisa and Face++ respectively.
Table 10: Success rate against online APIs.
dodging
T hdod
impersonation
T himp
5.3
ColorReco
98%
0.75
74%
0.80
Facevisa
96%
0.64
82%
0.74
6
Face++
93%
0.623
67%
0.691
Related Works
Our approach utilizes a synthesized dataset to fine-tune
our substitutes, for the purpose of approximating the
NLL property at PoIs. Synthesized data have also been
used in the prior research [17], to support completely different techniques and to different purpose. Specifically,
the prior research uses synthesized data to query the target model to train the substitute (1000 times to achieve
an 84.24% success rate in transferring adversarial examples). This is exact the attack scenario our query-free
approach is designed to avoid. Without communicating
with the target model, the only thing we can do is to
build a substitute as well-trained as the target, so as to
captures their common structural weaknesses to enhance
transferrability. Such an attack is found to be completely
feasible through simulating the target’s behaviors around
PoIs, based upon the NLL property we discovered and
additional images collected from the victims and the attackers.
Our approach also exploited the ensemble method.
The original ensemble-based approach is proposed by
Liu et al. [13]. Their work focus on the transferability
among DNN models with different structures, whereas
the data size is the factor that really matters in face recognition domain, as have been demonstrated before. Compared with them, our method can increase the transferability from a model trained with insufficient data to a
model trained with plenty of data, which goes beyond
their method’s capability. Another ensemble method is
proposed by Sarkar et al. [21]. They trained a DL model
to find “universal” perturbations fooling all their pre-
Attack on Industrial System
Commercial FR systems are often better trained and
more capable than the free FR APIs, which are mostly
used for online demo. Such industry-grade systems are
typically characterized by a large number of layers, and
being trained over a massive amount of data on clusters
of GPUs. The services they provide are not open to the
public and only available for purchase. In our research,
we obtained the commercial SDKs from SenseTime Ltd.
through our collaborations. SenseTime’s products are
known to be among the leading FR systems [25]. So the
system we analyzed represents the state-of-the-art in FR
technologies. It was trained over 20M photos for 1M individuals, using a ResNet-like model, though the details
of the structure are commercial secrets. The model tested
in our study was estimated to require at least 14,000
hours (50 epochs) to train, over our GPU server. By comparison, all 4 EXCIT substitutes used in our attack were
trained for 2,500 hours in total. With less than 1/5 of the
time spent on training the models, our approach achieved
a high success rate for the 100 individual selected for the
dodge attack and 100 pairs for the impersonation attack:
in the former case, 89% of transferability was achieved,
compared to 70% without the NLL enhancement, and the
in the latter, we raised the transferability from 11% to
12
trained target models. Particularly, their object function
combines both the sum of targets’ (mis)classification loss
and the scale of the finding perturbations. The difference
from their method is similar with above. In our attacking
scenario, their method may not work.
7
We also found that distillation does not work on EXCIT either: more specifically, We implemented the defense on our 300K target with temperature T = 20, and
ran four 100K substitutes to attack it. The result is
that distillation can only reduce our transferability from
89.6% to 88.8%, for dodging, and from 49% to 45.6%,
for impersonation.
Alternatively, we can consider to insert “secret” into
the commercial system. Specifically, train the system on
a custom dataset where all the photos are covered by a secret pattern. Since queries are not supposed to be made
to the target during an attack, the secret added to a commercial system could help mitigate the EXCIT threat. In
general, however, defense against adversarial learning is
known to be hard [1]. Further research is needed to find
an effective way to defeat our attack.
Discussion
Understanding NLL. Unlike existing cross-model attacks, EXCIT does not even interact with the target, so
there is no way for our approach to exploit the specific
defects of the target. The reason we can still find highly
transferable examples similar to the subject is that by
simulating better-trained models around PoIs, our approach is likely to discover some common (potentially
structural) defects fundamental to a certainly type of
DNN, and in the meantime, avoid exploring the subspace unlikely to contain adversarial examples, given the
reduction of training-specific weaknesses (e.g., lack of
sufficient data) around PoIs. Under an NLL-enhanced
substitutes, we could even discover transferable examples with changes restricted to given facial features: e.g.,
around the eyes (Appendix C), which allows the prior attack [23] (using printout glasses to evade detection) to
work in a query-free, black-box setting.
In the meantime, our understanding of transferability
is still limited. Still less clear are the questions such as
whether there exist adversarial examples inherent to certain DNN structures or even the fundamental design of
artificial neural networks. Further studies on these issues
are certainly important.
Our current definition of NLL describes a relation between Cosine distance and the L2 distance. However,
such a relation may not be general, particularly when it
comes to non-FR problems: as an example, we found
that the NLL-enhanced models still improved transferability but less significantly over Cifar10 [10], a dataset
for image classification. This could be attributed to the
unique features of FR: e.g., differences between two
faces can be added to to another one to form a new face,
which makes the “transitional” images easy to construct;
also FR tasks are characterized by a large number of
training categories, compared with other tasks (e.g., 1K
categories for ILSVRC vs. 600K entities for Megaface),
forcing the DNN models to map input images to a highdimensional sphere with a maximum space utilization.
All these features make NLL more effective for FR.
What is less clear, however, is how to expand the concept to enhance the transferability of other tasks, which
should be investigated in future research.
Cost of the attack. As mentioned earlier, training the
substitutes to attack SenseTime’s FR system took 2,500
hours on our server. An estimate cost for such resources
is about 10000 dollars on Amazon AWS. The computing time here can be shortened through parallelization,
since all 4 substitutes can be trained together. Also, the
computing cost could be reduced when the adversary attempts to impersonate multiple victims or hide multiple
subjects. In this case, only one set of substitutes need to
be trained over our dataset, which can later be augmented
with NLL for different subjects or subject-victim pairs to
support dodging or impersonation attacks.
8
Conclusion
In this paper, we present our new understanding of DNNbased FR systems, in terms of their vulnerability to the
transferable attack under a resource-constrained adversary. Our research shows that limited resources, particularly smaller training sets, can have a significant impact
on the effectiveness of the attack. This is important since
a real-world adversary typically cannot query the target
frequently and needs to build substitutes as capable as,
or even more powerful than the target model under his
limited resources. Narrowing such a resource gap, however, turns out to be feasible through a novel technique
we developed. Specifically, we found that the adversary
could make an effective use of the extra information (images) about subjects and victims in his possession, by approximating the relations of these PoIs with other images
in the training set characterized by a Near local linearity property we discovered. As a result, we can grossly
elevate the transferability in both a dodging and an impersonation attack by training NLL-enhanced models by
nearly 50% in attacking industry-grade systems. With
our new techniques and findings, still more effort needs
to be made to better understand transferability and mitigate the threat it poses.
Defense. Defensive distillation [19] has been demonstrated to be effective against most of previous attacks.
However, as pointed out by Carlini et al [2], a modified version of existing attacks will break them defense.
13
References
[12] L E C UN , Y., B OTTOU , L., B ENGIO , Y., AND
H AFFNER , P. Gradient-based learning applied to
document recognition. Proceedings of the IEEE 86,
11 (1998), 2278–2324.
[1] C ARLINI , N., AND WAGNER , D. Adversarial examples are not easily detected: Bypassing ten detection methods. arXiv preprint arXiv:1705.07263
(2017).
[13] L IU , Y., C HEN , X., L IU , C., AND S ONG ,
D.
Delving into transferable adversarial examples and black-box attacks. arXiv preprint
arXiv:1611.02770 (2016).
[2] C ARLINI , N., AND WAGNER , D. Towards evaluating the robustness of neural networks. In Security and Privacy (SP), 2017 IEEE Symposium on
(2017), IEEE, pp. 39–57.
[14] N ECH , A., AND K EMELMACHER -S HLIZERMAN ,
I. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (2017).
[3] FAN , H., C AO , Z., J IANG , Y., Y IN , Q., AND
D OUDOU , C. Learning deep face representation.
arXiv preprint arXiv:1403.2802 (2014).
[4] G OODFELLOW, I. J., S HLENS , J., AND S ZEGEDY,
C. Explaining and harnessing adversarial examples. stat 1050 (2015), 20.
[15] PAPERNOT, N., AND M C DANIEL , P. On the effectiveness of defensive distillation. arXiv preprint
arXiv:1607.05113 (2016).
[5] H E , K., Z HANG , X., R EN , S., AND S UN , J. Deep
residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 770–778.
[16] PAPERNOT, N., M C DANIEL , P., AND G OODFEL LOW, I. Transferability in machine learning: from
phenomena to black-box attacks using adversarial
samples. arXiv preprint arXiv:1605.07277 (2016).
[6] H UANG , G. B., R AMESH , M., B ERG , T., AND
L EARNED -M ILLER , E. Labeled faces in the wild:
A database for studying face recognition in unconstrained environments. Tech. rep., Technical Report 07-49, University of Massachusetts, Amherst,
2007.
[17] PAPERNOT, N., M C DANIEL , P., G OODFELLOW,
I., J HA , S., C ELIK , Z. B., AND S WAMI , A. Practical black-box attacks against machine learning.
In Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security
(2017), ACM, pp. 506–519.
[7] H UANG , S., PAPERNOT, N., G OODFELLOW, I.,
D UAN , Y., AND A BBEEL , P. Adversarial attacks on neural network policies. arXiv preprint
arXiv:1702.02284 (2017).
[18] PAPERNOT, N., M C DANIEL , P., J HA , S.,
F REDRIKSON , M., C ELIK , Z. B., AND S WAMI ,
A. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P),
2016 IEEE European Symposium on (2016), IEEE,
pp. 372–387.
[8] I OFFE , S., AND S ZEGEDY, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (2015), pp. 448–456.
[19] PAPERNOT, N., M C DANIEL , P., W U , X., J HA , S.,
AND S WAMI , A. Distillation as a defense to adversarial perturbations against deep neural networks.
In Security and Privacy (SP), 2016 IEEE Symposium on (2016), IEEE, pp. 582–597.
[9] J IA , Y., S HELHAMER , E., D ONAHUE , J.,
K ARAYEV, S., L ONG , J., G IRSHICK , R.,
G UADARRAMA , S., AND DARRELL , T. Caffe:
Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia (2014), ACM,
pp. 675–678.
[20] RUSSAKOVSKY, O., D ENG , J., S U , H., K RAUSE ,
J., S ATHEESH , S., M A , S., H UANG , Z., K ARPA THY, A., K HOSLA , A., B ERNSTEIN , M., B ERG ,
A. C., AND F EI -F EI , L. ImageNet Large Scale Visual Recognition Challenge. International Journal
of Computer Vision (IJCV) 115, 3 (2015), 211–252.
[10] K RIZHEVSKY, A., AND H INTON , G. Learning
multiple layers of features from tiny images.
[11] K RIZHEVSKY, A., S UTSKEVER , I., AND H IN TON , G. E. Imagenet classification with deep convolutional neural networks. In Advances in neural
information processing systems (2012), pp. 1097–
1105.
[21] S ARKAR , S., BANSAL , A., M AHBUB , U., AND
C HELLAPPA , R. Upset and angri: Breaking
high performance image classifiers. arXiv preprint
arXiv:1707.01159 (2017).
14
[22] S CHROFF , F., K ALENICHENKO , D., AND
P HILBIN , J. Facenet: A unified embedding for
face recognition and clustering. In Proceedings
of the IEEE Conference on Computer Vision and
Pattern Recognition (2015), pp. 815–823.
the target, called target simulators or simply simulators,
run substitutes to attack them and then collect the statistics about the relation between the features of the adversarial examples discovered in substitutes and the transferability of the examples. Given such a relation, the attacker can look at the features of an example to estimate
the likelihood that it could fool the target. This simple
approach, however, does not work in practice, as we do
not have the resources (e.g., a large number of images)
to build such powerful simulators.
Therefore in our research, we took a different path including three steps: firstly, we train simulators on the
data we have; secondly, we estimate the difference between simulators and the target, thirdly, we plus the estimated difference to our simulators’ outputs to predict
the target’s output. In this process, estimation the difference is challenge, cause we don’t know how the target
looks like. But we can know the scale of the training set
of the target model. Thus we built a function g(mα , mβ )
to estimate the difference between models trained on mα
photos and mβ photos. The detailed description of g(·) is
given in Appendix A. Specifically, our study shows that
when training data grows, the substitute becomes similar
to the target, and the impact of their data size difference
becomes less prominent. Next, plussing the average difference to every simulator’s output, we derive what the
target model would output for adversarial examples and
can choose the best one with the largest likelihood that it
will fool the target.
To build g(mα , mβ ) measuring the difference between
two models trained on mα photos and mβ photos, we
need to find a “bridge” to connect them. A nature idea is
leveraging their loss that measures how far way they are
from the perfect model and further implementing the triangle inequality to estimate the difference between themselves. While the classic softmax loss is inappropriate
here, cause it not satisfies the triangle inequality. Thus
we used the Cosine distance again. We define the Cosine
distance loss of a model R(·) as following:
[23] S HARIF, M., B HAGAVATULA , S., BAUER , L.,
AND R EITER , M. K. Accessorize to a crime: Real
and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM SIGSAC
Conference on Computer and Communications Security (2016), ACM, pp. 1528–1540.
[24] S IMONYAN , K., AND Z ISSERMAN , A. Very
deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556
(2014).
[25] S UN , Y., L IANG , D., WANG , X., AND TANG , X.
Deepid3: Face recognition with very deep neural
networks. arXiv preprint arXiv:1502.00873 (2015).
[26] S ZEGEDY, C., L IU , W., J IA , Y., S ERMANET, P.,
R EED , S., A NGUELOV, D., E RHAN , D., VAN HOUCKE , V., AND R ABINOVICH , A. Going deeper
with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition
(2015), pp. 1–9.
[27] S ZEGEDY, C., Z AREMBA , W., S UTSKEVER , I.,
B RUNA , J., E RHAN , D., G OODFELLOW, I., AND
F ERGUS , R. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013).
[28] TAIGMAN , Y., YANG , M., R ANZATO , M., AND
W OLF, L. Deepface: Closing the gap to humanlevel performance in face verification. In Proceedings of the IEEE conference on computer vision and
pattern recognition (2014), pp. 1701–1708.
[29] X U , W., E VANS , D., AND Q I , Y. Feature squeezing: Detecting adversarial examples in deep neural
networks. arXiv preprint arXiv:1704.01155 (2017).
[30] Z HOU , E., C AO , Z., AND Y IN , Q. Naive-deep face
recognition: Touching the limit of lfw benchmark
or not? arXiv preprint arXiv:1501.04690 (2015).
A
where
Lcos (R) = ∑ lcos (a, b, R)
a,b (
1 − |cos(R(a), R(b))|, same identity;
lcos (a, b, R) =
|cos(R(a), R(b))|, different identity.
And we count the mean and the standard deviation of
Cosine distance loss for models trained on different data
size. The results are showed on Fig 7. Now, we can infer
the mean (µβ ) and the standard deviation (δβ ) of the target model trained on mβ photos, according to the fitted
curve. Further, we assume the target model’s and simulators’ losses obey the normal distribution N (µβ , δβ2 )
and N (µα , δα2 ) respectively, and roughly estimate the
difference by assuming it also obey the normal distribu-
Transferability prediction
Given an adversarial example discovered, the attacker
needs to have some idea how likely the example could
also mislead the target model. Also in the presence of
multiple examples, the most promising one would be
given preference. One way to estimate the transferability
of an example is to train multiple models as capable as
15
.
0.12
0.1
0.08
10
12
14
16
log(# of photos)
data
fitted curve
0.12
grows quickly, moving the objective function away from
the optimality. Therefore, in the optimal situation, L2
should not exceed γ much.
In our research, we evaluated our approach using the
objective function when γ = 5, γ = 20 and γ = 30, and
exploiting 4 EXCITs trained on 600K photos of 100K
identities to attack the target trained on 1.5M photos of
300K identities. The results are presented in Fig 8 and
Table 12. As we can see, under various distances, the
adversarial examples found by our approach are always
much more transferable than the one without the NLL
enhancement. In the meantime, our approach tends to
pick up the examples away from the subject, given the
fact that the NLL property moves the decision boundary
of the substitute (with regard to the subject and the victim) closer to the ideal one, making it harder to find the
adversarial examples close to the subject’s image, though
once such an image is found, it is more likely to lead to
a successful attack.
0.1
0.08
0.06
10
12
14
16
log(# of photos)
Figure 7: Cosines distance loss of models trained on different size of data.
tion N (µβ − µα , δβ2 + δα2 ). Thus, for a certain simulator R(·), we can calculate out what is the likelihood that
1 − |cos(R(a), R(b))| plus the estimated difference will
surpass 0.5 for dodging attack or lower than 0.5 for impersonation attack.
Table 11: Statistics of Lcos (Rmα ) − Lcos (Rmβ ).
mβ
1541811
4019407
4019407
µ
0.0188
0.0371
0.0163
δ
0.1176
0.1225
0.1001
1
1
No NLL
NLL
0.8
Percentage
mα
62258
62258
1541811
0.6
0.4
0.6
0.4
0.2
0.2
0
0
5
10
15
20
1
No NLL
NLL
0.8
No NLL
NLL
0.8
Percentage
0.14
Percentage
0.14
data
fitted curve
Std of Cosine loss
Mean Cosine loss
0.16
0.6
0.4
0.2
0
15
20
25
30
25
30
35
40
Besides, in Table 11, we list true values of the distance
of some pairs of (mα , mβ ), and observe that, along with
the increase of data size, the simulator becomes similar
to the target, and the impact of their data size difference
becomes less prominent (small µ and δ ).
Figure 8: Distributions of modifications under different
γ.
B
Table 12: Impersonation Transferability of NLLenhanced Approach under Different Distances.
L2 distance
Performance in Difference Distance
Without querying the target, naturally the adversarial examples discovered by EXCIT, through simulating a “better” model, tend to be farther away from the subject.
What we want to know, however, is for a given distance
from the subject, whether the examples found by our approach still have a higher probability of success, compared with the attack without the NLL enhancement. For
this purpose, we need to modify the objective function
of the DNN to limit its search within a given distance (in
terms of L − 2 in our research) constraint, as follows:
minimize
γ=
with NLL
without NLL
C
5
6.3%
3.8%
L2 distance
20
51.3%
21.8%
30
74%
49.5%
Restrict the modification to certain region
The trivial method to restrict modifications within a certain region is to quench those derivatives out of the
region, while finding the adversarial examples. However, in this setting, finding adversarial examples becomes harder than before. So it is need to totally release the constrain on the magnitude of modifications.
We demonstrate two examples restricting modifications
around eyes on Fig 9. We observe that the modifications
become severe: the L2 distance between generated adversarial example and the original photo of the first case
is 24.63, and of the second case is 27.04.
exp(ktanh(w) − xk2 − γ) + f (w).
Here we use an exp function that penalizes the L2 distance when exceeding γ. More specifically, we calculated its derivative as follows:
exp(L2 − γ) ·
L2 distance
tanh(w) − x
5w tanh(w) + 1 · 5w f (w)
k tanh(w) − xk2
where L2 represents k tanh(w) − xk2 . As we can see here,
When L2 > γ, the component involving the exp function
16
Figure 9: Successful impersonation attacks within restricted region.
17
| 1cs.CV
|
Exploring an Infinite Space with Finite Memory Scouts
arXiv:1704.02380v2 [math.PR] 11 Apr 2017
Lihi Cohen1 , Yuval Emek2 , Oren Louidor3 , and Jara Uitto4
1
Technion, Israel. Email: [email protected]
Technion, Israel. Email: [email protected]
3
Technion, Israel. Email: [email protected]
4
Comerge AG, Switzerland. Email: [email protected]
2
Abstract
Consider a small number of scouts exploring the infinite d-dimensional grid with the aim of
hitting a hidden target point. Each scout is controlled by a probabilistic finite automaton that
determines its movement (to a neighboring grid point) based on its current state. The scouts,
that operate under a fully synchronous schedule, communicate with each other (in a way that
affects their respective states) when they share the same grid point and operate independently
otherwise. Our main research question is: How many scouts are required to guarantee that the
target admits a finite mean hitting time? Recently, it was shown that d + 1 is an upper bound
on the answer to this question for any dimension d ≥ 1 and the main contribution of this paper
comes in the form of proving that this bound is tight for d ∈ {1, 2}.
1
Introduction
Model and Contribution. Consider c scouts u1 , . . . , uc that explore the d-dimensional infinite
grid Zd . Each scout is controlled by a probabilistic finite automaton with (finite) state set S so
that at any given time n ∈ Z≥0 , the global system configuration is characterized by the scouts’
location vector Xn ∈ Zd×c and state vector Qn ∈ S c .1
The scouts sense their respective grid points through the following mechanism: At time n ∈ Z≥0 ,
each scout ui observes its local (binary) environment vector Eni ∈ {0, 1}S defined so that Eni (q) = 1
if and only if there exists some j 6= i such that Xnj = Xni and Qjn = q. In other words, scout ui can
sense the current state q ∈ S of another scout uj if they reside in the same grid point. Notice that
the environment Eni is fully determined by the configuration (Xn , Qn ).
The scouts’ actions are controlled by a (probability) transition function 2
Π : S × {0, 1}S × S × {−1, 0, +1}d → [0, 1]
defined so that Π ((q, e), (q ′ , ξ)) is the probability that at time n + 1, scout ui resides in state
i
Qin+1 = q ′ in grid point Xn+1
= x + ξ given that Qin = q, Eni = e, and Xni = x. The application
of the transition function Π for scout ui at time n is independent of its application for scout uj at
time n′ for any j 6= i or n 6= n′ .3
To complete the model specification, each scout ui is associated with an initial state q0i ∈ S
so that Qi0 = q0i . Furthermore, it is assumed that at time 0, all scouts share the same initial
location X0i = x0 ∈ Zd ; this is typically taken to be the origin, i.e., x0 = 0, however later on in the
technical sections, we also consider scouts with arbitrary initial locations. We refer to the 5-tuple
hc, S, x0 , q0 , Πi as a scout protocol.
The process (Xn , Qn )n≥0 will be referred to as a scout process (or a c-scout process). Given a
scout protocol P = hc, S, x0 , q0 , Πi, it is completely defined via
(X0 , Q0 ) = (xc0 , q0 )
and
P (Xn+1 − Xn = Ξ, Qn+1 = q | (Xm , Qm )nm=0 ) =
c
Y
i=1
Π (Qin , Eni ), (q i , Ξi ) a.s.4
Throughout, we use the boldfaced notation v to denote the vector v = (v 1 , . . . , v c ).
A function Π : A × B 7→ [0, 1], where A and B are countable non-empty sets, is called a (probability) transition
P
function or transition kernel if b∈B Π(a, b) = 1 for all a ∈ A.
3
The choice of introducing a single transition function Π for all scouts, rather than a separated transition function
Πi for each scout ui , does not result in a loss of generality, as the space set S can be partitioned into pairwise disjoint
subsets S = S 1 ∪ · · · ∪ S c , ensuring that Qin ∈ S i for every 1 ≤ i ≤ c and n ∈ Z≥0 .
4
Notice that both sides of the last equation are random variables. It should be interpreted as stating that for every
V
n, x0 , . . . , xn ∈ Zd×c , and q0 , . . . , qn ∈ S c , if the event n
m=0 (Xm , Qm ) = (xm , qm ) occurs with positive probability,
then for every Ξ ∈ Zd×c , and q ∈ S c , the probability, conditioned on that event, that Xn+1 = xn + Ξ and Qn+1 = q
Q
equals to ci=1 Π qni , ein , q i , Ξi , where ein is the environment of scout ui with respect to the configuration (xn , qn ).
1
2
1
for all n ≥ 0, Ξ ∈ Zd×c and q ∈ S c , where Ξi denotes the i-th (d-dimensional) component of Ξ.
The hitting time of grid point x ∈ Zd under P is defined as inf{n ≥ 0 : ∃i s.t. Xni = x}. The
scout protocol P is said to be effective if every grid point x ∈ Zd admits a finite mean hitting time.
It turns out that c = d + 1 scouts are sufficient for the design of an effective scout protocol on Zd
for every dimension d ≥ 1 (see [18, Section 3.2] and the related literature discussion of the present
paper). In this paper, we prove that c = d + 1 is also a necessary condition for d ∈ {1, 2} and
conjecture that this holds for d ≥ 3 as well.
Theorem. For d ∈ {1, 2}, any effective scout protocol on Zd requires at least d + 1 scouts.
Conjecture. For d ≥ 3, any effective scout protocol on Zd requires at least d + 1 scouts.
Related Work. Graph exploration appears in many forms and is widely studied in the
CS/Mathematics literature. In the general graph exploration setting, an agent (or a group thereof)
is placed in some node of a given graph and the goal is to visit every node (or every edge) by
traversing the graph along its edges.5 There are plenty of variants and modifications of the graph
exploration problem, where one natural classification criterion is to distinguish between directed
graph exploration [1, 14], where the edge traversals are uni-directional, and undirected graph exploration [6, 15, 17], where the edges can be traversed in both directions. The present paper falls
into the latter setting.
The graph exploration domain can be further divided into the case where the nodes are labeled
with unique identifiers that the agents can recognize [17, 29], and the case where the nodes are
anonymous [8, 11, 30]. The exact conditions for a successful exploration also vary between different
graph exploration works, where in some papers the agents are required to halt their exploration
process [15] and sometimes also return to their starting position(s) [5]. The setting considered in
the present paper requires neither halting nor returning to the initial grid point and the nodes
are anonymous (in fact, since our scouts are controlled by finite automata that cannot “read”
unbounded labels, unique node identifiers would have been meaningless).
A standard measurement for the efficiency of a graph exploration protocol is its time complexity [29]. When the agents are assumed to operate under a synchronous schedule, this measurement
typically counts the number of (synchronous) rounds until the exploration process is completed
in the worst case. This can be generalized to asynchronous schedules by defining a time unit as
the longest delay of any atomic action [8, 12]. A good example for a synchronous search problem
concerning the minimization of the worst-case search time is the widely studied cow path problem, introduced by Baeza-Yates et al. [7]. This problem involves a near-sighted cow standing at
a crossroads with w paths leading to an unknown territory. By traveling at unit speed, the cow
wishes to discover a patch or clover that is at distance d from the origin in as small time as possible.
5
In some research communities, it is common to refer to the exploring agents as robots, ants, or particles. Since the
agents considered in this paper differ slightly from those considered in the existing literature, we chose the distinctive
term scouts.
2
Baeza-Yates et al. proposed a deterministic algorithm called linear spiral search and showed that in
the w = 2 case, this algorithm will find the goal in time at most 9d. The classic cow path problem
was extended to the multiple cows setting by López-Ortiz and Sweet [28].
Another common measure for the efficiency of a graph exploration protocol is the size of the
memory that the agents can maintain [15, 24]. In the present paper, the execution is assumed to
be synchronous and the scouts are controlled by (probabilistic) finite automata whose memory size
is constant, independent of the distance to the target.
Graph exploration with agents controlled by finite automata has been studied, e.g., by Fraigniaud et al. [25] who showed that a single deterministic agent needs Θ(D log ∆) bits of memory to
explore a graph, where D stands for the diameter and ∆ stands for the maximum degree of the
input graph. Considerable literature on graph exploration by a finite automaton agent focuses on
a special class of graphs called labyrinths [9, 11, 16]. A labyrinth is a graph that can be embedded on the infinite 2-dimensional grid and has a set of obstructed cells that cannot be entered by
the agent. The labyrinth is said to be finite if the set of obstructed cells is finite and the term
exploration in the context of finite labyrinths refers to the agent getting arbitrarily far from its
(arbitrary) starting points. It is known that any finite labyrinth can be explored by a deterministic
finite automaton agent using 4 pebbles, where a pebble corresponds to a movable marker, and some
finite labyrinths cannot be explored with one pebble [10]. While the infinite 2-dimensional grid
studied in the present paper is a special (somewhat degenerate) case of a finite labyrinth and our
scouts are also controlled by finite automata, the goal of the scout protocols considered here is to
hit every grid point and we allow randomization.
An extensively studied special case of a single agent controlled by a probabilistic finite automaton is that of a (simple unbiased) random walk. Aleliunas et al. [2] proved that for every finite
(undirected) graph, the random walk’s cover time, i.e., the time taken to visit every node at least
once, admits a polynomial mean. Alon et al. [4] studied the extension of this classic model to
multiple agents. Among other results, they proved that in some graphs, the speed-up in the mean
cover time (as a function of the number of agents) is exponential, whereas in other graphs, it is
only logarithmic. Cooper et al. [13] investigated the case of c independent random walks exploring
a random r-regular graph, and analyzed the asymptotic behavior of the graph cover time and the
number of steps before any of the walks meet.
Random walks receives a lot of attention also in the context of infinite graphs. A classic result
in this regard is that a random walk on Zd is recurrent — namely, it reaches every point with
probability one — if and only if d ≤ 2; nevertheless, even in this case, it is only null-recurrent: the
mean hitting time of some (essentially all) points is infinite. This gives rise to another research
question asking whether the exploration process induced by c > 1 (non-interacting) random walks
on Zd is positive-recurrent, that is, the mean hitting time of all points is finite. It is relatively
well known (can be derived, e.g., from Proposition 4.2.4 in [27]) that for d = 1, the minimum
value of c that yields a positive-recurrent exploration process is c = 3, whereas for d = 2, the
3
exploration process induced by c random walks remains null-recurrent for any finite c. Cast in this
terminology, the research question raised in the present paper is how many agents are required to
turn the exploration process on Zd (dealing with d ∈ {1, 2}) into a positive-recurrent one, given
that the agents are augmented with a finite memory logic and local interaction skills.
The power of team exploration and the effect of the agents ability to communicate with each
other has been studied in various settings. Fraigniaud et al. [23] investigated the exploration of a tree
by c locally interacting mobile agents and constructed a distributed exploration algorithm whose
running time is O(c/ log(c)) times larger than the optimal exploration time with full knowledge of
the tree. They also showed that for some trees, in the absence of communication, every algorithm
is Ω(c) times slower than the optimal. Another good example for the advantages in the agents
communication can be found in the work of Bender et al. [8] showing that two cooperating robots
that can recognize when they are at the same node and can communicate freely at all times can learn
any strongly-connected directed graph with n indistinguishable nodes in expected time polynomial
in n.
Exploration processes by agents with limited memory occur also in nature. A prominent example for such a process is ant foraging that was modeled recently by Feinerman et al. [21, 20] as
an abstract distributed computing task involving a team of n ants that explore Z2 in search for
an adversarially hidden treasure. A variant of this model, where the ants are controlled by finite
automata with local interaction capabilities was studied in [19, 26].
While the focus of [19, 26], as well as that of [21, 20], is on the speed-up as n grows asymptotically, Emek et al. [18] investigated the smallest number of ants that can guarantee that the treasure
is found in finite time. They studied various different settings (e.g., deterministic ants, ants controlled by pushdown automata, asynchronous schedules), but when restricted to ants controlled by
probabilistic finite automata operating under a synchronous schedule, their model is very similar
to the one studied in the present paper. Cast in our terminology, they established the existence of
an effective scout protocol with 3 scouts on Z2 and their proof can be easily extended to design an
effective scout protocol with d + 1 scouts on Zd for any fixed d ≥ 1 (see [18, Theorem 3]). They also
claimed that there does not exist any effective scout protocol with 1 scout on Z2 (see [18, Theorem
7]), but the proof of this claim admits a significant gap;6 the impossibility result established in
the present paper regarding the 1-dimensional grid clearly subsumes that claim. For a survey on
generalized graph exploration problems, see, e.g., [22].
Paper Organization. The primary technical contribution in this paper is the proof that there
does not exist an effective scout protocol on Z2 with 2 scouts. This proof is presented in Section 2
with some necessary tools developed in Section 3. It turns out that the line of arguments leading
6
Specifically, the authors of [18], who has a non-empty intersection with the authors of the present paper, defined
in Section 5.1 of that paper a certain subset Zs of Z and claimed that every point in Zs has a finite mean hitting
time. This claim is not substantiated in [18] and we now understand that its proof is far more demanding than the
general intuition we had in mind when [18] was written.
4
to this proof also includes, in passing, a proof for the secondary result of this paper: there does not
exist an effective scout protocol on Z1 with 1 scout. For clarity, in Section 4, we refine the proof of
this secondary Z1 result from that of the primary Z2 case.
Two Scouts on Z2
2
Our main goal in this section is to establish the following theorem.
Theorem 2.1. Let (Xn , Qn )n≥0 be a scout process on Z2 with protocol P = h2, S, 0, q0 , Πi. Then
there exists some grid point x ∈ Z2 of which the expected hitting time is infinite, namely,
E inf{n ≥ 0 : Xn1 = x or Xn2 = x} = ∞ .
(1)
The proof of Theorem 2.1 is presented in a top-down fashion. Its main ideas and arguments
are introduced in Section 2.1. The argumentation relies on three non-trivial propositions which are
proved in Sections 2.3, 2.4, and 2.5. The proofs of these propositions rely on a generalization of the
scout process notion which is presented in Section 2.2. They also make use of some random walk
hitting time estimates which are relegated to section 3.
2.1
Top Level Proof
For the remaining of this section, let us suppose towards a contradiction that (1) does not hold.
The first step is to show that if this is the case, then the two scouts must meet infinitely often with
probability one and moreover, the distribution of the time between successive meetings must have
a stretched-exponentially decaying upper tail. Formally, let N0 := 0 and set
Nk := inf{n > Nk−1 : Xn1 = Xn2 },
k = 1, 2, . . . .
Proposition 2.2. Let (Xn , Qn )n≥0 be a scout process on Z2 with two scouts, state space S, transition function Π, initial position x0 = 0, and initial state q0 ∈ S 2 and suppose that (1) does not
hold. Then, with probability one, all Nk are finite. Moreover, there exists δ > 0 such that for all
k = 0, 1, . . . and u ≥ 0, almost surely
P (Nk+1 − Nk > u | QNk ) ≤
√
1 −δ u
e
.
δ
In view of Proposition 2.2, we now define for k = 0, 1, . . . , the random variables
Ak := QNk ,
2
1
,
= XN
Yk := XN
k
k
Rk+1 := Nk+1 − Nk ,
5
and R0 := 0
(2)
and observe that Ak is the states of the scouts’ automata at the time of their k-th meeting, Yk
is their position at this time, and Rk is the time that elapsed from the k − 1-st meeting. The
strong Markov property of the scout process then implies that (Yk , Ak , Rk )∞
k=0 is a Markov chain
2
2
on Z × S × Z≥1 . Moreover, thanks to the spatial homogeneity of the transition function of
(Xn , Qn )n≥0 , we further have for all k ≥ 0, almost surely
P (Yk+1 − Yk , Ak+1 , Rk+1 ) ∈ · | (Ym , Am , Rm )km=0 = P (Yk+1 − Yk , Ak+1 , Rk+1 ) ∈ · | Ak .
(3)
Notice that Rk is an upper bound on the maximal distance traveled by the scouts between their
k − 1-st and k-th meetings, namely,
∀i = 1, 2 , n ∈ [Nk−1 , Nk ] : max
In particular,
Xni − Yk−1 , Xni − Yk
≤ Rk .7
(4)
inf{k ≥ 0 : kYk − xk ≤ Rk+1 } ≤ inf{n ≥ 0 : Xn1 = x or Xn2 = x} .
It follows that if (1) is false, then for every x ∈ Z2 , we must have
E (inf{k ≥ 0 : kYk − xk ≤ Rk+1 }) < ∞ .
(5)
The process (Yk , Ak , Rk )k≥0 can be viewed as describing a single explorer on Z2 which has the
ability to explore a ball of radius Rk+1 around its location Yk for k = 0, 1, . . . Let us first define
such a process in a formal manner.
Definition 2.3. Let S be a finite non-empty state space, Π : S × S × Z2 × Z≥1 → [0, 1] a
(probability) transition function, x0 ∈ Z2 , and q0 ∈ S. An explorer process on Z2 with state space
S, transition function Π, initial position x0 and initial state q0 is a random process (Xn , Qn , Rn )n≥0
on Z2 × S × Z≥1 satisfying
(X0 , Q0 , R0 ) = (x0 , q0 , 0)
and for all n ≥ 1, ξ ∈ Z2 , q ∈ S, r ∈ Z≥1 ,
P (Xn+1 − Xn = ξ, Qn+1 = q, Rn+1 = r | (Xm , Qm , Rm )nm=0 ) = Π Qn , (q, ξ, r) a.s.
By (3), this definition applies to the process (Yk , Ak , Rk )k≥0 . Furthermore, by Proposition 2.2
and (4), the conditional distributions of both Yk+1 − Yk and Rk+1 given Ak have (at least) a
stretched-exponentially decaying upper tail (with deterministic constants). Proposition 2.4 will
now show that if, in addition, (5) holds, then such an explorer process must eventually get trapped
in a finite set of grid points.
7
Unless stated otherwise, the notation k · k is used to denote the ℓ∞ norm.
6
Proposition 2.4. Let (Xn , Qn , Rn )n≥0 be an explorer process on Z2 such that for some δ > 0 and
all n ≥ 0, u ≥ 0,
√
(6)
P (kXn+1 − Xn k + Rn+1 > u | Qn ) ≤ 1δ e−δ u a.s.
If for all x ∈ Z2 ,
E (inf{n ≥ 0 : kXn − xk ≤ Rn+1 }) < ∞ ,
(7)
there must exist a stopping time τ (for the explorer process) and a non-random r < ∞ such that
P τ < ∞ , ∀n ≥ τ : kXn − Xτ k < r = 1 .
(8)
Now, we set T := Nτ and observe that T is a stopping time for the process (Xn , Qn )n≥0 .
Therefore, Proposition 2.2, Proposition 2.4 and the assumed validity of (5) imply that there exist
r < ∞ and a stopping time T , such that with probability one, after time T < ∞, the two scouts
meet only inside a ball of radius r around the grid point XT1 = XT2 = Yτ . Moreover, they keep
meeting infinitely often and the times between successive meetings have finite means. Since this
implies, in particular, that whenever a scout is away from this ball, it evolves independently of the
other scout, it follows that each scout must always be in a grid point from which the mean time of
returning to the ball, as a single scout process, is finite. Proposition 2.5 will show that the set of
such grid points is a “small” subset of the whole grid.
To make things precise, given α̂ ∈ R2 with kα̂k2 = 1 and M > 0, let us define the thick ray of
width M in direction α̂ as
n
o
R(α̂, M ) := x ∈ R2 : |hx, α̂⊥ i| < M and hx, α̂i > −M ,
where α̂⊥ denotes (any) unit vector which is perpendicular to α̂ and hx, yi denotes the scalar
product between x and y. For what follows, if (Zn )n≥0 is a Markov chain on A and z0 ∈ A, then we
shall write Pz0 (·) for the probability measure under which Z0 = z0 . The same notation will apply
to the corresponding expectation.
Proposition 2.5. Let (Xn , Qn )n≥0 be a single scout process on Z2 with state space S and transition
function Π and let also r > 0. There exist m ∈ Z≥1 , unit vectors α̂1 , . . . , α̂m ∈ R2 and M < ∞,
such that if (x0 , q0 ) ∈ Z2 × S are such that
E(x0 ,q0 ) (inf{n ≥ 0 : kXn k < r}) < ∞
then
x0 ∈ D :=
m
[
i=1
R(α̂i , M ) .
(9)
(10)
en , Q
en )n≥0 be a single scout process on Z2
We can now finish the proof of the theorem. Let (X
with state space S and transition function Π as in the conditions of Theorem 2.1. Let also r be
7
given by Proposition 2.4 applied to the process (Yk , Ak , Rk )k≥0 . Applying Proposition 2.5 with
en , Q
e n )n≥0 and r, we obtain a subset D ⊆ R2 as defined in (10). We now claim that for i = 1, 2,
(X
P XTi +n ∈ Yτ + D , ∀n ≥ 0 = 1 .8
(11)
Indeed, for n ≥ 0 consider the first time after T + n such that scout i comes to within distance
smaller than r of Yτ , namely,
bni := inf m ≥ 0 : kXTi +n+m − Yτ k < r .
N
i
bni ≤ Nτ +n . Using the strong
Since Nτ +n ≥ T + n and kXN
− Yτ k < r, it follows that T + n + N
τ +n
Markov property, the almost surely finiteness of τ and Proposition 2.2 (or (6)), we then have
bni ≤ E (Nτ +n − Nτ − n) ≤
EN
=
n
X
k=1
n
X
E E Nτ +k − Nτ +k−1 QNτ +k−1
) N1
τ +k−1
E E(0,QN
k=1
< ∞,
where the inner expectation in the last line is with respect to the process (Xn , Qn )n≥0 .
On the other hand,
i
i
b i ≥ P Xi
b
∈
/
Y
+
D
·
E
EN
N
X
∈
/
Y
+
D
.
τ
τ
n
T +n
n
T +n
3−i
b i ≤ inf{m ≥ 0 : X i
In light of Proposition 2.4, we know that N
n
T +n+m = XT +n+m } with probability
one. From this fact and the strong Markov property, it follows that
bni XTi +n ∈
E N
/ Yτ + D
em − Yτ k < r} XTi +n ∈
= E E(X i ,Qi ) inf{m ≥ 0 : kX
/ Yτ + D ,
T +n
T +n
en , Q
en )n≥0 . But by Proposition 2.5,
where the inner expectation is just with respect to the process (X
under the conditioning the inner expectation is infinite almost surely. We therefore must have that
P(XTi +n ∈
/ Yτ + D) = 0. Summing over all n and using the union bound yields (11).
Finally, the validity of (11) for i = 1, 2 implies that with probability one,
Xni ∈ BT ∪ Yτ + D
∀i = 1, 2 and n = 0, 1, . . . ,
where henceforth Bρ is the closed ball of radius ρ around 0 in the ℓ∞ norm. Since T < ∞, the
random set BT ∪ Yτ +D is always a strict subset of Z2 . It follows that there must exist x ∈ Z2 such
that with positive probability, x will not be reached by either scout. This in turn shows that (1)
holds, in contradiction to what we have assumed in the first place, thus establishing Theorem 2.1.
8
If x ∈ Rd and A ⊆ Rd then x + A stands for the set {x + y : y ∈ A}.
8
2.2
Single and Generalized Scout Processes and Underlying Automaton
If (Xn , Qn )n≥0 is a single scout process, its environment vector En is always 0S and hence one can
effectively consider a reduced version Π′ : S × (S × {−1, 0, +1}d ) 7→ [0, 1] of the transition function
Π of the underlying protocol, where
Π′ q, (q ′ , ξ) = Π (q, 0S ), (q ′ , ξ) , q, q ′ ∈ S , ξ ∈ {−1, 0, +1}d .
The marginal process (Qn )n≥0 then becomes a Markov chain on S by itself (with transition function
P
given by (q, q ′ ) 7→ ξ Π′ q, (q ′ , ξ) ) and we shall refer to this process as the scout’s (underlying)
automaton. In particular, the irreducible classes of the automaton will be the irreducible states
classes of the Markov chain (Qn )n≥0 and the automaton will be called irreducible in this Markov
chain is such.
bn , Q
bn )n≥0 ,
In the sequel, we shall also need to consider a (single) generalized scout process (X
which is defined as the single scout process above, but with more general steps. Formally, for d ≥ 1,
the process, which now takes values in Rd × S, is defined exactly as before, only that the transition
b : S × (S × W) 7→ [0, 1], where W is some discrete nonempty subset
function now takes the form Π
bn )n≥0 , which is Markovian by itself,
of Rd . As in the case of a single scout process, the process (Q
will be referred to as the underlying automaton. The following lemma will be used more than once
in what follows.
bn , Q
bn )n≥0 be a generalized single scout process on Rd for d ≥ 1 with state space
Lemma 2.6. Let (X
b initial position x0 = 0 and some initial state q0 ∈ S. Suppose also that
S, transition function Π,
its automaton is irreducible and set
b n = q0 } , ζ := X
bT .
T := inf{n > 0 : Q
If P(ζ = 0) = 1, then there exist r ∈ (0, ∞) and a finite subset A ⊆ Br ⊂ Rd such that
and for all x ∈ A and u ∈ Z≥0 ,
with some C, C ′ > 0.
bn ∈ A = 1 .
P ∀n ≥ 0 : X
′
bn =
P ∀0 ≤ n ≤ u : X
6 x ≤ Ce−C u .
(12)
(13)
Proof. We will show that for each q ∈ S there exists xq ∈ Rd such that almost surely for all n,
bn = q ⇒ X
b n = xq .
Q
(14)
b n = q0 ⇒ X
bn = 0 .
Q
(15)
This will prove (12) with A := {xq : q ∈ S} and r = max{kxq k : q ∈ S}. To this end, observe first
that if ζ = 0, then with probability one
9
Suppose now that there are xq 6= x′q ∈ Zd and n1 , n2 ≥ 0 such that
bn = q , X
b n = xq > 0 , P Q
bn = q , X
b n = x′ > 0 .
P Q
1
1
2
2
q
(16)
bn )n≥0 is irreducible and hence recurrent, it follows from (15),(16) and the Markov property
Since (Q
that there exist m1 , m2 ≥ 0 such that
b m = q0 , X
bm = 0 > 0 , P(x′ ,q) Q
b m = q0 , X
bm = 0 > 0 .
P(xq ,q) Q
1
1
2
2
q
(17)
But then using the product rule with the first event in (16) and the second event in (17) and the
Markov property again, we get
bn +m = q0 , X
bn +m = xq − x′ > 0 .
P Q
1
2
1
2
q
Since xq − x′q 6= 0, this leads to a contradiction to (15) occurring with probability 1.
bn )n≥0 together
To show (13), we use the irreducibility and finiteness of the Markov chain (Q
with standard theory to claim that the hitting time of any q ∈ S has an exponentially decaying
upper tail. Then (13) follows by virtue of (14).
2.3
Two Scouts on Z2 Must Meet Frequently
In this section we prove Proposition 2.2. We shall do this gradually, first assuming that the two
scouts evolve independently as single scout processes and have irreducible automata (in the sense
discussed in beginning of Subsection 2.2). We shall then remove the irreducibility restriction and
finally consider the full (dependent) two-scout process.
Let therefore (Xn1 , Q1n )n≥0 and (Xn2 , Q2n )n≥0 be two independent single scout processes. For
i = 1, 2, suppose that scout process i has state space S i , transition function Πi , initial position
xi0 ∈ Z2 and initial state q0i ∈ S i . As before, we shall write Xn for (Xn1 , Xn2 ), Qn for (Q1n , Q2n ), etc.
We also define,
N := inf{n > 0 : Xn1 = Xn2 }
and for y ∈ Z2 ,
τy := inf{n ≥ 0 : Xn1 = y or Xn2 = y} .
The first lemma deals with the case of irreducible automata.
Lemma 2.7. Let (Xn1 , Q1n ) and (Xn2 , Q2n ) be two independent single scout processes on Z2 with
state spaces S and transition functions Π, starting from initial positions x0 and initial states q0 .
Suppose also that the automaton of each scout processes is irreducible. Then either
y ∈ Z2 : E(N ∧ τy ) = ∞
10
=∞
(18)
or for all u ≥ 0,
√
P(N > u) ≤ 1δ e−δ(
u−kx10 −x20 k)
.
(19)
for some δ > 0 which depends only on the transition functions Π.
Proof. We shall observe each scout i at times of successive returns to q0i as well as both scouts
together at times of successive simultaneous returns to states q0 = (q01 , q02 ). This will turn the
scouts processes into random walks.
Formally, let T0i = 0, T0∆ = 0 and for k = 1, . . . define:
i
: Qin = q0i } i = 1, 2 ,
Tki := inf{n ≥ Tk−1
∆
: Q1n = q01 , Q2n = q02 } .
Tk∆ := inf{n ≥ Tk−1
For n ≥ 0, and with e1 = (1, 0) ∈ Z2 , we also set:
Sni := hXTi i , e1 i i = 1, 2
n
Sn∆ := hXT1 ∆ − XT2 ∆ , e1 i .
,
n
n
Then for i = 1, 2, ∆ and k ≥ 1 we let:
i
i
, Rki = 2νki .
, νki := Tki − Tk−1
si0 := S0i , ζki := Ski − Sk−1
For i = 1, 2, the Markov property and spatial homogeneity of the process (Xni , Qin ) imply
that the triplets (ζki , νki , Rki )k≥1 are i.i.d. and consequently that (Sni , Tni )n≥0 is a random walk.
Since the process (Xn , Qn )n≥0 can be viewed as a single scout process on Z4 , the same applies to
(ζk∆ , νk∆ , Rk∆ )k≥1 and (Sn∆ , Tn∆ )n≥0 . Moreover, since the underlying scout automata are irreducible,
standard Markov chain theory implies that there exists C > 0 such that for i = 1, 2, ∆,
P(ν1i > u) ≤ C −1 e−Cu .
(20)
Then, since |ζ1i | ≤ R1i = 2ν1i , we also have
i
P(|ζ1i | + ν1i + R1i > u) ≤ δi e−δ u .
(21)
for some δi > 0.
For i = 1, 2, ∆, let the effective drift of walk i be defined as
di := (Eζ1i )/(Eν1i ) .
We next claim that
d∆ = d1 − d2 .
Setting Xn∆ := Xn1 − Xn2 , it is enough to show that for i = 1, 2, ∆, with probability one,
i
lim hXm
, e1 i/m = di .
m→∞
11
(22)
To this end, for all m ≥ 0, we set
i
Km
:= sup{n ≥ 0 : Tni ≤ m} .
By the law of large numbers, with probability one, for all ǫ > 0, there exists n0 such that for all
n > n0 ,
n(Eζ1i − ǫ) ≤ Sni ≤ n(Eζ1i + ǫ)
(23)
and
n(Eν1i − ǫ) ≤ Tni ≤ n(Eν1i + ǫ) .
i ,Ti
Since m ∈ [TK
i
Ki
m
m +1
(24)
) by definition, for all m large enough (24) implies
i
m/(Eν1i + 2ǫ) ≤ Km
≤ m/(Eν1i − ǫ) .
(25)
But then from (23),
m
Eζ1i + ǫ
Eζ1i − ǫ
i
≤
S
≤
m
.
i
K
m
Eν1i + 2ǫ
Eν1i − ǫ
Since this is true for all ǫ > 0, this shows that as m → ∞, almost-surely
i
i
i
i
SK
i /m → (Eζ1 )/(Eν1 ) = d .
m
(26)
At the same time,
i
i
i
≤ TK
hXm
, e1 i − SK
i
i
m +1
m
i
− TK
i .
m
Dividing by m both sides above, taking m → ∞ and using (24) and (25) we obtain
i
i
lim hXm
, e1 i/m − SK
i /m = 0 .
m→∞
m
Together with (26) this shows (22).
Care must be taken to handle the possible degeneracy of the steps of the walks (Sni , Tni )n≥0 , for
i = 1, 2, ∆. We therefore first assume that
P(ζ1∆ = 0) < 1 ,
(27)
and relegate the treatment of the complementary case to a later point in the proof. As for i = 1, 2,
if ζ1i /ν1i is a non-random constant, then we must have ζ1i = di ν1i with probability 1. Applying then
bni , Q
bin )n≥0 ,
Lemma 2.6 for the (single, irreducible one dimensional) generalized scout process (X
defined by setting
b i := X i − xi , e1 − di n
X
n
n
0
b i := Qi ,
Q
n
n
,
;
n = 0, . . . ,
bni | ≤ r i for all n ≥ 0. It follows that
we obtain r i ∈ (0, ∞) such that almost surely, |X
P ∀n ≥ 0 : hXni − xi0 , e1 i − di n ≤ r i = 1 ,
12
(28)
(29)
and accordingly we redefine (ζki , νki , Rki )k≥1 and (Sni , Tni )n≥0 as
∀k ≥ 1 : ζki := di , νki = 1 , Rki = r i
;
∀n ≥ 0 : Sni := si0 + di n , Tni = n .
(30)
Observe that under the new definition, (Sni , Tni )n≥0 is still a random walk, albeit with deterministic
i = m.
steps which trivially satisfy (21) for some δi > 0. Moreover Km
With the above definitions, the process (Sn∆ , Rn∆ )n≥0 is a look-around random walk of the type
considered in sub-section 3.1 and the processes (Sni , Tni , Rni )n≥0 for i = 1, 2, form a pair of timevarying look-around random-walks of the type considered in subsection 3.2. Moreover, by the
definition of Rni , we have
i
i
i
≤ RK
∀i = 1, 2, m ≥ 0 : hXm
, e1 i − SK
i
i
m
m +1
.
(31)
We now appeal to Proposition 3.8 with the random walks (Sn1 , Tn1 , Rn1 )n≥0 and (Sn2 , Tn2 , Rn2 )n≥0
and to Lemma 3.6 with the random walk (Sn∆ , Rn∆ )n≥0 . Let ρ > 0 be as given by the first proposition,
τρ be as in (67) for the walk Sn∆ and set Nρ∆ := Tτ∆ρ . Suppose first that P(Nρ∆ < ∞ , Nρ∆ < N ) > 0.
Then there exist n ≥ 0 and si1 ∈ Z for i = 1, 2, ∆ such that (70) holds and
P N > Nρ∆ = n , Sni = si1 : i = 1, 2, ∆ > 0 .
(32)
For such si1 , by Proposition 3.8, there exist −∞ < x < y < ∞ such that |y − x| > 2 and (71)
holds. But then for each z ∈ Z2 satisfying hz, e1 i ∈ [x, y] and kz − x10 k > n and kz − x20 k > n, by
the Markov property, the fact that N and Nρ∆ are stopping times and (31),
E N ∧ τz N > Nρ∆ = n , Sni = si1 : i = 1, 2, ∆
1
2
≥ n + E(s11 ,s21 ) min{σ, τ[x,y]
, τ[x,y]
} = ∞,
i
where τ[x,y]
is defined as in (69). In light of (32) shows (18).
Otherwise, P(N ≤ Nρ∆ ) = 1 and then by Lemma 3.6 (which is in force due to (27)) and (20),
for all u ≥ 0
P(N > u) ≤ P(Nρ∆ > u)
≤ Ps∆ (τρ∆ >
0
≤δ
√
u) +
√
−1 −δ( u−|s∆
0 |)
e
√
⌊ u⌋
X
m=1
∆
P νm
>
√
u
√
√
√
1
′
′
1
2
+ C ue−C u ≤ ′ e−δ ( u−kx0 −x0 k) ,
δ
for some C, C ′ > 0. This shows (19).
It remains to treat the case when (27) does not hold. In this case, we first replace e1 = (1, 0)
by e2 := (0, 1) in the entire argument. If (27) now holds, then we proceed as before and the proof
is complete. If not, then we must have (XT1 ∆ − XT2 ∆ ) − (x10 − x20 ) = 0 with probability 1. But then,
1
1
by Lemma 2.6 applied to the generalized scout process
bn , Q
b n )n≥0 := (Xn1 − Xn2 ) − (x10 − x20 ), (Q1n , Q2n )
(X
,
(33)
n≥0
13
there exists A ⊆ Z2 such that (12) and (13) hold.
′
If −(x10 − x20 ) ∈ A, then by (13) we have P(N > u) ≤ Ce−C u for all u ≥ 0, which in particular
shows (19). If not, then N = ∞ with probability one and hence almost surely
∀z ∈ Z2 : N ∧ τz = τz .
(34)
Thanks to Lemma 3.1 or Lemma 3.3 (depending on whether d1 6= 0 or d1 = 0) applied to the
look-around random walk (Sn1 , Rn1 )n≥0 , as defined above, there exist x1 ∈ R, α1 ∈ {−1, +1} and
C > 0 such that for all z ∈ Z2 with α1 (hz, e1 i − x1 ) > 0 and u ≥ 1
√
1
P ∀n ≤ u : Xn1 6= z ≥ P ∀n ≤ u : Sn1 − hz, e1 i > Rn+1
≥ C/ u .
(35)
By a similar argument, there exists there exist x2 ∈ R, α2 ∈ {−1, +1} and C ′ > 0 such that for
all z ∈ Z2 with α2 (hz, e2 i − x2 ) > 0 and u ≥ 1,
√
P ∀n ≤ u : Xn2 6= z ≥ C ′ / u .
(36)
It follows then by the independence of the scout processes that for all z ∈ Z2 satisfying α1 (hz, e1 i −
x1 ) > 0 and α2 (hz, e2 i − x2 ) > 0 and u ≥ 1,
P τz > u) ≥ CC ′ /u .
(37)
The tail formula for expectation then gives Eτz = ∞ for all such z. Since there infinitely many
such z-s and thanks to (34), we obtain (18).
Next, we remove the restriction to irreducible automata.
Lemma 2.8. Let (Xn1 , Q1n ) and (Xn2 , Q2n ) be two independent scout processes on Z2 with state spaces
S and transition functions Π, starting from x10 = x20 = 0 and initial states q0 . Then either
y ∈ Z2 : E(N ∧ τy ) = ∞
or there exists δ > 0 such that for all u ≥ 0.
P(N > u) ≤ 1δ e−δ
√
u
=∞
.
(38)
(39)
i ⊂ S i for some mi ≥ 1 be the recurrent irreducible classes of
Proof. For i = 1, 2, let S1i , . . . , Sm
i
i
automaton (Qn )n≥0 . Note that if q0i ∈ Sli for some l ∈ {1, . . . , mi } then (Xni , Qin ) is also a single
scout process with state space Sli and transition function Πil , which is the proper restriction of Πi
to Sli . Moreover, its automaton then is irreducible. Define,
i
i
σ i := inf{n ≥ 0 : Qin ∈ ∪m
l=1 Sl }
, i = 1, 2 ,
and set σ := σ 1 ∨ σ 2 . By standard Markov chain theory,
P(σ i > u) ≤ C −1 e−Cu , i = 1, 2,
14
which implies the same for σ. In particular σ, σ 1 and σ 2 are finite almost surely.
Suppose first that there exist n ≥ 0, l1 , l2 , x ∈ Z4 and q ∈ Sl11 × Sl22 such that
P N > σ = n , Xn = x , Q n = q > 0
(40)
and that (18) holds in Lemma 2.7 applied to (Xn , Qn )n≥0 as two independent scout processes with
transition functions Π1l1 , Π2l2 respectively and with x0 = x and q0 = q. Since N and σ are stopping
times, it then follows that for any y ∈ Z2 with kyk > n,
E(N ∧ τy | N > σ = n, Xn = x, Qn = q = n + E(x,q) (N ∧ τy ) = ∞ ,
(41)
Together with (41) this gives (38).
Otherwise, by Lemma 2.7, anytime (40) holds, we must also have (19) with some δ(l
2
1
which depends only the transition functions Πl and Πl . Setting
δ0 = min{δ(l
1 ,l2 )
1 ,l2 )
> 0,
: l1 = 1, . . . , m1 , l2 = 1, . . . , m2 }
we then have with some C > 0,
√
√
P(N > u) ≤ P N > u, σ ≤ u/4 + P σ > u/4
√
≤ E P(Xσ ,Qσ ) (N > u)1{σ≤√u/4} + C −1 e−C u
√
≤ δ0−1 e−δ0 (
√
u− u/2)
+ C −1 e−C
√
u
≤ δ−1 e−δ
√
u
,
where we have used the strong Markov property and the fact that kXσ1 −Xσ2 k ≤ 2σ. This shows (39).
Proof of Proposition 2.2. The proof follows by induction on k. For k = 1, since up to time N ,
the two-scout process evolves as two independent single scout processes, if y and q0 are such that
E(0,q0 ) N ∧ τy = ∞ for two independent scouts, then also
E(0,q0 ) τy ≥ E(0,q0 ) N ∧ τy = ∞ ,
(42)
for the two-scout process. Lemma 2.8 with (S 1 , Π1 ) = (S 2 , Π2 ) = (S, Π), and the assumption
that (1) does not hold, imply then that q0 must be such that (39) holds. Since N1 = N , this gives
the case k = 1.
Suppose now that (2) holds up to k − 1. In particular, this shows that Nk−1 < ∞ almost
surely. Conditioning on FNk−1 ,9 the strong Markov property and the spatial homogeneity of the
underlying processes imply that almost surely,
P Nk − Nk−1 ∈ · | FNk−1 = P(0,QN
)
k−1
9
N ∈· .
Throughout, for a stopping time Z, we use the standard notation FZ to denote the sigma-algebra generated by
Z.
15
If with positive probability QNk−1 = q0 for q0 such that (38) holds, then since the number of
vertices visited by both scouts up to time Nk−1 is finite, it follows as in (42), that (1) cannot hold.
We therefore must have for all u ≥ 0,
√
P Nk − Nk−1 > u | FNk−1 ≤ δ−1 e−δ u a.s.
The tower property for conditional expectation shows (2) .
2.4
One Explorer Must Eventually Get Trapped
In this section we prove Proposition 2.4. As in the case of a single scout process, if (Xn , Qn , Rn )n≥0
is an explorer process with state space S, then (Qn )n≥0 is a Markov chain on S, to be referred to
as the explorer’s automaton. We begin by assuming that this automaton is irreducible.
Lemma 2.9. Let (Xn , Qn , Rn )n≥0 be an explorer process on Z2 with state space S and transition
function Π, starting from x0 = 0 and some q0 ∈ S. Suppose also that its automaton is irreducible
and that there exists δ > 0 such that for all q ∈ S, n ≥ 0 and u ≥ 0,
If it holds that
√
P kXn+1 − Xn k + Rn+1 > u Qn ≤ 1δ e−δ u .
x ∈ Z2 : E inf{n ≥ 0 : kXn − xk ≤ Rn+1 } = ∞
(43)
< ∞,
(44)
then there must exist r > 0, which depends only on Π, such that
P ∀n ≥ 0 : kXn k < r = 1 .
Proof. Let T0 = 0 and for k = 1, . . . , set:
Tk := inf{n > Tk−1 : Qn = q0 }
ζk := XTk − XTk−1
νk := Tk − Tk−1
Rk′ :=
TX
k −1
(45)
Rn+1 .
n=Tk−1
Since the Markov chain (Qn )n≥0 is irreducible and hence recurrent, all times Tk are finite almost
surely. The Markov property then implies that the triplets (ζk , νk , Rk′ )k≥1 are i.i.d. Moreover, from
standard Markov chain theory, there exists C > 0 such that for all u ≥ 0
P(νk > u) ≤ C −1 e−Cu .
In addition (43) and the strong Markov property implies that for all m ≥ 1 and u ≥ 0,
√
P kXTk−1 +m − XTk−1 +m−1 k + RTk−1 +m > u ≤ δ−1 e−δ u .
16
Then for all k ≥ 1 and u ≥ 0, by the union bound,
√
⌊ u⌋
√
√ X
P kXTk−1 +m − XTk−1 +m−1 k + RTk−1 +m > u
P(kζk k + Rk′ > u) ≤ P νk ≥ u +
√
m=1
−C ′ u1/4
≤ C ue
≤
1 −δ′ u1/4
δ′ e
for some δ′ > 0.
Setting Sn = hXTn , e1 i for n ≥ 0, we observe that the process (Sn , Rn′ ) is of the type handled
by Lemma 3.5. Moreover, if (44) holds, then it is also true that
′
x ∈ Z : E inf{n ≥ 0 : |Sn − x| < Rn+1
} = ∞ < ∞,
Then by Lemma 3.5, we must have hζ1 , e1 i = 0 with probability one. Repeating the same with
Sn = hXTn , e2 i, we get hζ1 , e2 i = 0. Thus, we conclude that ζ1 = 0 with probability one. Invoking
then Lemma 2.6 for the process (Xn , Qn )n≥0 , which is, in particular, a generalized scout process,
the proof is complete.
The proof of Proposition 2.4 is now straightforward,
Proof. Let τ be the first time the Markov chain (Qn )n≥0 enters a recurrent class in S. By standard
Markov chain theory τ is is finite almost surely. Conditional on Fτ , by the Markov property,
(Xτ +n , Qτ +n , Rτ +n )n≥0 is an explorer process with an irreducible automaton. If (7) holds then
there could be at most τ random vertices x ∈ Z2 such that,
E inf{n ≥ 0 : kXτ +n − xk ≤ Rτ +n+1 } | Fτ = ∞ a.s ,
Thus by Lemma 2.9, there exists a finite R ∈ Fτ such that,
P ∀n ≥ 0 : kXτ +n − Xτ k < R | Fτ = 1 a.s.
Since the number of recurrent classes is finite, we may replace R above by a non-random r > 0.
Taking expectation, we recover (8).
2.5
One Trapped Scout Cannot Cover the Whole Space
In this section we prove Proposition 2.5. One more time, we start with the case of an irreducible
automaton.
Lemma 2.10. Let (Xn , Qn ) : n ≥ 1 be a single scout process on Z2 with state space S and
suppose that its automaton is irreducible. Let also r < ∞. There exist α̂ ∈ R2 with kα̂k2 = 1 and
M < ∞ such that if x0 ∈ R2 and q0 ∈ S are such that
E(x0 ,q0) inf n ≥ 0 : kXn k < r < ∞
(46)
then
x0 ∈ R(α̂, M ) .
17
Proof. Let x0 ∈ Z2 and q0 ∈ S and suppose that the scout process starts from position x0 and state
q0 . Defining Tk , ζk and νk as in (45), we have that (ζk , νk )k≥1 are i.i.d. and that
P(νk > u) ≤ C −1 e−Cu ,
(47)
for some C > 0. In particular,
Sn := x0 +
n
X
ζk = XTn , n = 0, . . . ,
k=1
is a random walk on
Z2
starting from x0 . Now define,
(
E(ζ1 )/kE(ζ1 )k2 if E(ζ1 ) 6= 0 ,
α̂ :=
0
if E(ζ1 ) = 0 ,
and let α⊥ be any unit vector which is perpendicular to α̂. Setting also for n, k ≥ 0,
ζkα := hζk , α̂i
, Snα := hSn , α̂i , xα0 := hx0 , α̂i ,
⊥
ζk⊥ := hζk , α⊥ i , Sn⊥ := hSn , α⊥ i , x⊥
0 := hx0 , α i ,
we see that both (Snα )n≥0 and (Sn⊥ )n≥0 are random walks on R with steps ζkα and ζk⊥ and initial
⊥
positions xα0 and x⊥
0 , respectively. Moreover, by definition Eζ1 = 0.
If P(ζ1α = 0) = 1, then Lemma 2.6 applied to the generalized scout process (hXn , α̂i, Qn )n≥0
implies the existence of r α < ∞ such that |hXn , α̂i − xα0 | < r α for all n ≥ 0 with probability 1. If
we therefore set
(
r + νn if P(ζ1α = 0) < 1 ,
α
Rn :=
r + r α if P(ζ1α = 0) = 1 ,
then (46) implies that
α
E inf{n ≥ 0 : kSnα k < Rn+1
} < ∞.
(48)
A similar argument shows that for some r ⊥ < ∞ and with
(
r + νn if P(ζ1⊥ = 0) < 1 ,
⊥
Rn :=
r + r ⊥ if P(ζ1⊥ = 0) = 1 ,
if (46) holds, then also
⊥
E inf{n ≥ 0 : kSn⊥ k < Rn+1
} < ∞.
(49)
Using Lemma 3.1 if Eζ1α > 0 or Lemma 3.4 if Eζ1α = 0, for the process (Snα , Rnα )n≥0 , noting
that (47) implies that condition (53) is in force, we see that for (48) to hold we must have
xα0 < M α ,
(50)
for some M α > 0. Similarly, since Eζ1⊥ = 0 and (47) holds, we may use Lemma 3.4 for the process
(Sn⊥ , Rn⊥ )n≥0 to obtain M ⊥ > 0 such that if (49) holds then
⊥
|x⊥
0|<M .
Combining (50) and (51) we have x0 ∈ R(−α̂, M ) for M := max{M α , M ⊥ } as desired.
18
(51)
Proof. For m ≥ 1, let S1 , S2 , ..., Sm be the recurrent irreducible state classes of the automaton
(Qn )n≥0 . If q0 ∈ Sl for some l ∈ {1, . . . , n}, then the process (Xn , Qn )n≥0 is a single scout process
with state space Sl and transition function Πl which is the proper restriction of Π to S l . Moreover,
such a scout process has an irreducible automaton. Therefore, by Lemma 2.10 there exists α̂l and
Ml such that if (9) holds then
x0 ∈ R(α̂l , Ml ) .
(52)
For any other q0 ∈ S, there exists l ∈ {1, . . . , m}, q ′ ∈ S l , n ≤ |S| and x′ ∈ Z2 with kx′ −x0 k ≤ n
such that
P(x0 ,q0 ) Xn = x′ , Qn = q ′ ) > 0
If (9) holds, then by the Markov property and (52) we must have that x′ ∈ R(α̂l , Ml ) which implies
that x0 ∈ R(α̂l , Ml + |S|). Taking M := maxl≤m Ml + |S| we obtain (10) as desired.
3
Random Walk Hitting Time Estimates
In this section we state and prove the random walk estimates, which are used in the proof of the
main theorems.
3.1
One Random Walk with a Stretched Exponential Look-Around
Fix δ > 0 and let (ζk , Rk ) : k = 1, . . . be a sequences of discrete i.i.d. random pairs, taking values
in R × [1, ∞) and satisfying
δ
P(|ζ1 | + R1 > u) ≤ 1δ e−u , u ≥ 0 .
(53)
Notice that we do not insist that for a given k, the random variables ζk , Rk are independent of each
other. We shall think of ζk as the spatial displacement of a random walk in the k-th step and of
Rk as the radius of a ball (interval) around the position of the walk at time k − 1, inside which the
walk is allowed to “peek” at this time. We therefore fix also s0 ∈ R and define for all n ≥ 0,
Sn := s0 +
n
X
ζk ,
k=1
The process (Sn , Rn )n≥0 will be referred to as a look-around random walk.
Lemma 3.1. Let (Sn , Rn )n≥0 be the look-around random walk process defined above and suppose
that (53) holds. If E(ζ1 ) > 0, then there exists r > 0, such that for all x < s0 − r,
P inf{n ≥ 0 : |Sn − x| ≤ Rn+1 } = ∞ > 0 .
(54)
Proof. Let µ := E(ζk ) > 0. By the Strong Law of Large numbers
surely. Therefore we may find m large enough such that
P Sn > µn/2 : n ≥ m) > 1/2 .
19
Sn
n
−→ µ as n → ∞ almost
(55)
On the other hand, by union bound, for possibly larger m we have,
P ∃n ≥ m : Rn+1 > n
1/3
∞
X
1 −nδ/3
≤
≤ 1/4 .
e
δ
n=m
(56)
Finally, since min{Sn − Rn+1 : n = 1, . . . , m} > −∞ with probability 1, it follows that we can find
x0 small enough such that for all x ≤ x0 ,
P |Sn − x| ≤ Rn+1 : n = 0, . . . , m ≤ 1/8 .
(57)
The event in (54) is implied by the intersection of the event on the left hand side of (55) with the
complements of the events on the left hand side of (56) and (57). The above shows that this has
positive probability as required.
Next we deal with the case E(ζ1 ) = 0.
Lemma 3.2. Let (Sn , Rn )n≥0 be the look-around random walk process defined above and suppose
that (53) holds. If E(ζ1 ) = 0, but P(ζ1 = 0) < 1, then there exists C > 0, such that for all u ≥ 1
and x ∈ R satisfying 0 < x − s0 < u1/4 ,
x − s0
P inf{n ≥ 0 : Sn + Rn+1 ≥ x} ≥ u ≤ C √ .
u
(58)
Proof. The probability on the left hand side of (58) is bounded above by
P inf{n ≥ 0 : Sn ≥ x} ≥ u
and the result follows immediately from Theorem 5.1.7 of [27].
The opposite direction is given by
Lemma 3.3. Let (Sn , Rn )n≥0 be the look-around random walk process defined above and suppose
that (53) holds. Suppose also that E(ζ1 ) = 0 and if P(ζ1 = 0) = 1, then P(R1 < M ) = 1 for some
M > 0. Then there exist r > 0 and C > 0, such that for all u ≥ 1 and x ≥ s0 + r,
C
P inf{n ≥ 0 : Sn + Rn+1 ≥ x} ≥ u ≥ √ .
u
(59)
Proof. For simplicity of the argument, we shall assume that ζ1 is supported on the integers. The
argument for a general (discrete, Real) distribution is similar. Without loss of generality we can
assume that s0 = −1 and x = r − 1, where r > 0 will be determined later. If P(ζ1 = 0) = 1,
then (59) will clearly hold, once r = M + 1. Otherwise, set
ηx := inf{n ≥ 0 : Sn ≥ x} .
From Theorem 5.1.7 of [27] we have that for all n ≥ 0 and x ∈ [0, n1/4 ],
√
√
C(x + 1)/ n ≤ P ηx > n ≤ C ′ (x + 1)/ n
20
(60)
At the same time from Theorem 7 of [3], we have that for all y ∈ [−n1/4 , −1],
P η0 > n , Sn = y ≤ C|y|/n3/2 .
Then for all k0 ≥ 1 and n > k0 ,
P η0 > n , ∃k ∈ [k0 , n] : Sk ≥ −(log k)3/δ
≤
n
X
−1
X
k=k0 y=−(log k)3/δ
≤C
n/2
X
k
−3/2
P η0 > k , Sk = y P η−y > n − k
9/δ −1/2
(log k)
n
+
k=k0
1/2
n−n
X
k=n/2
−1/3
≤ Cn−1/2 k0
n−3/2 (log n)9/δ (n − k)−1/2 + n1/2 n−3/2 (log n)9/δ
.
(61)
Now let K := inf k ≥ 1 : |ζm | + Rm+1 ≤ (log(m ∨ k))2/δ , ∀m ∈ [0, n] . Then for all k1 ≥ 2
and n > k1 ,
n/2
X
P η0 > n , K ∈ [k1 , n/2) =
η0 > n , K = k
k=k1
n/2
≤
X
P (log(k − 1))2/δ < min
m≤k−1
k=k1
≤C
n/2
X
ke−C
′ (log(k−1))2
k=k1
|ζm | + Rm+1 ≤ max |ζm | + Rm+1 ≤ (log k)2/δ , η0 > n
m≤k
′′
2
k(log k)2/δ + 1
√
≤ Cn−1/2 e−C (log(k1 )) .
n−k
(62)
Above to derive the one before last inequality, we have conditioned on Fk and used the Markov
property together with (53), (60), union bound and the fact that
max |ζm | + Rm+1 ≤ (log k)2/δ ⊆ Sk ≥ −k(log k)2/δ .
m≤k
Also by union bound and (53),
P(K > n/2) ≤ P |ζm | + Rm+1 > (log(n/2))2 : m ∈ [0, n]
′
2
≤ Cne−C (log n) ≤ e−C
′′ (log n)2
,
(63)
Combining (62) and (63) we get
′′
2
′
2
P η0 > n , K ≥ k1 ≤ n−1/2 e−C (log(k1 )) + e−C (log n) .
(64)
Subtracting the probabilities on the left hand sides of (61), (64) with k0 = k1 from the probability in (60) for x = 0, and using the derived upper bounds for the former and the lower bound
21
for the latter, we have
P Sm < −(log m)3/δ : m ∈ [k0 , n] , |ζm | + Rm+1 < (log(m ∨ k0 ))2/δ : m ∈ [0, n]
′′
2
′′′
2
−1/3
≥ n−1/2 C − C ′ k0
− e−C (log k0 ) − n1/2 e−C (log n)
We may now find k0 large enough such that for all n large enough, the right hand side above will
be at least Cn−1/2 for some C > 0. But the event on the left hand side implies that
Sm + Rm+1 ≤ x : m ∈ [0, n]
for all x > k0 (log k0 )2/δ . Setting r = k0 (log k0 )2/δ , this shows (59) for all x ≥ r as desired.
The following is an immediate consequence of Lemma 3.3.
Lemma 3.4. Let (Sn , Rn )n≥0 be the look-around random walk process defined above and suppose
that (53) holds, E(ζ1 ) = 0 and if P(ζ1 = 0) = 1, then P(R1 < M ) = 1 for some M > 0. Then there
exist r > 0, such that if |s0 − x| > r, then
E inf{n ≥ 0 : |Sn − x| ≤ Rn+1 } = ∞ .
(65)
Proof. By reversing the steps, it is enough to show the result for x ≥ s0 + r. But then, using the
tail formula for expectation, we sum (59) from u = 1 to ∞ and conclude (65) for all such x if r is
large enough.
Lemma 3.5. Let (Sn , Rn )n≥0 be the look-around random walk process defined above. Suppose
that (53) holds. Then if
we must have
x ∈ R : E inf{n ≥ 0 : |Sn − x| ≤ Rn+1 } = ∞
< ∞,
(66)
P(ζ1 = 0) = 1 .
Proof. If Eζ1 6= 0, then by Lemma 3.1 we have that (66) must be false. At the same time, if Eζ1 = 0
and P(ζ1 = 0) < 1, then by Lemma 3.4 the inequality in (66) must be again false. It follows that
P(ζ1 = 0) = 1 as required.
In the next two lemmas, the “look-around” feature of the random walk is not used. The first one
includes standard hitting time estimates for random walks. We omit the proof, as it is standard.
For ρ > 0, let
inf{n ≥ 0 : |Sn | > ρ}, if Eζ1 = 0 ,
τρ := inf{n ≥ 0 : Sn > ρ},
(67)
if Eζ1 > 0 ,
inf{n ≥ 0 : S < −ρ}, if Eζ < 0 .
n
1
Then,
22
Lemma 3.6. Let (Sn , Rn )n≥0 be the look-around random walk defined above and suppose that
P(ζ1 = 0) < 1. Then for all ρ > 0, there exists δ > 0, such that for all s0 ∈ R,
P τρ > u) ≤ δ−1 e−δ(u−|s0 |) .
The next lemma is the result of a standard application of the exponential Chebychev inequality.
Lemma 3.7. Let (Sn , Rn )n≥0 be the look-around random walk described above and suppose that
Eζ1 = s0 = 0. Then for all µ > 0, there exists δ > 0, such that if n ≥ 0 and y ≥ µn then,
P(Sn ≥ y) ≤ e−δy .
Proof. Let L(t) := log Eetζ1 be the log moment generating of ζ1 , whose existence and (Real) analyticity in a neighborhood of 0 is guaranteed by condition (53). Since L(0) = 0 and L′ (0) = Eζ1 = 0.
It follow by Taylor expansion of L around 0, that for any µ > 0, we may find t0 > 0 and δ0 > 0,
such that
L(t0 ) − µt0 < −δ0 .
Then for n and y as in the conditions of the lemma, by the exponential Chebychev inequality,
P Sn ≥ y ≤ exp{nL(t0 ) − nµt0 − (y − nµ)t0 }
≤ exp{−δ0 n − (y − nµ)t0 } ≤ exp{−δy} ,
where δ = min{t0 , δ0 /µ}.
3.2
Two Time-Varying Random Walks
In this sub-section we consider two random walks of the type considered in sub-section 3.1, only
that in addition, each walk also has a “time” component. The latter can be used to “synchronize”
between the two walks.
As in the previous subsection, we fix δ > 0 and for i = 1, 2, let (ζki , νki , Rki ) : k = 1, . . . be
two independent sequences of i.i.d. discrete random triplets, taking values in R × Z≥1 × [1, ∞) and
satisfying the following conditions:
δ
1. P(|ζ1i | + ν1i + R1i > u) ≤ 1δ e−u .
2. If ζ1i /ν1i is non-random, then P(ν1i = 1) = 1 and P(|R1i | < 1/δ) = 1.
Again, we do not insist that for a given i and k, the random variables ζki , νki , Rki are independent
of each other. We shall think of νki as the time it took walk i to make step k and of Rki as its
“look-around” radius (in the sense of the previous subsection). For i = 1, 2, we now fix also si0 ∈ R
and define the random walk (Sni , Tni ) : n = 0, 1, . . . by
Sni := si0 +
n
X
k=1
ζki , Tni :=
n
X
k=1
23
νki ;
n = 0, 1, . . . .
Setting also R0i = 0, the resulting process (Sni , Tni , Rni ) : n = 0, 1, . . . will be referred to as a
time-varying look-around random walk. Its effective drift is then given by
di := (Eζ1i )/(Eν1i ).
(68)
We now define various hitting times. First, to translate back “time”to number of steps, we set
for i = 1, 2 and m ≥ 0 the random variable,
i
Km
:= sup {k ≥ 0 : Tki ≤ m} .
Now, for a subset A ⊆ R, the hitting time of A by Sni is
i
i
τAi := inf m ≥ 0 : d(SK
i , A) ≤ Rk i
m +1
m
,
(69)
where for x ∈ R we use the usual definition d(x, A) := inf{kx − yk : y ∈ A}. The first time the
look-around balls (intervals) of the walks intersect, is defined via
2
1
2
1
σ := inf m ≥ 0 : SK
+ RK
≤ RK
.
1 − SK 2
2
1
m
m
m +1
m +1
1
2
∆
1
2
Finally, we also set s∆
0 := s0 − s0 and d := d − d .
The following proposition is the main product of this subsection.
Proposition 3.8. For i = 1, 2, let the time-varying look-around random walks (Sni , Tni , Rni ) : n =
0, 1 . . . be as defined above. Then, we may find ρ > 0, such that whenever:
|s∆
0 |>ρ
∆
s0 > ρ
s∆
0 < −ρ
if
if
if
d∆ = 0 ,
d∆ > 0 ,
d∆ < 0 ,
(70)
there exist x, y ∈ Z satisfying −∞ < x < y − 2 < ∞, such that
1
2
E min σ, τ[x,y]
, τ[x,y]
= ∞.
(71)
Proof. To simplify the arguments below, we shall assume that ζ1i , R1i are supported on Z. The
argument for general (discrete, Real) distributions is analogous. The proof follows by case-analysis
of d1 and d2 . By switching between walk 1 and 2 or negating the two walks simultaneously, we do
not lose any generality by considering only the cases:
• d1 ≥ 0, d2 ≤ 0,
• d1 > d2 > 0,
• d1 = d2 > 0.
24
If d1 ≥ 0 and d2 ≤ 0, then d∆ ≥ 0 and Eζ11 ≥ 0, Eζ12 ≤ 0. We define now for A ⊂ R, i = 1, 2, ∆
the hitting time of A by walk i as
i
NAi := inf n ≥ 0 : d(Sni , A) ≤ Rn+1
.
This is analog to τAi only that time is measured in steps.
1
2
Then, by Lemma 3.1 and Lemma 3.3, we may find ρ > 0 large enough such that if s∆
0 = s0 −s0 >
ρ, then there exist x, y ∈ Z such that s10 < x < y − 2 < s20 and for all u ≥ 0,
C
1
P N[x,y]
>u ≥ √ ,
u
C′
2
P N[x,y]
>u ≥ √ ,
u
C
1
P τ[x,y]
>u ≥ √ ,
u
C′
2
P τ[x,y]
>u ≥ √ .
u
i
i
Since τ[x,y]
≥ N[x,y]
, it follows that for all u ≥ 0, also
Now, it is not difficult to see that for such s∆
0 ,
1
2
σ ≥ min{τ[x,y]
, τ[x,y]
}.
It then follows that
1
2
1
2
E min{σ, τ[x,y]
, τ[x,y]
} = E min{τ[x,y]
, τ[x,y]
}.
Since the two walks are independent, by the tail formula for expectation,
1
2
E min{σ, τ[x,y]
, τ[x,y]
}
=
∞
X
1
P(τ[x,y]
u=1
≥
2
u)P(τ[x,y]
≥ u) ≥
∞
X
CC ′
u=1
u
= ∞.
Moving to the case d1 > d2 > 0. Here d∆ > 0. By the law of large numbers, for all ǫ > 0, we
may find n0 large enough, such that with probability at least 1 − ǫ for all n ≥ n0 , i = 1, 2, the
following holds:
n(Eζ i − ǫ) ≤ Sni ≤ n(Eζ i + ǫ) ,
n(Eν i − ǫ) ≤ Tni ≤ n(Eν i + ǫ) .
At the same time, by increasing n0 if needed, we have
∞
X
1/3
1/3
δ−1 e−δn ≤ e−Cn0 ≤ ǫ .
P ∃n ≥ n0 : Rni > n1/3 ≤
n=n0
Combining the last three displays, for all ǫ > 0, we may find m0 , such that with probability at least
1 − ǫ, for all m ≥ m0 ,
i
i
i
(72)
RK
m(di − ǫ) ≤ SK
i ≤ ǫm .
i ≤ m(d + ǫ) ,
m
m
i , |ζ i | + Ri , for all i = 1, 2 and n ≥ 1, for
By finiteness almost surely of the random variables Km
n
n
0
any m0 and ǫ > 0, we may find r > 0 large enough, such that with probability at least 1 − ǫ,
i +1
Km
0
X
n=1
|ζni | + Rni < r .
25
(73)
Combining the above, we first choose 0 < ǫ < min{d1 − d2 , d2 , 1}/5, then find m0 large enough
and finally r > 0 such that both (72) and (73) hold with probability at least 1 − ǫ. It follows that
2
if s∆
0 > 3r and x, y ∈ Z satisfy x + 2 < y < s0 − 2r, then with probability at least 1 − 2ǫ
1
2
τ[x,y]
= ∞ , τ[x,y]
= ∞, σ = ∞,
which, of course, implies (71) .
Turning to the hardest case d1 = d2 =: d > 0. For ρ > 2 to be determined later, let us assume
2
that s∆
0 > ρ and without loss of generality also that s0 = 0. For i = 1, 2 and n ≥ 0, set
Yni
:=
Sni
−
dTni
=
si0
+
n
X
k=1
ζki − dνki .
Let also z := s10 /2 and define
1
1
2
σ 1 := inf{n ≥ 0 : Yn1 − dνn+1
− Rn+1
≤ z} , σ 2 := inf{n ≥ 0 : Yn2 + Rn+1
≥ z} .
Observe that
Consequently,
σ1 > n
σ2 > n
1
1
1
SK m
,
1 − RK 1 +1 > dm + z : m = 0, . . . , Tn
m
2
2
2
.
⊆ SK m
2 + RK 2 +1 < dm + z : m = 0, . . . , Tn
m
⊆
1
σ > n , σ 2 > n ⊆ σ > Tn1 ∧ Tn2 ⊆ σ > n .
Similarly if x, y ∈ Z such that x + 2 < y < 0, then
2
1
2
1
2
σ 1 > n , σ 2 > n , N[x,y]
> n ⊆ σ > n , τ[x,y]
> n , τ[x,y]
> n = min{σ, τ[x,y]
, τ[x,y]
}>n .
Thanks to the independence between the two walks and the tail formula for expectation, to
show (71), it is therefore enough to prove
∞
X
n=0
2
P σ 1 > n)P(σ 2 > n , N[x,y]
> n = ∞.
(74)
To this end, first observe that (68) implies that Eζ1i − dν1i = 0. This makes Yni a random
walk with 0 drift starting from si0 . Moreover, by our assumptions on (ζ1i , ν1i , R1i ), the processes
(Yn1 , Rn1 + dνn1 )n≥0 and (Yn2 , Rn2 )n≥0 are look-around random walks, which satisfy the conditions in
Lemma 3.3. Therefore, by the lemma, if ρ > 0 is chosen large enough, we have
√
P(σ i > u) ≥ C/ u , u ≥ 0 , i = 1, 2 .
(75)
This handles the first probability in (74). For the second, we write
2
2
P σ 2 > n , N[x,y]
> n = P σ 2 > n) − P σ 2 > n , N[x,y]
≤n
26
(76)
√
By (75), the first term is lower bounded by C/ n once ρ is chosen large enough. We wish to show
√
now that the second term can be upper bounded by C/(2 n) be choosing then y small enough.
Thanks to Lemma 3.7, writing µ for Eζ12 , we have for all k ≥ 0 and y < 0,
2
2
≥ kµ/2 − y/2)
≤ y) ≤P(Sk2 ≤ kµ/2 + y/2) + P(Rk+1
P(Sk2 − Rk+1
(77)
′
≤ Ce−C (k−y) .
Therefore,
At the same time,
2
P N[x,y]
> n1/16 ≤
∞
X
k=n1/16
2
≤ y) ≤ e−Cn
P(Sk2 − Rk+1
1/16
.
2
> n1/4 ≤ P ∃k ≤ n1/16 + 1 : |ζk2 | + νk2 + Rk2 > n1/16
P ∃k ≤ n1/16 : |Sk2 | + |Tk2 | + Rk+1
≤ (n1/16 + 1)e−Cn
1/16
′ 1/16
≤ e−C n
.
It follows that the second term on the right hand side of (76) is bounded above by
−Cn−1/16
e
+
1/16
nX
y
X
1/4
n
X
k=1 w=−2n1/4 m=k
2
= w , Tk2 = m P σ 2 > n − k Y02 = w − dm ,
P Sk2 − Rk+1
(78)
where conditioning in the second probability is only formal and means that (Yn2 )n≥0 is redefined so
that Y02 = w − dm.
In the range of the sums, we can use Lemma 3.2 to upper bound the second probability by
C
z − w + dm
z − w + dm
√
√
≤ C′
.
n
n−k
Plugging this into (78) and removing some of the restrictions on the sums, we get as an upper
bound,
∞
X
−Cn−1/16
′ −1/2
2
E z − Sk2 + Rk+1
e
+C n
+ dTk2 1{S 2 −R2 ≤y} .
(79)
k
k+1
k=1
Using Cauchy-Schwartz and (77), the last expectation can be further bounded above by
2
C z 2 + E|Sk2 |2 + E|Rk+1
|2 + E|Tk2 |2
Consequently, (79) is bounded above by
e−Cn
−1/16
+ C ′ n−1/2 (z + 1)eC
′′ y
1/2
∞
X
2
P Sk2 − Rk+1
≤y
ke−C
k=1
′′′ k
≤ e−Cn
1/2
−1/16
≤ C(z + k)e−C
′ (k−y)
+ C ′ n−1/2 (z + 1)eC
′′ y
.
.
Choosing first ρ (and hence z) large enough and then y small enough, by (75) for i = 2 and (76),
we can indeed ensure that
√
2
P σ 2 > n , N[x,y]
> n ≥ C/ n .
Together with (75) for i = 1 we get (74) and hence also (71).
27
4
One Scout on Z1
The arguments leading to the proof of Theorem 2.1 can be adapted to yield a proof for the case
d = 1.
Theorem 4.1. Let (Xn , Qn )n≥0 be a single scout process on Z with protocol P = h1, S, 0, q0 , Πi.
Then there exists some grid point x ∈ Z of which the expected hitting time is infinite, namely,
E (inf{n ≥ 0 : Xn = x}) = ∞ .
(80)
Proof. Recalling that the automaton process (Qn )n≥0 is a Markov chain on S, we let S1 , . . . , Sl for
l ≥ 1, be its irreducible recurrent state classes. Suppose first that q0 ∈ Sm for some m ∈ {1, . . . , l},
let T0 = 0 and for k = 1, . . . , set
Tk := inf{n > Tk−1 : Qn = q0 } ,
ζk := XTk − XTk−1 ,
(81)
νk := Tk − Tk−1 .
By the Markov property and spatial homogeneity of the process (Xn , Qn )n≥0 , the pairs (ζk , νk )k≥1
are i.i.d. Moreover, standard Markov chain theory shows that
P(|ζk | + νk > u) ≤ C −1 e−Cu ,
(82)
for some C > 0. In particular, if we set for n ≥ 0,
Sn := x0 +
n
X
ζk = XTn ,
(83)
k=1
then (Sn )n≥0 is a random walk on Z. If P(ζ1 = 0) = 1, then we appeal to Lemma 2.6 to obtain
r > 0, such that P(|XT1 | < r) = 1 and accordingly we set Rn := r for all n ≥ 1. Otherwise, we set
Rn := νn . In both cases we then have
inf{n ≥ 0 : Xn = x} ≥ inf{n ≥ 0 : |Sn − x| ≤ Rn+1 } .
(84)
Now, in light of (82), the process (Sn , Rn )n≥0 is a look-around random walk of the type treated
in Subsection 3.1, starting from 0. Moreover, by definition if P(ζ1 = 0) = 1, then Rn is bounded
by a deterministic quantity. By employing Lemma 3.1 or Lemma 3.4, we conclude that there exist
infinitely many x ∈ Z such that
E inf{n ≥ 0 : Xn = x} ≥ E inf{n ≥ 0 : |Sn − x| ≤ Rn+1 } = ∞ .
(85)
In particular, this shows (80).
If q0 does not belong to any of Sm for m ∈ {1, . . . , l}, then we set
σ := inf n : Qn ∈ ∪lm=1 Sm
28
(86)
and let L be such that Qσ ∈ SL . Standard Markov chain theory gives that σ < ∞ almost surely and
hence L is well defined. Since σ is a stopping time, conditional on Fσ , the process (Xσ+n , Qσ+n )n≥0
is distributed as the original process starting from x0 = Xσ and q0 = Qσ . Moreover, since σ is
finite, the number of visited grid points x ∈ Z up to this time is finite. Repeating the argument
above with m = L, there is at least one (random, Fσ -measurable) x′ ∈ Z which was not visited up
to time σ and such that
E inf{n ≥ 0 : Xσ+n = x′ } Fσ = ∞ .
(87)
This readily implies (80) for some (deterministic) x ∈ Z.
Acknowledgments
The authors would like to thank Roger Wattenhofer and Tobias Langner for helpful discussions.
The work of O.L. was supported in part by the European Union’s - Seventh Framework Program
(FP7/2007-2013) under grant agreement no 276923 – M-MOTIPROX. The work of L.C. was supported by a Technion MSc. fellowship.
29
References
[1] S. Albers and M. Henzinger. Exploring unknown environments. SIAM Journal on Computing,
29:1164–1188, 2000.
[2] R. Aleliunas, R. M. Karp, R. J. Lipton, L. Lovasz, and C. Rackoff. Random walks, universal
traversal sequences, and the complexity of maze problems. In Proceedings of the 20th Annual
Symposium on Foundations of Computer Science (SFCS), pages 218–223, 1979.
[3] L. Alili and R. Doney. Wiener–hopf factorization revisited and some applications. Stochastics:
An International Journal of Probability and Stochastic Processes, 66(1-2):87–102, 1999.
[4] N. Alon, C. Avin, M. Koucký, G. Kozma, Z. Lotker, and M. R. Tuttle. Many random walks
are faster than one. Combinatorics, Probability & Computing, 20(4):481–502, 2011.
[5] I. Averbakh and O. Berman. A heuristic with worst-case analysis for minimax routing of two
travelling salesmen on a tree. Discrete Applied Mathematics, 68(1):17–32, 1996.
[6] B. Awerbuch and M. Betke. Piecemeal graph exploration by a mobile robot. Information and
Computation, 1999.
[7] R. A. Baezayates, J. C. Culberson, and G. J. Rawlins. Searching in the plane. Information
and computation, 106(2):234–252, 1993.
[8] M. Bender and D. Slonim. The power of team exploration: Two robots can learn unlabeled
directed graphs. In Proceedings of Foundations of Computer Science (FOCS), pages 75–85,
Nov 1994.
[9] M. Blum and D. Kozen. On the power of the compass (or, why mazes are easier to search than
graphs). In Proceedings of the 19th Annual Symposium on Foundations of Computer Science
(FOCS), pages 132–142, 1978.
[10] M. Blum and W. J. Sakoda. On the capability of finite automata in 2 and 3 dimensional space.
In Proceedings of the 18th Annual Symposium on Foundations of Computer Science (FOCS),
pages 147–161, 1977.
[11] L. Budach. Automata and labyrinths. Mathematische Nachrichten, pages 195–282, 1978.
[12] M. Chrobak, L. Gasieniec, T. Gorry, and R. Martin. Group search on the line. In International
Conference on Current Trends in Theory and Practice of Informatics, pages 164–176. Springer,
2015.
[13] C. Cooper, A. Frieze, and T. Radzik. Multiple random walks in random regular graphs. SIAM
Journal on Discrete Mathematics, 23(4):1738–1761, 2009.
[14] X. Deng and C. Papadimitriou. Exploring an unknown graph. Journal of Graph Theory,
32:265–297, 1999.
[15] K. Diks, P. Fraigniaud, E. Kranakis, and A. Pelc. Tree exploration with little memory. Journal
of Algorithms, 51:38–63, 2004.
[16] K. Döpp. Automaten in labyrinthen. Elektronische Informationsverarbeitung und Kybernetik,
7(2):79–94, 1971.
[17] C. A. Duncan, S. G. Kobourov, and V. S. A. Kumar. Optimal constrained graph exploration.
ACM Transactions on Algorithms (TALG), 2(3):380–402, 2006.
[18] Y. Emek, T. Langner, D. Stolz, J. Uitto, and R. Wattenhofer. How many ants does it take to
find the food? Theor. Comput. Sci., 608:255–267, 2015.
[19] Y. Emek, T. Langner, J. Uitto, and R. Wattenhofer. Solving the ANTS problem with asynchronous finite state machines. In Distributed Computing the 41st International Colloquium
on Automata, Languages, and Programming ICALP, pages 471–482, 2014.
[20] O. Feinerman and A. Korman. Memory lower bounds for randomized collaborative search and
implications for biology. In Proceedings of the 26th International Symposium on Distributed
Computing DISC, pages 61–75, 2012.
[21] O. Feinerman, A. Korman, Z. Lotker, and J. Sereni. Collaborative search on the plane without
communication. In Proceedings of ACM Symposium on Principles of Distributed Computing,
PODC, pages 77–86, 2012.
[22] F. V. Fomin and D. M. Thilikos. An annotated bibliography on guaranteed graph searching.
Theoretical Computer Science, 399(3):236–245, 2008.
[23] P. Fraigniaud, L. Gasieniec, D. R. Kowalski, and A. Pelc. Collective tree exploration. Networks,
48(3):166–177, 2006.
[24] P. Fraigniaud and D. Ilcinkas. Digraphs exploration with little memory. In Proceedings of the
21st Symposium on Theoretical Aspects of Computer Science (STACS), pages 246–257, 2004.
[25] P. Fraigniaud, D. Ilcinkas, G. Peer, A. Pelc, and D. Peleg. Graph exploration by a finite
automaton. Theoretical Computer Science, 345(2–3):331–344, 2005. Mathematical Foundations
of Computer Science 2004Mathematical Foundations of Computer Science 2004.
[26] T. Langner, J. Uitto, D. Stolz, and R. Wattenhofer. Fault-tolerant ANTS. In Prooceedings of
the 28th International Symposium on Distributed Computing DISC, pages 31–45, 2014.
[27] G. F. Lawler and V. Limic. Random Walk: A Modern Introduction. Cambridge University
Press, New York, NY, USA, 2010.
[28] A. López-Ortiz and G. Sweet. Parallel searching on a lattice. In Proceedings of the 13th
Canadian Conference on Computational Geometry, pages 125–128, 2001.
[29] P. Panaite and A. Pelc. Exploring unknown undirected graphs. In Proceedings of the 9th
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 316–322, 1998.
[30] H. Rollik. Automaten in planaren graphen. In K. Weihrauch, editor, Theoretical Computer
Science, volume 67 of Lecture Notes in Computer Science, pages 266–275. Springer Berlin
Heidelberg, 1979.
| 8cs.DS
|
c xxxx Society for Industrial and Applied Mathematics
Vol. xx, pp. x
February 23, 2017
Well-posed Bayesian Inverse Problems with Infinitely-Divisible and Heavy-Tailed
Prior Measures ∗
arXiv:1609.07532v2 [math.PR] 21 Feb 2017
Bamdad Hosseini
†
Abstract. We present a new class of prior measures in connection to `p regularization techniques when p ∈ (0, 1)
which is based on the generalized Gamma distribution. We show that the resulting prior measure
is heavy-tailed, non-convex and infinitely divisible. Motivated by this observation we discuss the
class of infinitely divisible prior measures and draw a connection between their tail behavior and
the tail behavior of their Lévy measures. Next, we use the laws of pure jump Lévy processes in
order to define new classes of prior measures that are concentrated on the space of functions with
bounded variation. These priors serve as an alternative to the classic total variation prior and result
in well-defined inverse problems. We then study the well-posedness of Bayesian inverse problems
in a general enough setting that encompasses the above mentioned classes of prior measures. We
establish that well-posedness relies on a balance between the growth of the log-likelihood function
and the tail behavior of the prior and apply our results to special cases such as additive noise models
and linear problems. Finally, we discuss some of the practical aspects of Bayesian inverse problems
such as their consistent approximation and present three concrete examples of well-posed Bayesian
inverse problems with heavy-tailed or stochastic process prior measures.
Key words. Inverse problems, Bayesian, Infinitely divisible, non-Gaussian, Bounded variation.
AMS subject classifications. 35R30, 62F99, 60B11.
1. Introduction. Gaussian prior measures are perhaps the most commonly used class of
priors in infinite-dimensional Bayesian inverse problems. While the Gaussian class is very
conveninet to use in both theory and practice, it has serious shortcomings in modelling of
certain types of prior knowledge such as sparsity. In this article we introduce some nonGaussian prior measures that are able to model parameters that are compressible or have
jump discontinuities. We will discuss our goals in more detail after a brief introduction to the
Bayesian framework for solution of inverse problems.
Consider the problem of estimating a parameter u ∈ X from a set of measurements y ∈ Y
where both X and Y are Banach spaces and y is associated with u through a model of the
form
(1.1)
y = G̃(u).
G̃ is a generic stochastic mapping that models the relationship between the parameter and the
observed data by taking the measurement noise into account (be it additive, multiplicative
etc). As an example, if the measurement noise is additive then we can write
G̃(u) = G(u) + η
∗
This work was supported in part by the Natural Sciences and Engineering Council of Canada.
Department of mathematics, Simon Fraser University, 8888 University Drive, Burnaby, BC, V5A 1S6,
Canada ([email protected]).
†
1
x–x
2
where G : X 7→ Y is the (deterministic) forward model and η is the (random) measurement
noise which is independent of u. We want to estimate the parameter u given a realization of
y. Since the map G may not be stably invertible this problem is in general ill-posed.
Here we consider the Bayesian framework for solution of such ill-posed problems. Recall
the infinite-dimensional version of Bayes’ rule [51] which is understood in the sense of the
Radon-Nikodym theorem [8, Thm. 3.2.2]:
(1.2)
dµy
1
(u) =
exp (−Φ(u; y))
dµ0
Z(y)
Z
where Z(y) =
exp(−Φ(u; y))dµ0 (u).
X
Here µ0 is the prior measure which reflects our prior knowledge of the parameter u, Φ(u; y)
is the likelihood potential that can be thought of as the negative log of the density of the data
conditioned on the parameter u and µy is the posterior measure on u. The posterior µy is, in
essence, an updated version of the prior µ0 that is informed by the data y.
The Bayesian approach has attracted a lot of attention in the last two decades [13, 36, 51].
Put simply, the unknown parameter u is modelled as a random variable and our goal is
to obtain a probability distribution µy on u that is informed by the data y and our prior
knowledge about u (modelled by the measure µ0 ). We can generate samples from the posterior
µy and if this measure is concentrated around the true value of the parameter, the sample
mean or median will be good estimators of the true value of the parameter.
The Bayesian approach is well-established in the statistics literature [14, 5] where it is
often applied in the setting where X, Y are finite-dimensional spaces. Here we take X to be an
infinite-dimensional Banach space, motivation by applications where the parameter u belongs
to a function space such as L2 or BV (the space of functions with bounded variation). Such
problems arise when the forward map involves the solution of a partial differential equation
(PDE) or an integral equation such as the examples in Section 5.
In practice we solve these problems by discretizing the forward model and approximating
the infinite-dimensional problem with a finite dimensional one. An important task is to ensure
that the finite dimensional approximation to the posterior measure remains consistent with
the infinite-dimensional posterior measure. For example, we require that the finite dimensional posterior converges to the (true) infinite-dimensional measure in the limit when the
discretization is infinitely fine. Ensuring this consistency is a delicate task. An example of an
inconsistent discretization of an infinite dimensional inverse problem was studied in [39] where
the authors demonstrated that the total variation prior loses its edge preserving properties in
the limit of fine discretizations. In order to resolve this issue we study the infinite-dimensional
inverse problem before constructing the discrete approximations.
In this article we set out to achieve the following goals:
G1. Construct a new class of infinitely divisible prior measures for recovery of compressible
parameters in connection to `p regularization techniques when p ∈ (0, 1).
G2. Present a systematic study of the class of infinitely divisible prior measures.
G3. Introduce an alternative to the classic total variation prior using the laws of pure Jump
Lévy process that is well-defined in infinite dimensions.
G4. Present a theory of well-posedness for Bayesian inverse problems that encompasses the
prior measures introduced under G1–G3.
3
Let us motivate some of these goals with an example.
Example 1. Suppose u ∈ Rn and the data y ∈ Rm is generated via the model
y = Au + η,
η ∼ N (0, σ 2 I)
where A ∈ Rn×m , σ > 0 is fixed and I is the m × m identity matrix. We wish to estimate u
given y. Here we are taking X = Rn , Y = Rm and the forward map has the form G(u) = Au.
Since η has a Gaussian density we can write the likelihood potential Φ(u; y) as:
Φ(u; y) =
1
kAu − yk22 .
2σ 2
Then, Bayes’ rule gives
1
dµy
1
2
(u) =
exp − 2 kAu − yk2 .
dµ0
Z(y)
2σ
Now define the prior measure via
(1.3)
dµ0
1
(u) = exp −kukpp
dΛ
U
where dΛ denotes the Lebesgue measure on Rn , k · kp denotes the usual `p (quasi-)norm in
Rn for p > 0 and U is the appropriate normalizing constant. Then the posterior µy can be
identified via its Lebesgue density as
dµy
1
1
2
p
(1.4)
(u) =
exp − 2 kAu − yk2 − kukp .
dΛ
Z(y)
2σ
The maximizer of the posterior density is referred to as the maximum a posteriori (MAP)
estimate. Formally, the MAP estimate of the posterior in (1.4) is given by
1
2
p
kAz − yk2 + kzkp .
uMAP = arg min
2σ 2
z∈Rn
For p ≥ 1 this optimization problem is convex and can be solved efficiently. Taking p = 1
results in the well-known `1 -regularization technique which is often used in the recovery of
sparse solutions. For values of p ∈ (0, 1) the resulting optimization problem is no longer
convex but it is a good model for recovery of sparse or compressible solutions [26, 42].
It is straightforward to check that the prior distribution (1.3) for p ∈ (0, 1) is non-convex
and heavy-tailed. However, we will see that this measure belongs to the much larger class of
infinitely divisible measures. Formally, a random variable ξ is infinitely-divisible if for every
P
1/n
1/n
n ∈ N its law coincides with the law of nk=1 ξk where {ξk } are i.i.d. random variables.
Thus, the above example is our first attempt at demonstrating the potential of infinitelydivisible prior measures (goals G1 and G2) that are introduced in Section 2. The connection
between sparse recovery and heavy-tailed or infinitely divisible priors has been observed in
the literature. Unser and Tafti [54] and Unser et al. [56, 55] study the sparse behavior of
stochastic processes that are driven by infinitely divisible force terms and advocate their use
4
in solution of inverse problems. A detailed discussion of some heavy-tailed prior distributions
such as generalizations of the student’s-t distribution and the `p -priors can also be found in
the dissertation [42]. Finally, Polson and Scott [47] and Carvalho et al. [16] propose a class
of hierarchical horseshoe priors that are tailored to the recovery of sparse signals.
In practice, solving a Bayesian inverse problem often refers to either identifying the posterior measure µy (such as in (1.4)) or extracting certain statistics from it such as the mean, the
variance, maximizer of the density etc. But before we can solve a Bayesian inverse problem
we need to know whether the problem is well-posed to begin with (point G4 above): Does µy
exist? Is it defined uniquely? Does it depend continuously on the data y? And finally, can
we approximate it in a consistent manner?
Later on we see that the well-posedness of a Bayesian inverse problem relies on the type
of prior measure µ0 that is chosen during the modelling step as well as certain properties of
the potential Φ. Well-posed Bayesian inverse problems were studied in [51, 20] with Gaussian
prior measures, in [21] with Besov priors, in [34] with convex prior measures and in [22, 52]
with heavy-tailed priors on separable Banach spaces. We note that our well-posedness results
in this article are closely related to those of [22]. The main difference is that our theory
does not rely on the assumption that the parameter space X is separable and we impose
slightly different conditions on the potential Φ. The non-separability condition is particularly
interesting when one takes X to be C α (the space of Hölder continuous functions) or BV ,
neither of which are separable. In Section 3 we introduce a class of prior measures that are
concentrated on BV and have piecewise constant samples (goal G3). This example is later
used in Section 5 as a prior measure in a deblurring problem.
1.1. Key definitions and notation. We gather here some key definitions and assumptions
that are used in the remainder of the article. We let R+ denote the positive real line [0, ∞)
and use the shorthand notation a . b when a and b are real valued functions and there exists
an independent constant C > 0 such that a ≤ Cb. Given two random variables ξ and ζ we
d
use the notation ξ = ζ to denote that they have the same laws (or distributions).
We use the shorthand notation {γk } to denote a sequence of elements {γk }∞
k=1 in a vector
space. The usual `p sequence spaces for p ∈ [1, ∞] are defined as the space of real valued
sequences {γk } such that k{γk }kp < ∞ where
k{γk }kp :=
∞
X
k=1
!1/p
p
|γk |
if
p ∈ [1, ∞)
and
k{γk }k∞ := sup |γk |.
k
Similarly, we define the k · kp norms of finite dimensional vectors. In particular k · k2 will
denote the usual Euclidean norm. Given a positive definite matrix Σ of size m × m, we define
the norm
kxkΣ := kΣ−1/2 xk2
for
x ∈ Rm .
Throughout the article we use Λ to denote the Lebesgue measure in finite dimensions.
Given a Borel measure µ on a Banach space X we define the spaces Lp (X, µ) for p ∈ [1, ∞)
as the space of µ-equivalent classes of functions h : X 7→ R such that |h|p is µ-integrable. We
also use the shorthand notation Lp (X) instead of Lp (X, Λ) whenever we are working with the
Lebesgue measure. Finally, if X is a Banach space, we use X ∗ to denote the topological dual
5
of X and BX (r) to denote the open ball of radius r > 0 in X that is centered at the origin.
The shorthand notation BX denotes the unit ball.
We shall consider the prior probability measure µ0 to be in the class of Borel probability
measures on X. In some cases we assume that the prior µ0 is Radon meaning that it is an
inner regular probability measure on the Borel sets of X. Furthermore, whenever we say that
µ is a probability measure on X we automatically mean that µ(X) = 1. Finally, throughout
this article we only consider complete probability measures in the following sense: If µ is a
Borel probability measure on X and A is a set of µ-measure zero then every subset of A also
has measure zero.
In this article we focus on the following notion of a well-posed Bayesian inverse problem:
Definition 1.1 (Well-posedness). Suppose that X is a Banach space and d(·, ·) 7→ R is a
metric on the space of Borel probability measures on X. Then for a choice of the prior
measure µ0 and the likelihood potential Φ, the Bayesian inverse problem given by (1.2) is
well-posed with respect to d if:
1. (Existence and uniqueness) There exists a unique posterior probability measure µy
µ0 given by Bayes’ rule (1.2).
0
2. (Stability) For every choice of > 0 there exists a δ > 0 so that d(µy , µy ) ≤ for all
y, y 0 ∈ Y so that ky − y 0 kY ≤ δ.
We will study the convergence of probability measures using the Hellinger and total variation metrics on the space of probability measures on X. For two probability measures µ1 and
µ2 that are absolutely continuous with respect to a third measure ν on X, the total variation
and Hellinger metrics are defined as
(1.5)
!2 1/2
r
r
Z
Z
1
dµ1 dµ2
1
dµ1
dµ2
dT V (µ1 , µ2 ) :=
−
dν and dH (µ1 , µ2 ) :=
−
dν .
2 X dν
dν
2 X
dν
dν
Both metrics are independent of the choice of the measure ν [8, Lem. 4.7.35]. Furthermore,
convergence in one of these metrics implies convergence in the other, due to the following
inequalities (see [8, Lem. 4.7.37] for a proof)
2d2H (µ1 , µ2 ) ≤ dT V (µ1 , µ2 ) ≤
(1.6)
√
8dH (µ1 , µ2 ).
However, one might prefer to work with the Hellinger metric as it relates directly to the error
in expectation of certain functions. Suppose that h ∈ L2 (X, µ1 ) ∩ L2 (X, µ2 ). Then using the
Radon-Nikodym theorem and Hölder’s inequality one can show (see [34, Sec. 1] for details)
(1.7)
Z
h(u)dµ1 (u) −
X
Z
Z
h(u)dµ2 (u) ≤ 2
X
2
Z
h (u)dµ1 +
X
2
h (u)dµ2
1/2
dH (µ1 , µ2 ).
X
For reasons that will become clear in Section 4, we prefer to study the well-posedness of
inverse problems using both the Hellinger and total variation metrics. The main difference is
in the restrictions that we need to impose on the prior µ0 in order to obtain a certain rate of
convergence for each metric.
6
2. Infinitely-divisible prior measures. We start by presenting a generalization of the prior
distribution (1.3) that was considered in Example 1. We show that this prior belongs to a
larger class of distributions that are closely related to `p regularization techniques. We shall
extend these distributions to measures on Banach spaces with an unconditional Schauder basis
and observe that they belong to the much larger class of infinitely-divisible (ID) measures (see
Definition 2.8). Motivated by this connection between `p regularization and ID priors, we
turn our attention to the ID class and discuss some of its properties. In particular, we study
the tail behavior of ID priors with respect to their Lévy measures (see Definition 2.10).
2.1. A class of shrinkage priors with compressible samples. When faced with the problem of recovering a sparse or compressible parameter we require the prior measure to reflect
the intuition that the solution to the inverse problem is likely to have only a few large modes
in some basis and the rest of the modes are negligible (see [42, Sec. 6.1]). Such prior distributions are often referred to as “shrinkage priors” and they have been the subject of extensive
research [47, 16, 28, 18, 17]. In this section we consider a few examples of shrinkage priors
that are closely related to `p regularization techniques.
Most of the existing literature on shrinkage priors is focused on finite dimensional problems
but we present an extension of these priors to infinite-dimensional Banach spaces. Since
compressibility is often considered with respect to a basis, it makes sense for us to consider a
parameter space that has a basis.
Given a parameter space X, or at least a subspace X̃ ⊆ X that has an unconditional
Schauder basis {xk }, we construct random variables of the form
(2.1)
u∼
∞
X
γk ξk xk
k=1
where {γk } is a fixed sequence of real valued coefficients that decay sufficiently fast and the
{ξk } are a sequence of independent real valued random variables that need not be identically
distributed. We will take the prior measure µ0 to be the law of the random variable u in (2.1).
We refer to such a prior measure µ0 as the product prior obtained from {γk } and {ξk }. This
construction of the prior is reminiscent of the Karhunen-Loéve expansion of Gaussian measures
[7, Thm. 3.5.1]. The following theorem gives sufficient conditions that ensure k·kX < ∞ µ0 -a.s.
Theorem 2.1.[34, Thm. 3.9] Suppose that X is a Banach space with an unconditional
Schauder basis and let u be as in (2.1). If {γk2 } ∈ `p and {Varξk } ∈ `q for 1 < p, q < ∞ so
that 1/p + 1/q = 1 (with p = 1 for the limiting case when q = ∞), then kukX < ∞ a.s. In
particular, if the {ξk } are i.i.d., Varξ1 < ∞ and {γk } ∈ `2 , then kukX < ∞ a.s.
We can also show that the product prior µ0 that is induced by (2.1) is Radon. Proof of
the next theorem follows the same approach as [34, Thm. 3.10(ii)] and is hence omitted.
Theorem 2.2. Let µ be the probability measure that is induced by the random variable u
given by (2.1) where {γk } and {ξk } satisfy the conditions of Theorem 2.1. Then µ is a Radon
probability measure on X if the random variables {ξk } are distributed according to Radon
probability measures on R.
Before going further we present a result on the second raw moment of product priors which
will be useful throughout the remainder of the article.
7
Theorem 2.3. Suppose that X is a Banach space with an unconditional Schauder basis
{xk } and let µ be the product prior obtained from {γk } ∈ `2 and {ξk } where ξk are i.i.d. and
Varξk < ∞. Then k ·P
kX ∈ L2 (X, µ).
Proof. Let uN = N
k=1 γk ξk xk then for M > N > 0 we have
Z
Z
Z
(kuM kX − kuN kX )(kuM kX + kuN kX ) dµ
kuM k2X dµ −
kuN k2X dµ =
X
X
X
By Theorem 2.1 we know that kukX < ∞ a.s. and so in the limit as M, N 7→ ∞, |(kuM kX −
kuN kX )| 7→ 2kukX and |(kuM kX − kuN kX )| 7→ 0 and so {kuN k2X } is Cauchy in L2 (X, µ).
We are now in position to discuss a few examples of shrinkage priors. Motivated by
Example 1, we define the class of `p -priors as follows:
Definition 2.4 (`p -prior). Let X be a Banach space with an unconditional Schauder basis
{xk }, then we say that
µ is an `p -prior on X if its samples can
P∞a Radon probability measure
2
be expressed as u = k=1 γk ξk xk where {γk } ∈ ` and {ξk } is an i.i.d. sequence of real valued
random variables with Lebesgue density
p
|t|p
(2.2)
ξk ∼
exp − p dΛ(t)
2αΓ(1/p)
α
p
where p ∈ (0, ∞) and α = Γ(1/p)/Γ(3/p).
Here Γ denotes the usual Gamma function. The distribution in (2.2) belongs to the larger
class of Generalized Normal distributions [44]. This class is also referred to as a Kotz-type
distribution [43] or a generalized Laplace distribution [37]. Here we will not use either of these
terms and simply refer to this distribution as the `p -distribution to emphasize its connection
to `p -regularization techniques. The random variables ξk have bounded moments of all orders
(see [44] or the discussions following the definition of the Gp,q -prior below), in fact
αs (1 + (−1)s )
s+1
s
E ξk =
Γ
for
s ∈ N.
2Γ(1/p)
p
In particular we have that Varξk = 1 and so it follows from Theorem 2.3 that the `p -prior
has bounded second moments.
Another, closely related class of priors to the `p -priors can be obtained by a symmetrization
of the Weibull distribution:
Definition 2.5 (Wp -prior).Let X be a Banach space with an unconditional Schauder basis
{xk }, then we say that
µ is a Wp -prior on X if its samples can
P∞a Radon probability measure
2
be expressed as u = k=1 γk ξk xk where {γk } ∈ ` and {ξk } is an i.i.d. sequence of real valued
random variables with Lebesgue density
p |t| p−1
|t|p
(2.3)
ξk ∼
exp − p dΛ(t),
α α
α
where p ∈ (0, ∞) and α = (2Γ(1 + 2/p))−1/2 .
The distribution of ξk is simply a symmetric version of the well-known Weibull distribution
[35], hence the name Wp . A straightforward calculation shows that Varξk = 1 and once again
it follows from Theorem 2.3 that the Wp -priors have bounded second moments.
8
Both the Wp and `p distributions reduce to the Laplace distribution when p = 1. For
p < 1 the `p distribution has non-convex level sets and puts a large portion of its mass close
to the axes (see Figure 1). This behavior becomes stronger for smaller p and suggests that
the `p -prior will incorporate sparse behavior as p 7→ 0.
The Wp distribution behaves very differently in comparison to the `p distribution. For
p < 1 the Wp distribution blows up at the origin (see Figure 1(a)). This means that the
Wp distribution puts more of its mass at the origin which leads us to believe that it must
incorporate stronger compressibility than the `p distribution.
Further insight into the behavior of the Wp -prior can be obtained by considering its MAP
point estimate in finite dimensions. Formally, using this prior in Example 1 gives rise to an
optimization problem of the form
(
)
n
X
1
2
p
uMAP = arg min
kAz − yk2 + kzkp + (1 − p)
log(|zk |) .
2
z∈Rn
k=1
Of course, the log term on right hand side is not bounded from below and so we cannot gain
much insight from this problem. However, we can consider a slightly modified version of this
optimization problem by introducing a small parameter > 0
)
(
n
X
1
2
p
kAz − yk2 + kzkp + (1 − p)
log( + |zk |) .
u = arg min
2
z∈Rn
k=1
Now if is small then the log term will heavily penalize any modes of the solution that are on
a larger scale than that of and so we expect that most of the modes of the solution u will
be on the scale of the small parameter . The stronger shrinkage of the posterior due to the
Wp -prior is also evident in Figure 2 where we compare a prototypical example of posteriors
that arise from the Wp and `p priors for solution of Example 1 in 2D. Here, we clearly see that
the W1/2 -prior results in a posterior that is highly concentrated around the axes compared to
the posterior that arises from `1/2 -prior which is more spread out. Note that in either case,
the posteriors are highly concentrated around the axes meaning that the map estimates as
well as most of the samples from these posteriors will incorporate sparsity.
Comparing the distributions (2.2) and (2.3) suggests the definition of a larger class of
priors that can interpolate between the `p and Wp -priors. To this end, we introduce a new
class of prior measures called the Gp,q -priors. The letter G is chosen due to the connection of
the one dimensional version of these measures to the generalized Gamma distribution [10].
Definition 2.6 (Gp,q -prior).Let X be a Banach space with an unconditional Schauder basis
{xk }, then we say that
µ is a Gp,q -prior on X if its samples can
P a Radon probability measure
2 and {ξ } is an i.i.d. sequence of real valued
be expressed as u = ∞
γ
ξ
x
with
{γ
}
∈
`
k
k
k=1 k k k
random variables with Lebesgue density
(2.4)
p
t
ξ1 ∼
2αΓ(q/p) α
q−1
t p
exp −
dΛ(t),
α
where p ∈ (0, ∞) and α = (Γ(q/p)/Γ((2 + q)/p))1/2 .
9
Figure 1: Contour plots of `p and Wp densities in 2D for different values of p.
Figure 2: A prototypical example of densities that arise in the solution of Example 1 in 2D
with the `1/2 (top row) and W1/2 priors (bottom row). From left to right columns: The
likelihood that arises from the additive Gaussian noise model, the prior densities and the
resulting posteriors. The densities are rescaled for better visualization.
Using the change of variables s =
Z
0
∞
tk
tp
βp
we see that for k, β ≥ 0
p
q−1
Z
t
β k+1
k+q
t
β ∞ k+q
−1
p
exp −
dΛ(t)
=
s
exp
(−s)
dΛ(s)
=
Γ
.
β
βp
p 0
p
p
Setting k = 0 leads us to the normalizing constant in the definition of the distribution in (2.4).
10
Furthermore, we obtain the following expression for the moments of the Gp,q distributions
E |ξ1 |s =
αs (1 + (−1)s )Γ ((s + q)/p)
2Γ (q/p)
s ∈ N.
In particular Varξ1 = α2 Γ((2 + q)/p)/Γ(q/p) = 1. Clearly, the `p prior is equivalent to Gp,1
and Wp is equivalent to Gp,p . Furthermore, the G1,q distribution coincides with a symmetrization of the the Gamma distribution. For q < 1 the distribution (2.4) will blow up at the origin
and so it will put a lot of its mass at zero. The Gp,q distributions belong to the class of ID
measures by the following theorem of Bondesson.
Theorem 2.7 ([10, Cor. 2]). All probability density functions on (0, ∞) of the form
p
π(t) =
αΓ(q/p)
q−1
p
t
t
exp −
α
α
are ID for q, α > 0 and 0 < p ≤ 1.
Later on we show that if the ξk are distributed according to an ID distribution then the
corresponding product prior on X will also be an ID probability measure. Then the Gp,q priors are also ID. This fact suggests the question of what other types of ID measures are
good models for compressibility? We know that heavy-tailed distributions such as the Cauchy
or Student’s t distributions are ID and they incorporate compressible samples as well. Then
there is much to be gained from the study of ID prior measures in Bayesian inverse problems.
To the best of our knowledge a thorough study of the compressible behavior of ID distributions
is still missing in the literature. The closest reference in this direction is the works of Unser
et. al. [54, 56, 55]. While we do not study the modelling of compressible parameters, we
recognize the potential impact of ID priors in this subject and so we dedicate the remainder
of this section to the study of ID priors.
2.2. Infinitely divisible priors. We begin by collecting some results on the class of ID
probability measures on Banach spaces. We only present the results that are needed in
our exposition and refer the reader to [41] for a detailed introduction to ID measures on
Banach spaces. Further reading can be found in the monograph [54] which contains a modern
treatment of ID probability measures on nuclear spaces and the books [3, 49, 50] that are
good references on the theory of ID measures in finite dimensions.
Recall that given a Borel probability measure µ on a Banach space X its characteristic
function µ̂ : X ∗ 7→ C is given by
Z
µ̂(%) =
exp(i%(u))dµ(u)
∀% ∈ X ∗ .
X
Characteristic functions play a crucial role in our discussion of ID measures in this section.
In what follows ν ∗n denotes the n-fold convolution of a measure ν with itself.
Definition 2.8 (ID measures [41]). A Radon probability measure µ on a Banach space X
is called an infinitely divisible measure if for each n ∈ N there exists a Radon probability
measure µ1/n so that µ = (µ1/n )∗n . Equivalently, the probability measure µ is ID if µ̂(%) =
(µ̂1/n (%))n , ∀% ∈ X ∗ .
11
Put simply, a real valued random variable ξ is distributed according to an ID measure if for
d P
every n ∈ N one can find a collection of i.i.d. random variables {ξk }nk=1 so that ξ = nk=1 ξk .
Examples of such distributions include Gaussian, Laplace, Gamma, log-normal, Cauchy and
Student’s-t. More examples can be found in the monograph [50] where ID distributions on R
are studied in detail. We note that an equivalent definition of an ID measure is given as the
law of a Lévy process terminated at unit time. However, we will not use this definition in order
to avoid the technicalities of dealing with Lévy processes but instead we refer the interested
reader to the monographs [46, 19] for further reading. The proof of the next theorem can be
found in [41, Sec. 5.1].
Theorem 2.9. Let µ be an ID probability measure on a Banach space X. Then
(i) µ̂(%) 6= 0 for all % ∈ X ∗ .
(ii) There exists a unique and continuous (in the dual norm) function ψ : X ∗ 7→ X so that
µ̂(%) = exp(ψ(%)) and ψ(0) = 0.
(iii) If µ is symmetric, i.e. µ(A) = µ(−A) for all Borel subsets A of X, then µ̂ is real valued
and positive.
(iv) For every n ∈ N the measures µ1/n are uniquely determined and µ̂1/n (%) = exp(n−1 ψ(%))
for all % ∈ X ∗ .
Furthermore, we define the function
Ψ(u, %) := exp(i%(u)) − 1 − i%(u)1BX (u)
∀u ∈ X, % ∈ X ∗ ,
where BX is the unit ball in X and 1BX is the characteristic function of the unit ball. We
recall the definition of a Lévy measure on a Banach space.
Definition 2.10 (Lévy Measure). A positive σ-finite Radon measure λ on X is called a Lévy
measure if and only if
1. λ({0})
= 0.
R
∗
2. X |Ψ(u,
R %)|dλ(u) < ∞ for every % ∈ X .
3. exp( X Ψ(u, %)dλ(u)) is the characteristic function of a Radon probability measure on
R for every % ∈ X ∗ .
We are now ready to present the celebrated Lévy-Khintchine representation theorem (see
[41, Sec. 5.7] for a proof):
Theorem 2.11 (Lévy-Khintchine representation). A Radon probability measure on a Banach
space X is infinitely divisible if and only if there exists an element m ∈ X, a (positive definite)
covariance operator R : X ∗ 7→ X and a Lévy measure λ, so that
Z
1
(2.5)
µ̂(%) = exp(ψ(%)) where ψ(%) = i%(m) − %(R(%)) +
Ψ(u, %)dλ(u).
2
X
Equivalently, µ is ID precisely when there exists
R a point mass δm , a Gaussian measure N (0, R)
and a Radon measure ν identified via ν̂(%) = X Ψ(u, ρ)dλ(u) so that
(2.6)
µ = δm ∗ N (0, R) ∗ ν.
The Lévy-Khintchine representation implies that the triple (m, R, λ) completely identifies an
ID measure µ and so we use the shorthand notation µ = ID(m, R, λ). To gain more insight
12
into the implications of the Lévy-Khintchine representation we recall the class of compound
Poisson random variables and their corresponding probability measures.
Definition 2.12 (Compound Poisson probability measure [41, Sec. 5.3]). Let η be a Radon
probability measure on a Banach space X and suppose that {uk } is a sequence of i.i.d. random
variables so that uk ∼ η. Also, let τPbe an independent Poisson random variable with rate
c > 0 taking values in Z+ . Then u = τk=0 uk is distributed according to a compound Poisson
probability measure denoted by CPois(c, η).
It is straightforward to check that the characteristic function of a compound Poisson
measure has the form
Z
\
∀% ∈ X ∗ .
CPois(c, η)(%) = exp c (exp(i%(u)) − 1) dη(u)
X
See [41, Prop. 5.3.1] for a proof of this formula along with the fact that CPois(c, η) is a Radon
measure on X.
Now let us return to the characteristic function of the probability measure ν that was
introduced in the Lévy-Khintchine representation (2.6)
Z
Z
ν̂(%) = exp
−i%(u) dλ(u) .
(exp(i%(u)) − 1) dλ(u) exp
(2.7)
X
BX
If 0 < λ(X) < ∞ then λ can be renormalized to define a probability measure λ̃ :=
Furthermore, we can define an element uλ ∈ X so that
Z
%(uλ ) = −
%(u)1BX (u)dλ(u)
∀% ∈ X ∗ .
1
λ(X) λ.
X
Putting these observations together with (2.7) gives the decomposition
(2.8)
ν = CPois(λ(X), λ̃) ∗ δuλ .
Therefore, from (2.6) we deduce that any measure µ = ID(m, R, λ) with λ(X) < ∞ can
be decomposed as
(2.9)
µ = (δm+uλ ) ∗ N (0, R) ∗ CPois(λ(X), λ̃).
In the remainder of this article we will restrict our attention to the case of ID measures with
λ(X) < ∞. Since the tail behavior of prior measures is of importance to our well-posedness
results in Section 4 we now present some results concerning the tail behavior of ID measures.
We begin with the notion of a submultiplicative function.
Definition 2.13 (Submultiplicative function).A non-negative, non-decreasing and locally bounded
function h : R 7→ R+ is called submultiplicative if it satisfies
h(t + s) ≤ Ch(t)h(s)
∀t, s ∈ R
with an independent constant C > 0
Our interest in the class of submultiplicative functions arises from the next theorem that
describes some of the properties of this class.
Theorem 2.14 ([49, Prop. 25.4]).
13
(i) The product of two submultiplicative functions is also submultiplicative.
(ii) If h is submultiplicative then so is (h(at + b))α for constants a, b ∈ R and α > 0.
(iii) The functions max{1, |t|} and exp(|t|β ) for β ∈ (0, 1) are submultiplicative.
We now present a theorem that relates the tail behavior of an ID measure to that of its
Lévy measure. This result was originally proven by Kruglov [38] for ID random variables on
R with Lévy measures that are not necessarily finite. Different generalizations of this result
to larger spaces are also available in the literature. For example, see [49, Thm. 25.3] for
extension to Rn and [46, Prop. 6.9] for Hilbert space valued Lévy processes. For the reader’s
convenience we briefly prove this result for Banach space valued ID random variables with
finite Lévy measures.
Theorem 2.15. Let X be a Banach space and λ be a Lévy measure so that 0 < λ(X) <
∞. Suppose that u ∼ µ = ID(m, R, λ), µ(X) = 1 and k · kX < ∞ µ-a.s. Then, given a
submultiplicative function h we have that h(k · kX ) ∈ L1 (X, µ) if h(k · kX ) ∈ L1 (X, λ).
Proof. Let u ∼ µ, then following Theorem 2.11 and the decomposition (2.9) above, we
know that there exists an element m̃ ∈ X and independent random variables w ∼ N (0, R)
1
and v ∼ CPois(λ(X), λ(X)
λ) so that
E h(kukX ) = E h(km̃ + w + vkX ) ≤ C 3 h(km̃kX )(E h(kwkX ))(E h(kvkX ))
where the inequality follows because of the triangle inequality and the fact that h is nondecreasing and locally bounded. Now by [49, Lem. 25.5] we know that there exist constants
a, b > 0 such that h(x) ≤ b exp(a|x|). Using this bound with the assumption that kukX < ∞
µ-a.s. along with Fernique’s theorem [7, Thm. 2.8.5] for Gaussian
measures on Banach spaces
R
1
implies that E h(kwkX ) < ∞. Now suppose that λ(X) X h(kukX )dλ(u) = U < ∞. Then
using the law of total expectation [6, Thm. 34.4], the fact that h is submultiplicative and v is
a compound Poisson random variable we get
! !
! !
N
N
X
X
E h(kvkX ) = E E h
vk
N ≤E Eh
kvk kX N
k=0
≤E
N
C E
N
Y
k=0
X
!
h kvk kX
!
=E
N
C
N
!
E h kvk kX
!
N
k=0
k=0
= E ((U C)N |N ) =
N
Y
∞ −λ(X)
X
e
(U Cλ(X))k
k=0
k!
< ∞.
By putting Theorems 2.15 and 2.14 together we immediately obtain the following corollary
concerning the moments of ID measures.
Corollary 2.16. Suppose that X is a Banach space and µ = ID(m, R, λ). If λ is a Lévy
measure on X so that 0 < λ(X) < ∞, µ(X) = 1 and k · kX < ∞ µ-a.s. then k · kX ∈ Lp (X, µ)
whenever k · kX ∈ Lp (X, λ) for p ∈ [1, ∞).
Another interesting case is when the Lévy measure λ is convex. Recall that a Radon
probability measure ν on X is said to be convex whenever it satisfies
ν(βA + (1 − β)B) ≥ ν(A)β ν(B)1−β
14
for β ∈ [0, 1] and all Borel sets A and B (see [34, 11] for more details about convex measures).
We are interested in convex measures since they have exponential tails under mild assumptions
[34, Thm. 3.6]. More precisely, if ν is a convex probability measure on X and k · kX < ∞
ν-a.s. then there exists a constant 0 < b < ∞ so that exp(bk · kX ) ∈ L1 (X, ν). Since the
exponential is a submultiplicative function, we immediately obtain the following corollary.
Corollary 2.17. Suppose that X is a Banach space and µ = ID(m, R, ν). If ν is a convex
probability measure on X, µ(X) = 1 and k · kX < ∞ a.s. under both ν and µ then there exists
a constant b > 0 so that exp(bk · kX ) ∈ L1 (X, µ).
At the end of this section we ask whether we would obtain an ID measure if we used
a sequence of ID random variables in order to generate a product prior. The answer to
this question is affirmative and serves as the proof of our claim that Gp,q -priors that were
introduced earlier belong to the class of ID probability measures.
Theorem 2.18. Let X be a Banach space with an unconditional Schauder basis {xk } and
let µ be the product prior that is obtained from {γk } ∈ `2 and an i.i.d. sequence {ξk } of real
valued random variables. Suppose that ξk ∼ ID(0, σ 2 , λ) where σ > 0 and λ is a symmetric
and finite Lévy measure on R such that max{1, | · |2 } ∈ L1 (R, λ). Then µ is a Radon ID
probability measure on X with characteristic function
"
#
∞
∞ Z
X
1X 2 2
2
µ̂(%) = exp −
σ γk %(xk ) +
(cos(γk %(xk )tk ) − 1)dλ(tk )
∀% ∈ X ∗ .
2
R
k=1
k=1
Proof. Since max{1, | · |2 } ∈ L1 (R, λ), the Lévy measure of the ξk has bounded second
moment and so by Corollary 2.16 we see that Varξk < ∞. Now it follows from Theorem 2.1
that k · kX < ∞ µ-a.s. since {γk } ∈ `2 . The fact that µ is Radon follows from Theorem 2.2.
Now we consider the characteristic function of µ. Using
R the definition of the characteristic
1 2 2
ˆ
function of µ and the fact that ξk (z) = exp − 2 σ z + R cos(tz − 1)dλ(z) , we can write
#
"
∞ Z
∞
∞
X
Y
1X 2 2
2
(cos(γk %(xk )tk ) − 1)dλ(tk ) .
µ̂(%) =
E exp (iγk ξk %(xk )) = exp −
σ γk %(xk ) +
2
R
k=1
k=1
k=1
{µN }∞
N =1
Now consider the sequence of measures
that are defined via
"
#
N
N Z
X
1X 2 2
µ̂N (%) = exp −
σ γk %(xk )2 +
(cos(γk %(xk )tk ) − 1)dλ(tk ) .
2
R
k=1
k=1
Each µN is ID given the fact that a P
finite sum of ID random variables is ID. Since the {xk }
2
2
2
are normalized and {γk } ∈ ` then ∞
k=1 γk %(xk ) < ∞. Furthermore, using the inequality
2
| cos(t) − 1| ≤ t we can write
Z
∞ Z
∞ Z
∞
X
X
X
2
2 2
2
2
(cos(γk %(xk )tk ) − 1)dλ(tk ) ≤
γk %(xk ) tk dλ(tk ) =
γk %(xk )
t2k dλ(tk )
k=1
R
k=1
R
k=1
R
But this term is also bounded since {γk } ∈ `2 , {xk } are normalized and max{1, |x|2 } ∈
L1 (R, λ). Then, µ̂N (`) 7→ µ̂(`) for all ` ∈ X ∗ and so the sequence µN converges weakly
to µ. Therefore, µ is also ID by [41, Thm. 5.6.2]. Observe that the Lévy measure of µ is
concentrated along the coordinate axes of the basis {xk }.
15
3. Stochastic process priors on BV. Total variation regularization is a classic technique
for recovery of blocky images in the variational approach to inverse problems [57, Ch. 8]. As
we mentioned earlier, it was shown in [39] that the TV-prior is not discretization invariant and
converges to a Gaussian measure in the limit of fine discretizations. In this section we consider
prior measures that are defined as laws of stochastic processes with jump discontinuities in
order to model discontinuous functions with bounded variation. The resulting prior measures
are defined directly on the function space BV and so our definition can get around the inconsistency that was observed in [39]. We emphasize that our construction of a BV -prior does
not disprove the result of [39] but provides a well defined alternative to the classic TV-prior.
We also note that our approach is not the only way to construct a well defined TV-prior. For
example, [58] presents a TV-prior that is absolutely continuous with respect to an underlying
Gaussian measure and results in a well-posed inverse problem.
Following [40, Ch. 13] we define the space BV (Ω) of functions of bounded variation on
an open set Ω ⊂ Rn as the space of functions u ∈ L1 (Ω) whose first order partial derivatives
are finite signed Radon measures i.e. for j = 1, 2, · · · , n there exist finite signed measures
Mj : B(Ω) 7→ R so that
Z
∂φ
u(t)
(t)dΛ(t) = −
∂tj
Ω
Z
φ(t)dMj (t),
We define the variation of u as
( n Z
X
φk (t)dMk (t)
V (u) := sup
k=1
∀φ ∈ Cc∞ (Ω).
Ω
)
φ∈
Cc∞ (Ω; Rn ), |φ(t)|
≤ 1, ∀t ∈ Ω ,
Ω
where Cc∞ (Ω; Rn ) denotes the space of vector valued smooth functions with compact support
in Ω. The space BV (Ω) is a Banach space when equipped with the norm
kukBV (Ω) := kukL1 (Ω) + V (u)
but it is not separable [12, Prop. 2.3]. There is a correspondence between the space BV ([0, 1])
and the space of functions with finite total variation in one dimensions. Recall that the total
variation of a function u : [0, 1] 7→ R is defined as
T V (u) := sup
(K
X
)
|u(tk ) − u(tk−1 )|
k=1
where the supremum is taken over all finite partitions 0 = t0 < t1 < t2 < · · · < tK = 1 of
the interval [0, 1] with K ∈ N. It is known that if T V (u) < ∞ then V (u) ≤ T V (u) and
every u ∈ BV ([0, 1]) has a right continuous representative with bounded total variation [40,
Thm. 7.2]. We prefer to work with the space BV and its corresponding norm rather than the
total variation functional since the former is readily defined in higher dimensions. We start
by constructing a prior measure on BV ([0, 1]) as the law of a pure Jump Lévy process. We
shall generalize our construction to BV (Ω) later in Section 3.2.
16
In order to define our prior measure we will use some well-known results from the theory
of Lévy processes (see [19] for an extensive introduction). Using the Lévy-Khintchine formula
for Lévy processes [19, Thm. 3.1] we identify a Lévy process u(t) via its characteristic function
E exp(isu(t)) = exp(tψ(s))
where
1
ψ(s) = ims − (σs)2 +
2
for
s ∈ R,
Z
exp(iξs) − 1 − isξ1{|ξ|≤1} (ξ)dλ(ξ).
R
Here, the constants m ∈ R and σ ≥ 0 are fixed and λ is a Lévy measure on R (see Definition 2.10). Similar to the case of ID measures, the characteristic triplet (m, σ, λ) uniquely
identifies the stochastic process u(t). Certain pathwise properties of u(t) can be inferred from
its characteristic triplet.
Theorem 3.1 ([19, Prop. 3.9]). Let u(t) be a Lévy process with characteristic triplet (m, σ, λ).
Then u(t) ∈ BV ([0, 1]) a.s. and E kukBV ([0,1]) < ∞ if
Z
|ξ|dλ(ξ) < ∞.
(3.1)
σ = 0 and
{|ξ|≤1}
Such a process is of the pure jump type. If in addition λ(R) < ∞ then u(t) is a compound
Poisson process with piecewise constant sample paths.
Thus, the law of a Lévy process u(t) that satisfies (3.1) coincides with a probability
measure that is supported on BV ([0, 1]). Let us denote this measure by µ. We wish to
use this measure as a prior within the Bayesian framework and achieve a well-posed inverse
problem. An important question at this point is whether µ is a Radon measure on BV ([0, 1])
since the Radon property can often simplify the well-posedness analysis. We will now show
that in the compound Poisson case, i.e. when λ(R) < ∞, the measure µ is tight and hence
Radon [2, Lem. 12.6]. To our knowledge this result does not hold for general choices of the
Lévy measure λ.
Recall Helly’s selection principle [29, Thm. 12] stating that a set A ⊂ BV ([0, 1]) is relatively compact if there exists a constant M > 0 so that
kwkL∞ ([0,1]) < M,
T V (w) < M,
∀w ∈ A.
Thus, to show that µ is a tight measure on BV ([0, 1]) we need to argue that for every > 0
there is an M > 0 so that
µ({w ∈ BV ([0, 1]) : kwkL∞ ([0,1]) ≥ M, T V (w) ≥ M }) < .
Now suppose that u(t) isRa compound Poisson process with characteristic triplet (0, 0, λ) such
that c = λ(R) < ∞ and R sdλ(s) < ∞ (We added the last condition to ensure that λ can be
normalized to define a probability measure with bounded expectation). Then, we can write
(3.2)
u(0) = 0,
u(t) =
τ (t)
X
k=1
ξk
for t ∈ (0, 1].
17
Here τ (t) is a Poisson process with intensity c i.e.
P(τ (t) = k) =
(ct)k
exp(−ct)
k!
and {ξk } is an i.i.d. sequence of random variables distributed according to the measure c−1 λ.
A few draws from such a process are given in Figure 3(a) when λ is a standard Gaussian.
Using this representation of u(t) and the law of total expectation [6, Thm. 34.4] we can write
(3.3)
E kukL∞ ([0,1]) = E sup
τ (t)
X
ξk ≤ E
t∈[0,1] k=1
E
N
X
!
|ξk | τ (1) = N
= cE |ξ1 | < ∞.
k=1
Furthermore, since the total variation of a piecewise constant function is simply the sum of
the the jump sizes we have
(3.4)
E T V (u) = E
τ (1)
X
|ξk | = cE |ξ1 | < ∞.
k=1
Now it follows from Markov’s inequality that for any M > 0
P(kukL∞ ([0,1]) ≥ M ) ≤
E kukL∞ ([0,1])
,
M
P(T V (u) ≥ M ) ≤
E T V (u)
.
M
A straightforward argument yields that for any choice of > 0 we can choose M > 0 large
enough so that P(kukL∞ ([0,1]) > M, T V (u) > M ) ≤ and so the measure µ, the law of u(t),
is tight (Radon) on BV ([0, 1]).
3.1. Combination with Gaussian processes. The compound Poisson process is a convenient model for functions with jump discontinuities. However, the fact that its sample paths
are piecewise constant can be too restrictive. In order to achieve a more flexible prior, that
can model piecewise continuous functions, we combine our compound Poisson processes with
a Gaussian process. If the sample paths of the Gaussian process are sufficiently regular then
the resulting prior measure will still be concentrated on BV ([0, 1]). The theory of Gaussian
processes is well developed and a detailed introduction can be found in the monograph [48].
Here we recall some basic results and only consider the case of a Gaussian process with C ∞
sample paths. Our approach can easily be generalized to less regular Gaussian processes by
choosing a different kernel [48, Sec. 4.2].
Let g(t) be a random function on [0, 1] so that for any finite collection of points {tk }nk=1
the random variables {g(tk )}nk=1 are jointly Gaussian. Furthermore, suppose that
(g(t1 ), · · · , g(tn ))T ∼ N (0, K)
where
Kk,j = κ(tk , tj ).
Here κ(r, s) := exp(−b|r − s|2 ) is the covariance kernel of g(t) and b > 0 is a fixed constant.
Under these assumptions g(t) is a mean zero Gaussian process and its samples are almost
surely in C ∞ ([0, 1]) (see [45, Sec. 2.5.4]). By definition, the Law of g(t) is a Gaussian measure
and since the kernel κ is positive definite and continuous it follows from the Karhunen-Loéve
18
theorem (see [27, Sec. 2.3] that the law of g(t) is supported on a Hilbert space and so it is a
Radon measure.
Now let us consider a compound Poisson process u(t) as in (3.2) in addition to the Gaussian
process g(t). Then the new process
(3.5)
v(t) = g(t) + u(t)
will have sample paths that are piecewise C ∞ with finitely many jumps. Furthermore, since
the laws of u(t) and g(t) are both Radon then the law of v(t) is also Radon. Examples of
draws from the process v(t) are given in Figure 3(b)
a)
b)
5
v(t)
u(t)
5
0
-5
0
-5
0
0.2
0.4
0.6
0.8
1
0
t
0.2
0.4
0.6
0.8
1
t
c)
d)
Figure 3: (a) Three samples from the compound Poisson process (3.2) with normal jumps
depicting the piecewise constant sample paths. (b) Samples from the v(t) process generated
using (3.5) by combining a smooth Gaussian process and an independent compound Poisson
process with normal jumps. (c,d) Two draws from the random field u(t) generated using (3.6)
by combining a smooth Gaussian field with a Poisson process.
3.2. Extension to higher dimensions. At the end of this section we discuss some possibilities for extension of the compound Poisson process priors to random fields in BV (Ω) for
compact domains Ω ⊂ Rn , for n = 2, 3, with Lipschitz boundary. Let g be a Gaussian process
on Ω with kernel κ(r, s) = exp(−|r − s|2 ) i.e.
(g(t1 ), · · · , g(tn ))T ∼ N (0, K),
Kk,j = κ(tk , tj )
for any collection t1 , · · · tn ∈ Ω.
Under these assumptions g ∈ C ∞ (Ω) a.s. Now consider the random field
τ (g + (t))
(3.6)
u(t) =
X
k=1
ξk ,
for t ∈ Ω
where g + (t) := max{0, g(t)}.
19
As before, τ is an independent Poisson random variable with rate c > 0 and {ξk } is a sequence of i.i.d. random variables distributed according to the probability measure c−1 λ. It
is straightforward to check that samples from u are piecewise constant functions on Ω that
jump along a finite number of the positive level sets of the field g. These level sets are chosen
by the poisson random variable τ . Since we assumed that g is in C ∞ (Ω) a.s. we expect that
the level sets of g are also smooth and so u is a piecewise constant function that jumps along
finitely many smooth curves (surfaces). We will now check whether the law of the field u is
indeed supported on BV (Ω).
In what follows we will occasionally suppress the dependence of different functions on t
to make the expressions more readable. Consider a test function φ ∈ Cc∞ (Ω; Rn ) so that
|φ(t)| ≤ 1. We can write
+
!
Z
Z
τ (g )
τ (max g + ) Z
j
X
X
X
u divφ dΛ(t) =
ξk divφ dΛ(t) =
ξk divφ dΛ(t)
Ω
Ω
Aj
j=1
k=1
k=1
with the convention that the sum inside the integral is set to zero whenever τ (max g + ) = 0.
The sets Aj := {t ∈ Ω τ (g + (t)) = j} are the subsets of Ω on which u is constant. Since τ is
almost surely finite, we can integrate by parts [24, Thm. 5.6] to get
τ (max g + ) Z
Z
u divφ dΛ(t) =
Ω
X
∂Aj
j=1
Jj (u)hφ, ϑj idΘn−1
(s).
j
Here ∂Aj is the boundary of Aj , ϑj is the unit outward normal on ∂Aj , Jj (u) is the jump of u
across ∂Aj going from Aj−1 to Aj and Θjn−1 are the (n − 1)-dimensional Hausdorff measures
on ∂Aj [24, Ch. 2]. Recall that the (n − 1)-dimensional Hausdorff measure of a simple curve
in 2D (resp. surface in 3D) coincides with its length (resp. surface area) [24, Sec. 3.4]. Since
the size of the jumps of u are a.s. finite we can write
Z
τ (max g + )
τ (max g + ) Z
X
X
u divφ dΛ(t) ≤
|ξk |
hφ, ϑj idΘn−1 (s)
Ω
j=1
k=1
≤
τ (max g + )
X
k=1
|ξk |
j
∂Aj
τ (max g + )
X
Θjn−1 (∂Aj )
j=1
The main question now is whether or not Θn−1
(∂Aj ) are finite a.s. This is solely a property
j
of the field g. We need a generalization of Rice’s formula in order to respond to this question.
Theorem 3.2 ([4, Thm. 6.8 and Prop. 6.12]). Let Ω be a compact set in Rn with n ≥ 1. Let
z : Ω 7→ R be a Gaussian random field so that z ∈ C 2 (Ω) a.s. and Varz(t) > 0 for all t ∈ Ω.
For a fixed constant b ∈ R, E Θn−1 ({t ∈ Ω|z(t) = b}) < ∞ if the pair (z(t), ∇z(t)) have a
joint density on Ω × Rn that is locally bounded.
∂g
It is well known [1, Thm. 2.2.2] that for j ∈ {1, · · · , n} the processes gj (t) := ∂t
(t) are
j
themselves Gaussian processes with kernels κj (r, s) =
∂2
∂rj ∂sj κ(r, s)
whenever the second order
20
partial derivative of the kernel exists. Using this fact it is straightforward to check that our
choice of the field g satisfies the assumptions of the above theorem. Therefore, Θjn−1 (∂Aj ) are
finite a.s. and we conclude that V (u) < ∞ a.s.
We show two samples from the random field of (3.6) on the box [0, 1]2 in Figure 3(c–d)
with standard normal jumps. The choice of the field g influences the shape of the discontinuity
curves of u. The main difficulty in proving u ∈ BV (Ω) is in showing that the level sets of the
underlying Gaussian field have finite length. Theorem 3.2 allows us to relax the regularity
assumptions on g and take Gaussian fields that are in C 2 (Ω) rather than C ∞ (Ω) but for less
regular fields it is not clear whether the level sets have finite length. Finally, one can show
that the law of the process in (3.6) is a Radon measure on BV (Ω). The method of proof
is identical to our argument for the 1D case except that now we need a different version of
Helly’s selection principle [24, Thm. 5.5] in order to construct compact sets in BV (Ω).
4. Well-posed Bayesian inverse problems. Recall from Section 1 that we are interested
in the problem of inferring a parameter u ∈ X from data y ∈ Y that are related via a generic
stochastic mapping G̃ that models the physical process that generates the data as well as the
measurement noise:
y = G̃(u).
In order to solve this problem we employ Bayes’ rule (1.2). In this section we collect certain
conditions on the prior measure µ0 and the likelihood potential Φ that result in well-posed
inverse problems. We consider a general enough setting that encompasses the heavy-tailed
priors of Section 2 and the stochastic process priors of Section 3. We assume that the parameter space X is a Banach space that is not necessarily separable (such as BV ) and the prior
measure µ0 is possibly heavy-tailed (such as the Gp,q -priors) and not necessarily Radon (such
as the law of the pure jump Lévy processes when λ is not finite).
The main results of this section are Theorems 4.2 and 4.3 that establish the existence,
uniqueness and stability of the posterior measure. We acknowledge that these theorems are
very similar to the results in [22, Sec. 4.1] and [52]. In comparison to these articles, we
impose slightly different assumptions on the potential Φ and assume that the space X is
not necessarily separable. We also note that under the assumption that the prior measure
µ0 is a Radon measure one can immediately generalize the result of [22] to non-separable
parameter spaces X using the fact that Radon measures on a Banach space are automatically
concentrated on a separable subspace.
Theorem 4.1 ([9, Thm. 7.12.4]). Let µ be a Radon probability measure on a Banach space
X. Then, there exists a reflexive and separable Banach space E embedded in X such that
µ(X \ E) = 0 and the closed balls of E are compact in X.
It is important to note that while this theorem guarantees the existence of the separable
space E it does not provide us with a method for identifying E or its norm. In the case of the
product priors of Section 2 one can argue that the measures are concentrated on a separable
Hilbert space but for the stochastic process priors of Section 3 it is no longer clear what the
space E is and so it is more convenient for us to analyze the inverse problem on the ambient
space X rather than passing to the space E.
We will present our well-posedness results using the total variation metric, since this metric
is less often used in previous works, and refer the reader to [22] for proofs using the Hellinger
21
metric that can easily be generalized to our setting by comparison to the proofs using the
total variation metric. Given the inequalities (1.6) we immediately see that well-posedness in
one of these metrics implies well-posedness in the other but the convergence rates will differ.
We start by presenting minimal assumptions on the likelihood potential and the forward
map and make our way to more specific cases of inverse problems such as problems with
linear forward maps. In a nutshell, as we put more restrictions on Φ we are able to relax our
assumptions on µ0 . In order to help with navigation through this section we present Table 1
that collects our main results and the key underlying assumptions.
Theorem/Corollary
Main assumptions
Theorem 4.2
Φ is locally bounded and Lipschitz in u.
Theorem 4.3
Φ satisfies Assumption 1
Corollary 4.4
Corollary 4.6
Corollary 4.8
Corollary 4.9
Corollary 4.11
type of result
µy
is welldefined
µy
depends
continuously
on y
Φ has polynomial growth in u and µ0 has finitely many
moments
Φ ≥ 0 in addition to Assumption 1
Y = Rm , measurement noise is additive and Gaussian,
prior is ID
Y = Rm , forward map is linear and bounded, measurement noise is additive and Gaussian
Y = Rm , forward map is linear and bounded, measurement noise is additive and Gaussian, Gp,q -prior
well-posedness
well-posedness
well-posedness
well-posedness
well-posedness
Table 1: Summary of the key theorems and corollaries of Section 4. In each case we identify
the key underlying assumptions as well as the type of final result.
We begin by identifying some conditions on Φ that allow us to use a very large class of
prior measures including those that are heavy-tailed.
Assumption 1. Suppose that X and Y are Banach spaces and the likelihood potential Φ :
X × Y 7→ R satisfies the following properties:
(i) (Lower bound in u): There is a positive and non-decreasing function f1 : R+ 7→ [1, ∞)
so that ∀r > 0, there is a constant M (r) ∈ R such that,
Φ(u; y) ≥ M (r) − log (f1 (kukX )) ,
∀u ∈ X and ∀y ∈ BY (r).
(ii) (Boundedness above): ∀r > 0 there is a constant K(r) > 0 such that
Φ(u; y) ≤ K(r),
∀u ∈ BX (r) and ∀y ∈ BY (r).
(iii) (Continuity in u): ∀r > 0 there exists a constant L(r) > 0 such that
|Φ(u1 ; y) − Φ(u2 , y)| ≤ L(r)ku1 − u2 kX ,
∀u1 , u2 ∈ BX (r) and y ∈ BY (r).
(iv) (Continuity in y): There is a positive and non-decreasing function f2 : R+ 7→ R+ so that
∀r > 0, there is a constant C(r) ∈ R such that
|Φ(u; y1 ) − Φ(u, y2 )| ≤ C(r)f2 (kukX )ky1 − y2 kY ,
∀y1 , y2 ∈ BY (r) and ∀u ∈ X.
22
Our first task is to establish the existence and uniqueness of the posterior measure.
Theorem 4.2. Suppose X is a Banach space, µ0 is a Borel probability measure on X and
let Φ satisfy Assumptions 1 (i), (ii) and (iii) with a function f1 ≥ 1. If f1 (k · kX ) ∈ L1 (X, µ0 )
then the posterior µy given by (1.2) is a well-defined Borel probability measure on X. If µ0 is
Radon then so is µy .
Proof. Our proof will closely follow the approach of [52, Thm. 4.3] and [34, Thm. 2.2].
Assumption 1(iii) implies the continuity of Φ on X which in turn implies that Φ(·, y) : X 7→ R
is µ0 -measurable. We will now show that the normalizing constant satisfies 0 < Z(y) < ∞
which proves that µy is well-defined. The assertion that µy inherits the Radon property of µ0
will then follow from the absolute continuity of µy with respect to µ0 [9, Lem. 7.1.11].
Following Assumption 1(i) we can write
Z
Z
f1 (kukX )dµ0 (u) < ∞.
exp(log(f1 (kukX )) − M )dµ0 (u) = exp(−M )
Z(y) ≤
X
X
We now need to show that the normalizing constant Z(y) does not vanish. It follows from
Assumption 1 that for R > 0
Z
Z(y) ≥
exp(−K)dµ0 (u) = exp(−K)µ0 (BX (R)).
BX (R)
However, µ0 (BX (R)) > 0 for large enough R. To see this consider the disjoint
P sets Ak :=
{u|kS− 1 ≤ kukX < k} for k ∈ N. The Ak are open and hence measurable and ∞
k=1 µ0 (Ak ) =
∞
µ0 ( k=1 Ak ) = µ(X) = 1. Then the measure of at least one of the Ak has to be nonzero.
We now establish the stability of Bayesian inverse problems with respect to perturbations
in the data. Similar versions of the following theorems are available for Gaussian priors in
[51], for Besov priors in [21], for convex priors in [34] and for heavy-tailed priors on separable
Banach spaces in [22, 52].
Theorem 4.3. Suppose that X is a Banach space, µ0 is a Borel probability measure on X
0
and Φ satisfies Assumptions 1(i), (ii) and (iv) with functions f1 , f2 . Let µy and µy be two
measures defined via (1.2) for any y and y 0 ∈ Y , both absolutely continuous with respect to
µ0 .
(i) If f2 (k · kX )f1 (k · kX ) ∈ L1 (X, µ0 ) then ∀r > 0, there exists a constant C(r) > 0 so that
0
dT V (µy , µy ) ≤ Cky − y 0 kY , ∀y, y 0 ∈ BY (r)
(ii) If instead (f2 (k·kX ))2 f1 (k·kX ) ∈ L1 (X, µ0 ) then ∀r > 0 there exists a constant C 0 (r) > 0
0
so that dH (µy , µy ) ≤ C 0 ky − y 0 kY .
Proof. We will only prove (i) and refer the reader to [22, Sec. 4.1] for the proof of (ii) that
will readily generalize to our setting. Consider the normalizing constants Z(y) and Z(y 0 ).
We have already established in the proof of Theorem 4.2 that neither of these constants will
0
vanish and they are both bounded. Thus the measures µy and µy are well-defined. Applying
the mean value theorem to the exponential function and using Assumptions 1(i), (iv) and the
23
assumption that f2 (k · kX )f1 (k · kX ) ∈ L1 (X, µ0 ) we obtain
Z
0
exp(−Φ(u; y))|Φ(u; y) − Φ(u; y 0 )|dµ0 (u)
|Z(y) − Z(y )| ≤
X
Z
≤
exp(log(f1 (kukX )) − M )Cf2 (kukX )dµ0 (u) ky − y 0 kY
(4.1)
X
Z
≤ C exp(−M )
f1 (kukX )f2 (kukX )dµ0 (u) ky − y 0 kY . ky − y 0 kY .
X
Following the definition of the total variation distance we have
Z
y
y0
2dT V (µ , µ ) =
Z(y)−1 exp(−Φ(u; y)) − Z(y 0 )−1 exp(−Φ(u, y 0 )) dµ0 (u)
X
Z
≤
Z(y)−1 exp(−Φ(u; y)) − Z(y 0 )−1 exp(−Φ(u, y)) dµ0 (u)
X
Z
0 −1
+ Z(y )
exp(−Φ(u; y)) − exp(−Φ(u, y 0 )) dµ0 (u) =: I1 + I2 .
X
Now using (4.1) we have
I1 = |Z(y)−1 − Z(y 0 )−1 |Z(y) =
Z(y)
|Z(y 0 ) − Z(y)| . ky − y 0 kX .
Z(y 0 )
Furthermore, using the mean value theorem, Assumption 1 (i) and (iv) we can write
Z
0
exp(−Φ(u; y)) − exp(−Φ(u, y 0 )) dµ0 (u)
Z(y )I2 =
ZX
≤
exp(−Φ(u; y)) Φ(u; y 0 ) − Φ(u, y) dµ0 (u)
X
Z
≤ C exp(−M )
exp(log(f1 (kukX )))f2 (kukX )dµ0 (u) ky − y 0 kY . ky − y 0 kY .
X
The main distinction between the choice of the metrics in Theorem 4.3 is that in order
to obtain the same rate of convergence in the Hellinger metric we need a (possibly) stronger
assumption regarding the integrability of f1 (kukX ) and f2 (kukX ). So far we encountered
conditions of the form (f2 (kukX ))p f1 (kukX ) ∈ L1 (X, µ0 ) for p ∈ {0, 1, 2}. Intuitively, these
conditions identify the interplay between the growth of Φ(u; y) as a function of kukX and the
tail behavior of the prior µ0 .
Corollary 4.4. Suppose that Φ satisfies the conditions of Assumption 1 with f1 (t) = f2 (t) =
max{1, |t|p } for p ≥ 0 and µ0 is a Borel probability measure on X. If µ0 has bounded raw
moments of degree up to d2pe then the Bayesian inverse problem (1.2) is well-posed in both
the total variation and Hellinger metrics.
Corollary 4.5. Suppose that µ0 is a Borel probability measure on X and exp(bk · kX ) ∈
L1 (X, µ0 ) for some constant b > 0. Then the Bayesian inverse problem (1.2) is well-posed in
24
both the total variation and Hellinger metrics whenever Φ satisfies the conditions of Assumption 1 with functions f1 , f2 that are polynomially bounded.
For the remainder of this section we will focus on specific classes of likelihood potentials
Φ which allow us to further relax our assumption regarding the tail behavior of µ0 . The rest
of our results follow from Theorems 4.2 and 4.3 but they are of great interest in practical
applications. We start with the case of additive noise models and consider linear inverse
problems afterwards.
4.1. The case of additive noise models. Additive noise models have a special place in
practical applications due to their convenience and flexibility [36]. Suppose that the data is
finite dimensional and, without loss of generality, take Y = Rm , m ∈ N. Now suppose that
y ∈ Y is related to the parameter u ∈ X via the model
(4.2)
y = G(u) + η
η ∼ π(y)dΛ(y)
where π(y) is the Lebesgue density of the measurement noise η and G : X 7→ Rm is the forward
map. It is straightforward to check that under these assumptions
(4.3)
Φ(u; y) = − log π(G(u) − y).
In particular if η ∼ N (0, Σ ) with an m × m positive definite matrix Σ then
(4.4)
Φ(u; y) =
1
2
k(G(u) − y)kΣ
.
2
Now if log π(y) ≤ 0 (which is clearly the case when η is Gaussian or Laplace) then Φ(u; y) will
satisfy Assumption 1(i) with the constant M = 0 and f1 (x) = 1. This observation will allow
us to relax our assumption on the tail behavior of the prior whenever the measurement noise
is additive.
Corollary 4.6. Suppose Y = Rm , X is a Banach space and Φ(u; y) ≥ 0 and it satisfies
Assumptions (ii) and (iv) with a function f2 . Suppose that the prior measure µ0 is a Borel
0
probability measure on X and let µy and µy be two measures defined via (1.2) for y and
y 0 ∈ Y . Then the posterior measure µy is well-defined and
0
(i) If f2 (k · kX ) ∈ L1 (X, µ0 ) then ∀r > 0, ∃C(r) > 0 so that dT V (µy , µy ) ≤ C(r)ky −
y 0 kY for all y, y 0 ∈ BY (r).
0
(ii) If f2 (k · kX ) ∈ L2 (X, µ0 ) then ∃C 0 (r) > 0 dH (µy , µy ) ≤ C 0 (r)ky − y 0 kY .
At this point it is natural to identify conditions on the distribution of the noise and the
forward operator that guarantee that the likelihood potential of (4.3) satisfies the conditions
of Assumption 1. We will address this when η is Gaussian but our approach can be generalized
to other types of additive noise models.
Theorem 4.7. Consider the additive noise model of (4.2) when η ∼ N (0, Σ ) and Σ is a
positive-definite matrix. Then the corresponding likelihood potential Φ(u; y) ≥ 0. Furthermore,
Φ satisfies the conditions of Assumption 1(iv) with f2 (x) = 1 + f˜(x) if there is a positive, nondecreasing and locally bounded function f˜ : R+ 7→ R+ so that
(i) ∃C > 0 for which kG(u)k2 ≤ C f˜(kukX ), ∀u ∈ X.
(ii) ∀r > 0, ∃K(r) > 0 so that kG(u1 ) − G(u2 )k2 ≤ K(r)ku1 − u2 kX for all u1 , u2 ∈ BX (r).
25
Proof. Since we assumed that η is Gaussian then the likelihood potential is of the form
(4.4). Then it is clear that Φ(u; y) ≥ 0 which immediately implies that Φ satisfies Assumption
1(i) with M = 0 and f1 (x) = 0. Now fix r > 0 and suppose that u ∈ BX (r) and y ∈ BY (r).
Define r̃ = max{r, C f˜(r)} and note that r̃ is bounded since we assumed that f˜ is locally
2 + kyk2 . r̃ 2 and so Φ satisfies Assumption
bounded. Therefore, we have Φ(u; y) ≤ kG(u)kΣ
Σ
1(ii).
Now we will show that Φ satisfies Assumption 1(iii) as well. Let r and r̃ be defined as above
and consider u1 , u2 ∈ BX (r) and y ∈ BY (r). Using the identity kak22 − kbk22 = ha − b, a + bi
for a, b ∈ Rm and the conditions (i) and (ii) of the theorem we obtain
Σ−1/2 (G(u1 ) − G(u2 )), Σ −1/2 (G(u1 ) + G(u2 ) − 2y)i
2|Φ(u1 ; y) − Φ(u2 ; y)| = hΣ
2
≤ kG(u1 )kΣ + kG(u2 )kΣ + 2kykΣ
k(G(u1 ) − G(u2 ))kΣ
≤ C(r̃)k(G(u1 ) − G(u2 ))kΣ ≤ 2K(r)ku1 − u2 kX .
Finally, fix r > 0 and consider y1 , y2 ∈ BY (r). Then using the same line of reasoning as
above, for any u ∈ X we can write
Σ−1/2 (y2 − y1 ), Σ−1/2 (2G(u) − y1 − y2 )i
2|Φ(u; y1 ) − Φ(u; y2 )| = hΣ
≤ (ky2 kΣ − ky1 kΣ + 2kG(u)kΣ ) k(y2 − y1 )kΣ
≤ C(r)(1 + f˜(kukX ))ky1 − y2 k2 .
By putting this result together with Theorem 2.15 and Corollary 4.6 we deduce the following corollary concerning the well-posedness of Bayesian inverse problems with ID priors.
Corollary 4.8. Let X be a Banach space and Y = Rm . Consider the additive noise model:
y = G(u) + η,
η ∼ N (0, Σ )
where Σ is a positive definite matrix and G : X 7→ Rm satisfies the conditions of Theorem 4.7
with a submultiplicative function f˜. Also, suppose that µ0 = ID(m, R, λ) where λ is a Lévy
measure such that λ(X) < ∞, µ0 (X) = 1 and k · kX < ∞ µ0 -a.s. Then the Bayesian inverse
problem (1.2) is well-posed if 1 + f˜(k · kX ) ∈ L1 (X, λ).
4.2. The case of linear inverse problems. We now assume that the likelihood potential
Φ has the form
1
2
Φ(u; y) : X × Rm 7→ R,
Φ(u; y) = kG(u) − ykΣ
2
where Σ is a positive definite matrix and G : X 7→ Rm is bounded and linear. This case is
of particular importance due to its occurrence in the Compressed Sensing literature [26] and
estimation of sparse parameters. In this case, we can further relax our conditions on the prior
measure µ0 and achieve well-posedness so long as the prior µ0 has bounded first moment.
Corollary 4.9. Let X be a Banach space and Y = Rm . Suppose that the forward map
G : X 7→ Rm is bounded and linear and consider the additive noise model
y = G(u) + η
where
η ∼ N (0, Σ )
and
Σ is positive definite.
26
Then the Bayesian inverse problem of identifying the posterior µy via (1.2) is well-posed in
both the Hellinger and total variation metrics if the prior µ0 is a Borel probability measure on
X and k · kX ∈ L1 (X, µ0 ).
Proof. Follows directly from Theorems 4.7 and 4.2 and Corollary 4.6(i).
Let us now return to the product priors of Section 2.1 and show that those measures result
in well-posed Bayesian inverse problems under the linear and additive noise assumptions.
Theorem 4.10. Let X be a Banach space with an unconditional Schauder basis {xk } and
take Y = Rm . Suppose that the measurement noise is additive and Gaussian and the forward
map G is P
bounded and linear. Furthermore, suppose that µ0 is a product prior with sample
2
paths u = ∞
k=1 γk ξk xk where {γk } ∈ ` and {ξk } are i.i.d. and Varξk < ∞. Then the inverse
problem (1.2) is well-posed in both the total variation and Hellinger metrics.
Proof. The fact that µ0 is a Radon probability measures on X follows from Theorem 2.2
and Theorem 2.1. Now if Varξk < ∞ then E ξk2 < ∞ as well and so it follows from Theorem
2.3 that k · kX ∈ L2 (X, µ). Then the assertion follows from Theorems 4.6 and 4.7.
Finally we turn our attention to the the Gp,q -priors of Section 2. The proof of the following
corollary follows directly from Theorem 2.3 and the fact that the Gp,q distributions in 1D have
bounded variance for 0 < p, q ≤ 1.
Corollary 4.11. Let X be a Banach space with an unconditional Schauder basis {xk } and
Y = Rm . Suppose that the measurement noise is additive and Gaussian and that the forward
map G is bounded and linear. Then the Bayesian inverse problem (1.2) is well-posed in both
the Hellinger and total variation metrics if µ0 is an Gp,q -prior with 0 < p, q < 1.
5. Practical considerations and examples. We now turn our attention to practical aspects of solving an inverse problem within the Bayesian framework. In the first part of this
section we discuss the problem of approximating the posterior measure via approximation of
the likelihood potential. Afterwards, we will present three concrete examples of Bayesian inverse problems with heavy-tailed priors that arise from practical problems in image deblurring
and ultrasound therapy.
5.1. Consistent approximation of the posterior. Up to this point we were concerned
with identifying prior measures µ0 that result in a well-posed Bayesian inverse problem for a
given likelihood potential Φ. However, in practice we cannot solve the inverse problem directly
on the infinite-dimensional Banach space. Therefore, we need to obtain a finite dimensional
approximation to the posterior measure µy which is, in some sense, consistent with the infinite
dimensional limit.
To this end, we will define the notion of consistent approximation of a Bayesian inverse
problem in the context of applications where one would discretize (1.2) by approximating the
likelihood potential Φ with a discretized version ΦN : X × Y 7→ R, akin to a finite element
discretization. We define the approximation µyN to µy via
Z
dµyN
1
(5.1)
=
exp(−ΦN (u; y))
where
ZN (y) =
exp(−ΦN (u; y))dµ0 (u).
dµ0
ZN (y)
X
Definition 5.1 (Consistent approximation[34]). The approximate Bayesian inverse problem
(5.1) is a consistent approximation to (1.2) for a choice of µ0 , Φ and ΦN if d(µy , µyN ) 7→ 0 as
|Φ(u; y) − ΦN (u; y)| 7→ 0. Here, d is either the total variation or the Hellinger metric.
27
This notion of a consistent approximation relates directly to practical applications. Suppose, for example, that we are interested in computing the expected value of a quantity h(u)
under the posterior µy but we can only compute the expectation under the approximation
µyN . If µyN is a consistent approximation in the Hellinger metric then we have, by the bound
(1.7), that if h ∈ L2 (X, µy ) ∩ L2 (X, µyN ) then
Z
Z
h(u)dµyN (u) ≤ CdH (µy , µyN ).
h(u)dµy (u) −
X
X
In what follows, we will provide sharper bounds on the rate of convergence of the distances
between µy and µyN under mild conditions.
Theorem 5.2. Assume that the measures µy and µyN are defined via (1.2) and (5.1), for
a fixed y ∈ Y and all values of N , and are absolutely continuous with respect to the prior
µ0 which is a Borel probability measure on X. Also assume that both Φ and ΦN satisfy
Assumptions 1(i) and (ii) with an appropriate function f1 , uniformly for all N and that there
exists a positive and non-decreasing function f3 : R+ 7→ R+ so that
|Φ(u; y) − ΦN (u; y)| ≤ f3 (kukX )ρ(N )
(5.2)
where ρ(N ) 7→ 0 as N 7→ ∞.
(i) If f3 (k · kX )f1 (k · kX ) ∈ L1 (X, µ0 ) then there exists a constant D > 0, independent of N
such that dTV (µy , µyN ) ≤ Dρ(N ).
(ii) If (f3 (k · kX ))2 f1 (k · kX ) ∈ L1 (X, µ0 ) then there exists a constant D0 > 0, independent of
N such that dH (µy , µyN ) ≤ D0 ρ(N ).
Proof. Our method of proof uses similar arguments as in Theorem 4.3 and so we will only
present it briefly for the total variation distance. Proof of part (ii) can also be found in [22]
for separable Banach spaces. The existence and uniqueness of the measures µy and µyN follows
from Theorem 4.2 for all values of N . Next, the mean value theorem, Assumption 1(i), (5.2)
and the assumption that f3 (k · kX )f1 (k · kX ) is µ0 -integrable give
Z
exp(−Φ(u; y))|Φ(u; y) − ΦN (u; y)|dµ0 (u)
|Z(y) − ZN (y)| ≤
X
Z
≤
exp(log(f1 (kukX )) − M )Cf3 (kukX )dµ0 (u) ρ(N ) . ρ(N ).
X
Furthermore, we have
Z
0
2dT V (µy , µy ) ≤
Z(y)−1 exp(−Φ(u; y)) − ZN (y)−1 exp(−Φ(u, y)) dµ0 (u)
X
Z
+ ZN (y)−1
|exp(−Φ(u; y)) − exp(−ΦN (u, y))| dµ0 (u) =: I1 + I2 .
X
It then follows in a similar manner to proof of Theorem 4.3 that I1 . ρ(N ) and I2 . ρ(N )
which gives the desired result.
We now consider a more specific setting where the prior measure µ0 has a product structure. Suppose that the likelihood potential Φ satisfies the Assumption 1 with some functions
28
f1 , f2 . Also, assume that the space X has an unconditional Schauder basis {xk } and consider
the sequence of spaces (XN , k · kX ) where XN = span{xk }N
k=1 . These are linear subspaces of
⊥
X and for each N ∈ N we have X = XN ⊕ XN , meaning that every u ∈ X can be written as
⊥
⊥
u = uN + u⊥
N where uN ∈ XN and uN ∈ XN .
Suppose that the prior measure µ0 has the product structure of (2.1) and assume that it
has sufficiently fast decaying tails so that the posterior measure µy is well-defined. Observe
that for every value of N the product prior can be factored as
(5.3)
µ0 = µN ⊗ µ⊥
N
⊥
where µN and µ⊥
N are Radon measures on XN and XN . It is natural for us to discretize the
potential Φ using a projection operator:
(5.4)
ΦN (u; y) := Φ(PN u; y)
where PN : X 7→ XN is defined by PN (u) = uN . Next, define the approximate posterior
measures µyN as in (5.1) using the above definition of ΦN . Under these assumptions, the µyN
will factor as (see [34, Sec. 4.1] for the details)
(5.5)
where
µyN = νN ⊗ µ⊥
N,
1
dνN
=
exp(−Φ(PN u; y)).
dµN
ZN (y)
In other words, the likelihood potential is only informative on the subspace XN and so by
comparing (5.3) and (5.5) we see that the approximate posterior µyN differs from the prior
⊥ . As an example, we now check
only on this subspace and it is identical to the prior on XN
whether this method for discretization of the posterior results in a consistent approximation
to µy in the additive Gaussian noise case.
Theorem 5.3. Consider the above setting where the posterior and the prior have the prescribed product structures and the XN are linear subpaces of X. Suppose that Φand ΦN are
given by
1
1
Φ(u; y) = kG(u) − yk22 ,
ΦN (u; y) = kG(PN u) − yk22
2
2
where PN : X 7→ XN is the projection operator that was defined before. Assume that the
following conditions are satisfied:
(a) ku − PN ukX ≤ kukX ρ(N ).
(b) kG(u)k2 ≤ C f˜1 (kukX )
∀u ∈ X.
˜
(c) kG(u1 ) − G(u2 )k2 ≤ f2 (max{(ku1 kX , ku2 kX })ku1 − u2 kX
∀u1 , u2 ∈ X.
Here ρ is a positive function such that ρ(N ) 7→ 0 as N 7→ ∞ and the functions f˜1 , f˜2 : R+ 7→
R+ are non-decreasing and locally bounded and f˜1 ≥ 1. Then
(i) If f˜1 (k · kX )f˜2 (k · kX ) ∈ L1 (X, µ0 ) then ∃D > 0 independent of N so that dT V (µy , µyN ) ≤
Dρ(N ).
(ii) If f˜1 (k · kX )f˜2 (k · kX ) ∈ L2 (X, µ0 ) then ∃D0 > 0 independent of N so that dH (µy , µyN ) ≤
D0 ρ(N ).
29
Proof. It follows from Theorem 4.7 that Φ and ΦN satisfy Assumption 1 uniformly in N
with M = 0, f1 (x) = 1. Then the measures µy and µyN are well-defined for all values of N ∈ N
by Theorem 4.2. Now it follows from our assumptions on G that
2
2
2|Φ(u; y) − ΦN (u; y)| = k(G(u) − y)kΣ
− k(G(PN u) − y)kΣ
Σ−1/2 (G(u) − G(PN u)), Σ −1/2 (G(u) + G(PN u) − 2y)i
= hΣ
2
≤ kG(u)kΣ + kG(PN u)kΣ + 2kykΣ
k(G(u) − G(PN u))kΣ
≤ C f˜1 (kukX )f˜2 (kukX )ku − PN ukX .
The claim will now follow by taking f3 (x) = f˜1 (x)f˜2 (x) and applying Theorem 5.2.
A few comments are in order concerning the previous theorem. First, the function ρ(N )
is independent of the forward map and the prior and depends solely on the topology of X.
Then the rate of convergence of µyN to µy depends directly on the rate of convergence of PN
to the identity map in the operator norm. Also, observe that in order to achieve the same
rate of convergence in the Hellinger metric as in the total variation metric, we need to impose
stronger tail assumptions on the prior µ0 .
5.2. Example 2: Deconvolution. We now turn our attention to a few concrete examples
of inverse problems with heavy-tailed or non-Gaussian prior measures. We begin with a
problem in deconvolution which is a classic example of a linear inverse problem with wide
applications in optics and imaging [57, 30]. This problem was also considered in [34] as an
example problem with a convex prior measure.
Let X = L2 (T) where T is the circle of radius (2π)−1 and let Y = Rm for a fixed integer
m. Suppose that η ∼ N (0, σ 2 I) where σ ∈ R is a fixed constant and I is the m × m identity
matrix. Let S : C(T) 7→ Rm be a bounded linear operator that collects point values of a
continuous function on a set of m points over T. Given a fixed kernel ϕ ∈ C 1 (T), define the
forward map G : X → Y as
Z
(5.6)
G(u) = S(ϕ ∗ u)
where
(ϕ ∗ u)(t) :=
ϕ(t − s)u(t)dΛ(s).
T
Now suppose that the data y is generated via y = G(u) + η and our goal is to estimate the
original image u given noisy point values of its blurred version. Note that our assumptions so
far imply a quadratic likelihood potential of the form (4.4)
It follows from Young’s inequality [32, Thm. 13.8] that (ϕ ∗ ·) : L2 (T) 7→ L2 (T) is a
bounded linear operator and furthermore, (ϕ ∗ u) ∈ C 1 (T) for all u ∈ L2 (T). Since pointwise
evaluation is a bounded linear functional on C 1 (T) then the forward map G : L2 (T) 7→ Rm is
bounded and linear. We will use the results of Section 4.2 to show this problem is well-posed.
We will take our prior measure to be in class of the product priors of Section 2.1. Consider
the functions
(
1 0 ≤ t ≤ 1/2,
1 0 ≤ t ≤ 1,
w̃(t) =
and
ṽ(t) = 1 1/2 ≤ t ≤ 1,
0 otherwise.
0 otherwise.
30
The function ṽ is the Haar wavelet and w̃ is its corresponding scaling function. Following [23,
Sec. 9.3], we can define the periodic functions
wjn (t) :=
X
φ(2j (t + l) − n),
vjn (t) :=
l∈Z
X
ψ(2j (t + l) − n),
l∈Z
as well as the functions
x1 (t) = w0,0 (t),
x2 (t) = v0,0 (t),
x2j +nj +1 (t) = vj,nj (t).
for j ∈ Z+ and nj = {0, 1, · · · , 2j − 1}. The {xk } form an orthonormal basis for L2 (T) and so
they can be used in the construction of a Gp,q -prior.
Now choose p, q ∈ (0, 1) and take the prior µ0 to be the Gp,q -prior generated by the wavelet
basis {xk } and the fixed sequence {γk } where
γ2j +n+1 = (1 + |2j+1 |2 )−1/2
∀n ∈ Z+ .
Clearly, {γk } ∈ `2 and so it follows from Theorem 2.1 that kukL2 (T) < ∞ a.s. Furthermore, we
know that the Gp,q -priors have bounded moments of order two. Putting this together with the
fact that the forward map G is bounded and linear we immediately obtain the well-posedness
of this inverse problem using Theorem 4.10.
5.3. Example 3: Deconvolution with a BV prior. We now formulate the deconvolution
problem of Example 2 with a prior measure that is supported on BV (T) using the stochastic
process priors of Section 3. Let u(t) for t ∈ [0, 1] denote a stochastic process such that
u(0) = 0,
Z
ût (s) exp t exp(iξs) − 1 dν(ξ)
s ∈ R.
R
where the measure ν = cN (0, 1) with a fixed constant c ∈ (0, ∞). Then u is a compound
Poisson process with piecewise constant sample paths and normal jumps. We can write
Pτ (t)
u(t) = k=1 ξk , where {ξk } is an i.i.d. sequence of standard normal random variables and
τ (t) is a Poisson process with intensity c. In Section 3 we saw that this process has piecewise
constant sample paths and its law is a Radon measure on BV ([0, 1]).
Let us denote the law of this process by µ̃0 . The next step is to use this measure to
define a new measure µ0 on BV (T). Take µ0 to be the law of the periodic versions of the
sample paths of the above compound Poisson process u on the interval [0, 1]. We can write
µ0 = µ̃0 ◦ T −1 where T : BV ([0, 1]) 7→ BV (T) is a bounded and linear operator. Thus, µ0 is a
Radon measure on BV (T). With an abuse of notation we use u to denote the corresponding
periodic processes on T. Since the convolution kernel ϕ ∈ C 1 (T) and BV (T) ⊂ L1 (T) then
the forward map G : BV (T) 7→ Rm (given by(5.6)) is well-defined, bounded and linear and
so the likelihood potential has the form (4.4) once more. We have shown, in Section 3 that
E kukBV (T) < ∞. Putting this together with the fact that G : BV (T) 7→ Rm is bounded and
linear we immediately obtain the well-posedness of this inverse problem via Corollary 4.9.
31
5.4. Example 4: Quadratic measurements of a continuous field. As our final example,
we will consider a problem with a non-linear forward map. Our goal is to estimate a continuous
field from quadratic measurements of its point values. This inverse problem was encountered
in [33] in recovery of aberrations in high intensity focused ultrasound treatment and it is
closely related to the phase retrieval problem [25, 31, 15]. Let X = C(T) and let {tk }nk=1 be
a collection of distinct points in T. Now define the operator
S : C(T) 7→ Rn
(S(u))j = u(tj )
j = 1, 2, · · · , n.
This operator collects point values of functions in C(T). Let {zk }m
k=1 be a fixed collection of
vectors zk ∈ Rn and define the forward map
G : C(T) 7→ Rm ,
(G(u))j := |zjT S(u)|2
for
j = 1, 2, · · · , m,
which collects quadratic measurements of the point values of a continuous function. We
complete our model of the measurements with an additive layer of Gaussian noise
y = G(u) + η,
η ∼ N (0, σ 2 I),
where σ > 0. Our goal in this problem is to infer the function u ∈ C(T) from the quadratic
measurements y.
A straightforward calculation shows that
kG(u)k2 ≤ K̃kS(u)k22 ≤ Kkuk2C(T) ,
(5.7)
where K̃, K > 0 are constants that are independent of u but depend on the zk . The last
inequality follows because pointwise evaluation is a bounded linear operator on C(T).
Furthermore, we have that for u1 , u2 ∈ C(T)
(G(u1 ) − G(u2 ))j = (zjT (S(u1 − u2 )))(zjT (S(u1 ) + S(u2 )))
≤ Dj (max{ku1 kC(T) , ku2 kC(T) })ku1 − u2 kC(T) .
Here, the constant Dj > 0 depends on zj . We can now use this bound to obtain
(5.8)
kG(u1 ) − G(u2 )k2 ≤ D(max{ku1 kC(T) , ku2 kC(T) })ku1 − u2 kC(T)
where the constant D > 0 will only depend on the Dj . Observe that the above bounds in
(5.7) and (5.8) imply that G satisfies the conditions of Theorem 4.7 with a function f˜(x) = x2 .
Therefore, that theorem implies that the likelihood function Φ for our problem will satisfy
Assumption 1 (iv) with f2 (x) = 1 + x2 . Now we use Corollary 4.6 to infer that well-posedness
can be achieved if we choose a prior measure µ0 for which f2 (k·k) = 1+k·k2C(T) ∈ L1 (C(T), µ0 ).
In order to construct such a prior measure µ0 we will consider a product prior with samples
of the form
X
u∼
γ k ξ k wk
where
wk (t) = (2π)−1/2 exp(2πikt).
k∈Z
The {wk } are simply the Fourier basis on T. Our plan is to construct the prior measure to
be supported on a sufficiently regular Sobolev space that is embedded in C(T). The reason
32
for going through the Sobolev space is the fact that C(T) does not have an unconditional
Schauder basis and so we cannot directly apply the methodology of Section 2.1.
To this end, we choose
γk = (1 + |k|2 )−3/2
k ∈ Z,
and suppose that the {ξk } are i.i.d. and ξ1 ∼ CPois(0, Lap(0, 1)) (recall Definition 2.12), where
Lap(0, 1) is the standard Laplace distribution on the real line with Lebesgue density π(x) =
1
2 exp(−|x|) which clearly has exponential tails and this, in turn, implies that Varξ1 < ∞.
Note that the random variables ξk have a positive probability of being zero and hence draws
from this prior will incorporate a certain level of sparsity. Observe that this is a different type
of sparsity in comparison to the Gp,q -prior. Samples from this compound Poisson prior have a
non-zero probability of having modes that are exactly zero. The samples from the Gp,q -prior
have a zero probability of having modes that are exactly zero and instead most of their modes
will concentrate in a neighborhood of zero.
The Sobolev space H 1 (T) is defined as
X
(1 + |k|2 )|hv, wk i|2 < ∞}
H 1 (T) := {v ∈ L2 (T) : kvk2H 1 (T) :=
k∈Z
where h·, ·i is the usual L2 (T) inner product. Now consider u ∼ µ0 then
X
kuk2H 1 (T) =
(1 + |k|2 )−1 |ξk |2 .
k∈Z
But {(1 + |k|2 )−1 } ∈ `1 and Var|ξk |2 < ∞ and so it follows from Theorem 2.1 that kuk2H 1 (T) <
∞ a.s. Now the Sobolev embedding theorem [53, Prop. 3.3] guarantees that kukC(T) < ∞ a.s.
and it follows from Theorem 2.3 that kuk2C(T) ∈ L1 (C(T), µ0 ).
6. Closing remarks. At the beginning of this article we set out to achieve four goals. We
introduced the new classes Gp,q , Wp and `p -prior measures in connection with `p regularization techniques (G1) and showed that these prior measures belong to the larger class of ID
measures. This motivated our study of the ID class as priors (G2). Afterwards, we introduced
another class of prior measures that were based on the laws of pure jump Lévy processes
(G3). Our goal here was to construct a well-defined alternative to the classic total variation
prior. Finally, we presented a theory of well-posedness for Bayesian inverse problems that was
general enough that it covered the new classes of prior measures that we had introduced (G4).
Our approach to well-posedness theory was to identify the minimal restrictions on the prior
measure given a choice of the likelihood potential Φ. A common theme in our results was
the trade-off between the tail decay of the prior and the growth of the likelihood potential.
As an example, we considered the setting where the likelihood had a quadratic form and the
forward map was linear. This example corresponds to linear inverse problems with additive
Gaussian noise that are of great interest in practice. We showed that in this simple setting
well-posedness can be achieved if the prior has moments of order one.
Finally, we considered some practical aspects of solving inverse problems with heavytailed or ID priors. We discussed consistent discretization of inverse problems and the use
of projections in discretization of the likelihood. Afterwards, we presented three concrete
33
examples of inverse problems that used heavy-tailed or ID prior measures. In particular, we
studied the well-posedness of a deconvolution problem with a Lévy process prior that was cast
on the non-separable space BV (T).
The results of this article open the door for the use of large classes of prior measures in
inverse problems and they can be extended in several directions. For example, we showed
that if the forward problem is linear and the measurement noise is Gaussian then one can
achieve well-posedness for priors that have poor tail behavior. Then many of the common
heavy-tailed priors can be used to model sparsity in the linear case. But it is not clear
which prior is the optimal choice and in what sense. Furthermore, given that the Compressed
Sensing literature is mainly focused on recovery of sparse signals from linear measurements,
it is interesting to study the implications of the Compressed Sensing theory in the setting
of Bayesian inverse problems. Throughout the article we mentioned the issue of sparsity on
several occasions but this is not the only setting where non-Gaussian priors can be useful.
For example, non-Gaussian priors can be used in modelling of constraints or in construction
of hierarchical models. Finally, a major issue when it comes to using non-Gaussian priors in
practice is that of sampling. For example, even in finite dimensions, the Gp,q -priors are far
from a Gaussian measure. Then we expect that Metropolis-Hastings algorithms that utilize
a Gaussian proposal will have poor performance in sampling from posteriors that arise from
Gp,q -priors. This issue will become worse as the diemension of the parameter space grows.
Therefore, new sampling techniques that are tailor made to these non-Gaussian priors are
needed if we wish to apply them in real world situations.
Acknowledgements. The author owes a debt of gratitude to Prof. Nilima Nigam for many
useful discussions and comments. We are also thankful to the reviewers for their suggestions
that helped improve this manuscript significantly.
REFERENCES
[1] R. J. Adler. The Geometry of Random Fields. Number 62 in Classics in Applied Mathematics. SIAM,
Philadelphia, 2010.
[2] C. D. Aliprantis and K. Border. Infinite dimensional analysis: a hitchhiker’s guide. Springer Science &
Business Media, New York, third edition, 2006.
[3] D. Applebaum. Lévy processes and stochastic calculus. Number 93 in Cambridge studies in advanced
mathematics. Cambridge University Press, Cambridge, 2009.
[4] J. M. Azaı̈s and M. Wschebor. Level sets and extrema of random processes and fields. John Wiley &
Sons, New Jersey, 2009.
[5] J. M. Bernardo and A. F. Smith. Bayesian theory. Wiley Series in Probability and Statistics. John Wiley
& Sons, New York, 2009.
[6] P. Billingsley. Probability and measure. Wiley Series in Probability and Mathematical Statistics. John
Wiley & Sons, New York, three edition, 2008.
[7] V. I. Bogachev. Gaussian measures, volume 62 of Mathematical Surveys and Monographs. American
Mathematical Society, Providence, 1998.
[8] V. I. Bogachev. Measure theory, volume 1. Springer, New York, 2007.
[9] V. I. Bogachev. Measure theory, volume 2. Springer, New York, 2007.
[10] L. Bondesson. A general result on infinite divisibility. The Annals of Probability, pages 965–979, 1979.
[11] C. Borell. Convex measures on locally convex spaces. Arkiv för Matematik, 12(1):239–252, 1974.
[12] G. Buttazzo, M. Giaquinta, and S. Hildebrandt. One-dimensional variational problems: an introduction.
34
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
Number 15 in Oxford Lecture Series in Mathematics and Its Applications. Oxford University Press,
Oxford, 1998.
D. Calvetti, J. P. Kaipio, and E. Somersalo. Inverse problems in the Bayesian framework. Inverse
Problems, 30(11):110301, 2014.
D. Calvetti and E. Somersalo. An introduction to Bayesian scientific computing: Ten lectures on subjective
computing, volume 2 of Surveys and Tutorials in the Applied Mathematical Sciences. Springer Science
& Business Media, New York, 2007.
E. J. Candes, T. Strohmer, and V. Voroninski. Phaselift: Exact and stable signal recovery from magnitude measurements via convex programming. Communications on Pure and Applied Mathematics,
66(8):1241–1274, 2013.
C. M. Carvalho, N. G. Polson, and J. G. Scott. The horseshoe estimator for sparse signals. Biometrika,
97(2):465–480, 2010.
I. Castillo, J. Schmidt-Hieber, and A. Van der Vaart. Bayesian linear regression with sparse priors. The
Annals of Statistics, 43(5):1986–2018, 2015.
I. Castillo and A. van der Vaart. Needles and straw in a haystack: Posterior concentration for possibly
sparse sequences. The Annals of Statistics, 40(4):2069–2101, 2012.
R. Cont and P. Tankov. Financial modelling with jump processes. Chapman & Hall/CRC Financial
mathematics series. CRC press LLC, New York, 2004.
S. L. Cotter, M. Dashti, and A. M. Stuart. Approximation of Bayesian inverse problems for PDEs. SIAM
Journal on Numerical Analysis, 48(1):322–345, 2010.
M. Dashti, S. Harris, and A. M. Stuart. Besov priors for Bayesian inverse problems. Inverse Problems
and Imaging, 6(2):183–200, 2012.
M. Dashti and A. M. Stuart. The Bayesian Approach to Inverse Problems. Springer International
Publishing, 2016.
I. Daubechies et al. Ten lectures on wavelets. Number 61 in CBMS-NSF Regional Conference Series in
Applied Mathematics. SIAM, Philadelphia, 1992.
L. C. Evans and R. F. Gariepy. Measure Theory and Fine Properties of Functions. Text Books in
Mathematics. CRC Press, New York, revised edition, 2015.
C. Fienup and J. Dainty. Phase retrieval and image reconstruction for astronomy. Image Recovery:
Theory and Application, pages 231–275, 1987.
S. Foucart and H. Rauhut. A mathematical introduction to compressive sensing. Applied and Numerical
Harmonic Analysis. Springer Sience & Business Media, New York, 2013.
R. G. Ghanem and P. D. Spanos. Stochastic finite elements: a spectral approach. Dover Publication Inc.,
New York, 2003.
P. Ghosh and A. Chakrabarti. Posterior concentration properties of a general class of shrinkage priors
around nearly black vectors. arXiv preprint at arXiv:1412.8161, 2014.
H. Hanche-Olsen and H. Holden. The Kolmogorov–Riesz compactness theorem. Expositiones Mathematicae, 28(4):385–394, 2010.
P. C. Hansen, J. G. Nagy, and D. P. O’leary. Deblurring images: matrices, spectra, and filtering. SIAM,
Philadelphia, 2006.
R. W. Harrison. Phase problem in crystallography. JOSA A, 10(5):1046–1055, 1993.
C. Heil. A basis theory primer: Expanded edition. Applied and Numerical Harmonic Analysis. Springer
Sicence & Business Media, New York, 2010.
B. Hosseini, C. Mougenot, S. Pichardo, E. Constanciel, J. M. Drake, and J. M. Stockie. A Bayesian
approach for energy-based estimation of acoustic aberrations in high intensity focused ultrasound
treatment. arXiv preprint arXiv:1602.08080, 2016.
B. Hosseini and N. Nigam. Well-posed Bayesian inverse problems: priors with exponential tails. 2016.
arXiv preprint at arxiv:1604.02575.
N. L. Johnson, S. Kotz, and N. Balakrishnan. Continuous univariate distributions, Volume 1: Models
and Applications. John Wiley & Sons, New York, second edition, 2002.
J. Kaipio and E. Somersalo. Statistical and computational inverse problems, volume 160 of Applied
Mathematical Sciences. Springer Sience & Business Media, New York, 2005.
S. Kotz, T. J. Kozubowski, and K. Podgorski. The Laplace distribution and generalizations: A revisit with
applications to communications, economics, engineering, and finance. Springer Science & Business
35
Media, New York, 2012.
[38] V. Kruglov. A note on infinitely divisible distributions. Theory of Probability & Its Applications,
15(2):319–324, 1970.
[39] M. Lassas and S. Siltanen. Can one use total variation prior for edge-preserving Bayesian inversion?
Inverse Problems, 20(5):1537, 2004.
[40] G. Leoni. A First Course in Sobolev spaces, volume 105 of Graduate Studies in Mathematics. 2009.
[41] W. Linde. Probability in Banach spaces: Stable and infinitely divisible distributions. John Wiley & Sons,
New York, 1986.
[42] F. Lucka. Bayesian inversion in biomedical imaging. PhD thesis, University of Muenster, december 2014.
[43] S. Nadarajah. The kotz-type distribution with applications. Statistics: A Journal of Theoretical and
Applied Statistics, 37(4):341–358, 2003.
[44] S. Nadarajah. A generalized normal distribution. Journal of Applied Statistics, 32(7):685–694, 2005.
[45] C. J. Paciorek. Nonstationary Gaussian processes for regression and spatial modelling. PhD thesis,
Carnegie Mellon University, May 2003.
[46] S. Peszat and J. Zabczyk. Stochastic partial differential equations with Lévy noise: An evolution equation
approach, volume 113 of Encyclopedia of Mathematics and its Applications. Cambridge University
Press, Cambridge, 2007.
[47] N. G. Polson and J. G. Scott. Shrink globally, act locally: Sparse Bayesian regularization and prediction.
Bayesian Statistics, 9:501–538, 2010.
[48] C. E. Rasmussen and W. C. K. I. Gaussian processes for machine learning. the MIT press, Cambridge,
2006.
[49] K.-i. Sato. Lévy processes and infinitely divisible distributions. Number 68 in Cambridge studies in
advanced mathematics. Cambridge university press, Cambridge, 1999.
[50] F. W. Steutel and K. Van Harn. Infinite divisibility of probability distributions on the real line. Pure and
Applied Mathematics. Marcel Dekker Inc., New York, 2003.
[51] A. M. Stuart. Inverse problems: a Bayesian perspective. Acta Numerica, 19:451–559, 2010.
[52] T. J. Sullivan. Well-posed bayesian inverse problems and heavy-tailed stable Banach space priors. 2016.
arXiv preprint at arxiv:1605.05898.
[53] M. E. Taylor. Partial Differential Equations I: Basic Theory, volume 115 of Applied Mathematical Sciences. Springer Science & Business Media, New York, second edition, 2011.
[54] M. Unser and P. Tafti. An introduction to sparse stochastic processes. Cambridge University Press,
Cambridge, 2013.
[55] M. Unser, P. Tafti, A. Amini, and H. Kirshner. A unified formulation of Gaussian vs. sparse stochastic
processes. Part II: Discrete-domain theory. IEEE Transactions on Information Theory, 60:3036–3051,
2011.
[56] M. Unser, P. Tafti, and Q. Sun. A unified formulation of Gaussian vs. sparse stochastic processes. Part
I: Continuous-domain theory. IEEE Transactions on Information Theory, 60:1945–1962, 2011.
[57] C. R. Vogel. Computational methods for inverse problems. SIAM, Philadelphia, 2002.
[58] Z. Yao, Z. Hu, and J. Li. A TV-Gaussian prior for infinite-dimensional Bayesian inverse problems and
its numerical implementations. Inverse Problems, 32(7):075006, 2016.
| 10math.ST
|
arXiv:1402.3532v2 [math.AC] 5 Jun 2014
REVERSE LEXICOGRAPHIC GRÖBNER BASES AND
STRONGLY KOSZUL TORIC RINGS
KAZUNORI MATSUDA AND HIDEFUMI OHSUGI
Abstract. Restuccia and Rinaldo proved that a standard graded K-algebra
K[x1 , . . . , xn ]/I is strongly Koszul if the reduced Gröbner basis of I with respect
to any reverse lexicographic order is quadratic. In this paper, we give a sufficient
condition for a toric ring K[A] to be strongly Koszul in terms of the reverse lexicographic Gröbner bases of its toric ideal IA . This is a partial extension of a result
given by Restuccia and Rinaldo. In addition, we show that any strongly Koszul
toric ring generated by squarefree monomials is compressed. Using this fact, we
show that our sufficient condition for K[A] to be strongly Koszul is both necessary
and sufficient when K[A] is generated by squarefree monomials.
Introduction
Herzog, Hibi, and Restuccia [9] introduced the notion of strongly Koszul algebras.
Let R be a standard graded K-algebra with the graded maximal ideal m. Then R is
said to be strongly Koszul if m admits a minimal system of generators u1 , . . . , un of
the same degree such that for any 1 ≤ i1 < · · · < ir ≤ n and for all j = 1, 2, . . . , r,
the colon ideal (ui1 , . . . , uij−1 ) : uij of R is generated by a subset of elements of
{u1 , . . . , un }. As inspired by this notion, Conca, Trung, and Valla [4] introduced the
notion of Koszul filtrations. A family F of ideals of R is called a Koszul filtration
if F satisfies (i) every I ∈ F is generated by linear forms; (ii) (0) and m are in
F ; and (iii) for each non-zero ideal I ∈ F , there exists J ∈ F with J ⊂ I such
that I/J is cyclic and J : I ∈ F . For example, if R is strongly Koszul, then
F = {(0)} ∪ {(ui1 , . . . , uir ) | 1 ≤ i1 < · · · < ir ≤ n, 1 ≤ r ≤ n} is a Koszul filtration
of R. The existence of a Koszul filtration of R is an effective sufficient condition for
R to be Koszul. Some classes of Koszul algebras which have special Koszul filtrations
have been studied, e.g., universally Koszul algebras [2] and initially Koszul algebras
[1]. On the other hand, it is important to characterize the Koszulness in terms of the
Gröbner bases of its defining ideal. It is a well-known fact that if R is G-quadratic
(i.e., its defining ideal has a quadratic Gröbner basis) then R is Koszul. Conca,
Rossi, and Valla [3] proved that, if R is initially Koszul, then R is G-quadratic.
Moreover, they and Blum gave a necessary and sufficient condition for R to be
initially Koszul in terms of initial ideals of toric ideals ([1, 3]).
Let A = {u1, . . . , un } be a set of monomials of the same degree in a polynomial ring
K[T ] = K[t1 , . . . , td ] in d variables over a field K. Then the toric ring K[A] ⊂ K[T ]
is a semigroup ring generated by the set A over K. Let K[X] = K[x1 , . . . , xn ] be
a polynomial ring in n variables over K. The toric ideal IA of K[A] is the kernel
of the surjective homomorphism π : K[X] −→ K[A] defined by π(xi ) = ui for each
1
1 ≤ i ≤pn. Then we have K[A] ≃ K[X]/IA . A toric ring K[A] is called compressed
[16] if in< (IA ) = in< (IA ) for any reverse lexicographic order <.
In this paper, we study Gröbner bases of toric ideals of strongly Koszul toric rings.
First, in Section 1, we give a sufficient condition for K[A] to be strongly Koszul in
terms of the Gröbner bases of IA (Theorem 1.2). We then have Corollary 1.3, i.e.,
if the reduced Gröbner basis of IA with respect to any reverse lexicographic order
is quadratic, then K[A] is strongly Koszul [15, Theorem 2.7]. On the other hand,
Examples 1.6 and 1.7 are counterexamples of [15, Conjecture 3.11] (i.e., counterexamples of the converse of Corollary 1.3). In Section 2, we discuss strongly Koszul
toric rings generated by squarefree monomials. We show that such toric rings are
compressed (Theorem 2.1). Using this fact, we show that the sufficient condition for
K[A] to be strongly Koszul in Theorem 1.2 is both necessary and sufficient when
the toric rings are generated by squarefree monomials (Theorem 2.3).
1. Gröbner bases and strong Koszulness
First, we give a sufficient condition for toric rings to be strongly Koszul in terms
of the reverse lexicographic Gröbner bases. We need the following lemma:
Lemma 1.1. Suppose that, for each 1 ≤ i < j ≤ n, there exists a monomial order
≺ such that, with respect to ≺, an arbitrary binomial g in the reduced Gröbner basis
of IA satisfies the following conditions:
(i) xi | in≺ (g) and xj 6 | in≺ (g) =⇒ g = xi xk − xj xℓ for some 1 ≤ k, ℓ ≤ n,
(ii) xj | in≺ (g) and xi 6 | in≺ (g) =⇒ g = xj xℓ − xi xk for some 1 ≤ k, ℓ ≤ n.
Then, K[A] is strongly Koszul.
Proof. Suppose that K[A] is not strongly Koszul. By [9, Proposition 1.4], there exists
a monomial uk1 · · · uks of a minimal set of generators of (ui ) ∩ (uj ) such that s ≥ 3.
Since uk1 · · · uks belongs to (ui ) ∩ (uj ), there exist binomials xk1 · · · xks − xi X α and
xk1 · · · xks −xj X β in IA . Let G be the reduced Gröbner basis of IA with respect to ≺.
Since xi X α − xj X β ∈ IA is reduced to 0 with respect to G, it follows that both xi X α
and xj X β are reduced to the same monomial m with respect to G. Suppose that
G
→ m and that xi divides in≺ (g). If xj divides
g ∈ G is used in the computation xi X α −
in≺ (g), then it follows that xk1 · · · xks − xi xj X γ belongs to IA . Thus, ui uj divides
uk1 · · · uks . This contradicts that uk1 · · · uks belongs to a minimal set of generators
of (ui ) ∩ (uj ). If xj does not divide in≺ (g), then g = xi xk − xj xℓ by assumption (i).
Hence, ui uk ∈ (ui ) ∩ (uj ) divides uk1 · · · uks . This contradicts that uk1 · · · uks belongs
to a minimal set of generators of (ui ) ∩ (uj ). Therefore, xi never appears in the
G
initial monomials of g ∈ G which are used in the computation xi X α −
→ m. Hence,
xi divides m. By the same argument, it follows that xj never appears in the initial
G
monomials of g ∈ G which are used in the computation xj X β −
→ m, and hence, xj
divides m. Thus, xi xj divides m, which means that ui uj divides uk1 · · · uks . This
contradicts that uk1 · · · uks belongs to a minimal set of generators of (ui ) ∩ (uj ).
2
Let G(I) denote the (unique) minimal set of monomial generators of a monomial
ideal I. Given an ordering xi1 < xi2 < · · · < xin of variables {x1 , . . . , xn }, let <rlex
denote the reverse lexicographic order induced by the ordering <.
Theorem 1.2. Suppose that, for each 1 ≤ i < j ≤ n, there exists an ordering xi1 <
xi2 < · · · < xin with {i1 , i2 } = {i, j}, such that any monomial in G(in<rlex (IA ))∩(xi2 )
is quadratic. Then, K[A] is strongly Koszul.
Proof. We may assume that xj < xi . By Lemma 1.1, it is enough to show that <rlex
satisfies conditions (i) and (ii) in Lemma 1.1. Let g be an arbitrary (irreducible)
binomial in the reduced Gröbner basis of IA with respect to <rlex .
Since xj is the smallest variable, xj does not divide in<rlex (g). Hence, <rlex satisfies condition (ii). Suppose that xi divides in<rlex (g). By the assumption for <,
deg(in<rlex (g)) = 2. Hence, g = xi xp − xq xr for some 1 ≤ p, q, r ≤ n. Since
xq xr <rlex xi xp , we have j ∈ {q, r}, and hence, <rlex satisfies condition (i).
As a corollary, in case of toric rings, we have a result of Restuccia and Rinaldo
[15, Theorem 2.7]:
Corollary 1.3. Suppose that the reduced Gröbner basis of IA is quadratic with
respect to any reverse lexicographic order. Then, K[A] is strongly Koszul.
−1
Example 1.4. Let K[An ] = K[s, t1 s, . . . , tn s, t−1
1 s, . . . , tn s]. Then, IAn is the kernel
of the surjective homomorphism π : K[X] −→ K[An ] defined by π(z) = s, π(xi ) =
ti s, and π(yi ) = t−1
i s. It is easy to see that K[An ] is isomorphic to
−1 −1
−1 −1
K[A±
G ] = K[s, t1 tn+1 s, . . . , tn tn+1 s, t1 tn+1 s, . . . , tn tn+1 s],
where A±
G is the centrally symmetric configuration [13] of AG associated with the star
graph G = K1,n with n + 1 vertices. By [13, Theorem 4.4], IAn is generated by F =
{xi yi − z 2 | i = 1, 2, . . . , n}. Then, the Buchberger criterion tells us that the set F ∪
{xi yi − xj yj | 1 ≤ i < j ≤ n} is a Gröbner basis of IAn with respect to any monomial
order (i.e., a universal Gröbner basis of IAn ). Thus, by Corollary 1.3, K[An ] is
strongly Koszul for all n ∈ N. Eliminating the variable z from F , by the same
−1
argument above, it follows that the toric ring K[Bn ] = K[t1 s, . . . , tn s, t−1
1 s, . . . , tn s]
is strongly Koszul for all n ∈ N. Note that K[Bn ] is isomorphic to some toric ring
generated by squarefree monomials.
Remark 1.5. A standard graded K-algebra R is said to be c-universally Koszul [6]
if the set of all ideals of R which are generated by subsets of the variables is a Koszul
filtration of R. Ene, Herzog, and Hibi proved that a toric ring K[A] is c-universally
Koszul if the reduced Gröbner basis of IA is quadratic with respect to any reverse
lexicographic order [6, Corollary 1.4]. However, it is known that a toric ring K[A] is
c-universally Koszul if and only if K[A] is strongly Koszul. See [14, Definition 7.2]
or [11, Lemma 3.18]. So, [6, Corollary 1.4] is equivalent to Corollary 1.3.
In Section 2, we will show that the converse of Theorem 1.2 holds when K[A] is
generated by squarefree monomials. However, the converse does not hold in general.
3
Example 1.6. It is known [9] that any Veronese subring of a polynomial ring is
strongly Koszul. Let K[A] be the fourth Veronese subring of K[t1 , t2 ], i.e., K[A] =
K[t41 , t31 t2 , t21 t22 , t1 t32 , t42 ]. Then IA is generated by the binomials
x3 x5 − x24 , x1 x3 − x22 , x23 − x2 x4 , x1 x5 − x2 x4 , x2 x3 − x1 x4 , x3 x4 − x2 x5 .
Let < be an ordering of variables such that xi1 < xi2 < xi3 < xi4 < xi5 with
{i1 , i2 } = {2, 4}. Since both x32 − x21 x4 and x34 − x2 x25 belong to IA , it is easy to see
that either x32 or x34 belongs to G(in<rlex (IA )) ∩ (xi2 ). Thus, IA does not satisfy the
hypothesis of Theorem 1.2.
On the other hand, the converse of Corollary 1.3 does not hold even if K[A] is
generated by squarefree monomials. Note that Examples 1.6 and 1.7 are counterexamples of [15, Conjecture 3.11].
Example 1.7. Let K[A] = K[t4 , t1 t4 , t2 t4 , t3 t4 , t1 t2 t4 , t2 t3 t4 , t1 t3 t4 , t1 t2 t3 t4 ], which is
the toric ring of the stable set polytope of the empty graph with three vertices.
Since any empty graph is trivially perfect (see also Example 2.2), K[A] is strongly
Koszul. See [10] for the details. The toric ideal IA is generated by the binomials
x1 x5 − x2 x3 , x1 x6 − x3 x4 , x1 x7 − x2 x4 , x5 x6 − x3 x8 , x6 x7 − x4 x8 , x5 x7 − x2 x8 ,
x1 x8 − x4 x5 , x2 x6 − x4 x5 , x3 x7 − x4 x5 .
Let < be an ordering x4 < x3 < x2 < x1 < x8 < x7 < x6 < x5 . Since, with respect
to <rlex, the initial monomial (i.e., the first monomial) of any quadratic binomial
above does not divide the initial monomial x2 x3 x8 of x4 x25 − x2 x3 x8 ∈ IA , we have
x2 x3 x8 ∈ G(in<rlex (IA )). Thus, the reduced Gröbner basis of IA with respect to <rlex
is not quadratic. Below, we show that Theorem 2.3 guarantees that IA satisfies the
hypothesis of Theorem 1.2. See also Example 2.2.
2. Strongly Koszul toric rings generated by squarefree monomials
In this section, we consider the case when K[A] is squarefree, i.e., K[A] is isomorphic to a semigroup ringp
generated by squarefree monomials. A toric ring K[A] is
called compressed [16] if in< (IA ) = in< (IA ) for any reverse lexicographic order <.
It is known that K[A] is normal if it is compressed.
Theorem 2.1. Suppose that K[A] is strongly Koszul. Then, the following conditions
are equivalent:
(i) K[A] is squarefree;
(ii) IA has no quadratic binomial of the form x2i − xj xk ;
(iii) K[A] is compressed.
In particular, any squarefree strongly Koszul toric ring is compressed.
Proof. First, (i) =⇒ (ii) is trivial. By [16, Theorem 2.4], we have (iii) =⇒ (i). Thus
it is enough to show (ii) =⇒ (iii).
Let K[A] be a strongly Koszul toric ring such that IA has no quadratic binomial of
the form x2i − xj xk . Suppose that an irreducible binomial f = x2i X α − xj X β belongs
to the reduced Gröbner basis of IA with respect to a reverse lexicographic order
4
<rlex and that xj is the smallest variable in f . Then, u2i U α belongs to (U α ) ∩ (uj ).
Since K[A] is strongly Koszul, by [9, Corollary 1.5], (U α ) ∩ (uj ) is generated by
the element in (U α ) ∩ (uj ) of degree ≤ deg(X α ) + 1. Hence, u2i U α is generated by
such elements. Thus, there exist binomials x2i X α − X α xk xℓ and x2i X α − xj X γ xℓ in
IA . Then, we have x2i − xk xℓ ∈ IA . By assumption, we have x2i − xk xℓ = 0, and
hence, i = k = ℓ. Thus, the binomial g = xi X α − xj X γ belongs to IA . Since xj is
the smallest variable in f , it follows that xi X α is the initial monomial of g. This
contradicts that f appears in the reduced Gröbner basis of IA with respect to <rlex.
Hence, K[A] is compressed.
Example 2.2. Let G be a simple graph on the vertex set V (G) = {1, . . . , d} with
the edge set E(G). A subset S ⊂ V (G) is said to be stable if {i, j} ∈
/ E(G)
Q for all
i, j ∈ S. For each stable set S of G, we define the monomial uS = td+1 i∈S ti in
K[t1 , . . . , td+1 ]. Then the toric ring K[QG ] generated by {uS | S is a stable set of G}
over a field K is called the toric ring of the stable set polytope of G. It is known that
• K[QG ] is compressed ⇐⇒ G is perfect ([12, Example 1.3 (c)], [8]).
• K[QG ] is strongly Koszul ⇐⇒ G is trivially perfect ([10, Theorem 5.1]).
Here, a graph G is said to be perfect if the size of maximal clique of GW equals to
the chromatic number of GW for any induced subgraph GW of G, and a graph G
is said to be trivially perfect if the size of maximal stable set of GW equals to the
number of maximal cliques of GW for any induced subgraph GW of G. (About the
standard terminologies of graph theory, see [5].) Since any trivially perfect graph is
perfect [7], these facts are consistent with Theorem 2.1. On the other hand, with
respect to some lexicographic order, the initial ideal of the toric ideal in Example
1.7 is not generated by squarefree monomials.
Using Theorem 2.1, we now show that the converse of Theorem 1.2 holds when
K[A] is squarefree:
Theorem 2.3. Suppose that K[A] is squarefree and strongly Koszul. Let 1 ≤ i <
j ≤ n, and let < be any ordering of variables satisfying
xj < xi < {xk | ui uk /uj ∈ K[A], k 6= j} < other variables.
Then, any monomial in G(in<rlex (IA )) ∩ (xi ) is quadratic.
Proof. Let G be the reduced Gröbner basis of IA with respect to <rlex . Suppose
that xi X α ∈ G(in<rlex (IA )) ∩ (xi ) is not quadratic. Then, there exists a binomial
g = xi X α − xj X β in G. Note that in<rlex (g) = xi X α is squarefree by Theorem 2.1.
Hence, X α is not divisible by xi . Moreover, since G is reduced, X α is not divisible
by xj .
Since g belongs to IA , it follows that ui U α = uj U β belongs to the ideal (ui ) ∩ (uj ).
Then, ui U α is generated by ui uk = uj uℓ ∈ (ui ) ∩ (uj ) for some 1 ≤ k, ℓ ≤ n. Thus,
there exist binomials xi X α − xi xk X γ and xi X α − xj xℓ X γ in IA . Then, we have
X α − xk X γ ∈ IA . If k ∈ {i, j}, then X α ∈ in<rlex (IA ). This contradicts xi X α ∈
G(in<rlex (IA )). Hence, k ∈
/ {i, j}. Then, 0 6= xi xk − xj xℓ ∈ IA and in<rlex (xi xk −
α
xj xℓ ) = xi xk . Since xi X is not divisible by xi xk , X α is not divisible by xk . In
particular, 0 6= X α − xk X γ ∈ IA and X α <rlex xk X γ . Since ui uk /uj = uℓ , xk belongs
5
to {xk | ui uk /uj ∈ K[A], k 6= j}. Thus, the smallest variable xm appearing in X α
belongs to {xk | ui uk /uj ∈ K[A], k 6= j}. Let um′ = ui um /uj . Then, xi xm − xj xm′
(6= 0) belongs to IA . Therefore, xi xm belongs to in<rlex (IA ) and divides xi X α , which
is a contradiction.
By Theorem 2.3, we can check whether a squarefree toric ring K[A] = K[u1 , . . . , un ]
is strongly Koszul by computing the reverse lexicographic Gröbner bases of IA at
most n(n − 1)/2 times.
Acknowledgement. This research was supported by the JST (Japan Science and
Technology Agency) CREST (Core Research for Evolutional Science and Technology) research project Harmony of Gröbner Bases and the Modern Industrial Society,
within the framework of the JST Mathematics Program “Alliance for Breakthrough
between Mathematics and Sciences.”
References
[1] S. Blum, Initially Koszul algebras, Beiträge Alg. Geom., 41 (2000), 455–467.
[2] A. Conca, Universally Koszul algebras, Math. Ann., 317 (2000), 329–346.
[3] A. Conca, M. E. Rossi and G. Valla, Gröbner flags and Gorenstein Algebras, Compositio
Math., 129 (2001), 95–121.
[4] A. Conca, N. V. Trung and G. Valla, Koszul property for points in project spaces, Math.
Scand., 89 (2001), 201–216.
[5] R. Diestel, Graph Theory, Fourth Edition, Graduate Texts in Mathematics, 173, Springer,
2010.
[6] V. Ene, J. Herzog and T. Hibi, Linear flags and Koszul filtrations, preprint.
arXiv:1312.2190v1[math.AC].
[7] M. C. Golumbic, Trivially perfect graphs, Discrete Math., 24 (1978), 105–107.
[8] J. Gouveia, P. A. Parrilo and R. R. Thomas, Theta bodies for polynomial ideals, SIAM J.
Optim., 20 (2010), 2097–2118.
[9] J. Herzog, T. Hibi and G. Restuccia, Strongly Koszul algebras, Math. Scand., 86 (2000),
161–178.
[10] K. Matsuda, Strong Koszulness of toric rings associated with stable set polytopes of trivially
perfect graphs, J. Algebra Appl., 13 (2014), 1350138 [11 pages].
[11] S. Murai, Free resolutions of lex ideals over a Koszul toric ring, Trans. Amer. Math. Soc. 363
(2011), 857–885.
[12] H. Ohsugi and T. Hibi, Convex polytopes all of whose reverse lexicographic initial ideals are
squarefree, Proceedings of Amer. Math. Soc., 129 (2001), No.9, 2541–2546.
[13] H. Ohsugi and T. Hibi, Centrally symmetric configurations of integer matrices, Nagoya Math.
J. (2013), to appear.
[14] I. Peeva, Infinite free resolutions over toric rings, in “Syzygies and Hilbert Functions” (I.
Peeva, Ed.), Lect. Notes Pure Appl. Math 254, Chapman & Hall/CRC, 2007, pp. 233–247.
[15] G. Restuccia and G. Rinaldo, On certain classes of degree reverse lexicographic Gröbner bases,
Int. Math. Forum (2007), no. 22, 1053–1068.
[16] S. Sullivant, Compressed polytopes and statistical disclosure limitation, Tohoku Mathematical
Journal 58 (2006), 433 – 445.
6
(Kazunori Matsuda) Department of Mathematics, College of Science, Rikkyo University, Toshima-ku, Tokyo 171-8501, Japan
E-mail address: [email protected]
(Hidefumi Ohsugi) Mathematical Sciences, Faculty of Science and Technology,
Kwansei Gakuin University, Sanda, Hyogo, 669-1337, Japan
E-mail address: [email protected]
7
| 0math.AC
|
Deep-Learnt Classification of Light Curves
A Mahabal∗ , K Sheth† , F Gieseke‡ , A Pai‡ , S G Djorgovski∗ , A J Drake§ , M J Graham∗ , and CSS/CRTS/PTF Teams
∗ Center
for Data-Driven Discovery, California Institute of Technology, Pasadena, CA, 91125
Institute of Technology Gandhinagar, Palaj, Gandhinagar, 382355, India
‡ Department of Computer Science, University of Copenhagen, Copenhagen, Denmark
§ Cahill Center for Astronomy and Astrophysics, California Institute of Technology, Pasadena, CA, 91125
E-mails: [email protected], [email protected], [email protected],
[email protected], [email protected], [email protected], [email protected]
arXiv:1709.06257v1 [astro-ph.IM] 19 Sep 2017
† Indian
Abstract—Astronomy light curves are sparse, gappy, and heteroscedastic. As a result standard time series methods regularly
used for financial and similar datasets are of little help and
astronomers are usually left to their own instruments and
techniques to classify light curves. A common approach is to
derive statistical features from the time series and to use machine
learning methods, generally supervised, to separate objects into
a few of the standard classes. In this work, we transform the
time series to two-dimensional light curve representations in
order to classify them using modern deep learning techniques.
In particular, we show that convolutional neural networks based
classifiers work well for broad characterization and classification.
We use labeled datasets of periodic variables from CRTS survey
and show how this opens doors for a quick classification of diverse
classes with several possible exciting extensions.
I. I NTRODUCTION
Astronomy has always boasted of big datasets. The data
holdings are getting even larger due to surveys that observe
hundreds of millions of sources hundreds of time. The observations are a time series of flux measurements called light
curves. The staple for discovery has been the flux variations
of individual astronomical objects as noted through such light
curves - that is where the science is. The large irregular
gaps in observing cadence makes classification challenging.
Traditionally statistical features have been derived from the
light curves in order to do follow-up classification (see, e.g.,
[1], [2], [3]). The features include standard statistical measures
like median, skew, kurtosis as well as specialized domain
knowledge based ones such as ‘fading profile of a single
peaked fast transient’. The standard features do not carry
special powers for classifying a varied set of objects. The
designer features are better for specific classes, but carry
with them a bias that does not necessarily translate to the
classification of a wider set.
In [4] we introduced a two-dimensional mapping of the
light curves based on the changes in magnitude (dm) over
the available time-differences (dt). In this work, we mold the
dm − dt mapping into an image format that is suitable as input
for convolutional neural networks (CNNs or ConvNets) [5].
By bringing to bear the machinery of CNNs we are able to
conjure a large number of features unimagined so far. We use
labeled sets to train the CNN as a classifier and following
validation we classify light curves from the Catalina RealTime Transient Survey (CRTS; [6], [7], [4], [8], [9]).
II. DATA
The three CRTS surveys span 33,000 sq. degrees encompassing light curves of close to half a billion sources. Of
these, the 0.7m CSS telescope yields ∼ 150 million light
curves. The light curves span well over ten years, and are
homogeneous in that all are collected using white-light without
a filter, and with an asteroid-searching cadence of four images
in 30 minutes. As is typical of the astronomical objects in
wide-area surveys a vast majority of these sources (> 90%)
are non-variable during the survey life-time and within the
typical ∼ 0.1 mag error-bars for CSS. The remaining sources variables - can be broadly classified as periodic and stochastic.
The irregularly spaced sparse light curves mean that often even
the periodically variable sources do not seem obviously so.
There is a third category, that of transients like supernovae
and flaring stars which exhibit enhanced activity over a short
period and otherwise a quiescent and relatively flat (within
error-bars) light curve.
Getting a large, uniform, well-labelled dataset is a challenge
in itself. In order to keep the problem simple during early
experiments, we start with a sample of ∼ 50k periodic variables from the CRTS North (CRTS-N) survey [8]. There are
17 classes represented in the sample. Ten of these have fewer
than 500 members. We exclude them from our experiments for
now and will include them in future studies. The numbers for
the remaining seven classes are given in Table I. These classifications have been carried out by humans mostly based on
calibrated light curves, their phased versions after periods were
determined, and some auxiliary information on the objects. A
little over 10% of these have spectroscopic confirmation of
the exact classification. As a result some misclassifications,
especially in a nearby class can not be ruled out, especially
for objects that are fainter and/or have fewer observations.
III. DMDT M APPINGS
A light curve consists of brightness variations as a function
of time. Besides the time (expressed here in days - MJD), and
Fig. 1. Part of a light curve is shown without error-bars to demonstrate dm
(dashed lines) and dt (dotted lines) values. Each pair of points in the light
curve leads to one dmdt pair. Three pairs are shown. These then populate the
dmdt grid of Fig. 2 and make the 23x24 dmdt-images in Figs. 3,4 etc.
Fig. 2. The dmdt grid associated with our fiducial dm and dt spacings. Most
labels near dt=0 and dm=0 are not printed due to crowding (see Table III
for the full list). Each unequal-area rectangle here translates to one of the
equal-area pixels in the 23x24 images used with CNNs (e.g. Figs. 3 and 4).
brightness (expressed here in the traditional inverse logarithmic scale - mags), we also have information about the error
in magnitudes.
For each pair of points in a light curve we determine
the difference in magnitude (dm) and the difference in
time (dt). This gives us p = n2 = n ∗ (n − 1)/2 points
for a light curve of length n (see Fig. 1). These points
are then binned for a range of dm and dt values. The
resulting binned 2D representation is our 2D mapping from
the light curve. The bin boundaries we have used are:
dm = ±[0, 0.1, 0.2, 0.3, 0.5, 1, 1.5, 2, 2.5, 3, 5, 8] mags and dt =
[1/145, 2/145, 3/145, 4/145, 1/25, 2/25, 3/25, 1.5, 2.5, 3.5,
4.5, 5.5, 7, 10, 20, 30, 60, 90, 120, 240, 600, 960, 2000, 4000]
days. The 23 × 24 bins are in approximately a semilogarithmic fashion so that frequent small magnitude
changes are distributed over many bins, and the infrequent
large magnitude changes can be combined together (see
Fig. 2). Similarly the more important rapid changes are well
represented, and the slower changes are clubbed together.
This is akin to histogram equalization, but not forced to be
exact. In case of dt these take into consideration the CSS
cadence of four images in 30 minutes. The image intensity,
i, is normalized by p to account for varying lengths of
original light curves, and stretched to fit between 0 and 255:
i = (255 ∗ nbin /p + 0.99999)int . Thus, a bin that does not
include a single point, now has an intensity of zero, and a
bin that had at least one point has an intensity of at least
1, and the bin that had the maximum number of points has
an intensity of at most 255. 255 is reached only when all
points are in one bin - the typical max values we have seen
in different classes are around 50. Unnormalized values go
to several thousand based on the length of the light curve.
The normalization also ensures that we can use our data to
fine-tune other pre-learned networks. For inclusion in the
training set, we require that the light curves contain at least
20 points.
The 2D representations - called dmdt-images hereafter reflect the underlying structure from variability of the source.
The dmdt-images are translation independent as they consider
only the differences in time. A light curve reflected about
the x-axis will provide a dmdt-image reflected about the yaxis. Thus the structure above and below the dm = 0 line is
discriminating for sources. In a sense the dmdt-images are like
a structure function without consideration to the error-bars.
The error-bars tend to be heteroskedastic, but are broadly a
function of magnitude, and unless an individual source varies
a lot, tend to be similar. In particular, the way the dmdttechnique works, the error-bars for neighboring points in the
mapped version are similar. Taking into consideration the
error-bars would be equivalent to smoothing along the y-axis
(dm). For now we ignore the error-bars (but see Section V).
Since the dmdt-image is a straightforward mapping of a
light curve and the CNN can bring out features hidden therein,
the proposed method opens up at least two important avenues.
(1) Implication for real-time classification of variables and
transients: A sparse light curve of recently discovered object
will correspond to a dmdt-image that is just a sparse version of
the dmdt-image that would be formed from the corresponding
non-sparse light curve. Since some of the unique features
are accentuated at discovery, the discriminative power would
already be encapsulated in the dmdt-image. (2) Transfer
learning: Training based on dmdt-images from one survey
can be used to classify dmdt-images from another survey. We
demonstrate this on a set of variables from CRTS-S a Southern
set corresponding to the main training set used here [10], and
corresponding PTF data [11].
There are three ways in which we experiment with the
setup to improve performance: (1) Change the dmdt bins for
optimality - these can be done based on the survey cadence, or
based on the classes being considered, (2) Change the layers
of the CNN depending on number of classes, size of training
sample, possible ways in which unbalancedness between the
TABLE I
N UMBER OF OBJECTS BELONGING TO THE SEVEN CLASSES THAT HAVE AT
LEAST 500 MEMBERS . T HE VARIABLE TYPES INCLUDE EW ( CONTACT
BINARIES ), EA ( DETACHED BINARIES ), THREE TYPES OF RR LYRAE ; AND
M IRA AND SEMI - REGULARS LUMPED INTO LPV. RS CV N ’ S ARE
ROTATING VARIABLES . T HUS BROADLY SPEAKING WE HAVE THREE
CLASSES : BINARIES , PULSATING , AND ROTATING . Class REFERS TO THE
NUMERIC LABELS USED IN [12].
Type
Class
Num
EW
1
30743
EA
2
4683
RRab
4
2420
RRc
5
5469
RRd
6
502
RSCVn
8
1522
(a) EW
LPV
13
512
(b) EA
classes is remedied (or not), and (3) modifying the light curve
to dmdt-image mapping to bring out features in the classes
being separated. We look at these in the next two sections.
IV. C ONVOLUTIONAL N ETWORKS
Recently, so-called deep learning techniques have become
very popular in machine learning and various application
domains. This is, in particular, the case for convolutional
neural networks (CNNs), which are a special type of artificial
neural networks (ANN) [13], [14]. In astronomy, CNNs have
been used for a few problems where structure in images is
obvious (e.g. [15] uses them for classifying galaxies based on
morphology and [16] for detecting supernovae). We provide
brief description of ANNs and CNNs here. More details can
be found in [13], [14], [5] etc.
An ANN is based on several layers. The overall input data
(e.g., images) are provided to the input layer. The output of
one layer serves as input for the next. The last layer forms the
output of the network. For instance, in a multiclass setting,
the output layer would contain one node per class. The layers
between the input and the output layer are the hidden layers.
For a standard ANN, the layers are fully connected, with each
node of one layer connected to all nodes of the next.
A CNN is an extension of classical ANNs and consists of
different types of layers. As before, we have an input and an
output layer. Further, the last layers before the output layer
are often standard fully connected layers, called dense layers.
The layers before these dense layers, however, are usually very
different from those of a standard ANN.
The most prominent types of layers are (a) convolutional
layers, (b) pooling layers, and (c) dropout layers: A convolutional layer usually consists of a small set of filters (e.g., 3x3,
5x5, . . . ), called kernels, and convolves every input image with
each of those kernels. Typically several such kernels are used
as filters in a given convolutional layer to match desired shapes
in the input images. This gives rise, for each kernel, to a new
representation of the input image. These representations are
called feature maps. The convolutional layers are used with
rectifiers to introduce non-linearity. A pooling layer decreases
the number of parameters of the network by aggregating
spatially-adjacent pixel values. One prominent type of pooling
layers is max-pooling that replaces patches of an input feature
map by the maximum value within each patch. While this
reduces the sizes of the feature maps, it also makes the network
more robust to small changes in the input data. Finally, a
(c) RRab
(d) RRc
(f) RS CVn
(e) RRd
(g) LPV
Fig. 3. Composite dmdt images for all classes obtained by stacking all
individual dmdt images of each class. Given that we can visually discriminate
between them, it should be clear that with purer base training samples image
based classifiers will be able to classify them easily.
dropout layer randomly omits hidden units by setting their
values to zero. Hence, the network cannot fully rely on them.
This helps prevent overfitting. One usually resorts to one or
more final dense layers based on a large number of nodes
connected to the previous layer. Such layers are unlike the
convolutional, pooling, and dropout layers, but the same as
for traditional neural networks. It is the depth provided by
these multiple layers, and the extensive mapping afforded by
them that has given rise to the name deep learning.
For approaching the classification task at hand, we considered a multi-layer CNN instance, called deep network.
Selecting good layers and parameters is, as yet, more an
art than science. An interesting research direction is the
use of Bayesian optimization for choosing the best network
hyperparameters. As a first step towards this, we considered a far simpler shallow network and were pleasantly
surprised by its equally good performance for the base allclass classifications compared to the performance of the deep
network. The code detailing the structure of both networks
is provided in Listing 1 and Listing 2, respectively. The
listings include type and size of layers, number and size of
kernels, and dropout fractions. We have used the theano framework (http://deeplearning.net/software/theano/) with lasagne
(https://lasagne.readthedocs.io/en/latest/) for our runs.
V. E XPERIMENTAL E VALUATION
Light curves were converted to dmdt-images, normalized
as described in Sec. II, and used as inputs to the network.
TABLE II
ACCURACIES FOR THE MULTICLASS CLASSIFICATION ( FIRST ROW ) AND
RECALL FOR THE BINARY CLASSIFICATIONS ( ALL OTHER ROWS ) ARE
SHOWN FOR BOTH SHALLOW AND DEEP CNN ON THE FIDUCIAL
dmdt- IMAGES . F1- SCORE AND M ATTHEW ’ S C OEFFICIENT ARE ALSO
PROVIDED FOR THE DEEP CNN ( BINARY CLASSIFICATIONS ).
C OMPARISON WITH RANDOM FORESTS WITH FEATURES IS IN THE LAST
COLUMN . T HOUGH NOT NORMALLY DONE FOR DEEP LEARNING , WE DID 5
RANDOM TRAIN - TEST SPLITS AND REPORT THE RANGE FOR THE A LL 7
NETWORKS AND RF. S UCH A PROCEDURE MAY MAKE MORE SENSE FOR
SMALLER DATASETS .
Classes
All7
1/2
1/4
1/5
1/6
1/8
1/13
2/4
2/5
2/6
2/8
2/13
4/5
4/6
4/8
4/13
5/6
5/8
5/13
6/8
6/13
8/13
CNNshallow
Recall
83.3 ± 0.5
97/76
99/67
97/57
99/35
99/23
100/86
97/96
96/99
99/98
97/92
100/96
7/88
93/71
92/91
98/94
99/9
96/77
100/88
83/93
100/97
98/97
Recall
83.2 ± 0.3
98/77.2
98.7/66.3
96.5/54.4
99.7/31
99.6/20
99.9/88.9
97.9/97.8
97.3/99.4
99.6/97.7
98.2/89.7
100/99
58.6/94.3
93.1/73.3
95.1/84.9
98.9/93.5
97.6/12.37
96.1/74.5
99.8/96.8
77.5/90.7
98.9/98.1
99.3/93
CNNdeep
F1-score
0.97/0.81
0.98/0.72
0.94/0.63
0.99/0.41
0.98/0.31
0.99/0.93
0.98/0.97
0.98/0.98
1/0.97
0.98/0.92
1/1
0.69/0.88
0.94/0.7
0.93/0.88
0.99/0.94
0.95/0.18
0.94/0.76
1/0.97
0.74/0.92
0.98/0.99
0.99/0.95
MCC
0.79
0.71
0.58
0.44
0.37
0.93
0.95
0.97
0.97
0.89
0.99
0.59
0.64
0.81
0.93
0.15
0.70
0.97
0.66
0.97
0.94
RF
Recall
82.8 ± 0.5
97/82
99/63
97/54
99/31
99/0
99/79
98/97
98/99
99/96
97/92
99/96
66/95
94/55
93/88
98/84
99/21
94/77
99/89
76/96
96/91
98/91
We used 500 training epochs for the deep network and 300
for the shallow network with 20% samples reserved for testing
for both configurations. We used a learning rate of 0.0002 and
the Adaptive Momentum (ADAM) algorithm to train all our
models. All our models have been trained on a single NVIDIA
GeForce GTX 560 graphics processing unit (GPU). It takes
∼ 5.5 and 42.3 seconds per epoch for the shallow and deep
networks respectively when trained on the CRTS-N training
set. Training RFs is one to two orders of magnitude faster after computing the features. Depending on how complex the
features are, the computing time can vary a lot. We used Red
Hat linux release 6.6, python 2.7.2.
A. Binary classification
We trained the network with pairs of classes as well as
with all seven classes together. When used in binary mode we
noticed poor performance when class 1 is involved (see Col. 2
of Table II). It is not unexpected since class 1 contains twothirds of all objects, and when paired with an individual class,
it overwhelms every other class easily. Class 13 reached an
accuracy of 89% - the highest - against class 1. These are the
Long period variables (LPVs) and the long-term structure is
likely getting picked up. In general the separation of all other
classes with class 13 was similarly far better than other binary
comparisons. Except in a few cases, the binary dmdt-classifier
did comparable or better than the corresponding feature-based
random forests (RF) classifier (see Table II and Sec. VI-A).
TABLE III
E XPERIMENTS WITH VARYING DMDT
Original binning (as outlined before):
dm = ±[0, 0.1, 0.2, 0.3, 0.5, 1, 1.5, 2, 2.5, 3, 5, 8] mags
dt = [1/145, 2/145, 3/145, 4/145, 1/25, 2/25, 3/25, 1.5, 2.5, 3.5, 4.5,
5.5, 7, 10, 20, 30, 60, 90, 120, 240, 600, 960, 2000, 4000] days
Average accuracy: 83%
New 1:
dm = ±[0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 1, 1.5, 2, 2.5, 3, 5, 8] mags
dt = [1/145, 2/145, 3/145, 4/145, 1/25, 2/25, 3/25, 1.5, 2.5, 3.5, 4.5,
5.5, 7, 10, 20, 30, 60, 90, 120, 240, 600, 960, 2000, 4000] days
Average accuracy: 84.5%
New 2:
dm = ±[0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 1, 1.5, 2, 2.5, 3, 5, 8] mags
dt = [1/145, 2/145, 3/145, 4/145, 1.5, 2.5, 3.5, 4.5, 5.5, 7, 20,
60, 120, 600, 960, 4000] days
Average accuracy: 84.3%
New 3:
dm = ±[0, 0.1, 0.2, 0.3, 0.5, 1, 1.5, 2, 2.5, 3, 5, 8] mags
dt = [1/145, 2/145, 3/145, 4/145, 1.5, 2.5, 3.5, 4.5, 5.5, 7, 20,
60, 120, 600, 960, 4000] days
Average accuracy: 82.7%
Listing 1. Convolutional Neural Network - deep
layers = [
InputLayer,
Conv2DLayer(64, size:3x3, rectify),
MaxPool2DLayer(2x2)),
DropoutLayer(0.1),
Conv2DLayer(128, size:5x5, rectify),
Conv2DLayer(256, size:5x5, rectify),
DenseLayer(512),
DropoutLayer(0.5),
DenseLayer(512),
DenseLayer(all, softmax),
]
Listing 2. Convolutional Neural Network - shallow network
layers = [
InputLayer,
Conv2DLayer(32, size:3x3, rectify),
DropoutLayer(0.1),
DenseLayer(128),
DropoutLayer(0.25),
DenseLayer(128),
DenseLayer(all, softmax),
]
B. Multi-class classification
When used in the 7-class mode, the dmdt-images produced
an average accuracy of 83%. This is remarkable in itself
given the sparse nature of the data and no fine tuning of the
parameters. The performance is comparable with feature-based
RF-classifier (see Table II last row, and Figs. 6 and 7). Class
1 still dominates to an extent but not as blatantly as in the
binary cases. It still leaves a lot to be desired if one wishes to
use it in real-time for light curves containing far fewer points,
and binary classification may be somewhat preferable.
TABLE IV
ACCURACY ( ROW 1) AND RECALL ( OTHER ROWS ) FOR MODELS TRAINED
USING THE SHALLOW NETWORK ON BACKGROUND SUBTRACTED IMAGES .
( A ) CNNc : PER CLASS BACKGROUND , ( B ) CNN: PSEUDO - CADENCE
BACKGROUND , ( C ) CNN pmax : MAX CADENCE BACKGROUND ESTIMATED
FROM THE TRAINING DATASET SUBTRACTED FROM EACH dmdt IMAGE IN
THE DATASET,( D ) CNNe : SMALL dm VALUES ELIMINATED ( HERE 6 ROWS
FROM OUR dmdt MODEL ).
Classes
All7
1/2
1/4
1/5
1/6
1/8
1/13
2/4
2/5
2/6
2/8
2/13
4/5
4/6
4/8
4/13
5/6
5/8
5/13
6/8
6/13
8/13
CNNc
99.1
99/100
99/100
99/98
99/99
100/97
100/94
100/100
99/100
100/98
99/99
100/100
100/100
100/100
99/100
100/98
100/98
99/100
99/100
100/100
100/100
99/100
CNN p
83.2
98/75
98/69
98/55
99/37
99/20
100/88
98/96
95/99
99/98
98/91
100/97
71/91
91/77
92/91
99/93
98/16
95/78
100/94
79/93
100/97
98/97
CNN pmax
83.1
98/76
98/70
97/57
99/33
99/24
100/88
98/96
95/99
99/97
98/90
100/96
68/91
96/52
94/90
98/93
99/10
95/80
99/95
82/92
100/97
98/100
CNNe
75.3
98/48
98/55
98/24
100/0
100/0
99/82
96/85
86/87
97/64
88/73
99/91
65/91
94/56
91/83
97/87
98/16
91/62
99/91
62/95
97/97
98/93
A dmdt-image can be said to be made of three components
(1) a static background, b, that results primarily from the
cadence of the survey. No matter what kind of astronomical
object there is, one will always find pairs of observations with
large dt and small dm, and peaks at specific dt depending on
the survey cadence, (2) a more specific background related to
the class-membership, ci , for the ith class. This is the kind of
dmdt one would get from a densely sampled prototype, and
(3) something like an individual signature, s, for each object.
dmdt–image = b + ci + s
We formulate the foreground-background separation problem by drawing parallels to video surveillance tasks. We
interpret each dmdt-image as a video frame, vectorize it and
then stack up all the dmdt-images columnwise to create a
matrix M. We then decompose this matrix into a sum of a
low rank matrix, L, and a sparse matrix, S. Each column of
the low rank matrix corresponds to the background in the
corresponding dmdt-image and each column of the sparse
matrix corresponds to the foreground of the corresponding
dmdt-image. L and S can be obtained by solving the following
expression where L is forced to be low-rank and S is forced
to be sparse:
MinkM − L − Sk2
L,S
C. Varying dmdt binning
Input image size of 23x24 is small for CNNs and the
training is relatively quick. As a result one can consider
finer binning in both dm and in dt. On the other hand, the
discriminating structure likely resides in a few smaller areas,
and one could use more granular binning. While it is desirable
to determine the binning for a given survey and classes under
consideration, such a systematic approach will need more
extensive work. Here we report some quick experiments. We
give below the variations we tried and the corresponding
results (see Table III).
We notice that finer dm bins improves performance a
little, and fewer dt bins do not seem to adversely affect
the performance. Each experiment is time consuming and we
continue our efforts to fine tune the parameters to get better
results. As stated earlier the current results are already usable.
Exploring these more is an obvious area of further advance.
D. Background Subtraction
How can we improve the classification further? Individuals
look at differences as well as similarities in order to classify
objects. Consider spectra: given other common things, one
asks if a particular line is too broad, or narrow, or missing, or
extra-intense. We do the same when looking at light curves.
With dmdt-images we have squared the number of points
and distributed them over a rectangle. If we could remove an
underlying, common background, the class-membership may
become more obvious. It is a non-trivial task at best. We
consider a few possibilities for determining such a background.
We use the Robust PCA algorithm [17] for decomposing the
matrix in this way. We use this particular method as it is one of
the earliest method which is a provable non-convex algorithm
in contrast to other works that rely upon convex relaxations
of the actual objective.
We determine backgrounds to subtract before training in a
few different ways: (1) for individual class backgrounds we
consider dmdt-images of just that class to form the matrix M.
Fig. 4 shows the class backgrounds. The class backgrounds
are not like the median images of stacked class images (see
Fig. 5). The difference partly springs from non-uniform lengths
of time-spans for individual light curves as well as differences
in maximum brightness variations over different time-scales.
We use the respective maximum background for each class. (2)
For a pseudo-cadence background we consider all our objects
together. We call this the pseudo-cadence background because
all our objects in the current set are periodic variables. In order
to not overwhelm this pseudo-background with a single class,
we take 500 samples of each type (or all, if the training sample
has fewer than 500). (3) As another possibility we ignored
the class imbalance, took all training samples, and used the
max from the background for subtraction from all training and
testing samples. (4) For a true background we will need to
consider dmdt-images for light curves of randomly selected
objects (or perhaps a large number of standard stars - stars
known not to vary). Since a vast majority of these objects will
be constant within error-bars over the entire time-span, the
corresponding dmdt-images will consist of a thin line along
the dm = 0 midrib. After the backgrounds are subtracted, the
CNN is trained on the foreground images.
(a) Class 1 (EW)
(b) Class 2 (EA)
Subtracting the class background provides far better results
as expected (see Table IV) since we use class information to
subtract a specific background even during testing, and in the
real world we are not be privy to this information. However it
does show that removing appropriate background accentuates
the different features between different classes. However. the
removal of the pseudo-cadence background is somewhat worse
than not removing any background. We mimicked cadence
background removal by blanking the middle 2,4,6 rows of the
dmdt-images, but they did not provide better results than not
removing the background either. Clearly, better modeling of
the background is required, a more time-consuming job that
we will be taking up in the near-future.
E. Transfer Learning
(c) Class 4 (RRab)
(d) Class 5 (RRc)
(f) Class 8 (RS CVn)
(e) Class 6 (RRd)
(g) Class 13 (LPV)
Fig. 4. Class backgrounds determined using the robust PCA method. The
max for each class is shown.
(a) Class 1 (EW)
(b) Class 2 (EA)
One of the real power of the dmdt-technique is its crosssurvey applicability. We used models trained with the CRTS-N
dmdt-images and tested them on CRTS-S dmdt-images with
same classes [10], but no overlapping objects, and with PTF
dmdt-images with a subset of the same objects as in the CRTSN sample. CRTS-S uses the same asteroid-finding cadence as
CRTS-N and also has an open filter. PTF used a more mixed
cadence with a greater emphasis on looking for explosive
events including a repeat cadence of 1, 3, 5 nights. We used
PTF data taken with the r filter. The results using both the
shallow and deep CNNs are given in Table V. The numbers
are not as good as with CRTS-N, but that is not unexpected.
In fact, for many classes, especially for CRTS-S, they are
better than one would naively expect. In case of PTF the
survey cadence is very different in addition to the aperture and
wavelength range and the results are somewhat worse. But the
very fact that they are still usable, and definitely a good starting
point indicates the merit of using such a technique. With
proper survey-based background subtraction the results should
improve further. The implications for domain adaptation are
obvious, especially with applicability to forthcoming surveys
like ZTF and LSST.
VI. D ISCUSSION
(c) Class 4 (RRab)
(d) Class 5 (RRc)
(f) Class 8 (RS CVn)
(e) Class 6 (RRd)
(g) Class 13 (LPV)
Fig. 5. Median dmdt-images for individual classes.
We have shown how to transform light curves to simple
dmdt-images for use with canned as well as simpler CNNs
for out-of-the-box classification of objects with performance
comparable to random forests, and without having to resort to
designing or extracting features, or other necessary evils like
dimensionality reduction. The internal features the CNN uses
need to be explored further using tools like deconvolutional
networks. That will make the results interpretable, and provide
insights. We have also shown various paths to take in order to
improve the results further e.g. background subtraction and
varying the dmdt bins. We have further demonstrated the
application of the technique to transfer learning and thereby
classifying objects from a completely different survey.
A. Comparison with RF
Random forests (RF) tend to be very versatile and difficult
to beat in performance. Hence we compare the performance of
TABLE V
ACCURACIES ( ROW 1) AND RECALL ( OTHER ROWS ) FOR MODELS
TRAINED ON dmdt- IMAGES FROM CRTS-N AND TESTED ON CRTS-S,
AND PTF FOR THE CNN SHALLOW AND DEEP NETWORKS .
Class
All7
1/2
1/4
1/5
1/6
1/8
1/13
2/4
2/5
2/6
2/8
2/13
4/5
4/6
4/8
4/13
5/6
5/8
5/13
6/8
6/13
8/13
CRT SS
shallow
deep
69.5
69.9
98/70
98/69
99/49
98/52
99/23
97/37
99/13
99/13
99/3
99/4
99/82
99/82
96/97
95/98
96/98
95/99
99/92
99/96
97/81
97/76
99/95
99/96
69/90
66/90
88/72
94/53
82/86
92/62
92/93
95/93
99/1
97/6
96/51
96/44
99/84
99/91
67/89
75/88
96/94
94/95
98/91
98/90
PTF
shallow
deep
66.5
66
95/44
95/45
95/38
96/32
98/8
97/13
99/4
98/23
86/63
88/48
75/83
64/80
85/94
80/96
70/94
73/96
99/51
99/75
28/97
55/94
85/76
61/83
80/59
57/85
98/11
97/8
41/97
72/91
70/89
79/97
99/0
93/5
46/92
80/53
84/83
81/89
3/99
15/97
80/89
78/85
77/67
58/89
Fig. 6. Confusion matrix for shallow CNN with fiducial dmdt-images. Note
the high misclassification between classes 5 and 6, both RR Lyrae.
our technique with random forests. We use the features given
in Table VI for our RF setup. The output is shown in Table II.
The features we have used are generic, and designer features
would provide better recall and precision for select classes.
We find the performance of the shallow CNN comparable
to that of unweighted RF. In a way this is remarkable since it
is akin to providing the light curve almost in its raw format
and getting a classification. For some classes the recall is low,
possibly due to the sparseness of the light curves. RFs have
provided better results when used with features based on nonsparse light curves [1], [18]. Combining the features with the
CNN to form a deep-and-wide network will likely provide
better performance than either.
TABLE VI
R ANDOM FOREST FEATURES . T HE FIRST THREE ARE NOT USED IN RF
WHEREAS THE REMAINING 18 ARE FAIRLY GENERIC FEATURES OFTEN
USED IN CLASSIFICATION E . G . [1], [2], [3]. F ORMULAE FOR THE
FEATURES ARE FROM htt p : //nirgun.caltech.edu :
8000/scripts/description.html#method summary
Feature
meanmag
minmag
maxmag
amplitude
beyond1std
flux percentile ratio mid20
flux percentile ratio mid35
flux percentile ratio mid50
flux percentile ratio mid65
flux percentile ratio mid80
linear trend
max slope
median absolute deviation
median buffer range percentage
pair slope trend
percent difference flux percentile
skew
small kurtosis
std
stetson j
stetson k
.
Formula
< mag >
magmin
magmax
0.5 ∗ (magmax − magmin )
p(|(mag− < mag >)| > σ )
( f lux60 − f lux40 )/( f lux95 − f lux5 )
( f lux67.5 − f lux32.5 )/( f lux95 − f lux5 )
( f lux75 − f lux25 )/( f lux95 − f lux5 )
( f lux82.5 − f lux17.5 )/( f lux95 − f lux5 )
( f lux90 − f lux10 )/( f lux95 − f lux5 )
b where mag = a * t + b
max(|(magi+1 − magi )/(ti+1 − ti )|)
med( f lux − f luxmed )
p(| f lux − f luxmed | < 0.1 ∗ f luxmed )
p( f luxi+1 − f luxi > 0; i = n − 30, n)
( f lux95 − f lux5 )/ f luxmed
µ3 /σ 3
µ4 /σ 4
σ
var j (mag)
vark (mag)
Fig. 7. Confusion matrix for random forest using features. The numbers are
given as percentages. Overall classification accuracy is 82.2%. Yellow, Green,
and Blue indicate successively larger percentages from 0 to 100.
B. Misclassified Sources
The confusion matrices (Figs. 6 and 7) during our various
experiments showed that for some classes large fractions of
objects were misclassified (see Fig. 8). We investigated the
light curves for some of these sources in order to identify
the source of errors. In some cases it was a genuine error
(wrong label) indicating that the network was working well.
In some other cases the misclassification was owing to a sparse
light curve indicating that in a handful of cases a smaller
number of features may be tilting the classifications one way or
another. In still other cases, the subclasses were just too close
for the technique to discern them apart just from the dmdt-
time cases with far fewer data points. We did a couple of tests
using error-bars to augment smaller classes, but that did not
work well. That needs to be explored further for reducing
the unbalancedness of the different classes. Also there is
the possibility of using Generative Networks to create large
simulated examples for different classes for understanding the
features that really separate different classes. The number of
possibilities is large – we invite others to explore them as well.
ACKNOWLEDGEMENTS
This work, and CRTS survey, was supported in part by the
NSF grants AST-0909182, AST-1313422, AST-1413600, and
AST-1518308, and by the Ajax Foundation. KS thanks IITGandhinagar and the Caltech SURF program.
R EFERENCES
Fig. 8. Histograms of misclassified objects for classes 2 (top) and 6 (bottom).
All class 6 (RRd) objects were misclassified, two-thirds of them being
classified as another type of RR Lyrae (RRc). 19% of class 2 (EA) objects
were misclassified, most of them as class 1 (EAs), and the distribution of
points in the misclassified set suggests that though there is a slight trend, it
is not the shortest light curves that were thus misclassified. That is true in
general for misclassifications in other classes as well.
(a) Foreground 1
(b) Foreground 2
Fig. 9. Two types of foregrounds were seen for class 1, suggesting a possible
split in the dataset. This dichotomy needs to be investigated further.
images based upon the light curves (e.g. RR Lyrae of different
types). In the case of EW and EA classes we suspect that
the technique may be teasing apart subclasses (e.g. based on
separation) or geometric dependence. For example, we noticed
two distinct kinds of foregrounds (see Fig. 9), and this needs
to be investigated further.
C. Future Work
We will explore various possibilities related to varying CNN
hyperparameters, improving background subtraction for more
reliable classification, expanding to more classes and surveys,
as well as identifying the misclassified sources. We will also
experiment to make the technique more useful in the real-
[1] J. W. Richards, D. L. Starr, N. R. Butler et al., “On Machine-learned
Classification of Variable Stars with Sparse and Noisy Time-series Data,”
ApJ, vol. 733, p. 10, May 2011.
[2] C. Donalek, A. Arun Kumar, S. G. Djorgovski et al., “Feature Selection
Strategies for Classifying High Dimensional Astronomical Data Sets,”
ArXiv e-prints, Oct. 2013.
[3] M. J. Graham, S. G. Djorgovski, A. J. Drake et al., “A novel variabilitybased method for quasar selection: evidence for a rest-frame ∼54 d
characteristic time-scale,” MNRAS, vol. 439, pp. 703–718, Mar. 2014.
[4] A. A. Mahabal, S. G. Djorgovski, A. J. Drake et al., “Discovery,
classification, and scientific exploration of transient events from the
Catalina Real-time Transient Survey,” Bulletin of the Astronomical
Society of India, vol. 39, pp. 387–408, Sep. 2011.
[5] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 5 2015.
[6] S. G. Djorgovski, C. Donalek, A. Mahabal et al., “Towards an Automated Classification of Transient Events in Synoptic Sky Surveys,”
ArXiv e-prints, Oct. 2011.
[7] A. J. Drake, S. G. Djorgovski, A. Mahabal et al., “First Results from
the Catalina Real-Time Transient Survey,” ApJ, vol. 696, pp. 870–884,
May 2009.
[8] A. A. Mahabal, C. Donalek, S. G. Djorgovski et al., “Real-Time
Classification of Transient Events in Synoptic Sky Surveys,” in New
Horizons in Time Domain Astronomy, ser. IAU Symposium, E. Griffin,
R. Hanisch, and R. Seaman, Eds., vol. 285, Apr. 2012, pp. 355–357.
[9] S. G. Djorgovski, M. J. Graham, C. Donalek et al., “Real-Time Data
Mining of Massive Data Streams from Synoptic Sky Surveys,” ArXiv
e-prints, Jan. 2016.
[10] A. J. Drake, S. G. Djorgovski, M. Catelan et al., “The Catalina Surveys
Southern periodic variable star catalogue,” MNRAS, vol. 469, pp. 3688–
3712, Aug. 2017.
[11] N. M. Law, S. R. Kulkarni, R. G. Dekany et al., “The Palomar Transient
Factory: System Overview, Performance, and First Results,” PASP, vol.
121, p. 1395, Dec. 2009.
[12] A. J. Drake, M. J. Graham, S. G. Djorgovski et al., “The Catalina
Surveys Periodic Variable Star Catalog,” ApJS, vol. 213, p. 9, Jul. 2014.
[13] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
Learning, ser. Springer Series in Statistics.
New York, NY, USA:
Springer New York Inc., 2001.
[14] K. P. Murphy, Machine Learning: A Probabilistic Perspective. The
MIT Press, 2012.
[15] S. Dieleman, K. W. Willett, and J. Dambre, “Rotation-invariant convolutional neural networks for galaxy morphology prediction,” MNRAS,
vol. 450, pp. 1441–1459, Jun. 2015.
[16] G. Cabrera-Vives, I. Reyes, F. Förster, P. A. Estévez, and J. Maureira,
“Supernovae detection by using convolutional neural networks,” in
2016 International Joint Conference on Neural Networks, IJCNN 2016,
Vancouver, BC, Canada, July 24-29, 2016, 2016, pp. 251–258.
[17] P. Netrapalli, U. N. Niranjan, S. Sanghavi, A. Anandkumar, and P. Jain,
“Non-convex robust PCA,” CoRR, vol. abs/1410.7660, 2014.
[18] P. Dubath, L. Rimoldini, M. Süveges et al., “Random forest automated
supervised classification of Hipparcos periodic variable stars,” MNRAS,
vol. 414, pp. 2602–2617, Jul. 2011.
| 1cs.CV
|
1
Energy- and Spectral- Efficiency Tradeoff for
D2D-Multicasts in Underlay Cellular Networks
arXiv:1712.00307v2 [cs.NI] 12 Dec 2017
Ajay Bhardwaj and Samar Agnihotri
School of Computing and Electrical Engineering, Indian Institute of Technology Mandi,
HP - 175 005, India
Email: ajay [email protected], [email protected]
Abstract
Underlay in-band device-to-device (D2D) multicast communication, where the same content is disseminated
via direct links in a group, has the potential to improve the spectral and energy efficiencies of cellular networks.
However, most of the existing approaches for this problem only address either spectral efficiency (SE) or energy
efficiency (EE). We study the trade-off between SE and EE in a single cell D2D integrated cellular network, where
multiple D2D multicast groups (MGs) may share the uplink channel with multiple cellular users (CUs). We explore
SE-EE trade-off for this problem by formulating the EE maximization problem with constraint on SE and maximum
available transmission power. A power allocation algorithm is proposed to solve this problem and its efficacy is
demonstrated via extensive numerical simulations. The trade-off between SE and EE as a function of density of
D2D MGs, and maximum transmission power of a MG is characterized.
Index Terms
Multicasting, Device-to-Device communication, energy-and-spectral efficiency trade-off, LTE-A networks.
I. I NTRODUCTION
Supporting ever increasing number of mobile users with data-hungry applications, running on batterylimited devices, is posing a daunting challenge to telecommunications community. Underlay device-todevice (D2D) communication, which allows physically proximate mobile users to directly communicate
with each other by reusing the spectrum and without going through the base station, holds promise to help
us tackle this challenge [1]. In a cellular network, underlay D2D communication offers opportunities for
spectrum reuse and spatial diversity which may lead to enhanced coverage, higher throughput, and robust
communication in the network [2]. Further, for applications such as weather forecasting, live streaming,
or file distribution, which may require the same chunk of data distributed to geographically proximate
group of users, D2D multicasting may provide better utilization of network resources compared to D2D
unicast or Base Station (BS) based multicast. However, extensive deployment of underlay D2D multicast
in a network may cause severe co-channel interference due to spectrum reuse and rapid battery depletion
of the multicasting D2D nodes due to higher transmit power to mitigate co- channel interference and
data-forwarding.
In cellular networks, two metrics namely spectrum efficiency (SE) and energy efficiency (EE) characterize how efficiently the spectrum and energy resources, respectively, are used. However, often conflicting
nature of these metrics may not allow for simultaneous maximization of both in a network. There exists
an extensive body of work that explores SE-EE trade-off in cellular networks [3], [4] and some work that
explore the nature of this trade-off for D2D communication [5], [6]. However, currently no systematic
study of the nature of such trade-off for multiple D2D multicasts in underlay cellular networks exists. To
the best of our knowledge, this letter provides the first such study.
Using stochastic geometry, we formulate a resource allocation problem that maximizes the EE of
multiple D2D-multicasts in underlay cellular networks with constraints on SE and maximum transmission
power. The formulated problem is non-convex, and is solved using the proposed heuristic gradient power
allocation algorithm. We establish the trade-off between EE and SE with various network parameters such
2
as density of D2D multicast transmitters, and maximum transmission power of MGs through numerical
simulations.
Organization: Section II introduces the system model. Section III and IV introduce the problem formulation
and the optimal power allocation algorithms, respectively. Performance evaluation is in Section V. Finally,
Section VI concludes the paper.
II. S YSTEM M ODEL
A D2D-integrated cellular network is considered, where multiple D2D-multicast groups (MGs) may
share the uplink channel with multiple cellular users (CUs). Let K = {1, 2, . . . , K} denote the total
number of orthogonal channels that can be shared by CUs and D2D MGs. Sharing of uplink channels is
assumed instead of downlink because of asymmetric uplink and downlink traffic loads [7], and also as the
eNB (evolved Node Base station) can handle interference effectively. The spatial distribution of CUs and
MGs on the k th channel is modeled as homogeneous Poisson Point Process Πc,k with density λkc , and Πg,k
with density λkg , respectively, in the Euclidean plane R2 . The proposed system model is an abstraction of
a system where a single cell is divided into small cells, and multiple CUs may share a single channel.
Let |Ug | be the number of receivers in the g th MG (|Ug | = 1 corresponds to unicast communication.) We
consider the variable number of receivers in every MG and assume that they always have data demands.
The transmission powers of CU and D2D MG transmitter (D2D-Tx) on the k th channel are denoted by
pc,k and pg,k , respectively. In addition, total transmission power of CUs and D2D MGs is PC and PG , i.e.
X|K|
X|K|
(1a),
(1b)
pc,k = PC ,
pg,k ≤ PG
k=1
k=1
For analysis, a reference receiver at the origin of cell (eNB for cellular and a typical D2D receiver
for D2D communication) is considered. The radio propagation channel gain between the ith transmitter
and the j th receiver, denoted as hi,j , is assumed to be Rayleigh faded, and independently and identically
exponentially distributed with unit mean, i.e. hi,j ∼ exp(1). Therefore, the received power at the j th
th
th
receiver is pj = pi hi,j d−α
node, and α is
i,j , whereas di,j denotes the distance between the i and the j
the path-loss exponent.
As we are considering the scenarios where a channel is shared by multiple CUs and multiple MGs,
thus, the rth (r ∈ Ug ) D2D-Rx experiences interference from co-channel CUs and other D2D MG-Tx.
Therefore, the signal-to-interference-plus-noise-ratio (SINR) at the rth (r ∈ Ug ) D2D receiver on the k th
channel is
pg,k hg,r,k d−α
k
P g,r
,
(2)
γg,r
= P
−α
pc,k hc,r,k dc,r +
pg0 ,k hg0 ,r,k d−α
g 0 ,r + N0
c∈Πc,k
g 0 ∈Πg,k
As the system model is interference limited, therefore, thermal noise N0 can be omitted, and hence (2)
becomes
hg,r,k d−α
g,r
k
γg,r =
,
(3)
Ic,r,k + Ig0 ,g,k
P
P
pc,k
pg,k
−α
where Ic,r,k = c∈Πc,k pg,k
hc,r,k d−α
j,r,k and Ig 0 ,g,k =
g 0 ∈Πg,k pg0 ,k hg 0 ,r,k dg 0 ,r,k . In a MG, transmission rate is
decided by channel conditions of the worst user [8], so, SIR, and the corresponding date rate, respectively,
are
hg,r,k d−α
g,r
k
(4)
γg = min
, Rgk = log2 1 + γgk
r∈Ug
Ic,r,k + Ig0 ,g,k
An outage event for a MG g occurs if the aggregate rate of the g th group falls below its target rate, Rgth .
Therefore, the outage probability of the g th MG is given by the following lemma.
3
Lemma 1: The outage probability of a D2D MG communicating on the k th shared channel is
#)
(
"
α2
p
c,k
+ λkg
,
Pr Rgk < Rgth = 1 − exp − χg,k λkc
pg,k
(5)
2/α
th
R∞
where χg,k = πd2g,r Γ 1 + α2 Γ 1 − α2 2(Rg /|Ug |) − 1
, Γ(x) = 0 tx−1 e−t dt denotes complete gamma
function.
Proof: Please refer to Appendix A
Similarly, the SIR of a CU transmitting on the k th channel is
γck
pc,k hc,b d−α
c,b
P
= P
−α
pc,k hc0 ,b,k dc0 ,b +
pg,k hg,b,k d−α
g,b
c∈Πc,k
(6)
g∈Πg,k
with the corresponding outage probability given by the following lemma.
Lemma 2: The outage probability of a CU on the k th channel, can be expressed as
(
"
α2 # )
p
g,k
Pr Rck < Rcth = 1 − exp − χc,k λkc + λkg
pc,k
(7)
2/α
th
where Rcth denotes the date rate threshold of CU and χkc = πd2c,b Γ 1 + α2 Γ 1 − α2 2(Rc ) − 1
.
Proof: Similar to the proof of Lemma 1.
These lemmas allow us to infer that
• The outage probability of D2D MGs increases with increase in MG geographical spread, dg,r , because
channel fading becomes too severe with increasing distance.
• The outage probability of D2D MGs decreases with decrease in densities of CUs and D2D MGs,
this is due to mitigation of co-channel interference.
• The outage of D2D MGs increases with increase in pc,k , because it creates more interference to D2D
transmission. While, higher pc,k decrease the outage of CUs.
III. O PTIMAL R ESOURCE A LLOCATION FOR D2D- M ULTICASTS IN U NDERLAY C ELLULAR
C OMMUNICATION
The average throughput, SEkg , of D2D mulicast communication on the k th channel, can be expressed as
[10] :
(8)
SEkg = |ug |λkg log2 1 + γgk Pr Rgk ≥ Rgth
The EE is defined as the ratio of average throughput to the power consumption [5]. As in [10], the
power consumption of D2D MGs which are communicating on the k th channel is λkg pg,k . Therefore, the
total energy efficiency of D2D multicast underlay network is
X|K|
X|K|
EEg =
EEkg =
SEkg /(λkg pg,k )
(9)
k=1
k=1
To ensure the high data rate to both CUs and D2D users, thresholds on outage probabilities of both the
D2D users and CUs are put as follows:
Pr Rck < Rcth ≤ Θc , and Pr Rgk < Rgth ≤ Θg ,
(10)
where Θc and Θg denote the outage thresholds for cellular and D2D MGs transmissions, respectively. The
transmit power in the k th channel should be less than the allowed upper bound for that channel, denoted
as pup
g,k . Thus, we have
0 ≤ pg,k ≤ pup
(11)
g,k
4
From (5), (7), and (10), feasible power region is
−ln (1 − Θg ) λkg
− k
λkc χg,k
λc
pg,k ≥ pc,k
!− α2
,
(12)
α
−ln (1 − Θc ) λkc 2
− k
(13)
λkg χc,k
λg
α
λkc 2
c)
= pc,k −ln(1−Θ
−
. Therefore, infimum pinf
g,k = max{0,
λk χc,k
λk
pg,k ≤ pc,k
α
λkg − 2
−ln(1−Θg )
−
and phigh
g,k
λkc χg,k
λkc
up
high
sup
supremum pg,k = min{pg,k , pg,k }
Let plow
g,k = pc,k
g
g
denote the upper and lower bound, respectively. With these
plow
g,k } and
tranformations, the EE optimization problem with constraints on SE and maximum transmission power
is formulated as:
X|K|
P : max EEg =
EEgk
(14)
pg,k
k=1
sup
s.t. pinf
g,k ≤ pg,k ≤ pg,k , and
X|K|
k=1
pg,k ≤ PG
The problem P is non-convex as we prove in Appendix B. From second constraint, one may infer that,
as the CUs are primary users, so to maintain their priority over MG users, they are assumed to transmit
at full power, i.e. Pc . While D2D MG users are secondary users, and interference creators for cellular
transmissions. Thus, they are assumed to transmit at lower powers, so that their sum power does not
overshoot the threshold PG .
IV. P OWER A LLOCATION F OR O PTIMAL R ESOURCE A LLOCATION
The objective function in optimization problem P is a summation of |K| functions. Let p∗g,k denote the
global maximal point where single function EEgk achieves its maximum, and if the sum power constraint
(1b) is removed, then pg,k are mutually independent. The EEg achieves the maximum value when every
EEgk achieves its maximum. Intuitively, finding power that maximizes the EEgk , is easier than solving P
as a whole. The power p∗g,k that maximizes the individual EEgk can be found using the following lemma.
Lemma 3: The value p∗g,k that maximizes the EEgk is
g α/2
2Z
sup
sup
pg,k ,
pg,k ≤ αk
g α/2
g α/2
2Zk
2Z
inf
p∗g,k =
, pg,k ≤ αk
≤ psup
g,k
α
g α/2
2Z
k
pinf ,
pinf
g,k
g,k ≥
α
2
where Zkg = χkg λkc (pc,k ) α
Proof: Please refer to Appendix C P
|K|
Now, two cases arise, Case 1: when k=1 p∗g,k ≤ PG
Then, the optimum power that is allocated is pg,k = p∗g,k , ∀k = 1, . . . , K, and the maximum value of
P|K|
energy efficiency is EEg = k=1 EEgk (p∗g,k ).
P|K|
Case 2: when k=1 p∗g,k > PG . Then, set pg,k = p∗g,k , ∀k = 1, . . . , K, and update the value of pg,k ,
such that power constraint (1b) is maintained, while causing the least reduction in EEg . Let pinst
g,k be the
∗
instant value, having initial value of pinst
=
p
.
Let
d
denotes
the
difference
between
maximum
available
g,k
g,k
power and sum of assigned power. δ and n denote the step size and number of steps, respectively, with
δ = d/n. Parameter n controls the balance between computational effort and the performance, its value is
assigned in accordance with convergence rate the error-tolerance requirements. As we are adjusting the
power value which gives maximum value of the function, therefore, the function value decreases with
5
k
inst
reducing power, i.e, EEgk pinst
g,k − δ < EEg pg,k . To satisfy the equality constraint (1b) while having
the least reduction in EEg , we need to adjust that pg,j , (j ∈ K) for which EEgj decreases the least after
decreasing the instantaneous transmit power.
j = argmin |EEgk pinst
− EEgk pinst
(15)
g,k
g,k − δ |
pinst
g,j
=
|K|
inst
pg,j −
δ
(16)
Iterating this process at least n times, leads to the sum power constraint (1b) be met, and a near-to-optimal
solution to (14) is achieved. The formal description is given in Algorithm 1. The computation complexity
of the proposed algorithm is O(n).
Algorithm 1: The proposed power allocation algorithm
Input: K, n, PG , PC , , Output: EEg
1 begin
2
Find p∗g,k from Lemma 3, and assign pg,k = p∗g,k
P
3
d = PG − K
k=1 pg,k
4
if d ≤ 0 then
P
k
∗
5
EEg = K
k=1 EEg pg,k
6
else
7
δ = d/n
P
8
while |PG − K
k=1 pg,k | ≥ do
k
inst
9
j = argmin|EEgk pinst
−
EE
p
−
δ
|
g
g,k
g,k
|K|
10
11
12
13
14
if pg,j + δ > psup
or pg,j − δ < pinf
g,j then
g,j
|K| = |K| ∩ j,
else
0
pg,j = pg,j − δ, and derj = |EEgj (pg,j ) |
return EEg =
PK
k=1
EEgk (pg,k )
V. S IMULATION R ESULTS
To explore the existence of trade-off between SE and EE for multiple D2D multicasts in underlay
cellular networks and exploit it to design optimal resource allocation schemes, we carried out extensive numerical simulations. Some of the simulation parameters are as follows. The CU density λkc =
[1e−4 , 1e−5 , 1e−4 , 1e−4 , 1e−4 ] and D2D density λkg = [1e−3 , 1e−4 , 1e−3 , 1e−3 , 1e−3 ] are considered, respectively. The CU density and D2D density are randomly chosen values. We used the following parameter
values: α = 3, |ug | = 3, pup
g,k = 15dBm, PG = 25dBm, Pc,k = 26dBm, Θc = 0.1, and Θg = 0.1. The value
of alpha has an impact on EE and SE, however, that is insignificant.
Fig. 1 depicts the behavior of EE with respect to required SE and the available D2D transmission power.
It can be observed that for a given power level, with increase in SE requirement (1 to 10 bps/Hz), EE
first increases then starts to decrease. This is because, for lower SE requirements, eNB tries to support
many MGs per channel until outage thresholds are not violated. This leads to increase in sum rate and
consequently, an increase in EE. While, for higher requirement of SE, eNB reduces the number of MG
transmitters sharing a channel until outage probability constrains are not fulfilled. Therefore, sum rate
decreases, consequently EE decreases. Similarly, increasing the transmission power of MG transmitter for
fixed SE, higher data rate may be supported for small range D2D communication, therefore, the sum rate
increases initially with power consumption. However, after some power threshold, MG transmitters start
Energy efficiency (EEg) (Mbps/J)
6
2.5
2
1.5
1
0.5
0
0
2
4
6
8
SE (bps/Hz)
10
Pmax
(dBm)
g
EE as a function of SE and D2D users transmitter power
Energy efficiency (EEg) (Mbps/J)
Fig. 1.
10
4
3
2
1
0
2
3
4
5
SE (bps/Hz)
6
5
7
8
9
10
Fig. 2.
30
26
22
18
14
1
2
6
7
8
9
10
4
−4
3 D2D density(λg)x10
EE as a function of SE and D2D users density
causing co-channel interference to CUs, and thus, CU outage probability starts increasing. To compensate
this, eNB reduces the number of MGs to fulfil CU outage probability thresholds. Therefore, decrease in
sum rate and consequently decrease in EE occurs. Fig. 2 depicts EE as a function of SE and D2D density.
It can be observed that, for a given SE requirement, with increasing D2D density, EE first increases
and then decreases. This is because, adding D2D users to the network (i.e increasing λg ), results in an
exponential increase in the average sum rate, and consequent increase in EE. However, in high density,
mutual interference starts increasing, and that limits the average sum rate per channel, leading to a decrease
in EE.
VI. C ONCLUSION
For underlay D2D multicast in cellular networks, we addressed the energy efficiency and spectral
efficiency trade-off. We assumed that multiple D2D-multicast group may share the channel with multiple
CUs. Exact expression of average sum-rate and its relation with energy consumption is derived by utilizing
the stochastic geometry. An energy optimization problem is formulated, having constraints on maximum
power, and outage data rate. Our results showed that EE has different behavior with available power and
spectral requirement. With increasing power, SE improves while EE initially increases then decreases.
Similarly, with increase in D2D MG density, EE initially increases and then decreases. Indeed, for the
EE, there is an optimal value of SE requirement that can be supported, which results in the maximal value
of EE for each value of transmitter power of MG.
R EFERENCES
[1] D. Feng, L. Lu, Y. Yuan-Wu, G. Li, G. Feng, and S. Li, “Device-to-device communications underlaying cellular networks,” IEEE Trans.
on Commun., vol. 61, no. 8, 2013.
7
[2] J. Liu, N. Kato, J. Ma, and N. Kadowaki, “Device-to-device communication in lte-advanced networks: A survey,” IEEE Commun. Surveys
& Tutorials, vol. 17, no. 4, 2015.
[3] J. Tang, D. K. So, E. Alsusa, and K. A. Hamdi, “Resource efficiency: A new paradigm on energy efficiency and spectral efficiency
tradeoff,” IEEE Trans. on Wireless Commun., vol. 13, no. 8, 2014.
[4] J. B. Rao and A. O. Fapojuwo, “An analytical framework for evaluating spectrum/energy efficiency of heterogeneous cellular networks,”
IEEE Trans. on Vehicular Techn., vol. 65, no. 5, 2016.
[5] H. Gao, M. Wang, and T. Lv, “Energy efficiency and spectrum efficiency tradeoff in the D2D-enabled hetnet,” IEEE Trans. on Vehicular
Techn., vol. 66, no. 11, 2017.
[6] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, “Energy efficiency and spectrum efficiency of multihop device-to-device communications
underlaying cellular networks,” IEEE Trans. on Vehicular Techn., vol. 65, no. 1, 2016.
[7] O. Onireti, F. Héliot, and M. A. Imran, “On the energy efficiency-spectral efficiency trade-off in the uplink of CoMP system,” IEEE
Trans. on Wireless Commun., vol. 11, no. 2, 2012.
[8] H. Meshgi, D. Zhao, and R. Zheng, “Joint channel and power allocation in underlay multicast device-to-device communications,” in
Proc. IEEE ICC, London, UK, June 2015.
[9] A. H. Sakr and E. Hossain, “Cognitive and energy harvesting-based d2d communication in cellular networks: Stochastic geometry
modeling and analysis,” IEEE Tran. on Commun., vol. 63, no. 5, 2015.
[10] Y. Kwon, T. Hwang, and X. Wang, “Energy-efficient transmit power control for multi-tier MIMO hetnets,” IEEE JSAC, vol. 33, no. 10,
2015.
A PPENDIX A
P ROOF OF L EMMA 1
where Tg,k
Pr Rgk < Rgth = 1 − Pr Rgk ≥ Rgth ∀g ∈ λkg
th
th
Rg
Rg
hg,r,k d−α
g,r
k
|Ug |
|Ug |
≥2
− 1 = 1 − Pr
−1
= 1 − Pr γg ≥ 2
Ic,r,k + Ig0 ,g,k
th
Rg
α
= 1 − Pr hg,r,k ≥ 2 |Ug |−1 dg,r (Ic,r,k + Ig0 ,g,k )
(
Rth
Y
g
pc,k
hc,r,k d−α
=1− E
exp − 2 |Ug | − 1 dαg,r
c,g
p
g,k
j∈Πc,k
Rth
)
Y
g
pg,k
×E
exp − 2 |Ug | − 1 dαg,r
hg0 ,g,k d−α
g 0 ,g
0
p
g ,k
g∈Πg,k
= 1 − LIc,r,k (hc,r,k ) Tg,k dαg,r LI 0 (h 0 ) Tg,k dαg,r ,
g ,g,k
g ,g,k
Rth
g
= 2 |Ug | − 1 . According to definition of Laplace transform, we have
"
2
#
2
pc,k α
2
2
α
k
2
α
LIc,r,k (hc,r,k ) Tg,k dg,r = exp −λg
πTg,k
dg,r Γ 1 +
Γ 1−
pg,k
α
α
LIg0 ,g,k (hg0 ,g,k )
"
= exp −λkg
Tg,k dαg,r
−λkg
Z
∞
−α
−Tg,k dα
g,r r
E (hg0 ,g,k ) 1 − e
2
#
2
pg,k α
2
2
α
Γ 1−
πTg,k
d2g,r Γ 1 +
0
pg ,k
α
α
= exp
dr
0
A PPENDIX B
P ROOF THAT THE FORMULATED PROBLEM P IS NON - CONVEX
2
Let Ykg = |ug |log2 1 + γgk exp −χkg λkg and Zkg = χkg λkc (pc,k ) α , then we can write EEgk as
α2 !
1
EEgk , (Ykg /pg,k ) exp −Zkg
pg,k
8
Taking double derivative of EEgk with respect to pg,k .
2/α
3+4/α
d2 EEgk
1
1
g
g
= 2Yk exp −Zk
dp2g,k
pg,k
pg,k
!
2(Zkg )2
Zkg 2
2/α
4/α
+ 3 pg,k +
× pg,k −
α α
α2
In this equation, the 1st and 2nd terms are greater than zero, therefore, we only focus on the last term.
2/α
Let ν = pg,k , then the last term can be written as
2(Zkg )2
Zkg 2
2
+3 ν+
V=ν −
α α
α2
The solutions ν1,k and ν2,k can be found as
√
νi,k = (Zkg /2α2 ) 2 + 3α ∓ α2 + 12α + 4 , i ∈ {1, 2}
V is positive on the interval
(0,ν1,k
)∪(ν2 , +∞),
d2 EEgk /dp2g,k
and negative in interval (ν1,k
, ν2,k ). Therefore,
α/2
α/2
α/2
α/2
!
!
is positive in interval 0, ν1,k ∪ ν2 , +∞ , and negative in interval ν1,k , ν2,k . Therefore, EEgk is
α
α α
α
2
2
2
2
convex on interval 0, ν1,k
∪ ν2,k
, +∞ , but concave on the interval ν1,k
, ν2,k
A PPENDIX C
P ROOF OF L EMMA 3
Take the derivative of EEgk with respect to pg,k ,
!!
d EEgk
2Zkg
Ykg
1
×
= 2 exp −Zkg
2/α
dpg,k
pg,k
α
pg,k
1
2/α
−1
pg,k
α/2
2Z g
From this equation, it can be observed that derivative of
is positive if pg,k lies in interval 0, αk
,
g α/2
α/2
2Z
2Zkg
, +∞ , and reaches the global maxima at pg,k = αk
.
and negative if pg,k lies in interval
α
α/2
2Z g
Therefore, three feasible region exist, EEgk is increasing monotonically in 0, αk
, decreasing
g α/2
g α/2
α/2
2Zk
2Zk
2Zkg
sup
,
+∞
,
and
having
maximal
point
at
p
=
.
If
p
≤
,
monotonically in
g,k
g,k
α
α
α
EEgk
then EEgk is an increasing function, and, it reaches the maximum value when p∗g,k = psup
g,k . Second region
g α/2
2Zk
k
is pinf
≤ psup
g,k ≤
g,k , then the maximal point of EEg is within feasible region. Therefore, optimal
α
g α/2
g α/2
2Zk
2Z
point is p∗g,k = αk
. Third region is, if pinf
, then EEgk decreases monotonically in the
g,k ≥
α
feasible region. Hence, the maximum value of EEg,k is achieved at p∗g,k = pinf
g,k
| 7cs.IT
|
Development and Evaluation of Two Learning-Based Personalized
Driver Models for Car-Following Behaviors
arXiv:1703.03534v1 [cs.SY] 10 Mar 2017
Wenshuo Wang1 , Student Member, IEEE, Ding Zhao2 , Junqiang Xi3 , David J. LeBlanc2 , and J. Karl Hedrick4
Abstract— Personalized driver models play a key role in
the development of advanced driver assistance systems and
automated driving systems. Traditionally, physical-based driver
models with fixed structures usually lack the flexibility to
describe the uncertainties and high non-linearity of driver behaviors. In this paper, two kinds of learning-based car-following
personalized driver models were developed using naturalistic
driving data collected from the University of Michigan Safety
Pilot Model Deployment program. One model is developed
by combining the Gaussian Mixture Model (GMM) and the
Hidden Markov Model (HMM), and the other one is developed by combining the Gaussian Mixture Model (GMM) and
Probability Density Functions (PDF). Fitting results between
the two approaches were analyzed with different model inputs
and numbers of GMM components. Statistical analyses show
that both models provide good performance of fitting while the
GMM–PDF approach shows a higher potential to increase the
model accuracy given a higher dimension of training data.
Index Terms— Personalized model, Learning-based driver
model, Gaussian mixture model, Hidden Markov model, Carfollowing behavior.
I. I NTRODUCTION
Understanding individual driver behaviors and development of personalized driver models are critical for active
safety control systems [1]–[3], vehicle dynamic performance
[4], and human-centered vehicle control systems [5]–[7],
eco-driving systems [8], and automated vehicles [9]. For
instance, a driver assistance system will be more effective if
the individual characteristics or/and driving styles can be incorporated [2]. Personalized driver models can be referred to
[9] “a driver model which can generate the output sequences
being as close as possible to what the individual driver would
have done in the same driving situation”. Lefevre et al.
[1], Butakov and Ioannou [2], [3] developed a personalized
driver model based on the Gaussian mixture model and then
applied to the advanced driver assistance systems (ADASs),
increasing the potential for more widespread acceptance and
use of ADASs.
*This work was supported by China Scholarship Council.
1 Wenshuo Wang is with the School of Mechanical Engineering, Beijing
Institute of Technology, Beijing, China, 100081, and with the Department
of Mechanical Engineering, University of California Berkeley, CA, 94720
USA. [email protected]
2 Ding Zhao and David J. LeBlanc are with the University of
Michigan Transportation Research Institute, Ann Arbor, MI 48109
[email protected]
3 Junqiang Xi is with the Department of Mechanical Engineering,
Beijing
Institute
of
Technology,
Beijing,
China,
100081.
[email protected]
4 J. Karl Hedrick is with the Department of Mechanical Engineering, University of California at Berkeley, Berkeley, CA 94720 USA.
[email protected]
Generally, the ways to establish a personalized driver
model can be grouped into two categories: physical-based
model and learning-based model. For the physical-based
model, formulations with unknown parameters are usually
used to describe the structure of driver’s driving behaviors such as car following, path following, lane change,
overtaking. The major benefit of the physical-based model
is that most model parameters have their specific physical
meanings, enabling them to be easily interpreted. For example, the intelligent driver model (IDM), optimal velocity
model (OVM) [10], and control-oriented car-following model
are popular physical-based models in the applications of
vehicle control [11]–[13] and traffic flow analysis [14]. The
model parameters can be identified using parameter estimation approaches [15], [16] such as least squares, Kalman
filter, stochastic parameter estimation, etc. The physicalbased approach can model driver’s basic behavior, however,
it is hard to model uncertainties and non-linearity because
of the uncertainty and diversity of individual driver’s behavior and driving environment. Fortunately, learning-based
models can be developed to overcome these issues. Popular
approaches have been developed to generate a learning-based
driver model such as stochastic switched AutoRegressive
eXogenous model (SS-ARX) [17], [18], hidden Markov
model [19], neural network [20], [21], and Gaussian mixture model [21]. These models are believed to represent
an individual driver’s driving characteristics and describe
the underlying source after correctly training. However, it
is difficult to explain the physical meaning of the model
parameters when learning-based models are directly utilized
to generate a highly nonlinear function for driver’s behavior
(e.g., decision-making and control). Butakov and Ioannou
[2] created a more explainable, flexible, and accurate driver
lane change model by combining the learning-based and
physical-based methods together, in which physical-based
model was used to mimic the driving behavior and the
learning-based model (i.e., GMM) was used to describe the
parameter distribution of the physical model.
In the above mentioned learning-based approaches, the
GMM method is usually chosen to establish driver model
due to its effectiveness of modeling driving tasks [2], [10],
[22]–[25]. However, limited works studied the comparison
between different learning-based approaches. In this paper,
two learning-based approaches were shown, in which the
influences of different combinations between parameters
(e.g., vehicle speed, range, relative speed.) with different
numbers of GMM component on model performance were
analyzed and discussed. This paper provides the systematic
∆𝑥
𝑣"
𝑣#
Host vehicle
Fig. 1.
zt = [∆xt , ∆vt , ∆v̇t , vth , ht ]> ∈ R5×1 are the current
states representing current driving situation at time t,
where ∆xt = xlt − xht is the relative distance between
the host vehicle and the leading vehicles, ∆vt = vtl −
vth is the relative speed between the host vehicle and
the leading vehicle, and ∆v̇t is the relative acceleration
between two vehicles.
The history of explanatory variables, z1:t , and acceleration sequences, ah1:t−1 , are taken as the model input.
The predicted vehicle acceleration is taken as the model
output. At each step t, the learned driver model generates
an acceleration aht . The general form of the proposed driver
model is presented as
•
Leading vehicle
An illustration of the car-following scenario.
re-examination, evaluation, and comparison of two learningbased approaches for modeling driver’s car-following behavior, and also helps researchers understand how many and
what parameters are more suitable to model a driver’s carfollowing behavior.
The structure of this paper is organized as follows. Section
II shows the problem formulation of driver’s car-following
behavior. Section III presents the basic methods for personalized driver model. Section IV shows the data collection
and data preprocessing. Section V discusses and analyzes
the experiment results.
II. P ROBLEM F ORMULATION
A. Personalized Driver Model
A specific definition of the personalized driver model is
given as: A personalized driver model can be referred to
a model that can generate or predict an individual driver’s
operating parameter (e.g., steering angle, throttle opening,
braking force.) or behavior (e.g., lane change, stop & go,
overtaking, decision-making with traffic light.) with the
same environment inputs, including traffic users (e.g., other
vehicles, bicycles, and pedestrians.), weather conditions, and
road conditions.
In this paper, we are going to investigate the personalized
driver models for car-following behaviors, which can generate a personalized longitudinal control signal sequence (i.e.,
acceleration).
B. Car-Following Scenario
The car-following behavior can be illustrated by Fig. 1.
We define the following variables to represent the relative
motion of the host vehicle and the vehicle located ahead in
the same lane as leading vehicle.
h h h >
• ξh = [xt , vt , t ]
∈ R3×1 is the state of the host
vehicle at time t, where xht ∈ R+ is the longitudinal
position of the host vehicle, vth is the longitudinal speed
of the host vehicle, and ht is the jerk of the host vehicle
defined as ht = v̈th .
l
l >
2×1
• ξl = [xt , vt ] ∈ R
is the state of the leading vehicle
l
at time t, where xt ∈ R+ is the longitudinal position of
the leading vehicle and vtl is the longitudinal speed of
the leading vehicle.
D(z1:t , ah1:t−1 ) : zt 7→ âht
(1)
The equation (1) is to generate an acceleration with the current input, zt , according to the history information, ξ1:t−1 =
[z1:t−1 , ah1:t−1 ], with the prediction step ∆t = 0.1 s.
III. M ETHODS
In this section, two learning-based approaches of modeling
a personalized driver car-following behavior are discussed,
i.e., the Gaussian Mixture Regression with the Hidden
Markov Model (GMR-HMM) and the Gaussian Mixture
Model with Probability Density Functions (GMM-PDF). To
understand the two approaches, the GMM, HMM, and PDF
are separately discussed in the following sections.
A. Gaussian Mixture Model
A set of d-dimension sequence, ξ = {ξi }N
i=1 with ξi ∈
R , can be encoded in a combination of N Gaussian
models. Assuming that the data in each component of GMM
is subject to a Gaussian distribution:
d×1
ξi ∼ Ni (µi , Σi )
(2)
where µi ∈ Rd×1 and Σi ∈ Rd×d is mean and covariance
of the ith Gaussian distribution Ni . For all data ξ, it can be
encoded by a Gaussian mixture model:
P(ξ; θ) =
N
X
i=1
N
X
πi Ni (ξ; µi , Σi )
1
d/2 |Σ |1/2
(2π)
i
i=1
1
> −1
× exp − (ξ − µi ) Σ (ξ − µi )
2
=
πi
(3)
where θ = {µP
i , Σi , πi }, i = 1, 2, . . . , N ; πi is the prior
N
probability and i=1 πi = 1.
For the car-following model, if we assign ξt = [zt , aht ],
the joint distribution between zt and aht can be rewritten as
P(zt , aht ; θ) ∼
N
X
i=1
πi Ni (zt , aht ; µi , Σi )
(4)
The parameter θ of (4) can be estimated by expectation maximization (EM) algorithm [2]. For the initial value (µ0 , Σ0 ) at
iteration step s = 0, we apply the k-means clustering method
to determine µ0 , and then calculate π0 . Thus, we can obtain
the estimated optimal parameter θ̂ until the log-likelihood
function is convergent or meets the maximum iteration steps
s ≥ smax , where the optimal objection for the log-likelihood
function is formulated as:
θ̂ = arg max L(θ) = arg max log(P(ξ; θ))
θ
θ
(5)
The number of GMM component can be determined by
Bayesian information criterion (BIC). Further, we also discussed the influences of numbers of GMM component on
training and tested the model performance.
Our goal is to generate a personalized acceleration sequence based on the learned driver model. With this purpose
in mind, two basic approaches are employed and discussed
as follows, i.e., HMM and PDF.
B. Hidden Markov Model
The joint distribution P(zt , aht ; θ) is encoded to generate
the output of the personalized driver model in a continuous
HMM of N states. Here, each component of GMM is
treated as a state of HMM. The HMM can be presented
by H(Π, Φ, µ, Σ), where Π = {πis=0 }N
i=1 is the initial
prior probability of being in state i, Φ = {φi,j }N
i,j is the
transitional probability from state i to j; µi and Σi are
the mean and the covariance matrix of the ith Gaussian
distribution of the HMM. Therefore, the input and output
components in each state of the HMM are defined as:
h
µi = [µzi , µai ]> ,
"
#
zah
Σzi
Σi t
Σi =
,
h
h
Σai z Σai
(6)
(7)
As such, the acceleration at time t can be estimated, given
the history information, ξ1:t−1 = [z1:t−1 , ah1:t−1 ] and the
observed state zt at time t, by using
âht =
N
X
h h
i
a
ah
αi (zt ) µi t + Σi t (Σzi t )−1 (zt − µzi t )
(8)
i=1
where αi (zt ) is the HMM forward variable, computed as the
probability of being in state i at time t, given by:
αj (zt−1 ) · φj,i · Ni (zt ; µzi , Σzi )
αi (zt ) = P P
N
N
α
(z
)
·
φ
· Nl (zt ; µzl , Σazl )
j
t−1
j,i
l=1
j=1
(9)
Here, the initial value at time t = 1 is computed by
P
N
j=1
πi N (z1 ; µzi , Σzi )
αi (z1 ) = PN
z
z
k=1 πk N (z1 ; µk , Σk )
C. Probability Density Function
The second approach to get the estimated output, âht is to
compute the value that can maximize the probability based
on the probability density function of the GMM, i.e.,
âht = arg max P(zt , ah ; θ̂)
ah ∈Ah
(10)
where Ah is the set of possible value that ah can reach and
θ̂ is the estimated parameter of the GMM using the collected
driving data on the basis of (5).
IV. E XPERIMENTS FOR DATA C OLLECTION
In this section, the data collection and the procedure of
data training and test are presented.
A. Data Collection
The data used in this paper is from the Safety Pilot Model
Deployment (SPMD) database [26]. It recorded naturalistic
driving of 2,842 equipped vehicles in Ann Arbor, Michigan
for more than two years. In the SPMD program, 98 sedans
are equipped with data acquisition system and MobilEyer
[13], [27], which provides: a) relative position to the lead
vehicle (range), and b) lane tracking measures about the
lane delineation both from the painted boundary lines and
the road edge. The error of range measurement is around
10% at 90 m and 5% at 45 m [28]. Data in two separate
months, October 2012 and April 2013, were downloaded
from the U.S. Department of Transportation website [29].
To ensure consistency of the used dataset, we apply the
following criteria to extracting the car-following events from
the entire datasets:
• ∆x ∈ [0.1 m, 120 m]
• Longitude ∈ [88.2, 82.0]
• Latitude ∈ [41.0, 44.5]
• Duration of car-following > 50 s
All the car-following events were detected from 76 drivers.
To the end, the number of entire purified car-following events
is 5,294.
B. Data Training Process
1) Preprocessing: For the jth driver, all the raw data, ξ j ,
were smoothed by a moving average filter with a window
size W = 10. The data for each single driver were evenly
divided into M groups and then M −1 groups were randomly
selected as the training data and the remaining group was
used to test the model, which is also called the leave-one-out
cross-validation method. Here, all the divided data groups for
each single driver meet the following conditions:
[
ξ j,p = ξ j and
p=1
\
ξ j,p = ∅, with p = 1, 2, · · · , M,
p=1
(11)
j,p
th
th
where
ξ
presents
the
p
group
of
data
for
the
j
driver,
S
T
and
are union and intersection, respectively; ∅ is the
empty set. In this paper, we set M = 20.
Errors of accelerations e7 [m/s2 ]
Mimimum of deceleration [m/s2 ]
0
-1
-2
-3
-4
-5
-6
-7
-8
3 & test
3 & train
4 & test
4 & train
5 & test
5 & train
6 & test
6 & train
0.16
0.14
0.12
0.1
0.08
0.06
-9
-10
GMM & HMM
0.18
0
5
10
15
20
25
Number of GMM components N
0
2
4
6
8
10
Maximum of acceleration [m/s ]
Fig. 2. The example of maximum and minimum accelerations for 75
drivers in our experiments.
2) Dimension of Model Inputs: We will investigate the
influence of different inputs on the model performance. For
the personalized driver model, the different input variables
are tested using the following combinations:
(1)
• zt
= [∆xt , ∆vt ]; ξ (1) = [zt , aht ]> ∈ R3×1 ;
(2)
• zt
= [∆xt , ∆vt , vth ]; ξ (2) = [zt , aht ]> ∈ R4×1 ;
(3)
• zt
= [∆xt , ∆vt , ∆v̇t , vth ]; ξ (3) = [zt , aht ]> ∈ R5×1 ;
(4)
• zt
= [∆xt , ∆vt , ∆v̇t , ht , vth ]; ξ (4) = [zt , aht ]> ∈
6×1
R .
(i)
where zt , i = 1, 2, 3, 4, represents the ith input. Here, we
default that the host vehicle speed, vth , and relative range,
∆xt , at current time t are the basic parameters for describing
a driver’s car-following behavior. In the training procedure,
the training data are ξ1:t−1 = [z1:t−1 , ah1:t−1 ].
3) Number of the GMM Components: Different numbers
of the GMM components will affect the model accuracy.
More components will cause the over-fitting problem, and
fewer components could not characterize the underlying
sources of data and will increase the prediction error. Therefore, N ∈ {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 15, 20,
25} are selected to investigate the influences of the GMM
components on model performance.
C. Data Testing Process
We will repeatedly run 10 times for each training dataset
of a driver participant and the average errors of 10 runs
is selected as the performance index to evaluate the model
performance. We run 10 times for each training dataset is
because the initial value used in (9) is generated by using
k-means cluster (k-MC) method in which the initial value
was randomly generated.
For the reachable region, Ah in (10), we set Ah =
{ah |ahmin ≤ ah ≤ ahmax }. The ahmin and ahmax can be
generated from the statistical information of each driver, as
shown in Fig. 2. In Fig. 2, 75 driver participants are included
and each point represents a driver. For most drivers, the ahmin
and ahmax are located at [−8, 8] m/s2 . Therefore, in our work,
for all drivers we set ahmin = −8 m/s2 and ahmax = +8
Errors of accelerations e7 [m/s2 ]
2
3 & test
3 & train
4 & test
4 & train
5 & test
5 & train
6 & test
6 & train
GMM & PDF
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0
5
10
15
20
25
Number of GMM components N
Fig. 3. The training errors and test errors for GMM+HMM approach and
GMM+PDF approach with different input dimensions.
m/s2 . Therefore, when inputing zt , we can obtain an locally
optimal corresponding estimated output âht using (10).
D. Performance Index
The average errors, ē, between the real value (aht ) and the
estimated value (âht ) are used as the performance index to
evaluate the presented methods and computed by
ē =
1
Z
tend
e(τ )dτ
tend 0
Z t
1
|âh − ahτ |dτ
=
tend 0 τ
(12)
where tend is the length of time-indexed test data. A smaller
(larger) value of ē indicates a better (undesirable) performance for the proposed approaches.
V. R ESULTS A NALYSIS
In this section, the training and test results with respect
to different input variables and numbers of GMM component based on two approaches, i.e., GMM+HMM and
GMM+PDF, are presented and discussed. To simplify the
description and show more clear, we take one of 75 driver
participants for example.
0.17
For the different number of GMM components, the training and test accuracy of the model will be different. More
components will decrease the training errors, but can result
in over-fitting problems and increase computational costs; inversely, fewer components can reduce computational efforts
but may induce larger errors. For example, Fig. 3 shows
the average errors of training and test results with different
numbers of GMM components using different approaches
for a driver. The horizontal and vertical axis are the number
of GMM component and average errors of acceleration, respectively. The number represents the dimension of training
data, as discussed in Section IV, B. For example, “5 &
train” represents the dimension of training input is 5, i.e.,
ξ = [∆xt , ∆vt , ∆v̇t , vth , aht ] = [zt , aht ]> ∈ R5×1 , and,
correspondingly, “5 & test” represents the input dimension
of test data is 4, i.e., zt = [∆xt , ∆vt , ∆v̇t , vth ].
1) GMM+HMM: Top plot in Fig. 3 shows the training and
test average errors of acceleration using the GMM+HMM
approach. It is obviously that the training errors are decreasing with the number of GMM components increasing.
The test errors are decreasing with the number of GMM
components increasing from 2 to 10; after that, the test errors
are increasing slightly.
2) GMM+PDF: Similarly, the bottom plot in Fig. 3 shows
the training and test errors of acceleration using GMMPDF approach. It can be concluded that the training errors
decreases and the test errors of acceleration increases while
the number of GMM increases.
0.16
B. Influence of Model Inputs
1) GMM+HMM: From the top plot of Fig. 3, we can
know that for different kinds of input by using GMM+HMM
approach, the training errors are decreasing with a higher
dimension of input, but not for the test errors. In addition,
for the test results using GMM+HMM approach while the
dimension of training data is 4, i.e., ξ = [∆xt , ∆vt , vth , aht ]> ,
we found that the estimation accuracy is better than others.
2) GMM+PDF: From the bottom plot of Fig. 3, it
can be seen that for different dimensions of training data
with the GMM+PDF approach, the training errors are decreasing with the dimension of training data increasing,
and the same case occurs for the test errors. For the
GMM+PDF approach, the estimation accuracy is the best
when the 6-dimension of training data is chosen, i.e., ξ =
[∆xt , ∆vt , ∆v̇t , ht , vth , aht ]> .
C. Comparison Between Two Methods
The comparison results between two methods are shown in
Fig. 4. It is obvious that for different dimensions of training
data (i.e., ξ ∈ Rd×1 , d = 3, 4, 5, 6), the GMM+HMM approach has a higher estimation accuracy than the GMM+PDF
approach. For the GMM+HMM method, the mean estimation
errors, ē, can be lower than 0.1, but for the GMM+PDF
method, ē is always larger than 0.1, even for different numbers of the GMM components and dimensions of training
data.
Errors of accelerations e7 [m/s2 ]
A. Influence of the GMM Component
GMM+PDF & GMM+HMM
3 & GMM+PDF
4 & GMM+PDF
5 & GMM+PDF
6 & GMM+PDF
3 & GMM+HMM
4 & GMM+HMM
5 & GMM+HMM
6 & GMM+HMM
0.15
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0
5
10
15
20
25
Number of GMM components N
Fig. 4.
The comparison between the GMM+HMM and GMM+PDF
approaches with different input dimensions.
2
2
GMM+HMM
GMM+PDF
Base line
Acceleration[m/s 2]
1.5
1
0.5
0
-2
740
745
750
0
-0.5
0.2
0
-1
-0.2
-1.5
-2
700
-0.4
780
750
785
800
790
795
850
800
900
Time[s]
Fig. 5.
The comparison of acceleration prediction between the
GMM+HMM and GMM+PDF approaches with N = 12 GMM components
and 4 input variables.
Fig. 5 shows the estimation results with two different
methods when the dimension of training data is 4 and the
number of components is 12. We note that the GMM + PDF
method has a higher potential to increase the model accuracy
given a higher dimension of training data.
VI. C ONCLUSIONS AND F UTURE W ORKS
This paper proposed and compared two personalized driver
models in car-following scenarios. The GMM+HMM method
(Gaussian mixture model + hidden Markov model) and the
GMM+PDF method (Gaussian mixture model + probability
density function) were used to fit large-scale naturalistic driving data to describe the uncertainties and nonlinearities of the
human behaviors. Different GMM components and training
data dimensions was tested out and their influences on the
model accuracy were analyzed. For training a personalized
car-following driver model, we found that:
• For the GMM + HMM method, a higher dimension of
the training data might not result in a higher estimation
accuracy. The preferred number of the GMM components is 10 ∼ 15 and the preferred dimension of training
•
•
data is 4, including host vehicle speed, relative range,
relative speed, and the acceleration of the host vehicle.
For GMM + PDF methods, a higher dimension of the
training data can slightly reduce the estimation errors of
acceleration but will increase the computational cost.
In the car-following case, the GMM + HMM method
can catch the underlying sources of naturalistic driving
data and shows a better prediction performance than
GMM + PDF method by about 27.3%.
The Gaussian mixture model is a popular tool to generate
a statistical model due to its flexibility and simplicity for
learning, but it is sensitive to outliers especially with small
numbers of data points. Also, due to the bounded nature of
driving behaviors, tails of the Gaussian distributions might
be shorter than required, which affects the fitting accuracy. In
the future work, we will take the bounded feature of driver
behaviors into consideration and develop a learning-based
bounded driver model.
R EFERENCES
[1] S. Lefèvre, A. Carvalho, Y. Gao, H. E. Tseng, and F. Borrelli, “Driver
models for personalised driving assistance,” Vehicle System Dynamics,
vol. 53, no. 12, pp. 1705–1720, Oct. 2015.
[2] V. A. Butakov and P. Ioannou, “Personalized driver/vehicle lane
change models for ADAS,” IEEE Transaction on Intelligent Transportation Systems, vol. 64, no. 10, pp. 4422 – 4431, Oct. 2015.
[3] V. A. Butakov and P. Ioannou, “Personalized driver assistance for
signalized intersections using V2I communication,” IEEE Transactions
on Intelligent Transportation Systems, vol. 17, no. 7, pp. 1910 –1919,
Jul. 2016.
[4] X. Fu and D. Soeffker, “Modeling of individualized human driver
model for automated personalized supervision,” SAE Technical Paper,
Tech. Rep., 2010.
[5] W. Wang, J. Xi, and J. Wang, “Human-centered feed-forward control
of a vehicle steering system based on a driver’s steering model,” in
2015 American Control Conference (ACC). IEEE, Jul. 2015, pp.
3361–3366.
[6] W. Wang, J. Xi, and H. Chen, “Modeling and recognizing driver
behavior based on driving data: a survey,” Mathematical Problems
in Engineering, vol. 2014, 2014.
[7] W. Wang, J. Xi, C. Liu, and X. Li, “Human-centered feed-forward
control of a vehicle steering system based on a driver’s path-following
characteristics,” IEEE Transactions on Intelligent Transportation Systems, DOI:10.1109/TITS.2016.26063.
[8] X. Xiang, K. Zhou, W.-B. Zhang, W. Qin, and Q. Mao, “A closedloop speed advisory model with driver’s behavior adaptability for ecodriving,” IEEE Transactions on Intelligent Transportation Systems,
vol. 16, no. 6, pp. 3313–3324, Dec. 2015.
[9] S. Lefèvre, A. Carvalho, and F. Borrelli, “A learning-based framework
for velocity control in autonomous driving,” IEEE Transactions on
Automation Science and Engineering, vol. 13, no. 1, pp. 32 – 42, Jan.
2016.
[10] C. Miyajima, Y. Nishiwaki, K. Ozawa, T. Wakita, K. Itou, K. Takeda,
and F. Itakura, “Driver modeling based on driving behavior and its
evaluation in driver identification,” Proceedings of the IEEE, vol. 95,
no. 2, pp. 427–437, 2007.
[11] S. Eben Li, K. Li, and J. Wang, “Economy-oriented vehicle adaptive
cruise control with coordinating multiple objectives function,” Vehicle
System Dynamics, vol. 51, no. 1, pp. 1–17, Jan. 2013.
[12] D. Zhao, H. Lam, H. Peng, S. Bao, D. J. Leblanc, and C. S.
Pan, “Accelerated Evaluation of Automated Vehicles Safety in
Lane Change Scenarios based on Importance Sampling Techniques,”
IEEE Transactions on Intelligent Transportation Systems, DOI:
10.1109/TITS.2016.2582208.
[13] D. Zhao, X. Huang, H. Peng, H. Lam, and D. J. Leblanc, “Accelerated
Evaluation of Automated Vehicles in Car-Following Maneuvers,”
arxivId 1607.02687, 2016.
[14] G.-h. Peng and R.-j. Cheng, “A new car-following model with the
consideration of anticipation optimal velocity,” Physica A: Statistical
Mechanics and its Applications, vol. 392, no. 17, pp. 3563–3569, Sep.
2013.
[15] P. J. Jin, D. Yang, and B. Ran, “Reducing the error accumulation in
car-following models calibrated with vehicle trajectory data,” IEEE
Transactions on Intelligent Transportation Systems, vol. 15, no. 1, pp.
148–157, 2014.
[16] M. Rahman, M. Chowdhury, T. Khan, and P. Bhavsar, “Improving
the efficacy of car-following models with a new stochastic parameter
estimation and calibration method,” IEEE Transactions on Intelligent
Transportation Systems, vol. 16, no. 5, pp. 2687–2699, 2015.
[17] S. Sekizawa, S. Inagaki, T. Suzuki, S. Hayakawa, N. Tsuchida,
T. Tsuda, and H. Fujinami, “Modeling and recognition of driving
behavior based on stochastic switched arx model,” IEEE Transactions
on Intelligent Transportation Systems, vol. 8, no. 4, pp. 593–606, Dec.
2007.
[18] O. Celik and S. Ertugrul, “Predictive human operator model to be
utilized as a controller using linear, neuro-fuzzy and fuzzy-arx modeling techniques,” Engineering Applications of Artificial Intelligence,
vol. 23, no. 4, pp. 595–603, Jun. 2010.
[19] M. C. Nechyba and Y. Xu, “On learning discontinuous human control
strategies,” International journal of intelligent systems, vol. 16, no. 4,
pp. 547–570, Apr. 2001.
[20] A. Khodayari, A. Ghaffari, R. Kazemi, and R. Braunstingl, “A
modified car-following model based on a neural network model
of the human driver effects,” IEEE Transactions on Systems, Man,
Cybernetics: Systems and Humans, vol. 42, no. 6, pp. 1440 –1449,
Nov. 2012.
[21] A. Wahab, C. Quek, C. K. Tan, and K. Takeda, “Driving profile
modeling and recognition based on soft computing approach,” IEEE
transactions on neural networks, vol. 20, no. 4, pp. 563–582, Apr.
2009.
[22] V. Butakov, P. Ioannou, M. Tippelhofer, and J. Camhi, “Driver/vehicle
response diagnostic system for vehicle following based on gaussian
mixture model,” in 2012 IEEE 51st IEEE Conference on Decision
and Control (CDC). IEEE, 2012, pp. 5649–5654.
[23] P. Angkititrakul, C. Miyajima, and K. Takeda, “Modeling and adaptation of stochastic driver-behavior model with application to car
following,” in Intelligent Vehicles Symposium (IV), 2011 IEEE. IEEE,
2011, pp. 814–819.
[24] D. Zhao, H. Peng, S. Bao, K. Nobukawa, D. J. LeBlanc, and C. S. Pan,
“Accelerated evaluation of automated vehicles using extracted naturalistic driving data,” in Proceeding for 24th International Symposium of
Vehicles on Road and Tracks, 2015.
[25] W. Wang, D. Zhao, J. Xi, and W. Han, “A learning-based approach
for lane departure warning systems with a personalized driver model,”
arXiv preprint arXiv:1702.01228, 2017.
[26] D. Bezzina and J. R. Sayer, “Safety Pilot: Model Deployment
Test Conductor Team Report,” no. June, 2014. [Online]. Available:
http://safetypilot.umtri.umich.edu/
[27] J. Harding, G. Powell, R. Yoon, J. Fikentscher, C. Doyle, D. Sade,
M. Lukuc, J. Simons, and J. Wang, “Vehicle-to-vehicle communications: Readiness of v2v technology for application,” Tech. Rep., 2014.
[28] G. Stein, O. Mano, and A. Shashua, “Vision-based ACC with a single
camera: bounds on range and range rate accuracy,” pp. 120–125, 2003.
[29] RDE Data Environment, “Safety Pilot Model Deployment
Data.” [Online]. Available: https://www.its-rde.net/data/showds?
dataEnvironmentNumber=10018
| 3cs.SY
|
1
Multiple Instance Fuzzy Inference Neural Networks
Amine B. Khalifa, Hichem Frigui, Member, IEEE,
Multimedia Research Lab
CECS Department
University of Louisville
Louisville, KY 40292, USA
arXiv:1610.04973v1 [cs.NE] 17 Oct 2016
Abstract
Fuzzy logic is a powerful tool to model knowledge uncertainty, measurements imprecision, and vagueness.
However, there is another type of vagueness that arises when data have multiple forms of expression that fuzzy logic
does not address quite well. This is the case for multiple instance learning problems (MIL). In MIL, an object is
represented by a collection of instances, called a bag. A bag is labeled negative if all of its instances are negative,
and positive if at least one of its instances is positive. Positive bags encode ambiguity since the instances themselves
are not labeled. In this paper, we introduce fuzzy inference systems and neural networks designed to handle bags of
instances as input and capable of learning from ambiguously labeled data. First, we introduce the Multiple Instance
Sugeno style fuzzy inference (MI-Sugeno) that extends the standard Sugeno style inference to handle reasoning
with multiple instances. Second, we use MI-Sugeno to define and develop Multiple Instance Adaptive Neuro Fuzzy
Inference System (MI-ANFIS). We expand the architecture of the standard ANFIS to allow reasoning with bags
and derive a learning algorithm using backpropagation to identify the premise and consequent parameters of the
network. The proposed inference system is tested and validated using synthetic and benchmark datasets suitable for
MIL problems. We also apply the proposed MI-ANFIS to fuse the output of multiple discrimination algorithms for
the purpose of landmine detection using Ground Penetrating Radar.
I. I NTRODUCTION
Fuzzy inference is a powerful modeling framework that can handle computing with knowledge uncertainty
and measurements imprecision effectively [1]. It has been successfully applied to a wide range of problems,
mainly in system modeling and control [2]–[4]. Most of the proposed fuzzy inference methods gained
success because of their ability to leverage expert knowledge to identify the model parameters [5]. This
practice simplifies system design and ensures that the knowledge base (if-then rules) used by the system is
easy to interpret [6].
More recently, fuzzy inference has increasingly been applied to more advanced applications, such as
content-based information retrieval [7], image segmentation [8], image annotation [9], pattern recognition
[10], recommender systems [11], and multiple classifier fusion [12]. The aforementioned applications are
more challenging as they require extensive knowledge base to accommodate for various scenarios. Since
this diverse knowledge base cannot be fully captured by domain experts, data-driven techniques are typically
used to identify and learn the inference system’s parameters [13], [14]. One such technique is the Adaptive
Neuro-Fuzzy Inference System (ANFIS) [15]. ANFIS is a universal approximator that combines the learning
and modeling power of neural networks and fuzzy logic into an adaptive inference system. It is a hybrid
intelligent system and it provides a systematic approach to jointly learn the optimal input space partition
(rules) and the optimal output parameters using supervised learning.
Typically, in supervised learning, access to large labeled training datasets improves the performance of
the devised algorithms by increasing their robustness and generalization capabilities. Nowadays, access to
such large datasets is becoming more convenient. However, for a supervised leaning method to benefit from
this data, it need to be carefully preprocessed, filtered, and labeled. Unfortunately, this process can be too
Email: [email protected]
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may
no longer be accessible.
2
tedious as the vast portion of the collected data is unstructured, labeled ambiguously and at a coarse level.
An alternative and a relatively new framework of learning that tackles the inherent ambiguity better than
supervised learning, is the Multiple Instance Learning (MIL) paradigm [16].
A. Multiple Instance Learning
Unlike standard supervised learning, in MIL, an object is not represented by a simple data point, but
rather by a collection of instances, called a bag. Each bag can contain a different number of instances.
A bag is labeled negative if all of its instances are negative, and positive if at least one of its instances
is positive1 . Positive bags can encode ambiguity since the instances themselves are not labeled. Given a
training set of labeled bags, the goal of MIL is to learn a concept that predicts the labels of training data
at the instance level and generalizes to predict the labels of testing bags and their instances [17]. We refer
to this definition as the standard MIL assumption. Multiple MIL paradigms have been proposed [18], but
for simplicity we focus our formulation on the standard MIL assumption.
The MIL is a well known problem that has been studied for the last 20 years, it was first formalized by
Dietterich et al. [19] providing a solution to drug activity prediction. Ever since, it has increasingly been
applied to a wide variety of tasks including content-based information retrieval [20], drug discovery [21],
pattern recognition [22], image classification [23], region-based image categorization [24], image annotation
[25], object tracking [26] and time series prediction [16]. In general, MIL can be applied in two contexts
of ambiguity: “polymorphism ambiguity” and “part-whole ambiguity” [27]. In polymorphism ambiguity,
an object can have multiple forms of expression in the input space and it is not known which form is
responsible for the object label. Whereas, in part-whole ambiguity, an object can be broken into several
parts represented by different feature vectors in the input space. However, only few parts are responsible for
the object label [28]. Polymorphism Ambiguity arise more often in applications related to chemistry and
bioscience. The original MIL application of drug discovery [16], [17] is a case of polymorphism ambiguity.
Part-whole Ambiguity is more common in pattern recognition problems. For example, in image annotation
features are usually extracted locally (from patches) while the labels, or tags, are only available gloablly at
the image level. Another closely related application is object detection. In this application, objects of interest
may cover only a limited region of the image, the rest could be other objects or background. Traditional
supervised learning requires identifying image patches containing the object of interest only and labeling
them. As indicated by Viola et al. [29], placing bounding boxes around objects is an inherently ambiguous
task. Thus, to avoid the tedious task of object segmentation and annotation, the problem of object detection
can be addressed using an MIL paradigm. To illustrate the need for MIL further, in the following we analyze
how a multiple instance (MI) representation can be applied to image classification. More details about MIL
taxonomy have been reported by Amores [30].
Consider the simple example of classifying images that contain “sky”. Using an MIL approach, each
training image is represented by a bag of instances where each instance corresponds to features extracted
from a region of interest. These regions could be obtained by segmenting the image or simply by dividing
it into patches. A multiple instance representation is well suited for this purpose because only few regions
may contain the object of interest (sky), that is the positive class. Other patches will be from background
or other classes. This representation is illustrated in Figure 1. Traditional single instance learning are based
on instance level (patch-level) labels and would require each image region to be correctly segmented and
labeled prior to learning.
B. Fuzzy Inference Systems
A Fuzzy Inference System (FIS) is a paradigm in soft computing which provides a means of approximate
reasoning [31]. A FIS is capable of handling computing with knowledge uncertainty and measurements
imprecision effectively [1]. It performs a non-linear mapping from an input space to an output space by
1
Note that positive bags may also contain negative instances.
3
Fig. 1. Example of an image represented as a bag of 12 instances. Each instance correspond to a feature vector (e.g., color, texture) extracted
from one patch. The bag is labeled “sky” because at least one of its instances is sky. However, many other instances are not “sky”. Labels at
the instance level are not available.
deriving conclusions from a set of fuzzy if-then rules and known facts [32]. Fuzzy rules are condition/action
(if-then) rules composed of a set of linguistic variables (e.g. image patch). Each variable is assigned a
linguistic term (e.g. red, green, blue). For instance, the following rules could be used to identify patches
from the image in Figure 1:
• If patch is blue and texture is smooth then region is sky.
• If patch is blue and patch position is upper half then region is sky.
Typically, a FIS is composed of 5 components: (1) a Fuzzification unit that assigns a membership degree
to each crisp input dimension in the input fuzzy sets; (2) a Knowledge Base characterized by fuzzy sets of
linguistic terms; (3) a Rule Base containing a set of fuzzy if-then rules; (4) an Inference unit that performs
fuzzy reasoning; and (5) a Deffuzification unit that generates crisp output values. FIS has proven to be very
effective in various applications [2]–[4], [33]–[40]. However, it is not applicable to cases where objects are
represented by multiple instances.
C. Motivations For Multiple Instance Fuzzy Inference
There are two major limitations that prevent using standard FIS methods with multiple instance data.
First, due to the absence of labels at the instance level, we cannot use standard FIS learning methods to
construct the knowledge base. Second, we need an effective mechanism to aggregate instances’ confidences
and infer at the bag level. The above limitations are due mainly to the inherent architecture of fuzzy inference
systems. The standard inference systems reason with individual instances. First, the system’s input is an
individual instance. Second, the rules describe fuzzy regions within the instances space. Third, the output
of the system corresponds to the fuzzy inference using a single instance. Fourth, labels of the individual
instances are required when using learning techniques to identify the parameters of the system. In summary,
traditional fuzzy inference systems cannot be used effectively within the MIL framework.
To address the above limitations, we introduce two FIS designed to handle reasoning with bags of
instances and capable of learning form ambiguously labeled data. The first one, called Multiple InstanceSugeno (MI-Sugeno) extends the standard Sugeno system [41]. The second one, called Multiple InstanceANFIS (MI-ANFIS) extends the standard ANFIS [15] system and uses MI-Sugeno rules. We report results
on various experiments and discuss the advantages of using our proposed methods over closely related MIL
4
algorithms such as Multiple Instance Neural Networks [42] (MI-NN) and Multiple Instance RBF Neural
Networks [43] (RBF-MIP).
II. M ULTIPLE I NSTANCE F UZZY I NFERENCE
In the following, let Bp be a bag of Mp instances with the jth instance denoted as xpj ∈ RD with
elements x(p,j,k) corresponding to features, i.e.,
xp1
x(p,1,1)
x(p,1,2) . . . x(p,1,D)
xp2 x(p,2,1)
x(p,2,2) . . . x(p,2,D)
=
.
Bp =
(1)
.
.
..
..
...
..
..
.
.
xpMp
x(p,Mp ,1) x(p,Mp ,2) . . . x(p,Mp ,D)
Note that the number of instances can vary between bags (Mp depends on Bp ). A bag is labeled positive
if at least one of its instances is positive, and negative if all of its instances are negative.
A. Multiple Instance Sugeno Style Fuzzy Inference
To adapt Sugeno inference to problems where objects are described by multiple instances, we propose
a multiple instance Sugeno inference (MI-Sugeno) system that uses multiple instance fuzzy if-then rules.
Recall that a fuzzy if-then rule is expressed as
if x is A then y is C
(2)
where A and C are fuzzy sets on universes of discourse X and Y , respectively. The rule in (2) combines
the fuzzy propositions (x is A, y is C) into a logical implication abbreviated as A → C with membership
function µA→C (x, y). The rule is defined using a premise part that is a single instance fuzzy proposition.
To generalize the rule in (2) to MI data, we define a multiple instance fuzzy rule as:
if Bi is A then y is C ⇐⇒ if
Mi
_
(xij is A) then y is C
(3)
j=1
where as in (2), A and C are fuzzy sets on the universes of discourse X and Y , respectively. In (3), Bi is a
bag of instances xij as defined
WMi in (1), and Mi is the number of instances in Bi . The premise part of a multiple
instance fuzzy rule (i.e., j=1 (xijWis A) ) is a multiple instance proposition, whereas the consequent part is
a traditional proposition. In (3), is a joint operator that can be any T-conorm (maximum, algebraic sum,
bounded sum, etc.). The reason behind using a T-conorm for combining individual instances’ responses,
goes back to the standard MIL assumption [16], [17] which states that a bag is positive if and only if
one or more of its instances are positive. Thus, the bag-level class label is determined by the disjunction
of the instance-level class labels. We note that the T-conorm can be designed to handle a broader set of
non-standrad MIL problems, for example to allow the inference process to assign a higher degree of belief
to bags with more than one positive instance.
The proposed MI-Sugeno uses multiple instance fuzzy rules with a consequent part that is described by
means of a function C that maps a bag of instances to a crisp numerical value. Specifically, we define a
multiple instance sugeno rule as:
i
R (Bp ) :
Mp
_
( If x(p,j,1) is Ai1 and x(p,j,2) is Ai2 . . . and x(p,j,D) is AiD ), then
j=1
(4)
oi = C(xp1 · bi , xp2 · bi , . . . , xpMp · bi )
In (4), bi = bi0 , ..., biD is a set of polynomial coefficients. When the polynomial coefficients bi are first
order, the MI-Sugeno fuzzy model is called first order, and zero order when the polynomial coefficients
5
Fig. 2.
Illustration of the proposed multiple instance Sugeno fuzzy inference system with 2 rules.
are zero order.
Figure 2 illustrates the proposed MI-Sugeno system and its fuzzy inference mechanism to derive the output,
o, in response to a bag of M instances for the simple case of two rules. The premise part of the rules
evaluates all the bag’s instances simultaneously. The inference starts by the fuzzification of instances xpm of
input bag Bp . Fuzzification assigns a membership degree to each input instance dimension in the rules input
fuzzy sets. In Figure 2, instance xpm activates the ith input fuzzy set of the jth rule by a degree of truth
w(m,i,j) . Next, an implication process is executed to combine the activations of the instances within the bag
resulting in the activation of the rules’ output with different degrees. In this example, we use a simple min
operator, and the output of rule Rj will be partially activated by a degree wmj = mink=1,...,D w(m,k,j) . The
wmj (truth instances) are combined in the premise part using the max T-conorm, resulting in the activation
of rule Rj by a degree wj = maxm=1,...,M {wmj }. To evaluate the consequent part, first the linear response of
each instance is computed, i.e., xpj ·bi . Then, a function C is used to compute the final output by combining
the instances’ responses. Many functions could be used and the choice should be domain-specfic. The output
of each rule, o1 and o2 , are crisp values. As in the traditional Sugeno fuzzy inference system, the overall
output of the system is obtained by taking the weighted average of the rules’ outputs.
The consequent part of the proposed MI-Sugeno style inference system is inspired by the work of Ray
and Page on multiple instance regression [44]. In their work, the authors proposed a regression framework
for predicting bags’ labels. This formulation allows the linear coefficients bi and the parameters of the
combining function C to be learned using optimazation techniques, as we will show in section II-B.
Similar to traditional fuzzy inference, the premise part of a multiple instance rule defines a local fuzzy
region within the instance space, and the consequent part describes the characteristics of the system’s
output within each region. More specifically, in problems, a local region describes a positive concept (also
called target concept), and the output of a rule represents the degree of “positivity” of the instances in that
target concept. A target concept is a region in the instances’ feature space that includes as many instances
from positive bags as possible and as few instances from negative bags as possible.
The Sugeno fuzzy model [41] was the first attempt at learning fuzzy rules from training data. It has
6
Fig. 3.
Architecture of the proposed Multiple Instance Adaptive Neuro-Fuzzy Inference System
been used to develop the standard ANIFS which combines the representation power of fuzzy inference and
learning capability of neural networks to learn the rules. In the next section, we will use our MI-Sugeno
to develop a multiple instance extension of ANFIS (MI-ANFIS).
B. MI-ANFIS: A Multiple Instance Adaptive Neuro-Fuzzy Inference System
Let Bi be a bag of Mi instances as defined in (1). For simplicity, we introduce our MI-ANFIS for the
case of two rules. The generalization to an arbitrary number of rules is trivial. The MI-ANFIS with two
Sugeno rules can be described as:
1
R (Bp ) :
Mp
_
( If x(p,j,1) is A(1,1) and x(p,j,2) is A(1,2) , . . . , and x(p,j,D) is A(1,D) ),
j=1
then f1 = C(xp1 · b1 , xp2 · b1 , . . . , xpMp · b1 )
R2 (Bp ) :
Mp
_
(5)
( If x(p,j,1) is A(2,1) and x(p,j,2) is A(2,2) , . . . , and x(p,j,D) is A(2,D) ),
j=1
then f2 = C(xp1 · b2 , xp2 · b2 , . . . , xpMp · b2 )
Figure 3 illustrates the proposed MI-ANFIS architecture. As in the traditional ANFIS, nodes at the same
layer have similar functions. We denote the output of the ith node in layer l as O(l,i)
Layer 1 is an adaptive layer, it calculates the degree to which an input instance satisfies a quantifier
A. Every node evaluates the membership degree µA(k,j) of an input instance, x, in the fuzzy set
A(k,j) . Generally, µA(k,j) is a parameterized membership function (MF), for example a Gaussian
MF with
−(x − ckj )2
µA(k,j) (x) = exp(
).
(6)
2σkj 2
In (6), ckj and σkj are the mean and variance of the gaussian function, and are referred to as the
premise parameters.
7
Layer 2 is a fixed layer where every node computes the product of all incoming inputs. It evaluates the
degree of truth of proposition instances, or simply, “truth instances”. The output of this layer is
computed using:
D
Y
O(2,i) = r
=
µA (x(p,i[Mp ],j) ).
(7)
i/Mp ,i[Mp ]
j=1
i/Mp
,j
where
is a ceiling operator, and i[Mp ] is 1 + ((i − 1) mod Mp ). As in the traditional ANFIS,
any T-norm can replace the product as the node function in this layer.
Layer 3 is a new addition when compared to the traditional ANFIS. Every node in this layer aggregates
the truth instances (within each bag) of the previous layer by means of a smooth T-conorm. In
this paper, we use a “softmax” function (Sα ):
sof tmaxα (x1 , x2 , . . . , xn ) = Sα (x1 , x2 , . . . , xn ) =
n
X
x · eαxi
Pin αx .
j
j=1 e
i=1
(8)
In (8), α determines the behavior of softmax. As α approaches ∞, softmax approaches the max
operator. When α = 0, it calculates the mean. As α approaches −∞, softmax approaches the
min operator. The outputs of this layer are the firing strength of each input bag in each multiple
instance fuzzy rule. i.e.,
Mp
).
(9)
O(3,i) = wi = Sα ({r(i,j) }j=1
Layer 3 is also a fixed layer.
Layer 4 is a fixed layer. Every node in this layer calculates the normalized firing strength of each
rule, i.e.,
wi
(10)
O(4,i) = wi = P|O3 | .
w
j
j=1
where |O3 | is the number of rules.
Layer 5 is an adaptive layer. Every node i in this layer computes the output of the ith multiple instance
rule using
O(5,i) = wi C(xp1 · bi , xp2 · bi , . . . , xpMp · bi ).
(11)
|O |
The parameters {bi }i=13 are referred to as the consequent parameters. The only constraint on
C is that it has to be a smooth function to allow for optimization techniques to be applied. We
use the “softmax” as the combining function for this layer. In this case, (11) is equivalent to:
O(5,i) = wi Sα (xp1 · bi , xp2 · bi , . . . , xpMp · bi ).
(12)
note that the constant α here is not necessary the same as in Layer 3.
Layer 6 is a fixed layer with a single node labeled Σ. It computes the overall output of the system
using
|O3 |
|O3 |
X
X
wi Sα (xp1 · bi , xp2 · bi , . . . , xpMp · bi ).
O(6,1) =
O5,i =
(13)
i=1
i=1
To learn the parameters of the proposed MI-ANFIS network, we propose a generalization to the basic
learning algorithm presented by Jang [45]. Our variation is different from the ANFIS standard backpropagation learning rule due to the additional layers (Layers 3 and 5) and the use of “softmax” function (in (9)
and (11)). Thus, all update equations need to be rederived.
BackPropagation Learning Rule: we assume that we have N training bags, B = {Bp | p = 1, . . . , N },
and their corresponding labels T = {tp | p = 1, . . . , N }. After presenting the pth training bag, we compute
its squared error measure:
Ep = (tp − Op )2 .
(14)
8
In (14), tp is the desired bag output, and Op is the computed output of the network when presented with
training bag Bp . Recall that labels at the instances level are not available and errors can be computed only
at the bag level.
The overall error measure of the network after presenting all N bags is
E=
N
X
Ep .
(15)
p=1
To develop the gradient descent optimization on E, we compute the error rate for the pth training bag at
each output node O(l,i) . This error rate ε(l,i) (where 1 ≤ l ≤ 6 indicates the MI-ANFIS layer) is defined as
εl,i =
∂Ep
.
∂O(l,i)
(16)
At the output node, we have
ε(6,1) =
∂Ep
∂Ep
=
= −2(tp − Op ).
∂O(6,1)
∂Op
(17)
For non-output nodes (i.e. internal nodes, l < 6), we derive the error rate using the chain rule
ε(l,i)
∂Ep
=
=
∂O(l,i)
Card(l+1)
X
h=1
∂Ep ∂O(l+1,h )
.
∂O(l+1,h) ∂O(l,i)
(18)
where Card(l + 1) refers the number of nodes at layer l + 1.
Next, we seek to minimize the network error with respect to the premise parameters {ckj , σkj | 1 ≤ k ≤
|O |
|O3 |, 1 ≤ j ≤ D}, and with respect to the consequent parameters {bi }i=13 .
The error rate with respect to a generic parameter θ can be computed using
X ∂Ep ∂O∗
∂Ep
=
.
(19)
∗ ∂θ
∂θ
∂O
∗
O ∈G
where G is the set of nodes whose outputs depend on θ.
Using(15), the total error rate is given by
N
∂E X ∂Ep
=
.
∂θ
∂θ
p=1
(20)
Update Rule For Premise Parameters: First we compute the error rate for the premise parameters ckj
and σkj using
Mp
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂Ep X
∂Ep
=
.
(21)
∂ckj
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂ckj
i=1
and,
Mp
X
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂Ep
∂Ep
=
.
∂σkj
∂O
∂σ
kj
(1,i+[(k−1)D+(j−1)]M
)
p
i=1
(22)
9
Using the chain rule defined in (18), it can be shown that (see derivation in Appendix A)
P|O3 |
wl − wk
∂Ep
k
k
k
= −2(tp − Op ) × Sα (xp1 · b , xp2 · b , . . . , xpMp · b ) × l=1
2
P
∂ckj
|O3 |
l=1 wl
Mp
X
i
eαr(k,(i+(k−1)Mp )) h
Mp
1 + α r(k,(i+(k−1)Mp )) − Sα ({r(k,m) }m=1
)
PMp αr
(k,m)
m=1 e
i=1
D
Y
×
x
µA
(p,(i+(k−1)Mp )[Mp ],d)
×
d=1,d6=j
(i+(k−1)Mp )/Mp
(23)
,d
!
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )
× exp(−
) .
×
2
2
σkj
2σkj
The center parameters ckj are then updated using
4ckj = −η
∂E
.
∂ckj
(24)
where η is the learning rate.
The update formula for σkj can be derived in a similar manner. It can be shown that
P|O3 |
∂Ep
wl − wk
k
k
k
= −2(tp − Op ) × Sα (xp1 · b , xp2 · b , . . . , xpMp · b ) × l=1
2
P
∂σkj
|O3 |
w
l
l=1
Mp
X
i
eαr(k,(i+(k−1)Mp )) h
Mp
1 + α r(k,(i+(k−1)Mp )) − Sα ({r(k,m) }m=1
)
PMp αr
(k,m)
m=1 e
i=1
D
Y
µA
x
×
(p,(i+(k−1)M
)[M
],d)
p
p
×
d=1,d6=j
(i+(k−1)Mp )/Mp
(25)
,d
!
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
×
× exp(−
) .
3
2
σkj
2σkj
The MF’s width, σkj , are then updated using
4σkj = −η
∂E
.
∂σkj
(26)
Update Rule For Consequent Parameters: The error rate for the consequent parameters {bi = {bi0 , ..., biD }, i =
1 . . . |O3 |} is defined as
∂Ep
∂Ep ∂Ep
∂Ep
=
, i ,..., i .
(27)
i
i
∂b
∂b0 ∂b1
∂bD
where,
∂Ep
∂Ep ∂O(5,i)
=
, f or j = 1, . . . , D.
i
∂bj
∂O(5,i) ∂bij
(28)
10
Using (18), it can be shown that (see Appendix B)
Mp
N
N
∂E X ∂Ep X X
wi
=
=
i
∂bij
∂b
j
m=1
p=1
p=1
×
h
x(p,m,j)
Mp
X
1
P
Mp
h=1
2
exp(α(xph · bi − xpm · bi ))
!
Mp
i
X
exp(α(xph ·bi −xpm ·bi ) − xpm ·bi
exp(α(xph ·bi −xpm ·bi )α(x(p,h,j) −x(p,m,j) )
.
h=1
h=1
(29)
The consequent parameters are then updated using
4bij = −η
∂E
.
∂bij
(30)
Equations (24), (26) and (30) can be used to update ckj , σkj and bij parameters either on-line, bag by bag (
we want to emphasis here that the on-line learning is not achieved instance by instance, but rather bag by
bag), or off-line in batch mode after presentation of the entire data.
The proposed MI-ANFIS learning algorithm is summarized in Algorithm 1.
Algorithm 1 MI-ANFIS Basic Learning Algorithm
Inputs: B: the set of training bags.
T : labels of the training bags.
M : the number of instances in each bag.
α: the constant used in the “softmax” function.
η: the learning rate.
Emax : number of epochs.
: minimum parameters change value.
Outputs: bi : the sets of consequent parameters.
ci : the set of membership functions’ centers.
σ i : the set of membership functions’ widths.
Initialize bi , ci , and σ i .
repeat
Update bi using (30) and bi(new) = bi(old) + 4bi .
Update ci using (24) and ci(new) = ci(old) + 4ci .
Update σ i using (26) and σ i(new) = σ i(old) + 4σ i .
until max(kbi(new) − bi(old) k, kci(new) − ci(old) k, kσi(new) − σi(old) k) < or N umber of epochs > Emax
return bi , ci , σ i
III. P REVENTING OVERFITTING : RULE D ROPOUT
Neural networks with large number of parameters are susceptible to overfitting. MI-ANFIS is no exception, particularly when using large number of multiple instance fuzzy rules and relatively small training
datasets. In such scenario, some rules could co-adapt to the training data and degrade the network ability to
generalize to unseen examples. In this section, we present a technique, known as Dropout, used to prevent
overfitting and rules’ co-adaptation.
Dropout is a regularization method that was introduced by Hinton et al. [46] to alleviate the serious problem
of overfitting in deep neural networks. Over the years, many methods have been developed to reduce
overfitting, including using a validation dataset to stop the training as soon as the performance gets worse,
adding weight penalties using L1 and L2 regularization, or artificially augmenting the training dataset using
11
label-preserving transformations. However, as noted by Hinton [46], the best way to regularize a fixed-size
model is to average the predictions of all possible settings of the parameters weighted by its posterior
probability given the training data. This can be achieved by combining the predictions of an exponential
number of models. Combining several models with different architectures have the advantage of better
generalization and per consequence better testing performance. While generating an ensemble of models is
trivial, training them all is prohibitively expensive.
Generally, Dropout works by setting to 0 the output of each node in a given layer with probability 1 − p
(p typically equals 0.5), during training. Nodes that are dropped out do not contribute to the parameters
updates. During testing, all nodes are used but the outputs are weighted by the probability p. Following
this strategy, every time a new training example is presented, the network samples and trains a different
architecture. In other words, Dropout trains an ensemble of networks (2N networks, N being the number of
nodes) simultaneously leading to an important speedup in training time as compared to traditional ensemble
methods. Figure 4 and Figure 5 illustrate the Dropout model.
Fig. 4. Dropout neural network model. (a) is a standard neural network. (b) is the same network after applying dropout. Doted lines indicate
a node that has been dropped. (source [46])
Fig. 5. Illustration of Dropout application. (a) a node is dropped with probability 1 − p at training time. (b) at test time the node is always
present and its outputs are weighted by p. (source [46])
In this paper, we propose to adopt the Dropout strategy to regularize MI-ANFIS networks. Typically,
overfitting occurs in MI-ANFIS networks when a set of multiple instance rules co-adapt to the provided
data early during the training process and prevent the remaining rules from learning. Thus, degrading the
12
network’s generalization capability. While the Dropout technique could be applied to MI-ANFIS as is (given
the inherited neural network nature of the architecture), care should be exercised when selecting nodes to
include in the list of the randomly dropped out nodes. MI-ANFIS nodes are different from that of standard
neural networks as they are grouped into rules to model and express linguistic terms. Simply dropping
few nodes from a given rule can change its role and could severely handicap the fuzzy inference process.
Hence, Dropout should be executed differently. In deep neural nets, Dropout is applied to selected layers
(vertically), for MI-ANFIS, we propose to apply Dropout on a rule by rule basis (i.e., horizontally). Either
the whole rule is included, or the whole rule is dropped. This can be achieved by applying Dropout to
Layer 5 (see Figure 6), i.e., setting to zero the output of the “to be dropped out” rules. We will refer to
this derived technique as “Rule Dropout”. Using a Rule Dropout strategy to train MI-ANFIS networks is
approximatively equivalent to sampling and training 2R (R is the number of rules) ensemble of networks.
Let p be the probability with which a rule is present, formally, Rule Dropout is applied to Layer 5 during
training as follows
O5,i = hi wi Sα (xp1 · bi , xp2 · bi , . . . , xpMp · bi ),
(31)
hi ∼ Bernoulli(p)
(32)
where
is a Bernoulli random variable with probability p of being 1. During testing, Layer 5 output is scaled by
p, i.e., O3,i = pwi Sα (xp1 · bi , xp2 · bi , . . . , xpMp · bi ). Figure 6 illustrates our MI-ANFIS network with 3
multiple instance fuzzy rules where, at a given iteration, rule 2 has been dropped out..
Deriving the new update equations for MI-ANFIS parameters requires taking into consideration the added
Bernoulli random variable, hi . It is straightforward to show that the new gradients with respect to premise
and consequent parameters are given by
∂Ep
= −2(tp − Op )
∂ckj
P|O3 |
wl − wk
× hk × Sα (xp1 · b , xp2 · b , . . . , xpMp · b ) × l=1
P|O3 | 2
l=1 wl
k
×
k
k
Mp
X
i
eαrk,(i+(k−1)Mp ) h
Mp
1 + α rk,(i+(k−1)Mp ) − Sα ({rk,m }m=1
)
PMp αr
k,m
m=1 e
i=1
D
Y
×
µA
xp,(i+(k−1)Mp )[Mp ],d
(i+(k−1)Mp )/Mp
d=1,d6=j
×
,d
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )
2
σkj
!
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
× exp(−
) .
2
2σkj
(33)
13
Fig. 6.
Illustration of Rule Dropout application. Doted lines indicate a rule that has been dropped.
and,
∂Ep
= −2(tp − Op )
∂σkj
P|O3 |
wl − wk
× hk × Sα (xp1 · b , xp2 · b , . . . , xpMp · b ) × l=1
P|O3 | 2
l=1 wl
k
×
k
k
Mp
X
i
eαrk,(i+(k−1)Mp ) h
Mp
1 + α rk,(i+(k−1)Mp ) − Sα ({rk,m }m=1
)
PMp αr
k,m
m=1 e
i=1
D
Y
×
µA
x
p,(i+(k−1)Mp )[Mp ],d
(i+(k−1)Mp )/Mp
d=1,d6=j
×
,d
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
3
σkj
!
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
× exp(−
) .
2
2σkj
(34)
14
In a similar manner,
N
X
∂E
=
−2(tp − Op )
i
∂bj
p=1
× hi wi
Mp
X
m=1
1
P
Mp
h=1
exp(α(xph · bi − xpm · bi ))
2
(35)
Mp
h
X
× xpmj
exp(α(xph · bi − xpm · bi )
h=1
Mp
X
i
− xpm · b
i
exp(α(xph · bi − xpm · bi )α(xphj − xpmj ) .
h=1
As it can be seen, equations (33), (34), and (35) will get zeroed when the rule is dropped out (i.e., hk = 0
and hi = 0). Thus, its premise and consequent parameters are not updated.
IV.
EXPERIMENTAL RESULTS
A. Synthetic Data
To illustrate the proposed multiple instance fuzzy inference and its ability to learn from data without
instance-level labels, first, we use a simple 2-Dim synthetic dataset. This data were generated from a
distribution of two positive contexts with centers at (0.5,0.5) and (1.5,1.5), and with a fixed standard
deviation. These centers are marked with squares in Figure 7, and the circles around the centers indicates
regions within 1 standard deviation. These regions are considered the two target concepts. From each positive
concept we generated 50 bags. Each bag has a random number, between 2 and 10, of instances. Each bag
from concept 1 (or 2) will have at least one instance close to target concept 1 (or 2). We also generated 50
negative bags randomly from non concept regions. Negative bags will have all of their instances outside
both target concepts. In Figure 7, instances from negative bags are shown as “.”, and instances from positive
bags are shown as “+” or “M” depending on the underlying concept. In Figure 7, we highlight one bag
from Concept 1 by circling all of its instances. As it can be seen, one of its instances is within one standard
deviation region of target concept 1 while the other instances are scattered around. We should emphasize
here that the centers of the target concepts in Figure 7 are unknown and not used by the learning algorithm.
They are shown here for illustration and validation purposes only.
1) MI-ANFIS Rules Learning: In the following, we show that the MI-ANFIS Learning Algorithm (Algorithm 1) is capable of identifying positive concepts as well as their corresponding multiple instance
fuzzy rules. To initialize the premise parameters, we partition the instances’ space into 6 partition generated
randomly 2 . We use the partitions’ centers as initial centers for the Gaussian MFs, and we initialize all
standard deviation parameters to a default value of 0.5.
The initial fuzzy sets (MFs) of the rules, before training, are displayed in Figure 8 in dashed lines. As
it can be seen, the initial 6 partitions simply cover random quadrants of the 2D instance space (if no label
information is used, as in this case, data would appear to have uniform distribution (refer to Figure 7)).
The learned fuzzy sets after convergence are shown in Figure 8 in bold lines. As it can be seen, the system
has correctly identified the positive concepts, and at the same time identified irrelevant rules (MI-Rule 1,
MI-Rule 3 and MI-Rule 5) and assigned low output values to each, −0.3, −0.06 and −0.12 respectively.
B. Benchmark Datasets
To provide a quantitative evaluation of the proposed MI-ANFIS, we apply it to five benchmark data
sets commonly used to evaluate MIL methods: The MUSK1, MUSK2 [19], and Fox, Tiger, and Elephant
2
A grid or manual partitioning could also be used
15
Fig. 7. Instances from positive and negative bags drawn from data that have 2 concepts. The center of each target concept is indicated by
a square and the circles indicate the region within 1 - standard deviation from the mean. Instances from negative bags are shown as “.”, and
instances from positive bags are shown as “+” or “M”. Instances from one sample positive bag are circled.
from the COREL data set [47]. MUSK1 and MUSK2 data sets consist of descriptions of molecules and
the objective is to classify whether a molecule smells musky [48]. In these data sets, each bag represents a
molecule and instances within each bag represent the different low-energy conformations of the molecule.
Each instance is characterized by 166 features. MUSK1 has 92 bags, of which 47 are positive, and MUSK2
has 102 bags, of which 39 are positive. The other data sets (Fox, Tiger, and Elephant), classify whether an
image contains the corresponding animal. Each data set consists of 200 images (bags): 100 positive images
containing the target animal and 100 negative images containing other animals. Each image is represented
as a set of patches (instances) and each patch is in turn represented by a 230 dimensional feature vector
describing its color, texture and shape information. We note that the three data sets are independent and used
as binary classification problem (positive v.s. negative). Table I summarizes the characteristics of the five
data sets. It is to be noted that for each benchmark data set, PCA was applied to reduce the dimensionality
of the features in order to speedup MI-ANFIS training and increase the interpretability of the generated
multiple instance fuzzy rules.
For each experiment, we construct a zero-order MI-ANFIS with a given number of multiple instance rules.
For MI-ANFIS the number of rules is not critical. It should be large enough to cover the diverse regions of
the input space and the multiple concepts. If the specified number of rules is too large, some will vanish
as illustrated in Figure 8 for the example with the synthetic data. Also, larger number of rules leads to
slower training. We use Gaussian MFs to describe the input fuzzy sets. For initialization, we use the FCM
algorithm to cluster the instances of the positive bags into a number of clusters equal to the number of fuzzy
16
Fig. 8. Learned MFs after convergence of MI-ANFIS training algorithm. Initial MFs before training are marked with dashed lines. Learned
MFs are shown in sold bold lines.
TABLE I.
B ENCHMARK DATA SETS
Data set
dim.(PCA)
No. Bags
Positive
Negative
MUSK1
166(25)
92
47
45
No.Instances
2 → 40
MUSK2
166(25)
102
39
63
1 → 1044
Fox
230(10)
200
100
100
2 → 13
Tiger
230(10)
200
100
100
1 → 13
Elephant
230(10)
200
100
100
2 → 13
rules, and we initialize MFs’ centers as the clusters centers. MI-ANFIS was trained and tested using ten
fold cross validation. Table II summarizes all parameters used in training the MI-ANFIS (parameters were
manually selected using trial and error). We note that the reason behind using larger standard deviations
for MUSK1 and MUSK2 datasets is the higher dimensionality of these data sets. We expect the sparsity
to increase with the dimensions of the feature space, so we set the standard deviations to larger values to
allow the initial rules to cover the entirety of the input space.
First, to illustrates the advantage of using MI-ANFIS over the traditional ANFIS we compare these
two algorithms on the two MUSK data sets. Since ANFIS cannot learn from ambiguously labeled data,
for the sake of comparison, we consider the naive MIL assumption where all instances from positive
bags are considered positive and all instances from negative bags are considered negative. We refer to this
implementation as Naive-ANFIS. The results are summarized in Table III where the performance is reported
in terms of prediction accuracy averaged over all 10 cross validation sets (% of correct ± standard deviation).
17
TABLE II.
Parameter
TABLE III.
MI-ANFIS T RAINING PARAMETERS
MUSK1
MUSK2
FOX
Tiger
Elephant
No. of MI Rules
6
3
2
4
3
No. of Inputs
25
25
10
10
10
MF’s σ
100
100
10
10
10
Output parameters
1s
1s
1s
1s
1s
Softmax’s α
1
1
1
1
1
Learning rate
0.1
0.1
0.1
0.1
0.1
C OMPARISON OF MI-ANFIS
PREDICTION ACCURACY ( IN PERCENT ) TO
NAIVE -ANFIS ON THE BENCHMARK DATA SETS .
Algorithms
MUSK1
MUSK2
Fox
Tiger
Elephant
MI-ANFIS
93.49
±0.76
90.58
±1.31
66.4
±2.77
84.5
±0.61
86.97
±1.10
Naive-ANFIS
67.82
±4.04
79.43
±5.04
58.70
±1.35
77.70
±0.83
82.2
±0.83
As it can be seen, MI-ANFIS outperforms Naive-ANFIS significantly. This is because inaccurately labeled
instances within the positive bags were used for training the Naive-ANFIS. The difference in performance
between MI-ANFIS and Naive-ANFIS is greater for MUSK1 and MUSK2 because of the greater number
of instances per bag (more ambiguousity).
TABLE IV.
C OMPARISON OF MI-ANFIS PREDICTION ACCURACY ( IN PERCENT ) TO OTHER METHODS ON THE BENCHMARK DATA
R ESULTS FOR 3 TOP PERFORMING METHODS ARE SHOWN IN BOLD FONT. W E USE REPORTED RESULTS , N/A INDICATED THAT A
GIVEN ALGORITHM WAS NOT APPLIED TO THAT DATA SET.
SETS .
Algorithms
MUSK1
MUSK2
Fox
Tiger
Elephant
MI-ANFIS
93.49
±0.76
90.58
±1.31
66.4
±2.77
84.5
±0.61
86.97
±1.10
MILES [49]
86.3
87.7
N/A
N/A
N/A
APR [19]
92.4
89.2
N/A
N/A
N/A
DD [21]
88.9
82.5
N/A
N/A
N/A
DD-SVM [50]
85.8
91.3
N/A
N/A
N/A
EM-DD [51]
84.8
84.9
56.1
72.1
78.3
Citation-KNN [52]
92.4
86.3
N/A
N/A
N/A
MI-SVM [47]
77.9
84.3
57.8
84.0
81.4
mi-SVM [47]
87.4
83.6
58.2
78.4
82.2
MI-NN [53]
88.0
82.0
N/A
N/A
N/A
Bagging-APR [54]
92.8
93.1
N/A
N/A
N/A
RBF-MIP [43]
91.3
±1.6
90.1
±1.7
N/A
N/A
N/A
BP-MIP [42]
83.7
80.4
N/A
N/A
N/A
RBF-Bag-Unit [55]
90.3
86.6
N/A
N/A
N/A
MI-kernel [56]
88.0
89.3
60.3
84.2
84.3
PPPM-kernel [57]
95.6
81.2
60.3
80.2
82.4
MIGraph [56]
90.0
90.0
61.2
81.9
85.1
miGraph [56]
88.9
90.3
61.6
86.0
86.8
ALP-SVM [58]
86.3
86.2
66.0
86.0
83.5
MIForest [59]
85.0
82.0
64.0
82.0
84.0
Table IV compares the performance of the proposed algorithm to state of the art MIL algorithms on the
benchmark data sets.
Overall, MI-ANFIS is comparable to other MIL algorithms. In fact, on all tested data sets, MI-ANFIS
ranked consistently among the top three. For MUSK1, PPPM-kernel [57] performed the best (95.6%), but
this algorithm did not perform as well for the other sets. For MUSK2 Bagging-APR [54] achieved the best
18
TABLE V.
B ENCHMARK DATA SETS
Data set
dim.(PCA)
No. Bags
Positive
Negative
No.Instances
Elephant
230(10)
200
100
100
2 → 13
accuracy, as reported by [49]. MI-ANFIS achieved the best average performance for the Fox and Elephant
data sets, and second best performance after the miGraph [56] and ALP-SVM [58] methods for the Tiger
data set.
In order to demonstrate the gain in generalization acquired by MI-ANFIS when utilizing Rule Dropout,
we train an MI-ANFIS architecture for binary classification with and without Rule Dropout on a multiple
instance dataset sampled from COREL [47]. The dataset classify whether an image contains an elephant or
not, and consists of 200 images (bags): 100 positive images containing the target animal and 100 negative
images containing other animals. Each image is represented as a set of patches (instances) and each patch
is in turn represented by 230 features describing color, texture and shape information. Before training, we
applied PCA to reduce the dimensionality of the features to 10 dimensions to speedup MI-ANFIS. Table
V summarizes the dataset characteristics. Two MI-ANFIS networks composed of 15 rules each, with one
network employing Rule Dropout (with p = 0.7, this hyper-parameter was selected based on trial and error),
were trained on 90% of the data, and the remaining 10% was used for testing (split was done randomly).
Figure 9 shows the training and testing errors for both networks during 100 epochs. As it can be seen,
without Rule Dropout, starting at epoch 20, testing performance begins to degrade while training error
continues to decrease. In other words, overfitting begins to occur. Typically, using a cross validation data
set, this point can be detected and training would be stopped. However, this assumes that a cross validation
data is available (or training data is large enough to be split into training and testing) and more important
that the cross validation data is representative of the testing data. On the other hand, when using Rule
Dropout, overfitting is significantly reduced and MI-ANFIS achieved better testing performance at the end
of the training phase. Even though, when using Rule Dropout the training and testing error rates oscillate
(due to the randomness of the dropout process), overall MI-ANFIS achieved 0.1123 testing SSE with Rule
Dropout compared to 0.1451 testing SSE without Rule Dropout.
C. Application To Landmine Detection
In this section, we report the results of applying the proposed Multiple Instance Inference to fuse the
output of multiple discrimination algorithms for the purpose of landmine detection using Ground Penetrating
Radar (GPR). GPR data collected at different locations and different dates were used to train and test the
proposed MI-ANFIS. The alarm collection covers 319 encounters of various anti-tank mines with high
metal content (ATHM) and 422 encounters of various anti-tank mines with low metal content (ATLM).
The vehicle-mounted GPR sensor collects 3-dimensional data as the vehicle moves (Figure 10). The 3dimensions correspond to the spatial location on the ground (down-track, cross-track, and depth) and is
shown in Figure 11. Figure 11(b) shows a 2-D views of (depth, down-track) and (depth, cross-track) slices
of GPR data. As it can be seen, the target signature does not extend over all depth values. Thus, one global
feature vector may not discriminate between mines and clutter effectively. To overcome this limitation,
most classifiers developed for this application extract multiple features from small overlapping windows at
multiple depths. In the following, we assume that each training alarm (3-D data cube) has been divided
into 15 overlapping (depth wise) patches. Each patch is processed by 2 discrimination algorithms. These
algorithms are based on the Edge Histogram Descriptor (EHD) [60]. The first one, called EHDDT, extracts
features from each 2-D (down-track, depth) patch. The second discrimination algorithm, called EHDCT,
extracts information for the 2-D (Cross-Track, depth) patch. In addition, auxiliary features are synthesized
from each patch. In particular, “SignatureWidth” in the Down-track direction and “SignatureWidth” in the
Cross-Track direction are used to capture the effective width of the hyperbolic shape within each patch.
These auxiliary features are intended to provide contextual information that can support the relevance of
the EHDDT and/or EHDCT. As a result, each alarm is represented by a Bag of 15 instances and each
19
Fig. 9.
Fig. 10.
Training and testing errors for two MI-ANFIS networks with and without Rule Dropout.
Vehicle mounted GPR system for detecting buried Landmines.
instance is a 4-dimensional feature vector. Each bag is labeled as positive (has a target) or negative (non
target), but labels at the instance level are not available. The X-Y Ground truth information about the target
is available (using GPS and known target position on calibration lanes). However, the depth position cannot
be easily identified as it depends on target size, burial depth, soil type, and other environmental conditions.
Manually extracting the depth location can be very tedious. Similarly, during testing, it is not trivial how
to combine partial confidence values from the multiple windows. Therefore, the MIL paradigm is suitable
to solve this problem.
20
(a) 3D GPR Raw data.
Fig. 11.
(b) 2-D views of the raw GPR data.
3-dimensional and 2-dimensional raw GPR data.
We construct a zero-order MI-ANFIS (constant consequent parameters) having 5 multiple instance rules,
and employing Gaussian MFs to describe the input fuzzy sets. To initialize the system’s parameters, first,
we use the FCM algorithm to cluster the instances that belong to positive bags into 5 clusters, and we
initialize the MFs’ centers as the clusters’ centers. Then, we initialize the standard deviations of the input
MFs and the output parameters to 1.
After initialization, we run MI-ANFIS basic learning algorithm (Algorithm 1) to jointly learn a fuzzy
description of the positive concepts as well as optimal rules’ output. Figure 12 is a graphical representation
of the 5 multiple instance rules prior to running the optimization process (dotted line curves) and the learned
rules after training (continuous curves). The fuzzy sets of the rules’ antecedents describe the location and
the extent of the positive concepts in the 4-D instance feature space. The rules’ consequent values can
be interpreted as an assessment of the “positivity” of each learned concept. For instance, the MI-ANFIS
learned the following two positive concepts to describe targets:
R1 : If EHDDT is M edium and EHDCT is M edium and W idthDT is High and W idthCT is High then o1 = 1.15.
R2 : If EHDDT is M edium and EHDCT is Low and W idthDT is High and W idthCT is High then o2 = 0.94.
1) Results: The proposed fusion method was trained and tested using 10-folds cross validation. Figure
14 displays the ROCs (averaged over the 10 folds). To provide a quantitative evaluation of the proposed
multiple instance fuzzy inference fusion method, we compare its performance to a fusion method based
on the standard Mamdani [12] and standard ANFIS [61]. Since the standard Mamdani and AFNIS cannot
learn from partially labeled data, an expert is used to label all instances of all bags within the training data.
We also compare MI-ANFIS performances to a naive MIL implementation of Mamdani (NaiveMamdani)
and ANFIS (NaiveANFIS) where all instances from positive bags are considered positive and all instances
from negative bags are considered negative.
As it can be seen in Figure 14, MI-ANFIS performed better than the standard ANFIS on the large dataset,
and as expected NaiveANFIS performed worse. The standard ANFIS performed better at low FAR (False
Alarms Rate), the reason is that strong Mines are easy to identify manually and in this case, the ground
truth helps. However, weaker mine signatures are not as easy to localize, so the truth may not be as accurate
and can degrade the performance. Overall, MI-ANFIS outperformed all presented fusion approaches and
21
Fig. 12.
MI-ANIFS fusion rules before (dotted lines) and after training (solid lines).
Fig. 13.
A plot of MI-ANFIS RMSE during 150 training epochs.
22
the individual discriminators (EHDDT and EHDCT). This is due to the ability of MI-ANFIS to overcome
labeling ambiguity by learning meaningful concepts.
Fig. 14. Comparison of the individual discriminators, the proposed MI-ANFIS, Mamdani, ANFIS, NaiveMamdani, and NaiveANFIS fusion
methods. Note that the Mamdani and ANFIS systems have the advantage of instance-level labels on training data.
As in standard ANFIS, we cannot prove convergence of the algorithms. However, in all conducted
experiments MI-ANFIS converged in less than 150 epochs. Figure 13 plots the root mean squared error
(RMSE) vs. the training epoch number.
V. R ELATED W ORK
MI-ANFIS deals with ambiguity by introducing the novel concept of truth instances: when carrying
reasoning using a bag of instances at Layer 2 (Figure 3), a proposition will not only have one degree of
truth, it will have multiple degrees of truth (rij ), we call truth instances. Thus, effectively encoding the
third vagueness component of ambiguity and increasing the expressive power of traditional fuzzy logic. In
addition to effectively model ambiguity, MI-ANIFS has the inherited capability of assessing the prediction
quality by outputting soft values. For example, depending on the α parameter of Softmax in Layer 3, MIANFIS can assign higher outputs to bags with more than one positive instance. Thus, giving the end user
a way to assess the positiveness of a given bag.
Learning positive target concepts from ambiguously labeled data has been the core task of various MIL
algorithms (e.g. Diverse Density [16]). MI-ANFIS has proven that it can learn positive concepts effectively
while jointly providing a fuzzy representation of such regions. The fuzzy representation is combined into
meaningful and simple multiple instance rules that can be easily visualized and interpreted.
Compared to previously proposed multiple instance neural networks, such as Multiple Instance Neural Networks [42] (MI-NN) and Multiple Instance RBF Neural Networks [43] (RBF-MIP), MI-ANFIS advantage
is the use of multiple instance fuzzy logic to learn a fuzzy representation of true positive concepts. MI-NN
only learns standard neural network weights that do not carry any information regarding target concepts.
On the other hand, while standard RBF neural networks have been shown to be equivalent to zero order
23
traditional Sugeno systems under certain constraints [62], thus, capable of learning a fuzzy representation
of the inputs, RBF-MIP networks have different architecture and they do not employ adaptive radial basis
functions in the first layer. Instead, they represent the inputs by computing their distances to clusters of
training bags. This latter method is computationally expensive and its success depends greatly on the quality
of the training data as it takes into consideration all the training examples which may include wrongly
(noisily) labeled bags. RBF-MIP learns only discriminative regions of the bags space and does not learn
true positive concepts. Moreover, MI-ANFIS learning algorithms can be updated to support a wide range
of loss functions (criterions) such as cross entropy [63], maximum margin [64], etc. MI-NN is designed
to use a handcrafted loss function which is largely responsible for the multiple instance behavior of the
system and cannot be changed without substantially changing the architecture of MI-NN. This could be
disadvantageous if MI-NN is to be used to solve multiple instance - multiple class classification problems.
VI. C ONCLUSIONS
In this paper, we have introduced a new framework to accomplish fuzzy inference with multiple instance
data. Our work generalizes Sugeno fuzzy inference style to reason with multiple instances, the new inference
style is called MI-Sugeno. We then used MI-Sugeno to develop MI-ANFIS, a novel neuro-fuzzy architecture
that extends the standard Adaptive Neuro-Fuzzy Inference System (ANFIS) to reason with bags of instances
in order to solve multiple instance learning problems. We developed a BackPropagation learning algorithm
and showed that the proposed system is capable of learning meaningful concepts from ambiguously labeled
data.
MI-ANFIS deals with ambiguity by introducing the novel concept of truth instances: when carrying reasoning using a bag of instances at Layer 2 (Figure 3), a proposition will not only have one degree of truth,
it will have multiple degrees of truth (rij ), we call truth instances. Thus, effectively encoding the third
vagueness component of ambiguity and increasing the expressive power of traditional fuzzy logic.
Learning positive concepts from ambiguously labeled data has been the core task of various MIL algorithms.
MI-ANFIS has proven that it can learn positive concepts effectively while jointly providing a fuzzy
representation of such regions. The fuzzy representation is combined into meaningful and simple multiple
instance rules that can be easily visualized and interpreted.
Using synthetic and benchmark data sets we showed that the proposed Multiple Instance Fuzzy Inference
is comparable to state of the art MI machine learning algorithms. We also used our framework for a
real application and applied it to fuse the output of multiple discrimination algorithms for the purpose of
landmine detection using Ground Penetrating Radar.
In situations where overfitting is imminent, for example when using relatively smaller datasets to learn very
large MI-ANFIS networks, we proposed a regularization technique, we called Rule Dropout, and showed
that it could be used to train MI-ANFIS systems with better generalization.
In future work, we intend to develop a multiple class version of MI-ANFIS to be used to solve multiple
class classification problems. In addition, we will conduct a detailed analysis of MI-ANFIS convergence.
A PPENDIX A
DERIVATION OF PREMISE PARAMETERS UPDATE RULES
From equations (21) and (22) the error rate for the premise parameters ckj and σkj are defined as following
Mp
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂Ep
∂Ep X
=
.
∂ckj
∂O
∂c
kj
(1,i+[(k−1)D+(j−1)]M
)
p
i=1
and,
Mp
X
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂Ep
∂Ep
=
.
∂σkj
∂O
∂σ
kj
(1,i+[(k−1)D+(j−1)]M
)
p
i=1
24
Using the chain rule defined in (18), we have
∂Ep
∂O(1,i+[(k−1)D+(j−1)]Mp )
=
∂O(2,i+(k−1)Mp )
∂O(6,1) ∂O(5,k) ∂O(4,k)
∂O(3,k)
∂Ep
×
×
×
×
×
.
∂O(6,1) ∂O(5,k) ∂O(4,k) ∂O(3,k) ∂O(2,i+(k−1)Mp ) ∂O(1,i+[(k−1)D+(j−1)]Mp )
(36)
Hence, (21) is equivalent to
∂O(6,1) ∂O(5,k) ∂O(4,k)
∂Ep
∂Ep
=
×
×
×
∂ckj
∂O(6,1) ∂O(5,k) ∂O(4,k) ∂O(3,k)
"
#
Mp
X
∂O(2,i+(k−1)Mp )
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂O(3,k)
×
×
×
. (37)
∂O(2,i+(k−1)Mp ) ∂O(1,i+[(k−1)D+(j−1)]Mp )
∂ckj
i=1
From (17), we have
∂Ep
= −2(tp − Op ).
∂O(6,1)
(38)
P 3|
∂O(6,1)
∂( |O
i=1 O(5,i) )
=
= 1.
∂O(5,k)
∂O(5,k)
(39)
It is also straightforward to show that,
and,
∂(wk Sα (xp1 · bk , xp2 · bk , . . . , xpMp · bk ))
∂O(5,k)
=
= Sα (xp1 · bk , xp2 · bk , . . . , xpMp · bk ). (40)
∂O(4,k)
∂(wk )
Continuing with the derivation, we have
wi
P|O3 |
∂ P|O3 |
∂O(4,k)
∂wk
wl − wk
l=1 wl
(41)
=
=
= l=1
2 .
P
∂O(3,k)
∂wk
∂wk
|O3 |
l=1 wl
and,
∂O(3,k)
∂O(2,i+(k−1)Mp )
Mp
i
∂Sα ({r(k,j) }j=1
)
eαr(k,(i+(k−1)Mp )) h
Mp
1 + α r(k,(i+(k−1)Mp )) − Sα ({r(k,m) }m=1 ) . (42)
=
= PMp αr
(k,m)
∂r(k,(i+(k−1)Mp ))
m=1 e
The details of the derivation of the “softmax” function details can be found at [21].
∂O(2,i+(k−1)Mp )
Next, we need to compute ∂O(1,i+[(k−1)D+(j−1)]M
. We have,
)
p
O(2,i+(k−1)Mp ) =
D
Y
d=1
µA
(i+(k−1)Mp )/Mp
x(p,(i+(k−1)Mp )[Mp ],d) .
(43)
,d
and,
O(1,i+[(k−1)D+(j−1)]Mp ) = µA(k,j) (x(p,(i+(k−1)Mp )[Mp ],d) ).
(44)
Thus,
∂O(2,i+(k−1)Mp )
∂O(1,i+[(k−1)D+(j−1)]Mp )
∂
QD
d=1
µA
=
(i+(k−1)Mp )/Mp
x(p,(i+(k−1)Mp )[Mp ],d)
!
,d
∂ µA(k,j) (x(p,(i+(k−1)Mp )[Mp ],d) )
=
D
Y
d=1,d6=j
µA
(i+(k−1)Mp )/Mp
x(p,(i+(k−1)Mp )[Mp ],d) .
,d
(45)
25
Finally, we have
∂µA(k,j) (x(p,(i+(k−1)Mp )[Mp ],j) )
∂O(1,i+[(k−1)D+(j−1)]Mp )
=
=
∂ckj
∂ckj
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
×
exp(−
). (46)
2
2
σkj
2σkj
Substituting the derivatives in (37), we obtain (25).
The update formula for σkj can be derived in a similar manner. It can be shown that
P|O3 |
∂Ep
wl − wk
k
k
k
= −2(tp − Op ) × Sα (xp1 · b , xp2 · b , . . . , xpMp · b ) × l=1
2
P
∂σkj
|O3 |
l=1 wl
Mp
X
i
eαr(k,(i+(k−1)Mp )) h
Mp
)
1 + α r(k,(i+(k−1)Mp ) ) − Sα ({r(k,m) }m=1
PMp αr
(k,m)
m=1 e
i=1
D
Y
×
µA
x
(p,(i+(k−1)Mp )[Mp ],d)
×
d=1,d6=j
(i+(k−1)Mp )/Mp
(47)
,d
!
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
(x(p,(i+(k−1)Mp )[Mp ],j) − ckj )2
×
× exp(−
) .
3
2
σkj
2σkj
A PPENDIX B
DERIVATION OF CONSEQUENT PARAMETERS UPDATE RULE
The error rate for the consequent parameters is defined in equations (27) and (28). Next, we compute
using the previously defined chain rule in (18), and obtain
∂Ep
∂O(5,i)
∂O(6,1)
∂Ep
∂Ep
=
×
.
∂O(5,i)
∂O(6,1)
∂O(5,i)
(48)
∂Ep
= −2(tp − Op ).
∂O(6,1)
(49)
∂O(6,1)
= 1.
∂O(5,i)
(50)
From (17), we have
And from (50), we have
Continuing with the derivation, we have
Mp
∂(wi Sα (xp1 · bi , xp2 · bi , . . . , xpMp · bi ))
∂O(5,i)
∂ X xpm · bi exp(αxpm · bi )
=
= i wi
PMp
i
∂bij
∂bij
∂bj
h=1 exp(αxph · b )
m=1
Mp
Mp
X
X
∂ xpm · bi exp(αxpm · bi )
= wi
= wi
P
PM p
i
i)
Mp
∂b
exp(αx
·
b
j
ph
h=1
m=1
m=1
1
2
i−x
i ))
exp(α(x
·
b
·
b
ph
pm
h=1
×
h
x(p,m,j)
Mp
X
h=1
Mp
i
X
i
exp(α(xph ·b −xpm ·b ) − xpm ·b
exp(α(xph ·bi −xpm ·bi )α(x(p,h,j) −x(p,m,j) ) .
i
i
h=1
(51)
Thus, the overall error rate with respect to the consequent parameter bij is given according to (20) in equation
(29).
26
ACKNOWLEDGMENT
This work was supported in part by U.S. Army Research Office Grants Number W911NF-13-1-0066
and W911NF-14-1-0589. The views and conclusions contained in this document are those of the authors
and should not be interpreted as representing the official policies, either expressed or implied, of the Army
Research Office, or the U.S. Government.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
L. A. Zadeh, A theory of approximate reasoning (AR). Berkeley: Electronics Research Laboratory, College of Engineering, University
of California, 1977.
Zadeh, “Outline of a new approach to the analysis of complex systems and decision processes,” Systems, Man and Cybernetics, IEEE
Transactions on, no. 1, pp. 28–44, 1973.
C.-W. Xu and Y.-Z. Lu, “Fuzzy model identification and self-learning for dynamic systems,” Systems, Man and Cybernetics, IEEE
Transactions on, vol. 17, no. 4, pp. 683–689, July 1987.
R. Jager, H. Verbruggen, and P. Bruijin, “Fuzzy inference in rule-based control systems,” in Intelligent Systems Engineering, 1992., First
International Conference on (Conf. Publ. No. 360), Aug 1992, pp. 232–237.
C.-C. Lee, “Fuzzy logic in control systems: fuzzy logic controller. ii,” Systems, Man and Cybernetics, IEEE Transactions on, vol. 20,
no. 2, pp. 419–435, Mar 1990.
J. Casillas, Interpretability issues in fuzzy modeling. Springer, 2003, vol. 128.
C.-Y. Chiu, H.-C. Lin, and S.-N. Yang, “A fuzzy logic cbir system,” in Fuzzy Systems, 2003. FUZZ ’03. The 12th IEEE International
Conference on, vol. 2, May 2003, pp. 1171–1176 vol.2.
A. Othman, H. Tizhoosh, and F. Khalvati, “Efis 2014;evolving fuzzy image segmentation,” Fuzzy Systems, IEEE Transactions on, vol. 22,
no. 1, pp. 72–82, Feb 2014.
S. Hajimirza and E. Izquierdo, “Gaze movement inference for implicit image annotation,” in Image Analysis for Multimedia Interactive
Services (WIAMIS), 2010 11th International Workshop on, April 2010, pp. 1–4.
H. Kwan and L. Y. Cai, “Supervised fuzzy inference network for invariant pattern recognition,” in Circuits and Systems, 2000. Proceedings
of the 43rd IEEE Midwest Symposium on, vol. 2, 2000, pp. 850–854 vol.2.
M. Adnan, M. Chowdury, I. Taz, T. Ahmed, and R. Rahman, “Content based news recommendation system based on fuzzy logic,” in
Informatics, Electronics Vision (ICIEV), 2014 International Conference on, May 2014, pp. 1–6.
A. Ben Khalifa and H. Frigui, “Fusion of multiple algorithms for detecting buried objects using fuzzy inference,” Proc. SPIE, vol.
9072, pp. 90 720V–90 720V–10, 2014. [Online]. Available: http://dx.doi.org/10.1117/12.2051217
M.-Y. Chen and D. Linkens, “Rule-base self-generation and simplification for data-driven fuzzy models,” in Fuzzy Systems, 2001. The
10th IEEE International Conference on, vol. 1, 2001, pp. 424–427.
D. Saletic, “On data-driven procedure for determining the number of rules in a takagi-sugeno fuzzy model,” in Computer as a Tool,
2005. EUROCON 2005.The International Conference on, vol. 2, Nov 2005, pp. 1132–1135.
J.-S. R. Jang, C.-T. Sun, and E. Mizutani, “Neuro-fuzzy and soft computing-a computational approach to learning and machine intelligence
[book review],” Automatic Control, IEEE Transactions on, vol. 42, no. 10, pp. 1482–1484, 1997.
O. Maron, “Learning from ambiguity,” Ph.D. dissertation, Massachusetts Institute of Technology, 1998.
T. G. Dietterich, R. H. Lathrop, and T. Lozano-Prez, “Solving the multiple instance problem with axis-parallel rectangles,” Artificial
Intelligence, vol. 89, no. 12, pp. 31 – 71, 1997. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0004370296000343
E. Alpaydın, V. Cheplygina, M. Loog, and D. M. Tax, “Single-vs. multiple-instance classification,” Pattern Recognition, vol. 48, no. 9,
pp. 2831–2838, 2015.
T. G. Dietterich, R. H. Lathrop, and T. Lozano-Pérez, “Solving the multiple instance problem with axis-parallel rectangles,” Artificial
intelligence, vol. 89, no. 1, pp. 31–71, 1997.
C. Zhang, X. Chen, and W.-B. Chen, “An online multiple instance learning system for semantic image retrieval,” in Multimedia Workshops,
2007. ISMW ’07. Ninth IEEE International Symposium on, Dec 2007, pp. 83–84.
O. Maron and T. Lozano-Pérez, “A framework for multiple-instance learning,” in Proceedings of the 1997 Conference on Advances in
Neural Information Processing Systems 10, ser. NIPS ’97. Cambridge, MA, USA: MIT Press, 1998, pp. 570–576. [Online]. Available:
http://dl.acm.org/citation.cfm?id=302528.302753
A. Karem and H. Frigui, “A multiple instance learning approach for landmine detection using ground penetrating radar,” in Geoscience
and Remote Sensing Symposium (IGARSS), 2011 IEEE International. IEEE, 2011, pp. 878–881.
R. Rahmani and S. A. Goldman, “Missl: Multiple-instance semi-supervised learning,” in Proceedings of the 23rd international conference
on Machine learning. ACM, 2006, pp. 705–712.
Y. Chen, J. Bi, and J. Wang, “Miles: Multiple-instance learning via embedded instance selection,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 28, no. 12, pp. 1931–1947, Dec 2006.
C. Yang, M. Dong, and F. Fotouhi, “Region based image annotation through multiple-instance learning,” in Proceedings of the 13th
annual ACM international conference on Multimedia. ACM, 2005, pp. 435–438.
27
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
B. Babenko, M.-H. Yang, and S. Belongie, “Robust object tracking with online multiple instance learning,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 33, no. 8, pp. 1619–1632, 2011.
S. H. T. Andrews, “Multiple-instance learning via disjunctive programming boosting,” Advances in neural information processing systems.,
no. 16, pp. 65–72, 2004.
B. Babenko, “Multiple instance learning: algorithms and applications,” View Article PubMed/NCBI Google Scholar, 2008.
C. Zhang, J. C. Platt, and P. A. Viola, “Multiple instance boosting for object detection,” in Advances in neural information processing
systems, 2005, pp. 1417–1424.
J. Amores, “Multiple instance classification: Review, taxonomy and comparative study,” Artificial Intelligence, vol. 201, pp. 81–105,
2013.
P. L. Lanzi, W. Stolzmann, and S. W. Wilson, Learning classifier systems: from foundations to applications. Springer, 2000, no. 1813.
O. Cordn, “A historical review of evolutionary learning methods for mamdani-type fuzzy rule-based systems: Designing interpretable
genetic fuzzy systems,” International Journal of Approximate Reasoning, vol. 52, no. 6, pp. 894 – 913, 2011.
E. Mamdani, “Application of fuzzy algorithms for control of simple dynamic plant,” Electrical Engineers, Proceedings of the Institution
of, vol. 121, no. 12, pp. 1585–1588, December 1974.
R. Babuka and H. Verbruggen, “An overview of fuzzy modeling for control,” Control Engineering Practice, vol. 4, no. 11, pp. 1593 –
1606, 1996.
M. Mizumoto, “Fuzzy controls under various fuzzy reasoning methods,” Information Sciences, vol. 45, no. 2, pp. 129 – 151, 1988.
[Online]. Available: http://www.sciencedirect.com/science/article/pii/0020025588900370
C.-C. Lee, “Fuzzy logic in control systems: fuzzy logic controller. i,” Systems, Man and Cybernetics, IEEE Transactions on, vol. 20,
no. 2, pp. 404–418, Mar 1990.
M. Sugeno and T. Yasukawa, “A fuzzy-logic-based approach to qualitative modeling,” Fuzzy Systems, IEEE Transactions on, vol. 1,
no. 1, pp. 7–, Feb 1993.
R. Yager and D. Filev, “Unified structure and parameter identification of fuzzy models,” Systems, Man and Cybernetics, IEEE Transactions
on, vol. 23, no. 4, pp. 1198–1205, Jul 1993.
E. Tacker, “Modeling stabilization policies in financial systems,” in Decision and Control including the 16th Symposium on Adaptive
Processes and A Special Symposium on Fuzzy Set Theory and Applications, 1977 IEEE Conference on, Dec 1977, pp. 194–194.
P. Singh, S. Bhanot, and H. Mohanta, “Optimized adaptive neuro-fuzzy inference system for ph control,” in Advanced Electronic Systems
(ICAES), 2013 International Conference on, Sept 2013, pp. 1–5.
T. Takagi and M. Sugeno, “Fuzzy identification of systems and its applications to modeling and control,” Systems, Man and Cybernetics,
IEEE Transactions on, vol. SMC-15, no. 1, pp. 116–132, 1985.
Z.-H. Zhou and M.-L. Zhang, “Neural networks for multi-instance learning,” Proceedings of the International Conference on Intelligent
Information Technology, Beijing, China, pp. 455–459, 2002.
M.-L. Zhang and Z.-H. Zhou, “Adapting rbf neural networks to multi-instance learning,” Neural Processing Letters, vol. 23, no. 1, pp.
1–26, 2006.
S. Ray and D. Page, “Multiple instance regression,” in ICML, vol. 1, 2001, pp. 425–432.
J.-S. Jang, “Anfis: adaptive-network-based fuzzy inference system,” Systems, Man and Cybernetics, IEEE Transactions on, vol. 23, no. 3,
pp. 665–685, 1993.
N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from
overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.
S. Andrews, I. Tsochantaridis, and T. Hofmann, “Support vector machines for multiple-instance learning,” in Advances in neural
information processing systems, 2002, pp. 561–568.
Y. Li, D. M. Tax, R. P. Duin, and M. Loog, “Multiple-instance learning as a classifier combining problem,” Pattern Recognition, vol. 46,
no. 3, pp. 865–874, 2013.
Y. Chen, J. Bi, and J. Z. Wang, “Miles: Multiple-instance learning via embedded instance selection,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 28, no. 12, pp. 1931–1947, 2006.
Y. Chen and J. Z. Wang, “Image categorization by learning and reasoning with regions,” The Journal of Machine Learning Research,
vol. 5, pp. 913–939, 2004.
Q. Zhang and S. A. Goldman, “Em-dd: An improved multiple-instance learning technique,” in Advances in neural information processing
systems, 2001, pp. 1073–1080.
J. Wang, “Solving the multiple-instance problem: A lazy learning approach,” in In Proc. 17th International Conf. on Machine Learning,
2000.
J. Ramon and L. De Raedt, “Multi instance neural networks,” in Proceedings of the ICML-2000 workshop on attribute-value and
relational learning, 2000, pp. 53–60.
Z.-H. Zhou and M.-L. Zhang, “Ensembles of multi-instance learners,” in Machine Learning: ECML 2003. Springer, 2003, pp. 492–502.
A. Bouchachia, “Multiple instance learning with radial basis function neural networks,” in Neural Information Processing. Springer,
2002, Conference Proceedings, pp. 440–445.
Z.-H. Zhou, Y.-Y. Sun, and Y.-F. Li, “Multi-instance learning by treating instances as non-iid samples,” in Proceedings of the 26th annual
international conference on machine learning. ACM, 2009, pp. 1249–1256.
28
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
H.-Y. Wang, Q. Yang, and H. Zha, “Adaptive p-posterior mixture-model kernels for multiple instance learning,” in Proceedings of the
25th international conference on Machine learning. ACM, 2008, pp. 1136–1143.
P. V. Gehler and O. Chapelle, “Deterministic annealing for multiple-instance learning,” in International conference on artificial intelligence
and statistics, 2007, pp. 123–130.
C. Leistner, A. Saffari, and H. Bischof, “Miforests: multiple-instance learning with randomized trees,” in Computer Vision–ECCV 2010.
Springer, 2010, pp. 29–42.
H. Frigui and P. Gader, “Detection and discrimination of land mines in ground-penetrating radar based on edge histogram descriptors
and a possibilistic k-nearest neighbor classifier,” Trans. Fuz Sys., vol. 17, no. 1, pp. 185–199, Feb. 2009.
A. B. Khalifa and H. Frigui, “Fusion of multiple landmine detection algorithms using an adaptive neuro fuzzy inference system,” in
Geoscience and Remote Sensing Symposium (IGARSS), 2014 IEEE International. IEEE, 2014, pp. 3148–3151.
J.-S. Jang and C.-T. Sun, “Functional equivalence between radial basis function networks and fuzzy inference systems,” Neural Networks,
IEEE Transactions on, vol. 4, no. 1, pp. 156–159, Jan 1993.
M. Yi-de, L. Qing, and Q. Zhi-Bai, “Automated image segmentation using improved pcnn model based on cross-entropy,” in Intelligent
Multimedia, Video and Speech Processing, 2004. Proceedings of 2004 International Symposium on. IEEE, 2004, pp. 743–746.
H. Li, T. Jiang, and K. Zhang, “Efficient and robust feature extraction by maximum margin criterion,” Neural Networks, IEEE Transactions
on, vol. 17, no. 1, pp. 157–165, 2006.
| 3cs.SY
|
Locally compact groups acting on trees, the type I
conjecture and non-amenable von Neumann algebras
arXiv:1610.00884v2 [math.GR] 29 Nov 2016
Cyril Houdayer and Sven Raum
Abstract. We address the problem to characterise closed type I subgroups of the
automorphism group of a tree. Even in the well-studied case of Burger–Mozes’
universal groups, non-type I criteria were unknown. We prove that a huge class
of groups acting properly on trees are not of type I. In the case of Burger-Mozes
groups, this yields a complete classification of type I groups among them. Our
key novelty is the use of von Neumann algebraic techniques to prove the stronger
statement that the group von Neumann algebra of the groups under consideration
is non-amenable.
1
Introduction
In discrete and topological group theory, groups acting on trees are important examples thanks to
Bass-Serre theory [35]. In particular, the discovery of Bruhat-Tits theory [35, 6] describing rank one
reductive algebraic groups over non-Archimedean fields as groups acting on semi-regular trees provides
strong motivation to study general closed subgroups of Aut(T ), the automorphism group of a tree.
In contrast to Bruhat-Tits buildings of higher rank [46], semi-regular trees host a bigger variety of
interesting groups, some of whose basic properties are not yet understood. An intriguing problem
asking us to prove surprising parallels between reductive algebraic groups and closed subgroups of
Aut(T ) is posed by the type I conjecture.
Conjecture. Let T be a locally finite tree and assume that G ≤c Aut(T ) is a closed subgroup acting
transitively on the boundary ∂T . Then G is a type I group.
Here, a locally compact group G is called a type I group if every unitary representation of G generates
a type I von Neumann algebra. This is one equivalent definition of type I groups provided by [20,
Theorem 2, page 592]. Bernstein and Kirillov termed “tame” those algebraic groups and Lie groups
that are type I – in contrast to “wild” groups. In this context, type I or tameness results are derived
from a positive solution to the admissibility conjecture. The notion of type I groups bears its relevance from representation theory. Loosely speaking, type I groups are precisely those locally compact
groups all of whose unitary representations can be written as a unique direct integral of irreducible
representations, thus reducing the study of arbitrary unitary representations to considerations about
irreducible unitary representations. Prominent examples of type I groups are provided by reductive algebraic groups over non-Archimedean fields [3, 22] (see also the introduction of [4]), adelic reductive
groups [12], semisimple connected Lie groups [24, Theorem 8.1] and nilpotent connected Lie groups
last modified on November 30, 2016
Cyril Houdayer’s research was supported by ERC Starting Grant GAN 637601.
Sven Raum’s research leading to these results has received funding from the People Programme (Marie Curie Actions)
of the European Union’s Seventh Framework Programme (FP7/2007-2013) under REA grant agreement n°[622322].
MSC classification: Primary 20E08; Secondary 22D10, 46L45
Keywords: Groups acting on trees, type I groups, free product von Neumann algebras
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
[16, Théorème 1]. However, only very few results confirming the type I conjecture beyond rank one
algebraic groups are known, all of them being based on combinatorial considerations for the special
class of groups satisfying Tits’ independence property [40]. See [29], [30], [1] and [10].
From the theory of algebraic groups, natural examples of non type I groups, such as most adelic
nilpotent groups are known [27]. For groups acting on trees the situation looks worse, since tools
from Lie theory and from algebraic groups are not available in the generality of groups acting on trees.
There is one small class of groups for which non-type I results are known and it lies at the far opposite
end of boundary transitive groups. Already in the 60’s Thoma proved in [39] that virtually abelian
groups are the only discrete groups of type I, which completely clarifies type I questions for discrete
groups acting on trees. In the rich spectrum between discrete groups and boundary transitive groups
acting on trees, however, up to now very little is known about representation theory. This is despite
the fact that this class contains very natural examples, such as Burger-Mozes groups associated with
non 2-transitive permutation groups [7]. Astonishingly, up to now there is no result available that
provides examples of non-discrete non-type I groups acting on trees. Recent attempts to approach
this problem by classical methods [11] did not yield the desired conclusion even for the best understood
examples of Burger-Mozes groups. In this article, we take a new point of view and employ operator
algebraic methods, proving that a huge class of groups acting on trees is not of type I.
Theorem A. Let T be a locally finite tree and G ≤c Aut(T ) a closed non-amenable subgroup acting
minimally on T . If G does not act locally 2-transitive, then G is a not a type I group.
An action of a group G on a tree T is called minimal, if T is the smallest non-empty G-invariant subtree
of T . The action G ↷ T of a group on a tree is called locally 2-transitive, if for every vertex v ∈ V(T )
the action of the point stabiliser Gv on adjacent geometric edges of v is 2-transitive. See Section 2.3
for more explanations. The fact that we are able to prove a non-type I result in the generality of
Theorem A, insinuates the possibility of characterising those groups acting on trees that are of type I.
In fact, a non-compact closed subgroup G ≤c Aut(T ) is boundary transitive if and only if it is n-locally
transitive for every n in the sense of [7]: for every vertex v of T and every n the stabiliser Gv acts
transitively on spheres of radius n around v . Since G is locally 2-transitive if and only if it is 2-locally
transitive, this notion provides a clear transition between groups acting not locally 2-transitively and
boundary transitive groups. We hence pose the following problem, going beyond the type I conjecture.
Problem 1. Among closed subgroups of Aut(T ), characterise those which are of type I.
G ≤c Aut(T ) is . . .
Statement
Expectation/Result
boundary transitive
Type I conjecture
G is type I
(n − 1)-locally transitive,
but not n-locally transitive
open
G is not type I
not 2-locally transitive
Theorem A
G is not type I
Problem 1: non-amenable groups acting minimally on T
The operator algebraic perspective introduced in this article reduces the problem to extend Theorem A to general non-boundary transitive groups to considerations in representation theory.
2
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Burger–Mozes groups [7], also known as universal groups acting on trees, form a particularly
interesting class of examples of closed subgroups of Aut(T ). After choice of a permutation group
F ≤ Sn , Burger–Mozes construct groups U(F ) and index two subgroups U(F )+ acting on the n-regular
tree in such a way that their local action around vertices is prescribed by F . These groups U(F )+
attract particular interest of the totally disconnected group community, since they provide concrete
examples of abstractly simple and compactly generated non-discrete groups [8, 9, 2, 36]. Applying
Theorem A and combining it with known type I results [1, 10], we give a complete characterisation of
type I groups in this important class of examples.
Theorem B. Let T be a locally finite tree and G ≤c Aut(T ) a closed vertex transitive subgroup with
Tits’ independence property acting minimally on ∂T . Then G is a type I group if and only if G is
locally 2-transitive.
In particular, if F ≤ Sn is a permutation group, then the Burger-Mozes groups U(F ) and U(F )+
are type I groups if and only if F is 2-transitive.
We prove Theorem A with operator algebraic methods. The possibility to apply operator algebraic
methods to study totally disconnected groups in general and groups acting on trees in particular has
been previously suggested by the second author. Positive results exploiting the additional flexibility
provided by this idea can be found in [32] and [33]. A locally compact group is of type I if and only
if its maximal group C∗ -algebra C∗max (G) is a type I C∗ -algebra in the sense of [20]. Further, it is a
well-known fact for operator algebraists that every type I C∗ -algebra is amenable. This line of thoughts
suggests to study non-amenability of operator algebras associated with groups acting on trees. Since
amenability of C∗max (G) implies amenability of the group von Neumann algebra L(G), Theorem A is
an immediate consequence of the following operator algebraic result, which is the main result of the
present article. Its proof is based on the possibility to reduce considerations about amalgamated free
products of groups to plain free products of von Neumann algebras, for which clear non-amenability
criteria are available.
Theorem C. Let T be a locally finite tree and G ≤c Aut(T ) a closed non-amenable subgroup acting
minimally on T . If G does not act locally 2-transitive, then L(G) is non-amenable.
Although we want the type I conjecture to be understood as the main motivation for our present
work, our von Neumann algebraic techniques allow us to prove other non-amenability criteria. We
single out the class of groups acting properly and not edge-transitively on a tree T , but which not
necessarily embed as subgroups of Aut(T ). If G ↷ T , we denote by G + ≤ G the subgroup of typepreserving elements, which has at most index two.
Theorem D. Let T be a tree and G ↷ T a proper action of a locally compact group. Let X = G + /T
be the quotient graph and note that π1 (X) is a free group. Under either of the following sets of
assumptions, L(G) is non-amenable.
• rank π1 (X) ≥ 2.
• rank π1 (X) = 1 and G is non-amenable.
• π1 (X) = 0, T is thick, X is finite but not an edge and G is non-amenable.
A tree is called thick, if each of its vertices has valency at least three.
3
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
While in the case of a discrete group Γ, the group von Neumann algebra L(Γ) is amenable if and
only if Γ is amenable, it is even an open problem to provide general non-amenability criteria for the
maximal group C∗ -algebra of a non-discrete group. A result demonstrating the surprising difficulty of
this problem is provided by Connes [13, Corollary 7], who shows that the maximal group C∗ -algebra
of a connected locally compact separable group is amenable. Only Lau-Paterson were able to provide
a non-amenability criterion of general nature, although their assumption of inner amenability is very
strong [25]. Our work contributes to the understanding of further non-amenability criteria.
Theorem E. Let T be a locally finite tree and G ≤c Aut(T ) a closed non-amenable subgroup acting
minimally on T . If G does not act locally 2-transitive, then C∗max (G) is not nuclear.
In line with the previous explanations and the success of operator algebraic methods applied to
groups acting on trees, it is natural to pose the following problem, parallel to Problem 1.
Problem 2. Characterise closed subgroups G ≤c Aut(T ) for which L(G) is amenable. For which
groups among these is C∗max (G) amenable?
Acknowledgements
The authors thank the Mittag-Leffler Institute and the organisers of the workshop “Classification
and dynamical systems II: Von Neumann Algebras” as well as the Mathematisches Forschungsinstitut
Oberwolfach and the organisers of the workshop “C∗ -algebras” for providing excellent working conditions. Part of our work on this project was completed during these workshops. The second author
thanks the University of Münster for the excellent working conditions during the return phase of his
Marie Curie Fellowship, when big parts of this work were done.
2
Preliminaries
In the proceeding extensive preliminaries we provide readers with either operator algebraic or group
theoretic background with the necessary background to follow the main Sections 3, 4, 5 and 6.
2.1
Locally compact groups
In this article we are working in the setting of topological groups and their morphisms. This means
that a homomorphism between topological groups is understood to be continuous and isomorphisms
of topological groups are continuous bijective group homomorphisms with a continous inverse.
If G is a locally compact group, we write ∫G f (x)dx for integration against a left Haar measure.
Here, the function f on G can take values in any Banach space, thanks to the theory of Bochner
integrals. We refer the reader to [15] for these and other basics about locally compact groups.
The following theorem characterises totally disconnected locally compact groups. It is well-known
to people working in group theory, but we give a short proof for the convenience of the reader.
Theorem 2.1 (TG 39 in [44]). Let G be a locally compact group. Then G is totally disconnected if
and only if its identity admits a basis of neighbourhoods consisting of compact open subgroups.
4
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Proof. If G admits a basis of neighbourhoods consisting of compact open subgroups, then it is clear
that the connected component of e is {e}. So G is totally disconnected.
Assume that G is totally disconnected and let U ⊂ G be a compact open neighbourhood of the
identity. We will find a compact open subgroup of U. Let m ∶ G × G → G be the multiplication map.
Since {e} × U ⊂ m−1 (U), for every g ∈ U there is a neighbourhood Vg × Ug ⊂ m−1 (U) of (e, g). Since U
is compact, we hence find identity neighbourhoods V1 , . . . , Vn ⊂ G and open sets U1 , . . . , Un ⊂ G such
that Vi Ui ⊂ U for all i ∈ {1, . . . , n} and U = ⋃i Ui . Putting V ∶= ⋂i (Vi ∩ Vi −1 ), we obtain a non-empty
open symmetric set V ⊂ U such that V U ⊂ U. We conclude that the group K = ⋃k∈N V k ⊂ U is a
compact open subgroup of G lying in U.
The unimodular part. We denote the modular function of a locally compact group G by ∆G ∶ G → R>0 .
The modular function of totally disconnected groups is nicely behaved. If K ≤ G is a compact
open subgroup of a locally compact group, then ∆G ∣K = ∆K ≡ 1, shows that the kernel of ∆G is
open. In this case, we write G0 ∶= ker ∆G for the unimodular part of G.
2.2
Permutation groups
An action of a topological group on a set is called a permutation action. A permutation group is
a group G with a fixed faithful permutation action G ↷ X. We usually write G ≤ Sym(X) for a
permutation group.
If G ↷ X is a permutation action and S ⊂ X, we denote by FixG (S) = {g ∈ G ∣ ∀s ∈ S ∶ gs = s} the
pointwise stabiliser of S. In case S = {s} is a one-element set, we also write FixG (S) = Gs .
Definition 2.2. Let G ↷ X be a permutation action. We say that G acts 2-transitively, if Gx ↷ X ∖{x}
is transitive for every x ∈ X.
Remark 2.3. The notion of 2-transitivity for G ↷ X slightly defers from the usual definition. If ∣X∣ ≥ 3,
then it is equivalent to the assumption that for each pairs x1 ≠ x2 and y1 ≠ y2 in X there is some
g ∈ G such that gxi = yi for i ∈ {1, 2}. Only in case ∣X∣ = 2, our definition says that the trivial action
is 2-transitive, while it does not satisfy the usual definition.
We chose to adopt our notion of 2-transitivity to obtain clean formulations of all theorems about
groups acting on trees, for which otherwise the vertices of degree two need a cumbersome separate
treatment, complicating the theorems’ statements.
For an arbitrary topological group G and an open subgroup H ≤ G, the action G ↷ G/H is a
permutation action. The next lemma is a reformulation of the well-known fact that a 2-transitive
permutation group is primitive.
Lemma 2.4. Let H ≤ G be an open subgroup of a topological group. If ∣H/G/H∣ ≤ 2, then H ≤ G is a
maximal subgroup.
Proof. Assume that there is a proper inclusion of open subgroups H ≤ H̃ ≤ G of the topological group
G. Then H̃ ⊂ G is a H-biinvariant set, so that H/G/H = (H/H̃/H) ∪ (H/(G ∖ H̃)/H). Since H ≤ H̃ is
a proper inclusion, ∣H/H̃/H∣ ≥ 2 and ∣H/G/H∣ ≥ 3 follows. This proves the lemma.
5
Von Neumann algebras of groups acting on trees
2.3
by Cyril Houdayer and Sven Raum
Groups acting on trees
We follow Serre’s formalism of undirected trees [35]. A graph X is a set of vertices V(X) with a set of
(directed) edges E(X) as well as maps o, t ∶ E(X) → V(X) and a involutive operation taking opposite
edges e ↦ e such that e ≠ e, o(e) = t(e) and t(e) = o(e) for all e ∈ E(X). If X, Y are graphs, a graph
homomorphism ϕ ∶ X → Y is a pair of maps ϕV ∶ V(X) → V(Y ) and ϕE ∶ E(X) → E(Y ) such that
tY ○ ϕE = ϕV ○ tX and oY ○ ϕE = ϕV ○ oX .
Segments and paths. The standard segment of length n is written [0, n]. Its set of vertices is
V([0, n]) = {0, . . . , n}
and its edges are pairs
E([0, n]) = {(i , i + 1) ∣ i ∈ {0, n − 1}} ∪ {(i , i − 1) ∣ i ∈ {1, . . . , n}}
with o(i , j) = i , t(i , j) = j and (i , j) = (j, i ) for all (i , j) ∈ E([0, n]). A path in a graph X is graph
homomorphism [0, n] → X. We set o(s) = s(0) and t(s) = s(n) for a path s ∶ [0, n] → X.
Trees. A graph X is connected if there is a path between pairs of vertices in X. A circuit in X
is the image of an injective path s ∶ [0, n] → X with o(s) = t(s) for some n ≥ 1. A tree
is a connected non-empty graph without circuits. Let T be a tree. For v ∈ V(T ) we write
E(v ) = {e ∈ E(T ) ∣ o(e) = v } for the neighbouring edges of v . We call T locally finite if E(v ) is
finite for all v ∈ V(T ). We call T thick if E(v ) contains at least three elements for all v ∈ V(T ).
Automorphisms of a tree. The group Aut(T ) of graph automorphisms of a tree T naturally identifies
with the subgroup Aut(T ) = {g ∈ Sym(V(T )) ∣ v ∼ w ⇔ g(v ) ∼ g(w )} and thus inherits a
totally disconnected group topology, which is uniquely defined by declaring vertex stabilisers
open subgroups of Aut(T ). An action of a topological group G on a tree T is a continuous
group homomorphism G → Aut(T ). If T is locally finite, then vertex stabilisers are compact in
Aut(T ). If T is a tree and G ↷ T is an action, then the following statements are equivalent.
• Gv ≤ G is compact for all v ∈ V(T ).
• G ↷ T is proper.
If further, G ≤ Aut(T ) embeds as a subgroup, then both previous statements are equivalent to
G ≤ Aut(T ) being closed.
Locally 2-transitive actions. A group action G ↷ T on a tree is called locally 2-transitive if for every
vertex v ∈ V(T ) the natural action Gv ↷ E(v ) is 2-transitive.
Type-preserving automorphisms. An element g ∈ Aut(T ) is called type-preserving if 2 ∣ d(gv , v )
for all v ∈ V(T ). Denote by Aut(T )+ ≤ Aut(T ) the group of type-preserving automorphisms.
Partitioning V(T ) is two classes by v ∼ w if and only if 2 ∣ d(v , w ), we obtain a quotient map
V(T ) ↦ {0, 1}. Since Aut(T ) preserves this partition, we obtain a map Aut(T ) ↦ S2 , whose
kernel is Aut(T )+ . This shows that Aut(T )+ ≤ Aut(T ) is an open subgroup of index at most
two. If G ↷ T is a group action on a tree, we denote by G + the inverse image of Aut(T )+ under
the action map G → Aut(T ) and call G ↷ T type-preserving if G = G + .
6
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Note that if G ↷ T is proper, then also the type-preserving part G + ≤ G acts properly, because
the restriction of a proper action to a closed subgroup remains proper.
Minimal actions on trees. A group action G ↷ T on a tree is called minimal, if T is the smallest
non-empty G-invariant subtree of T .
Ends of a tree. The standard ray [0, ∞) is a tree with vertices V([0, ∞)) = N and edges
E([0, ∞)) = {(i , i + 1) ∣ i ∈ N} ∪ {(i , i − 1) ∣ i ∈ N≥1 }
with o(i , j) = i , t(i , j) = j and (i , j) = (j, i ) for all (i , j) ∈ E([0, ∞)). A geodesic ray in a tree
T is an injective graph homomorphism [0, ∞) → T . Two geodesic rays are called equivalent, if
after shifting they eventually agree. Formally, ξ ∼ ξ′ if there are n0 ∈ N and m ∈ Z such that
ξ(n + m) = ξ′ (n) for all n ≥ n0 . An end of T is an equivalence class of geodesic rays in T .
Hyperbolic elements. The standard two-sided geodesic (−∞, ∞) is a tree with vertices V((−∞, ∞)) =
Z and edges E((−∞, ∞)) = {(i , i +1) ∣ i ∈ Z}∪{(i , i −1) ∣ i ∈ Z}. The origin and target functions are
o(i, j) = i and t(i , j) = j. The opposite edge of (i , j) ∈ E((−∞, ∞)) is (i , j) = (j, i ). A (two-sided)
geodesic in a tree T is an injective graph homomorphism (−∞, ∞) → T . An element g ∈ Aut(T )
is called hyperbolic if it neither fixes a vertex nor an edge (formally: a set {e, e} ⊂ E(T )). For
every hyperbolic element g ∈ Aut(T ) there is a unique two-sided geodesic ξ in T which is setwise
fixed. The unique ` ∈ N such that g ○ ξ(n) = ξ(n + `) for all n ∈ Z is called the translation length
of g.
The following result characterises amenable groups acting on trees.
Theorem 2.5 (Theorem 1 in [28]). Let T be a locally finite tree and G ≤c Aut(T ) a closed subgroup.
Then G is amenable if and only if one of the following statements holds
• G fixes a vertex.
• G stabilises an edge.
• G fixes a point in ∂T .
• G stabilises a pair of points in ∂T .
2.4
Bass-Serre theory
Bass-Serre theory as described in [35] (see in particular Section 5 in there) provides a natural way to
study groups acting on trees G ↷ T by means of the quotient graph G/T together with vertex and
edge stabilisers. The general fundamental assumption of Bass-Serre theory is that G ↷ T must act
without inversions, i.e. if g ∈ G fixes a geometric edge of T , then it fixes both its ends. It follows
from the definition that every type-preserving action satisfies this assumption. Bass-Serre theory was
originally built for discrete groups, not taking into account topologies. Its extension to topological
groups however is straight forward, as we will clarify at the end of this section.
Graphs of groups. A graph of groups is a graph X with vertex groups (Gv )v ∈V(X) and edge group
(Ge )e∈E(X) as well as inclusions Ge ↪ Gt(e) such that Ge = Ge . We denote this graph of groups
by (G, X) for short.
7
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Fundamental group of a graph of groups. If (G, X) is a graph of groups, then Bass-Serre theory
provides a tree T — called universal covering of (G, X) — with an action of a group π1 (G, X)
on T , such that X = π1 (G, X)/T and (G, X) is obtained by considering vertex and edge stabilisers
of lifted edges from X to T . This construction provides a one-to-one correspondence between
isomorphism classes of graphs of groups and groups acting on trees. See Theorem 13 in [35].
If G ↷ T is a group acting on a tree with quotient graph X = G/T , we will use the convenient
notation (G, X) for the graph of groups obtained from this action.
Contractions of subtrees. See [35, p.46ff]. If (G, X) is a graph of groups and (G, Y ) is a subgraph,
then π1 (G, Y ) can be naturally identified with a subgroup of π1 (G, X). Contracting Y ≤ X
to a vertex, we obtain a graph X/Y . The contraction can be naturally turned in a graph of
groups such that the vertex group of the contracted vertex Y ∈ V(X/Y ) is π1 (G, Y ). We
denote this graph of groups by (G, X/Y ). Now we have the identity of fundamental groups
π1 (G, X) = π1 (G, X/Y ) extending uniquely the natural inclusion of vertex and edge stabilisers
of (G, X) into π1 (G, X/Y ).
Semi-direct product decomposition. See [35, p. 45, exercise]. If (G, X) is a graph of groups, then
the universal cover T̃ of X in the usual sense can be naturally turned into a tree of groups whose
vertex and edge groups are isomorphic to vertex and edge groups of X. We denote this tree of
groups by (G, T̃ ) and call it the covering tree of groups of (G, X). If Γ = π1 (X) is the usual
fundamental group of the graph X, then the action of Γ by Deck transformations on T̃ induces
an action on π1 (G, T̃ ) and we obtain a natural isomorphism π1 (G, X) ≅ π1 (G, T̃ ) ⋊ Γ.
Graphs of topological groups. If T is a tree and G ↷ T an action (which is understood to be continuous) of a topological group, then Bass-Serre theory naturally applies and is compatible with
the topology of G. Denote by X = G/T the quotient graph and by (G, X) the associated graph
of groups. In this context, vertex and edge stabilisers of G ↷ T are topological groups and
inclusion homomorphisms are continuous and open. Since G as a topological group is uniquely
determined by the abstract group G together with the topology on vertex stabilisers, it makes
sense to speak about graphs of topological groups.
Definition 2.6. A graph of topological groups is a graph of groups (G, X) with the structure of
a topological group on each vertex and edge stabiliser such that inclusion homomorphisms are
continuous and open.
Based on Bass-Serre theory and Serre’s “dévissage” it is not difficult to prove that the fundamental group of a graph of topological groups carries a unique group topology turning the
inclusion of vertex groups into continuous and open homomorphisms. All previously mentioned
constructions and statements remain valid in the topological setting. For later use, we remark
in particular that the semi-direct product decomposition π1 (G, X) ≅ π1 (G, T̃ ) ⋊ π1 (X) for a
graph of topological groups (G, X) and the covering tree of groups (G, T̃ ) of (G, X) gives rise
to an embedding of π1 (G, T̃ ) as an open subgroup of π1 (G, x). We fix the following notation:
a locally compact amalgamated free product is an amalgamated free product with an open locally compact amalgam. As previously discussed a locally compact amalgamated free product is
naturally a locally compact group.
8
Von Neumann algebras of groups acting on trees
2.5
by Cyril Houdayer and Sven Raum
Von Neumann algebras
Let H be a complex Hilbert space and B(H) the *-algebra of all bounded linear operators on H. The
topology of pointwise convergence on B(H) is called the strong operator topology.
Definition 2.7. A von Neumann algebra is a unital strongly closed *-subalgebra of B(H) for some
Hilbert space H.
The σ-weak topology. Since the norm topology on B(H) is finer than the strong operator topology,
every von Neumann algebra is naturally a Banach space. By a result of Sakai [34], a von
Neumann algebra admits a unique isometric predual M∗ , that is a Banach space satisfying
(M∗ )∗ ≅ M isometrically. The weak-*-topology on M is called the σ-weak topology. A positive
linear map (in particular a *-homomorphism) ϕ ∶ M → N between von Neumann algebras is
called normal if it is σ-weakly continuous.
Traces and finite von Neumann algebras. A positive functional τ ∶ M → C on a von Neumann algebra
is called a trace if τ (xy ) = τ (y x) for all x, y ∈ M. A von Neumann algebra is called finite if
it admits a faithful family of normal traces, that is a family (τi )i of normal traces such that
τi (x ∗ x) = 0 for all i implies x = 0.
Factors. A factor is a von Neumann algebra M with trivial centre Z(M) ∶= M ∩ M ′ = C1. If M is an
infinite dimensional factor with a non-zero trace, then M is called a II1 factor. The non-zero
trace on a II1 factor is unique up to normalisation.
Positive elements. We denote by M + = {x ∗ x ∣ x ∈ M} the set of all positive elements in a von Neumann
algebra M. A linear map ϕ ∶ M → N between von Neumann algebras is called positive if
ϕ(M + ) ⊂ N + .
Conditional expectations. If N ⊂ M is an inclusion of von Neumann algebras, a conditional expectation
E ∶ M → N is a projection of norm one. It is called normal if it is σ-weakly continuous. It satisfies
E(n1 mn2 ) = n1 E(m)n2 for all n1 , n2 ∈ N and all m ∈ M.
Weights. A weight on a von Neumann algebra M is an additive and positive homogeneous map
ϕ ∶ M + → R≥0 ∪ {∞}. We say that ϕ is faithful, if ϕ(x) = 0 implies x = 0 for every x ∈ M + .
The weight ϕ is called normal if supi ϕ(xi ) = ϕ(supi xi ) for every bounded ascending net (xi ) of
positive elements in M. Here supi xi denotes the smallest upper bound for the net (xi )i . One
calls nϕ = {x ∈ M ∣ ϕ(x ∗ x) < ∞} the set of 2-integrable elements. If ϕ is a normal weight and
nϕ ⊂ M is σ-weakly dense, then ϕ is called semifinite. A normal faithful semifinite weight is
abbreviated to an nfsf weight.
Modular automorphism group. If ϕ is an nfsf weight on a von Neumann algebra M, the set nϕ with
the scalar product ⟨x, y ⟩ ∶= ϕ(y ∗ x) can be completed to a Hilbert space L2 (M, ϕ) on which M is
faithfully represented via left multiplication. The map S ∶ x ↦ x ∗ on nϕ ∩ n∗ϕ ⊂ L2 (M, ϕ) defines
a conjugate linear closable unbounded operator, whose polar decomposition is denoted by S =
J∆1/2 . For every t ∈ R, the operator ∆it is a well-defined unitary on L2 (M, ϕ). Tomita-Takesaki
theory [38] says that the conjugation (Ad ∆it )t∈R defines a one-parameter automorphism group
of B(L2 (M, ϕ)) that preserves M. Its restriction to M is denoted by (σtϕ )t and it is called the
modular automorphism group of ϕ.
9
Von Neumann algebras of groups acting on trees
2.6
by Cyril Houdayer and Sven Raum
Group von Neumann algebras
We refer the reader to [15] for an introduction to locally compact groups, their representations and
convolution algebras. Let G be a locally compact group and λG ∶ G → U(L2 (G)) its left-regular
representation. It satisfies (λG (g)f )(x) = f (g −1 x) for all f ∈ Cc (G) and g, x ∈ G. The group von
Neumann algebra of G is by definition
L(G) ∶= {λG (g) ∣ g ∈ G}′′ ⊂ B(L2 (G)) .
We usually write ug = λG (g) for the canonical unitaries in L(G). They span an isomorphic copy of
CG, to which we refer without explicitly mentioning λG . Von Neumann’s bicommutant theorem says
that L(G) is the strong and the σ-weak closure of the set CG. After choice of a left Haar measure
on G, the Bochner integral provides a natural *-homomorphism L1 (G) → L(G) ∶ f ↦ ∫G f (g)λG (g)dg,
which we will also denote by λG . If no confusion is possible, we write L1 (G) ⊂ L(G) instead of
λG (L1 (G)) ⊂ L(G).
The convolution algebra Cc (G) is a left Hilbert algebra in the sense of [38, Chapter VI.1]. After
choice of a left Haar measure it defines a nfsf weight ϕ on L(G) that satisfies ϕ(f ) = f (e) for all
f ∈ Cc (G) ⊂ L(G). This weight is called a (left) Plancherel weight of L(G). It satisfies ϕ(g ∗ ∗f ) = ⟨f , g⟩
for all f , g ∈ Cc (G) ⊂ L(G)∩L2 (G). The modular autormorphism group of ϕ satisfies σtϕ (ug ) = ∆it
G (g)ug
for all g ∈ G. If G is a discrete group, the Plancherel weight associated with the counting measure
extends to the natural normal trace τ ∶ L(G) → C satisfying τ (ug ) = δe,g for all g ∈ G.
The next proposition is well-known and clarifies the relation between the group von Neumann
algebras of a locally compact group and its closed subgroups. It can be found for example as Theorem A
of [23].
Proposition 2.8. Let H ≤ G be a closed subgroup of a locally compact group. Then the group
homomorphism H ∋ h ↦ λG (h) ∈ U(L(G)) extends to a unique injective normal *-homomorphism
L(H) → L(G).
Proof. Denote by A(G) Eymard’s Fourier algebra [19, Chapitre 3], which is a Banach algebra densely
spanned by continuous positive type functions with compact support in G. By Theorem 3.10 of [19] we
have L(G)∗ = A(G), i.e. there is an isomorphism L(G) ≅ A(G)∗ carrying the σ-weak topology onto the
weak-*-topology. This isomorphism identifies ug ∈ L(G) with the evaluation functional evg ∈ A(G)∗
for all g ∈ G.
Since H ≤c G is a closed subgroup, every compactly supported function of positive type on G
restricts to a compactly supported function of positive type on H. So Proposition 3.4 in [19] shows
that the restriction gives rise to well-defined map A(G) → A(H). By Theorems 1a and 1b of [23] (see
also Theorem 4.21 of [26]), every element of A(H) can be extended to an element of A(G). This
shows surjectivity of the restriction map A(G) → A(H). It follows that the dual map A(H)∗ → A(G)∗
is injective. In view of the first paragraph this finishes the proof of the proposition.
Averaging projections. Applied to a compact subgroup K ≤ G of a locally compact group, the previous
proposition shows that the Bochner integral pK ∶= ∫K λG (k)dk ∈ L(G) defines a projection. Here
we integrate against the Haar probability of K. It is the image of 1K ∈ Cc (K) ⊂ L(K) ⊂ L(G).
This projection is called averaging projection associated with K ≤ G.
If H ≤o G is an open subgroup, the inclusion L(H) ⊂ L(G) from Proposition 2.8 admits a natural
conditional expectation. Also this fact is well-known. It follows from Theorem 3.1(a) in [21] in the
special case M = C1 and ϕ = 1H . We give a short proof only for the readers convenience.
10
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Proposition 2.9. Let H ≤ G be an open subgroup of a locally compact group. Then the embedding
H ≤ G extends to a unique injective normal *-homomorphism L(H) ⊂ L(G). Further, there is a unique
normal conditional expectation E ∶ L(G) → L(H) satisfying E(ug ) = 1H (g)ug for all g ∈ G.
Proof. The fact that h ↦ λG (h) extends to a unique embedding L(H) ↪ L(G) is the content of
Proposition 2.8. Let us construct E. Denote by ϕ ∶ L(G)+ → [0, +∞] a Plancherel weight on
L(G). The dense subalgebra Cc (G) ⊂ L(G) consists of ϕ-integrable elements and ϕ(f ) = f (e) for
all f ∈ Cc (G). Since H ≤ G is open, we have Cc (H) ⊂ Cc (G). Further, Cc (H) ⊂ L(H) is a σweakly dense subalgebra, implying that ϕ is semifinite on L(H). Further, L(H) is σ ϕ -invariant. By
Takesaki’s theorem [37] there is a unique normal conditional expectation E ∶ L(G) → L(H) satisfying
ϕ(E(x)) = ϕ(x) for all x ∈ mϕ . For f ∈ Cc (H) ⊂ Cc (G) we have E(f ) = f . For f ∈ Cc (G ∖ H) and
g ∈ Cc (H), we have g ∗ f ∈ Cc (G ∖ H) and ϕ(gE(f )) = ϕ(gf ) = 0. So E(f ) = 0. This shows that
E∣Cc (G) is the restriction map Cc (G) → Cc (H). If g ∈ G ∖ H, then ug is a σ-weak limit of elements in
Cc (G ∖ H), so that E(ug ) = 0 follows. This proves existence of E. Uniqueness follows from the fact
that CG ⊂ L(G) is σ-weakly dense.
In case K ⊴ G is a compact subgroup of a locally compact group, the group von Neumann algebras
L(G) and L(G/K) can also be compared in a natural way. This is the content of the next well-known
proposition.
Proposition 2.10. Let G be a locally compact group and K ⊴ G a compact normal subgroup. Then
the averaging projection p associated with K defines a central projection in L(G) such that pL(G) ≅
L(G/K). In particular, L(G) is non-amenable, if L(G/K) is non-amenable.
Proof. Recall that we can write p = ∫K uk dk as a Bochner integral against the Haar probability measure
of K. We have
⟨ug pug∗ ξ, η⟩ = ∫ ⟨ugkg −1 ξ, η⟩dk
K
= ∫ ⟨u g k ξ, η⟩dk
K
= ∫ ⟨uk ξ, η⟩dk
K
= ⟨pξ, η⟩ ,
for all g ∈ G and ξ, η ∈ L2 (G). The third equality follows from the fact that the Haar measure on K is
invariant under the conjugation action of G. So p ∈ L(G) ∩ CG ′ = Z(L(G)).
Note that (pξ)(g) = ∫K ξ(kg)dk for all ξ ∈ Cc (G) ⊂ L2 (G), so that pL2 (G) = L2 (G)K follows.
Consider the map V ∶ L2 (G/K) → L2 (G) defined by (V f )(g) = f (gK) for f ∈ Cc (G/K). Since
K ⊴ G is compact and normal, V is well-defined and isometric. A short calculation shows that
V L2 (G/K) = L2 (G)K = pL2 (G), meaning that V V ∗ = p. So V ∶ L2 (G/K) → L2 (G)K is a unitary.
Denoting the canonical unitaries in L(G/K) by vgK , gK ∈ G/K, another calculation on the dense
subset Cc (G/K) verifies that pug V = V vgK for every g ∈ G. This shows V ∗ pL(G)pV = L(G/K).
Since x ↦ px is a conditional expectation (even a *-homomorphism) from L(G) onto pL(G) ≅
L(G/K), it follows from Proposition 2.15 that non-amenability of L(G/K) implies non-amenability of
L(G).
11
Von Neumann algebras of groups acting on trees
2.6.1
by Cyril Houdayer and Sven Raum
Hecke (von Neumann) algebras
On the level of group algebras, there is a replacement for the quotient G/K of a locally compact
group G by a compact normal subgroup K ⊴ G, even if we drop the assumption of normality. This
replacement is provided by Hecke algebras.
Definition 2.11. Let G be a totally disconnected group and K ≤ G a compact open subgroup. Then
(G, K) is called a Hecke pair. Let p = pK ∈ Cc (G) be the averaging projection associated with K.
Then Cc (G, K) ∶= pCc (G)p is called the Hecke algebra of the pair (G, K) and L(G, K) ∶= pL(G)p is
called the Hecke von Neumann algebra of the pair (G, K).
Remark 2.12. By a result of Tzanev [41] our definition of a Hecke algebra and a Hecke von Neumann
algebra agree with the usual definitions. That is, Cc (G, K) is the set of all compactly supported
K-biinvariant functions in Cc (G) and L(G, K) is the von Neumann algebra closure of Cc (G, K) in its
representation on `2 (K/G).
We will need the following formula for the dimension of a Hecke algebra in later applications.
Proposition 2.13. Let (G, K) be a Hecke pair. Then dim Cc (G, K) = ∣K/G/K∣.
Proof. We write p = 1K ∈ Cc (G) for the averaging projection associated with K ≤ G. If KgK ∈ K/G/K,
then pug p = 1KgK ∈ Cc (G, K). Further, it is clear that these elements generate Cc (G, K) as a linear
space. Let ϕ ∶ Cc (G) → C be (the linear extension of) a Plancherel weight on Cc (G) ⊂ L(G). For
KgK ≠ KhK, we have ϕ((pug p)∗ puh p) = (1K ∗ 1gK ∗ 1h−1 K )(e) = 0, since e ∉ KgKh−1 K. This shows
that the elements pug p are pairwise orthogonal in L2 (G) ⊃ Cc (G). In particular, (pug p)KgK∈K/G/K is
a linearly independent family in Cc (G, K). This shows dim Cc (G, K) = ∣K/G/K∣.
2.6.2
Group factors
The following criterion describes discrete groups whose group von Neumann algebra is a factor. In
the well-known proof, we make use of the right-regular representation ρG ∶ G → U(L2 (G)) of a locally
compact group G, which satisfies (ρG (g)f )(x) = f (xg) for all f ∈ Cc (G) and g, x ∈ G.
Proposition 2.14. Let Γ be a discrete group. Then L(Γ) is a factor if and only if every non-trivial
conjugacy class in Γ is infinite. If Γ is non-trivial, then L(Γ) is a II1 factor.
Proof. If Γ has a non-trivial finite conjugacy class c ⊂ Γ, then x ∶= ∑g∈c ug satisfies ug xug∗ = x for all
g ∈ Γ. So x ∈ Z(L(G)) is a witnesses that L(Γ) is not a factor.
Assume that every conjugacy class of Γ is infinite. The map L(Γ) ∋ x ↦ xδe ∈ `2 (Γ) is faithful,
since xδg = ρg −1 xδe for all g ∈ Γ and the vectors δg , g ∈ Γ are a basis of `2 (Γ). So if x ∈ Z(L(G)), it
suffices to show that xδe ∈ Cδe . We have
(xδe )(ghg −1 ) = ⟨xδe , δghg −1 ⟩
= ⟨xδe , λG (g)ρG (g)δh ⟩
= ⟨λG (g)∗ ρG (g)∗ xδe , δh ⟩
= ⟨xλG (g)∗ ρG (g)∗ δe , δh ⟩
= ⟨xδe , δh ⟩
= (xδe )(h) ,
12
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
for all g, h ∈ Γ. Hence xδe is constant on conjugacy classes. Since xδe is also 2-summable and every
non-trivial conjugacy class of Γ is infinite, it follows that xδe ∈ Cδe indeed.
If Γ is a non-trivial icc group, then it is infinite. So L(Γ) is an infinite dimensional factor. Since Γ
is discrete, there is the natural trace on L(Γ) showing that it is a II1 factor.
2.6.3
Amenable von Neumann algebras
A von Neumann algebra M ⊂ B(H) is called injective, if there is some (not necessarily normal)
conditional expectation E ∶ B(H) → M. Following the suggestion of Connes [14], we refer to this class
of von Neumann algebras as amenable von Neumann algebras.
Proposition 2.15. If N ⊂ M is an inclusion of von Neumann algebras with conditional expectation
and M is amenable, then N is amenable. In particular, if M is an amenable and finite von Neumann
algebra, then every von Neumann subalgebra of M is amenable.
Proof. From the definitions, the first part of the proposition follows on the nose. We only have
to prove that every von Neumann subalgebra of a finite von Neumann algebra admits a conditional
expectation. This follows from Takesaki’s theorem [37] and the fact that the modular automorphism
group of a trace is trivial.
We fix the following important consequence of Proposition 2.15.
Proposition 2.16. Let H ≤ G be an open subgroup of a locally compact group. If L(H) is nonamenable, then also L(G) is non-amenable.
Proof. Assume that L(H) is non-amenable. Proposition 2.9 tells us that there is a natural embedding
L(H) ↪ L(G) with a normal conditional expectation L(G) → L(H). We can apply Proposition 2.15,
in order to conclude that L(G) is non-amenable.
Let M be a II1 factor, k ∈ N>0 p ∈ Mk (C) ⊗ M a non-zero projection. Then p(Mk (C) ⊗ M)p is
called an amplification of M. Its isomophism class depends only on t ∶= (Tr ⊗ τ )(p), where Tr denotes
the non-normalised trace of Mk (C) and τ is the unique trace of M. Hence, we write M t for this
amplification.
We also need the following simple stability result for amenable II1 factors.
Proposition 2.17. Let M be a II1 factor and t > 0. Then M is amenable if and only if M t is amenable.
Proof. Fix an amenable von Neumann algebra M ⊂ B(H) and a conditional expectation E ∶ B(H) →
M. Then id ⊗ E ∶ B(K)⊗B(H) → B(K)⊗M is a conditional expectation witnessing amenability of
B(K)⊗M. If p ∈ M is a non-zero projection and p ⊥ = 1 − p is its orthogonal complement, then
M ∋ x ↦ pxp ∈ pMp ⊕ Cp ⊥ is a conditional expectation. So Proposition 2.15 implies amenability of
pMp ⊕ Cp ⊥ and hence of pMp . These arguments show that every amplification of M is amenable.
Further, M = (M t )1/t , so that the proposition follows.
The next theorem is classic and a proof can be found in Theorem 2.5.8 of [5].
Theorem 2.18. Let Γ be a discrete group. Then L(Γ) is amenable if and only if Γ is amenable.
13
Von Neumann algebras of groups acting on trees
2.6.4
by Cyril Houdayer and Sven Raum
Free group factors and non-amenable free products of von Neumann algebras
Let M1 , M2 be von Neumann algebras with fixed faithful normal states ϕi ∈ Mi∗ . The free product von
Neumann algebra (M1 , ϕ1 ) ∗ (M2 , ϕ2 ) is described in Chapters 1.6 and 2.5 of [45]. It is the unique
von Neumann algebra M generated by isomorphic copies of M1 and M2 together with a normal state
ϕ on M satisfying the freeness condition ϕ(x1 ⋯xn ) = 0 for all x1 , . . . , xn ∈ M1 ∪ M2 satisfying xi ∈ Mji ,
ϕji (xi ) = 0 with ji ≠ ji+1 for i ∈ {1, . . . , n − 1}. If no confusion is possible, we write M = M1 ∗ M2 for
the free product von Neumann algebra.
In this section, we briefly explain the following result due to Dykema.
Theorem 2.19 (See Theorem 4.6 of [17]). Let M, N be hyperfinite tracial von Neumann algebras
such that dim M, dim N ≥ 2 and dim M + dim N ≥ 5. Then M ∗ N is a non-amenable von Neumann
algebra.
Let Fn denote some non-abelian free group. Then L(Fn ) is called a free group factor. For any
k ∈ N>0 and any non-zero projection p ∈ Mk (C) ⊗ L(Fn ), the compression p(Mk (C) ⊗ L(Fn ))p is called
an interpolated free group factor. These von Neumann algebras were introduced independently in [18]
and [31], where among other things it was proven that the isomorphism class of p(Mk (C) ⊗ L(Fn ))p
n−1
only depends on t ∶= (Tr⊗τ
+ 1, where Tr denotes the non-normalised trace on Mk (C) and τ is the
)(p)2
canonical trace on L(Fn ). We hence write L(Ft ) for this von Neumann algebra.
Proposition 2.20. Interpolated free group factors L(Ft ), t > 1 are non-amenable.
Proof. Let t > 1 be real. By Proposition 2.14 and Theorem√2.18 we know that L(Fn ) is a non-amenable
II1 factor. So Proposition 2.17 shows that L(Ft ) = L(F2 )
1
t−1
is non-amenable.
Now Theorem 2.19 is a consequence of the following result, which is stated explicitly in the
literature.
Theorem 2.21 (See Theorem 4.6 of [17]). Let M, N be hyperfinite tracial von Neumann algebras
such that dim M, dim N ≥ 2 and dim M + dim N ≥ 5. Then a direct summand of M ∗ N is isomorphic
to some interpolated free group factor.
2.6.5
Amalgamated free product von Neumann algebras
If N ⊂ M is an inclusion of von Neumann algebras with a normal faithful conditional expectation
E ∶ M → N, we write M ⊖ N = {x ∈ M ∣ E(x) = 0}. Given two von Neumann algebras M1 , M2 with a
common von Neumann subalgebra N and normal faithful conditional expectations Ei ∶ Mi → N, there
is an amalgamated free product von Neumann algebra (M1 , E1 ) ∗N (M2 , E2 ) described in Chapter 3.8
of [45]. It is the unique von Neumann algebra M generated by isomorphic copies of M1 and M2 such
that M1 ∩ M2 = N in M as well as a normal conditional expectation E ∶ M → N obeying the freeness
condition E(x1 ⋯xn ) = 0 for all elements x1 , . . . , xn ∈ M1 ∪ M2 with xi ∈ Mji ⊖ N and ji ≠ ji+1 for all
i ∈ {1, . . . , n − 1}. Compare with Proposition 2.5 in [42].
Proposition 2.22. Let G = G1 ∗H G2 be a locally compact amalgamated free product. Then the inclusions L(G1 ), L(G2 ) ⊂ L(G) induce an isomorphism L(G) ≅ L(G1 ) ∗L(H) L(G2 ) where the amalgamated
free product is taken with respect to the natural conditional expectations.
14
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Proof. Denote by E ∶ L(G) → L(H) the normal conditional expectation associated by Proposition 2.9
with the open subgroup H ≤ G. It satisfies E(ug ) = 1H (g)ug for all g ∈ G. Denote by Ej ∶ L(G) → L(Gj )
the natural conditional expectations for j ∈ {1, 2}.
We want to apply Proposition 2.5 of [42] to conclude the proof. In order to do so we only need
to verify the freeness condition for L(Gj ) ⊂ L(G) with respect to E. Note that if g1 , . . . , gn ∈ G1 ∪ G2
with gi ∈ Gji ∖ H and ji ≠ ji+1 for all i ∈ {j1 , . . . , jn−1 }, then g1 ⋯gn ∈ G ∖ H. This implies E(ug1 ⋯ugn ) = 0.
Let x1 , . . . , xn ∈ L(G1 ) ∪ L(G2 ) with xi ∈ L(Gji ) ⊖ L(H) and ji ≠ ji+1 for all i ∈ {1, . . . , n − 1}. Since
CGj ⊂ L(Gj ) is strongly dense for j ∈ {1, 2}, Kaplansky’s density theorem provides us with bounded
α→∞
nets (xα,i )α in CGji for all i ∈ {1, . . . , n} such that xα,i → xi strongly. Write xα,i = ∑g∈Gi cg,α,i ug .
j
Since Eji (xi ) = 0, we have yα,i ∶= xα,i − Eji (xα,i ) → xi strongly. Since (yα,i )α is a bounded net, we also
obtain yα,1 ⋯yα,n → x1 ⋯xn strongly and hence also σ-weakly. We have yα,i = ∑g∈Gj ∖H cg,α,i ug , so that
i
E(yα,1 ⋯yα,n ) = 0 for all α by our initial remark on E. It follows that E(x1 ⋯xn ) = 0 by normality of
E.
3
Basic non-amenability results for group von Neumann algebras
In this section we provide the basic non-amenability results for group von Neumann algebras, which
are going to be used in Section 4. By means of Bass-Serre theory, all non-amenability questions we
face, can be answered with the next Lemmas 3.1 and 3.3.
Lemma 3.1. Let K ≤ G, H be two locally compact groups with a common compact open subgroup.
If ∣K/G/K∣ ≥ 3 and K ≤ H is a proper subgroup, then L(G ∗K H) is non-amenable.
Proof. Since K is a compact open subgroup of G and H, we have K ≤ G0 , H0 . So G0 ∗K H0 ≤ G ∗K H
is an open subgroup. So by Proposition 2.16 it suffices to prove that L(G0 ∗K H0 ) is non-amenable.
If ∣K/G0 /K∣ ≤ 2, then G0 follows compact. Hence G0 ⊴ G is a compact open normal subgroup,
showing that G = G0 is unimodular. So also ∣K/G/K∣ ≤ 2, which is a contradiction. We conclude
that ∣K/G0 /K∣ ≥ 3. Similarly, if K = H0 then H contains a compact open normal subgroup, and hence
H = H0 is unimodular. So K = H, which is a contradiction. This shows that K ≤ H0 is a proper
inclusion.
From now on assume that G, H are unimodular groups satisfying the assumptions of the lemma.
By Proposition 2.22, there is a natural isomorphism L(G ∗K H) ≅ L(G) ∗L(K) L(H) =∶ M. Write
p = pK ∈ L(K) for the averaging projection over K. Let ϕ be the Plancherel weight on M normalised
to satisfy ϕ(p) = 1. Then
pMp ⊃ pL(G)p ∗pL(K)p pL(H)p = pL(G)p ∗Cp pL(H)p .
We have dim pL(G)p ≥ ∣K/G/K∣ ≥ 3 by Proposition 2.13 and pL(H)p ≠ Cp, since dim pL(H)p ≥
∣K/H/K∣ ≥ 2. Since G, H are unimodular, pL(G)p and pL(H)p are tracial von Neumann algebras.
We can find unital hyperfinite von Neumann subalgebras NG ⊂ pL(G)p and NH ⊂ pL(H)p such that
dim NG ≥ 3 and NH ≠ Cp. So Dykema’s Theorem 2.19 applies to show that NG ∗Cp NH is nonamenable. Since NG ∗Cp NH is a non-amenable von Neumann subalgebra of the finite von Neumann
algebra pL(G)∗L(K) L(H)p, Proposition 2.15 says that also the latter is a non-amenable von Neumann
algebra. We conclude that a corner of M is non-amenable, and hence M is non-amenable by the same
proposition.
15
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Remark 3.2. Lemma 3.1 can be alternatively proved without reducing to the unimodular setting, if
we employ Ueda’s [43]. We prefer however to present a proof of Lemma 3.1 based on more classical
theorems on free product von Neumann algebras.
Lemma 3.3. Let G be the fundamental group of one of the following graphs of groups (G, X).
(i) X =
with compact edge groups and all inclusions proper, except for possibly one
inclusion into the vertex group of the middle vertex.
(ii) X a graph with at least three terminal edges e, f , g and terminal vertices x = t(e), y = t(f ), z =
t(g) such that Ge , Gf , Gg are compact and the inclusions Ge ↪ Gx , Gf ↪ Gy and Gg ↪ Gz are
proper.
Then L(G) is non-amenable.
Proof. Consider case (i) first. The statement that L(G) is non-amenable is equivalent to showing
that L(K1 ∗L1 K2 ∗L2 K3 ) is non-amenable if K1 , K2 , K3 are locally compact groups, L1 ≤ K1 , K2 is
a proper compact open subgroup, L2 ≤ K2 is some compact open subgroup and L2 ≤ K3 is a proper
compact open subgroup. We have
L(K1 ∗L1 K2 ∗L2 K3 ) ≅ (L(K1 ) ∗L(L1 ) L(K2 )) ∗L(L2 ) L(K3 ) ,
by Proposition 2.22. Since L1 ≤ K1 , K2 is proper, the group K1 ∗L1 K2 is non-compact. So
∣L2 /(K1 ∗L1 K2 )/L2 ∣ = ∞. Since also L2 ≤ K3 is a proper inclusion, Lemma 3.1 applies to show
that L(G) is non-amenable.
We consider case (ii). Let Y ⊂ X be the graph formed by removing the vertices x, y , z and the
edges e, f , g from X. Let H = π1 (G, Y ). Then G is the fundamental group of the contraction (G, Z)
given as
Gx
Ge
Gf
H
Gy
Gg
Gz
If one of the inclusions Ge ↪ H, Gf ↪ H or Gg ↪ H is proper, the first part of the lemma applies to
show that the group von Neumann algebra of an open subgroup of G is non-amenable. Indeed, by
symmetry we may assume that Ge ↪ H is proper. Since Ge ↪ Gx and Gf ↪ Gy are proper inclusions
by assumption, case (i) applies to Gx ∗Ge ∗H ∗Gf Gy , which is an open subgroup of G. It follows that
L(G) is non-amenable using Proposition 2.16.
If Ge = Gf = Gg = H, then H is compact and G = Gx ∗H Gy ∗H Gz follows from Serre’s dévissage.
The inclusions H ↪ Gx , Gy , Gz are all proper, so that (i) applies to show that L(G) is non-amenable.
4
Groups acting properly on trees
In this section we consider several criteria for non-amenability of L(G) for locally compact groups
acting properly on trees. In case G ≤ Aut(T ) is a subgroup of the automophisms of a locally finite
16
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
tree, properness of the action is easily seen to be equivalent to closedness G ≤c Aut(T ). Our nonamenability criteria for L(G) are organised according to the rank of the free group π1 (G/T ). An
increasing number of extra assumptions for π1 (G/T ) of lower rank is required. For the rest of this
section, we fix the setting of a proper action G ↷ T of a locally compact group on a tree.
Naturally, L(G) is non-amenable, if π1 (G/T ) is a non-abelian free group.
Proposition 4.1. Let T be a tree and G ↷ T a proper action of a locally compact group.
rank π1 (G + /T ) ≥ 2, then L(G) is non-amenable.
If
Proof. Since G + ≤ G is an open subgroup of index at most two, it suffices by Proposition 2.16 to
show that L(G + ) is non-amenable. We may hence from now on assume that the action of G on T is
type-preserving.
We write X = G/T . Let S ⊂ X be a maximal subtree of X. Denote by (G, Y ) the contraction of
(G, X) along S and denote the unique vertex of Y by y . Then π1 (Y ) ≅ π1 (X) is a non-abelian free
group by assumption. Let (G, T̃ ) be the covering tree of groups of Y . Then
G ≅ π1 (G, X) ≅ π1 (G, Y ) ≅ π1 (G, T̃ ) ⋊ π1 (Y ) ,
as described in Section 2.4. We identify G = π1 (G, T̃ ) ⋊ π1 (Y ) via this natural isomorphism.
First assume that π1 (G, T̃ ) is compact. We denote it by K. Let p = pK ∈ L(G) be the averaging
projection associated with p. We have pL(G)p ≅ L(G/K) by Proposition 2.10. Further, G/K ≅ π1 (Y )
is a discrete non-amenable group, so that Theorem 2.18 shows that L(G/K) is non-amenable. So
L(G) has a non-amenable corner, implying that it is non-amenable itself.
Now we assume that π1 (G, T̃ ) is non-compact. In this case we denote it by H. Since edge
stabilisers of (G, T̃ ) are compact and H is non-compact, there is some proper inclusion of an edge
group into a vertex group of (G, T̃ ). Since (G, T̃ ) arises from the universal covering T̃ of Y , there
is also some edge e ∈ E(Y ) such that the inclusion Ge ≤ Gy is non-trivial. Since π1 (Y ) has rank at
least two, there is another edge f ∈ E(Y ) such that e, e, f , f are pairwise different edges in Y . The
subgraph of Y having the vertex y and the set of edges {e, e, f , f } lifts to a 4-regular subtree R of
T̃ in which all lifts ẽ of e with target ỹ define proper inclusions Gẽ ≤ Gỹ . We consider the following
subgraph Z of R, where the lifts of e in Z as well as their target vertices are marked red.
Z=
We obtain an open subgroup π1 (G, Z) ⊂ π1 (G, T̃ ), to which Lemma 3.3 (ii) applies. So L(π1 (G, Z))
is non-amenable, implying that also L(G) is non-amenable by Proposition 2.16.
Also if π1 (G/T ) is a non-trivial group, we obtain a convincing criteria for non-amenability of L(G).
In fact, non-amenability of G and L(G) are equivalent in this case.
Proposition 4.2. Let T be a tree and G ↷ T a proper action of a non-amenable locally compact
group. If rank π1 (G + /T ) = 1, that is π1 (G + /T ) ≅ Z, then L(G) is non-amenable.
17
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Proof. Since G + ≤ G is an open subgroup of index at most two, it suffices by Proposition 2.16 to
show that L(G + ) is non-amenable. We may hence from now on assume that the action of G on T is
type-preserving.
Write X = G/T . We distinguish several cases.
Case 1. Assume that X has no vertex of degree 1. Then X is a circuit. Let T̃ be the covering
tree of X. It can be identified with the Cayley graph Cay(Z, {−1, 1}). Since (G, T̃ ) is the covering
tree of a circuit, there is p ∈ N such that for all n ∈ N
(G(n,n+1) ≤ Gn ) = (G(n+p,n+1+p) ≤ Gn+p )
(G(n,n+1) ≤ Gn+1 ) = (G(n+p,n+1+p) ≤ Gn+1+p ) .
If Gn = G(n,n+1) for all n ∈ Z or Gn+1 = G(n,n+1) for all n ∈ Z, then π1 (G, T̃ ) = lim Gn is an inductive
Ð→
limit of compact groups. Since G ≅ π1 (G, T̃ ) ⋊ Z is non-amenable, this is a contradiction. So there
are m, n ∈ Z such that G(m,m+1) ≤ Gm and G(n,n+1) ≤ Gn+1 are proper inclusions. Shifting indices by p,
we may find m < n < o ∈ Z such that
G(m,m+1) ≤ Gm
is proper,
G(n,n+1) ≤ Gn+1
is proper,
G(o,o+1) ≤ Go+1
is proper.
Fixing m, n ∈ Z with such properties, we can assume o > n to be minimal with these properties. Let
H ∶= ⟨G(m,m+1) , Gm+1 , Gm+2 , . . . , Gn , G(n,n+1) ⟩ .
Further note that ⟨G(n+1,n+2) , Gn+2 , Gn+3 , . . . , Go , G(o,o+1) ⟩ = G(n+1,n+2) by minimality of o. We obtain
that
⟨Gm , Gm+1 , . . . , Go+1 ⟩ ≅ Gm ∗G(m,m+1) Gm+1 ∗G(m+1,m+2) ⋯ ∗G(o,o+1) Go+1
≅ Gm ∗G(m,m+1) H ∗G(n,n+1) Gn+1 ∗G(n+1,n+2) ⋯ ∗G(o,o+1) Go+1
≅ Gm ∗G(m,m+1) H ∗G(n,n+1) Gn+1 ∗G(o,o+1) Go+1 .
This is an open subgroup of G. If either H ≠ G(n,n+1) or H ≠ G(m,m+1) , then Lemma 3.1 applies to
Gm ∗G(m,m+1) H ∗G(n,n+1) Gn+1 and shows that its group von Neumann algebra is non-amenable. So also
L(G) is non-amenable by Proposition 2.16. In case G(n,n+1) = H = G(m,m+1) , we have
Gm ∗G(m,m+1) H ∗G(n,n+1) Gn+1 ∗G(o,o+1) Go+1 ≅ Gm ∗H Gn+1 ∗G(o,o+1) Go+1
and H ≤ Gm , Gn+1 as well as G(o,o+1) ≤ Go+1 are proper inclusions. So Lemma 3.3 (i) applies to show
that the group von Neumann algebra of Gm ∗H Gn+1 ∗G(o,o+1) Go+1 is non-amenable.
Case 2. Assume that X has some vertex of degree 1. Let v ∈ V(X) have degree 1 and let e ∈ E(X)
be the unique edge satisfying t(e) = v . If Ge = Gv , then any lift of v to T is a terminal vertex. We
may hence remove v and e from X without changing G. This either reduces to Case 1, or it provides
us with a vertex v ∈ V(X) of degree 1 and an edge e ∈ E(X) with t(e) = v such that Ge ≤ Gv is a
proper inclusion. Let (G, T̃ ) be the covering tree of groups of (G, X). Then (G, T̃ ) takes the form
18
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
x
y
z
f
g
h
where x, y , z ∈ V(T̃ ) are lifts of v and f , g, h ∈ E(T̃ ) are lifts of e. The inclusions Gf ≤ Gx , Gg ≤ Gy
and Gg ≤ Gz are proper, since they are isomorphic with Ge ≤ Gv . So Lemma 3.1 (ii) applies and says
that π1 (G, T̃ ) has a non-amenable group von Neumann algebra. Since π1 (G, T̃ ) ≤ π1 (G, T̃ ) ⋊ Z ≅ G
is an open subgroup, Proposition 2.16 implies that L(G) is non-amenable.
As can be expected, the case π1 (G/T ) = 0 becomes the most subtle. This is due to the fact that
there are many edge transitive closed type I subgroups of Aut(T ). Their group von Neumann algebras
are in particular amenable. We obtain a non-amenability result in this case, which is sufficient for all
applications presented in this article.
Proposition 4.3. Let T be a thick tree and G ↷ T a proper action of a non-amenable locally compact
group such that G + /T is finite and satisfies π1 (G + /T ) = 0. If G + /T is not an edge, then L(G) is
non-amenable.
Proof. Since G + ≤ G is an open subgroup of index at most two, it suffices by Proposition 2.16 to
show that L(G + ) is non-amenable. We may hence from now on assume that the action of G on T is
type-preserving.
We write X = G/T . Let v ∈ V(X) be a terminal vertex of X and ṽ ∈ V(T ) a lift of v . If e ∈ E(X) is
the unique edge satisfying t(e) = v , then ∣Gv /Ge ∣ = ∣E(ṽ )∣ ≥ 3, since T is thick. In particular Ge ≤ Gv is
a proper inclusion. So if X has at least three terminal edges, then Lemma 3.3 (ii) applies to show that
G = π1 (G, X) has a non-amenable group von Neumann algebra. Otherwise, X is a finite segment,
which we can identify with the standard segment [0, n] for some n ∈ N>0 . Since G does not act edge
transitively, we have n ≥ 2. We distinguish different cases.
Case 1. We have a proper inclusion G(0,1) ≤ G1 or G(n−1,n) ≤ Gn−1 . By symmetry we may assume
that G(n−1,n) ≤ Gn−1 is a proper inclusion. Put
H = G1 ∗G(1,2) ⋯ ∗G(n−2,n−1) Gn−1 .
Then G = G0 ∗G(0,1) H ∗G(n−1,n) Gn with G(0,1) , G(n−1,n) compact and with proper inclusions G(0,1) ≤ G0
and G(n−1,n) ≤ H as well as G(n−1,n) ≤ Gn . So Lemma 3.3 (i) applies to show that L(G) is nonamenable.
Case 2. We have G(0,1) = G1 or G(n−1,n) = Gn−1 . By symmetry we may assume that G(0,1) = G1 .
Let k ∈ N be maximal with the property that
G0 ≥ G(0,1) = G1 ≥ G(1,2) = G2 ≥ ⋯ ≥ G(k−1,k) = Gk .
We know that G(n−1,n) ≤ Gn is a proper inclusion, implying that k ≤ n − 1. So
G ≅ G0 ∗G(0,1) G1 ∗G(1,2) ⋯ ∗G(n−1,n) Gn ≅ (G0 ∗G(k,k+1) Gk+1 ) ∗G(k+1,k+2) ⋯ ∗G(n−1,n) Gn .
We will show that the open subgroup G0 ∗G(k,k+1) Gk+1 ≤ G has a non-amenable group von Neumann
algebra. Thanks to Proposition 2.16, this will finish the proof.
19
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Let i ∈ {1, . . . , k}. If v ∈ V(T ) is a lift of i ∈ V([0, n]) and e, f ∈ E(v ) are lifts of (i −1, i ), (i , i +1) ∈
E(k), respectively, then
E(v ) ≅ Gv /Ge ⊔ Gv /Gf = {Ge } ⊔ Gv /Gf .
Since ∣E(v )∣ ≥ 3, it follows that ∣Gi /G(i,i+1) ∣ = ∣Gv /Gf ∣ ≥ 2. So Gi ≥ G(i,i+1) is a proper inclusion for all
i ∈ {1, . . . , k}. Since also G0 ≥ G(0,1) is a proper inclusion, we have the chain of proper inclusions
G0 G(0,1) = G1 G(1,2) .
This shows that G(k,k+1) ≤ G0 is not a maximal subgroup.
So Lemma 2.4 shows that
∣G(k,k+1) /G0 /G(k,k+1) ∣ ≥ 3. We checked all conditions to apply Lemma 3.1 to G0 ∗G(k,k+1) Gk+1 , finishing
the proof of the proposition.
We end this section, by a non-amenability result for edge transitive groups G ↷ T . A condition on
the local action of G ↷ T around a vertex ensures non-amenability of L(G).
Proposition 4.4. Let T be a thick tree and G ↷ T a proper action of a locally compact group.
Assume that G + is edge transitive but not locally 2-transitive. Then L(G) is non-amenable.
Proof. Consider the open subgroup G + ≤ G of index at most two. Note that G + ↷ T is not locally
2-transitive, since G ↷ T is not locally 2-transitive. By Proposition 2.16 it hence suffices to show
that L(G + ) is non-amenable. We may hence from now on assume that the action of G on T is
type-preserving and G/T is an edge.
Since G is not locally 2-transitive, there is some v ∈ V(T ) such that Gv ↷ E(v ) is not 2-transitive.
Let e ∈ E(v ) and w = t(e). Bass-Serre theory says that G ≅ Gv ∗Ge Gw , since G is edge transitive and
type-preserving. Since Gv ↷ E(v ) is transitive, we have a Gv equivariant identification E(v ) ≅ Gv /Ge .
Since Gv ↷ E(v ) is not 2-transitive, we further have ∣Ge /Gv /Ge ∣ = ∣Ge /E(v )∣ = 1+∣Ge /(E(v )∖{e})∣ ≥ 3.
Note also that Ge ≤ Gw is a proper inclusion, since ∣Gw /Ge ∣ = ∣E(w )∣ ≥ 3. Now Lemma 3.1 applies to
show that L(G) is non-amenable.
5
Proof of Theorems C and D
To start this section let us note that Theorem D simply summarises Propositions 4.1, 4.2 and 4.3.
We will thus devote the rest of this section to the proof of Theorem C.
Lemma 5.1. Let T be a locally finite tree such that Aut(T ) is not virtually abelian and acts minimally
on T . Then T has infinitely many ends.
Proof. We assume that T has only finitely many ends and deduce a contradiction. If T has no end,
then it is finite and Aut(T ) is a finite group, hence virtually abelian. So T has at least one end. If
T has exactly one end, then it contains a unique maximal geodesic ray. This ray is pointwise fixed
by Aut(T ), which contradicts minimality of Aut(T ) ↷ T . If T has exactly 2 ends, then Aut(T )
setwise fixes the unique two-sided infinite geodesic of T . By minimality of Aut(T ) ↷ T , it follows that
T ≅ Cay(Z, {−1, 1}). Then Aut(T ) ≅ D∞ is a dihedral group, which is virtually abelian. This shows
that T has at least 3 ends. Let F = {(x, y ) ∩ (y , z) ∩ (z, x) ∣ x, y , z pairwise different ends of T }.
Since T has only finitely many ends, F is finite. Further, its definition makes it clear that F is
Aut(T )-invariant, contradicting minimality of Aut(T ) ↷ T . This finishes the proof of the lemma.
20
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Lemma 5.2. Let T be a tree with at least some vertex of degree 3 and such that Aut(T ) acts
minimally on T . Then there is a thick tree S such that
• V(S) ⊂ V(T ),
• V(S) is Aut(T )-invariant, and
• the restriction map Aut(T ) → Sym(V(S)) induces an isomorphism of topological groups
Aut(T ) ≅ Aut(S).
Further,
• if G ≤c Aut(T ) is a closed subgroup acting minimally on T , then also G ↷ S is minimal and,
• G acts locally 2-transitively on T if and only if it acts locally 2-transitively on S.
Proof. We define
V(S) = {v ∈ V(T ) ∣ deg(v ) ≥ 3}
and
E(S) = {s ∶ [0, n] ↪ T ∣ n ≥ 1, deg(s(0)), deg(s(n)) ≥ 3, ∀i ∈ {1, . . . , n − 1} ∶ deg(s(i )) = 2} ,
with origin o(s) = s(0) and target t(s) = s(n). It is clear that S is a non-empty graph.
We first show that S is a tree. To this end we prove that S is connected and that every circuit in
S has backtracking. Let v , w ∈ V(S). There is some injective path s ∶ [0, n] ↪ T such that s(0) = v
and s(n) = w . Let B = {i ∈ {1, . . . , n − 1} ∣ deg(s(i )) ≥ 3}. If B = ∅, then s ∈ E(S) is an edge between
v and w . Otherwise let i1 < ⋯ < ik be an enumeration of B. Put i0 ∶= 0 and ik+1 ∶= n. Set sj ∶= s∣[ij ,ij+1 ]
for j ∈ {0, . . . , k}. Then sj ∈ E(S) (after identifying [ij , ij+1 ] ≅ [0, ij+1 − ij ]) for all j ∈ {0, . . . , k}. We
have
o(s0 ) = s0 (0) = s(0) = v ,
t(sk ) = sk (n) = s(n) = w ,
and
t(sj ) = sj (ij+1 ) = s(ij+1 ) = sj+1 (ij+1 ) = o(sj+1 ) ,
for all all j ∈ {0, . . . , k − 1} .
This shows that s0 , . . . , sk define a chain of edges connecting v and w in S. So S is connected.
Let now s0 , . . . , sk ∈ E(S), sj ∶ [ij , ij+1 ] ↪ T define a circuit in S. Define s ∶ [0, ik+1 ] → T as the
path that agrees with sj on [ij , ij+1 ]. Then s is a circuit in T since,
o(s) = s0 (0) = o(s0 ) = t(sk ) = sk (ik+1 ) = t(s) .
Since T is a tree, there is some l ∈ [0, ik+1 − 2] such that s((l, l + 1)) = s((l + 1, l + 2)). Since sj is an
injective path for all j ∈ {0, . . . , k}, we must have l = ij − 1 for some j ∈ {1, . . . , k}. This means that
sj and sj+1 are injective paths in T all of whose non-terminal vertices have degree 2 and such that the
last edge of sj is the conjugate of the first edge of sj+1 (i.e. sj ((ij − 1, ij )) = sj+1 ((ij , ij + 1))). This
implies sj = sj+1 . So s0 , . . . , sk has backtracking and we conclude that S is a tree.
If g ∈ Aut(T ), v ∈ V(T ), then deg(gv ) = deg(v ), so that gV(S) = V(S) follows. Denote by
Res ∶ Aut(T ) → Sym(V(S)) the restriction homomorphism. We show that Res(Aut(T )) ⊂ Aut(S).
Assume v , w ∈ V(S) are adjacent in S. Then there is s ∈ E(S), s ∶ [0, n] ↪ T such that s(0) = v ,
s(n) = w . Since gs ∈ E(S), with (gs)(0) = g(s(0)) = gv and (gs)(n) = g(s(n)) = gw , we also have
21
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
gv ∼ gw in S. This shows that Res(g) is a bijective graph homomorphism. Since Res(g −1 ) = Res(g)−1 ,
it follows that Res(g) ∈ Aut(S).
We show that Res is injective. Assume that Res(g) = idS for some g ∈ Aut(T ). Then g∣V(S) =
idV(S) . Since V(S) is Aut(T )-invariant and Aut(T ) ↷ T is minimal by assumption, T is the convex
closure of V(S). So g = idT .
We show that Res is surjective. Let h ∈ Aut(S). We want to define β(h)(s(i )) ∶= (hs)(i ) for all
s ∈ E(S), s ∶ [0, n] ↪ T and i ∈ {0, . . . , n}. We first prove that this gives rise to a well-defined map
β(h) ∶ V(T ) → V(T ). If v ∈ V(S) ⊂ V(T ) and s ∈ E(S), s ∶ [0, n] → T satisfies s(i ) = v for some
i ∈ {0, 1, . . . , n}, then i ∈ {0, n}. We obtain
(hs)(0) = o(hs) = h(o(s)) = hv
or
(hs)(n) = t(hs) = h(t(s)) = hv ,
respectively.
So the image β(h)v = hv is independent of the choice of s. If v ∈ V(T ) ∖ V(S), then there is some
s ∈ E(S) containing v , that is, writing s ∶ [0, n] ↪ T , there is some i ∈ {1, . . . , n −1} such that v = s(i ).
This follows from the fact that the convex closure of V(S) is T . If s, s ′ ∈ E(S) both contain v , then
s ′ = s or s ′ = s. Take s ∈ E(S), s ∶ [0, n] ↪ T and i ∈ {1, . . . , n − 1} such that s(i ) = v . Then
(hs)(n − i ) = (hs)(n − i ) = (hs)(i )
showing that β(h)v is well-defined. Note that β(h−1 ) = β(h)−1 , so that we obtain a map β ∶ Aut(S) →
Sym(V(T )). Also note that β is a group homomorphism. If h ∈ Aut(S) and v , w ∈ V(T ) are adjacent,
then there is s ∈ E(S), s ∶ [0, n] ↪ T and i ∈ {0, . . . , n − 1} such that s(i ) = v , s(i + 1) = w . Then
β(h)v = (hs)(i ) ∼ (hs)(i + 1) = β(h)w .
This shows that β(h) is a graph homomorphism. Since also β(h)−1 = β(h−1 ) is a graph homomorphism, we find that β ∶ Aut(S) → Aut(T ). It is clear that β(h)∣V(S) = h for all h ∈ Aut(S), which
shows that Res ○β = idAut(S) . So Res is surjective.
By construction Res is a continuous map. Also β is continuous as it can be easily checked on a
neighbourhood basis of the identity in Aut(S). So Res is an isomorphism of topological group.
Now assume that G ≤c Aut(T ) is a group acting minimally on T . We will show that G ↷ S is also
minimal. To this end, take v ∈ V(S) and g ∈ G. Then Res(g)v = gv . Further, the construction of S
shows that [v , gv ]T ∩ V(S) = [v , gv ]S ∩ V(S). So a vertex of V(S) is in the convex closure of Gv
inside T if and only if it is in the convex closure of Res(G)v inside S. This suffices to conclude that
G ↷ S is minimal.
Now let us consider local 2-transitivity of G ↷ T and G ↷ S. For every v ∈ V(S) the map
ES (v ) ∋ s ↦ s((0, 1)) ∈ ET (v ) is a bijection. Let us denote its inverse by Resv ∶ ET (v ) → ES (v ).
Then Res(g) Resv (e) = Resgv (ge) for all g ∈ Aut(T ), v ∈ V(T ) and e ∈ ET (v ). Together with the
observation that Res(G)v = Res(Gv ), this directly implies that G ↷ T is locally 2-transitive if and only
if G ↷ S is locally 2-transitive. This finishes the proof of the lemma.
The following lemma is well-known. Its proof can be found for example as Lemma 2.4 in [9]
Lemma 5.3. Let T be a locally finite tree and let G ≤c Aut(T ) be a closed subgroup acting minimally
on T . Then G is compactly generated if and only if G acts cocompactly on T .
22
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
We now come to the major reduction result necessary to apply results from Section 4 in the proof
of Theorem C.
Proposition 5.4. Let T be a locally finite tree with infinitely many ends. Let G ≤c Aut(T ) be a closed
non-amenable subgroup acting minimally on T . Then there is an open non-amenable subgroup H ≤ G,
a compact normal subgroup K ⊴ H and a locally finite thick tree S such that H/K ≤c Aut(S) acts
minimally, cocompactly and in a type-preserving way on S.
If G ↷ T is not locally 2-transitive, then also H/K ↷ S can be chosen to be not locally 2-transitive.
Proof. Since T is a locally finite tree with infinitely many ends and G ≤c Aut(T ) is a non-amenable
subgroup acting minimally on T , it contains a hyperbolic element and T is the convex closure of
all translation axes of hyperbolic elements in G. Let H ≤ G be an open non-amenable compactly
generated subgroup. In case G ↷ T is not locally 2-transitive, there is v ∈ V(T ) with deg(v ) ≥ 3
and Gv ↷ E(v ) is not 2-transitive. Since G ↷ T is minimal, the convex closure of translation axes
of hyperbolic elements in G equals T . Adding finitely many elements to a compact generating set of
H, we may hence assume that E(v ) lies in the convex closure T ′ of all translation axes of hyperbolic
elements in H. Note that H ↷ T ′ is minimal by construction. The fix group K = FixG (T ′ ) ∩ H is
compact and normal in H. It is the kernel of the map H → Aut(T ′ ). We obtain the closed subgroup
H/K ≤c Aut(T ′ ). Since Gv ↷ E(v ) is not 2-transitive, also Hv ↷ E(v ) is not 2-transitive. Further,
this action factors through (H/K)v , since E(v ) ⊂ E(T ′ ). We thus find that H/K ↷ T ′ is not locally
2-transitive in case G ↷ T is not locally 2-transitive.
We apply Lemma 5.2 to T ′ to obtain a thick tree S such that
• V(S) = {v ∈ V(T ′ ) ∣ deg(v ) ≥ 3},
• V(S) is Aut(T ′ )-invariant,
• the restriction map Res ∶ Aut(T ′ ) → Sym(V(S)) induces an isomorphism of topological groups
Aut(T ′ ) ≅ Aut(S),
Further, Lemma 5.2 says that since H/K ↷ T ′ acts minimally, H/K ↷ S has the same property. Also
if H/K ↷ T ′ is not locally 2-transitive, then H/K ↷ S has the same property. Lemma 5.3 applies to
show that H/K acts cocompactly on S. If H/K ↷ S is not type-preserving, we may replace H by an
index two subgroup of itself in order to guarantee also this property. Note in particular, that H/K ↷ S
remains minimal, since squares of hyperbolic elements are type-preserving. This finishes the proof of
the proposition.
We are now ready to combine our results from Section 4 with Proposition 5.4 in order to prove
our main theorem of this article.
Proof of Theorem C. Let T be a locally finite tree and G ≤c Aut(T ) a closed non-amenable subgroup
acting minimally on T . Assume that G does not act locally 2-transitively on T . Since G is not
amenable, Aut(T ) is not virtually abelian. So Lemma 5.1 implies that T has infinitely many ends.
Applying Proposition 5.4, we find an non-amenable open subgroup H ≤ G a compact normal subgroup
K ⊴ H and a thick tree S such that H/K ≤c Aut(S) acts minimally cocompactly type-preservingly and
not locally 2-transitively on S. In particular, H/K is non-amenable. Further H/K ↷ S is proper, since
H/K ≤c Aut(S) is closed. So the results of Section 4 apply to show that L(H/K) is non-amenable.
Since Proposition 2.10 says that L(H/K) is a corner of L(H), also the latter von Neumann algebra
is non-amenable. Since H ≤ G is open, also L(G) follows non-amenable by Proposition 2.16. This
finishes the proof of the theorem.
23
Von Neumann algebras of groups acting on trees
6
by Cyril Houdayer and Sven Raum
Applications to type I groups and to Burger–Mozes groups
In this section we will prove Theorems A and B.
6.1
Type I groups
Definition 6.1. Let M be a von Neumann algebra. We say that M is a type I von Neumann algebra
if for every projection p ∈ M there is some q ≤ p (i.e. pq = q) such that qMq is abelian.
A locally compact group G is called a type I group, if every unitary representation of G generates
a type I von Neumann algebra.
The following description of type I von Neumann algebras is well-known and provides the reader
unfamiliar with this von Neumann algebraic notions with some orientation.
Proposition 6.2. A von Neumann algebra M is of type I if and only if there is a cardinal κ and (possibly
empty) measure spaces Xω , ω ≤ κ such that M ≅ ⊕ω≤κ L∞ (Xω ) ⊗ B(Hω ), where Hω is a Hilbert space
with an orthonormal basis of cardinality ω.
With this characterisation at hand, we see that every type I von Neumann algebra is amenable.
Corollary 6.3. Every type I von Neumann algebra is amenable.
We can now proceed to the proof of our main theorem’s first application.
Proof of Theorem A. This follows immediately from Theorem C and Corollary 6.3.
6.2
Applications to Burger-Mozes groups
The following property is the foundation of combinatorial considerations about type I groups acting
on trees.
Definition 6.4. Let T be a locally finite tree. If e ∈ E(T ) is an edge in T , then the graph T without
e is a disjoint union of two trees, which we call the half trees emerging from e.
A closed subgroup G ≤c Aut(T ) has Tits’ independence property if for all edges e ∈ E(T ) with half
trees h1 , h2 emerging from e there is a decomposition FixG (e) = FixG (h1 ) × FixG (h2 ).
An important class of examples enjoying Tits’ independence property are Burger–Mozes groups.
Definition 6.5 (Burger–Mozes [7]). Let n ≥ 3 and T be the n-regular tree. A legal colouring of T
is a map l ∶ E(T ) → {1, . . . , n} such that l(e) = l(e) for all e ∈ E(T ) and l∣E(v ) is a bijection for every
v ∈ V(T ). Given a legal colouring l of T , we define the local action of g ∈ Aut(T ) at v ∈ Aut(T ) by
σ(g, v ) ∶= l ○ g ○ l∣−1
E(v ) ∈ Sym({1, . . . , n}) = Sn .
If F ≤ Sn is given, we define the Burger-Mozes groups by
U(F ) ∶= {g ∈ Aut(T ) ∣ ∀v ∈ V(T ) ∶ σ(g, v ) ∈ F }
and their type-preserving subgroups
U(F )+ ∶= U(F ) ∩ Aut(T )+ .
24
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
Note that the definition of U(F ) and U(F )+ a priori depends on the choice of a legal colouring.
However, the fact that a legal colouring is unique up precomposition with a tree automorphism shows
that U(F ) and U(F )+ are independent of this choice up to conjugation by a tree automorphism. Since
Aut(T )+ ≤ Aut(T ) has index 2, also U(F )+ ≤ U(F ) has index 2. In this context, note that our definition
of U(F )+ as type-preserving part of U(F ) in general differs from the subgroup ⋁e∈E(T ) U(F )e from
BM, which could be trivial. However, these two definitions agree in case F is transitive and generated
by point-stabilisers.
Thanks to Tits’ independence property, U(F )+ is abstractly simple, if F is transitive and generated
by point-stabilisers. Burger–Mozes groups are an important class of examples in the theory of totally
disconnected groups.
Actually Burger-Mozes groups account for a large class of groups having Tits’ independence property, as it is demonstrated by the following theorem. Its statement did not yet appear in the literature,
and we add it for the reader’s convenience. The proof combines known results from Burger–Mozes
[7] and Bank-Elder-Willis [2].
Theorem 6.6. Let T be a locally finite tree and G ≤c Aut(T ) a closed vertex and edge transitive
group with Tits’ independence property. Let F ≤ Sn be permutation isomorphic with the image of Gv
in Sym(E(v )). Then G = U(F ) for a suitable colouring of T .
Proof. Since G is edge transitive, it is locally transitive. So Proposition 3.2.2 of [7] applies to show
that there is a suitable legal colouring of T for which the inclusion G ≤ U(F ) holds. Theorem 5.4 of
[2] says that
G = {g ∈ Aut(T ) ∣ ∀v ∈ V(T ) ∃h ∈ G ∶ g∣B1 (v ) = h∣B1 (v ) } = U(F ) .
This finishes the proof.
The following result says that the type I conjecture holds for vertex transitive groups with Tits’
independence property. Note that non-compact boundary transitive groups are edge transitive. So the
previous theorem shows that Theorem 6.7 applies exactly to Burger-Mozes groups.
Theorem 6.7 (Olshanskii [30], Amann [1], Ciobotaru [10, Theorem 3.5]). Let T be a locally finite tree and G ≤c Aut(T ) a closed subgroup acting transitively on vertices of T . Assume that G has
Tits’ independence property. If G acts transitively on the boundary ∂T , then G is a type I group.
In order to formulate a converse to this theorem, which is the content of our Theorem B, we need
to characterise boundary transitivity of groups with Tits’ independence property. The next lemma is
essentially contained in the ideas of Burger–Mozes’ [7, Lemma 3.1.1]. It also appeared as Proposition
15 in [1]. We claim no originality, but give a full proof for the convenience of the reader.
Lemma 6.8 (Compare with Burger–Mozes [7]. See also Proposition 15 in [1]). Let T be a locally finite tree that is not a line nor a vertex and let G ≤c Aut(T ) be a closed vertex transitive group
with Tits’ independence property. Then G is boundary transitive if and only if G is locally 2-transitive.
Proof. Since G is vertex transitive, it is non-compact. So Lemma 3.1.1 in [7] shows that if G is
transitive on the boundary, then G is locally 2-transitive.
In order to prove the converse we appeal to Lemma 3.1.1 [7] again and have to show that for every
v ∈ V(T ) and every n ∈ N the action of Gv on ∂Bn (v ) is transitive. Since G ↷ T is vertex transitive,
25
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
T is a homogeneous tree and its degree is at least three, since T is not a line nor a vertex. Let
x, y ∈ ∂Bn (v ) and let r ∶ [0, n] → T , s ∶ [0, n] → T be the unique geodesics satisfying o(r ) = o(s) = v ,
t(r ) = x and t(s) = y . We inductively show the existence of g1 , . . . , gn ∈ Gv such that (gi r )(i ) = s(i ) for
all i ∈ {1, . . . , n}. Since G is locally 2-transitive and T is homogeneous of degree at least three, G also
acts locally transitively. So there is some g1 ∈ Gv such that g1 r (1) = s(1). Assume that g1 , . . . , gi have
been constructed for i < n. Let h1 , h2 be the two half-trees emerging from the edge e ∶= (s(i −1), s(i )).
The notation can be fixed by assuming s(i −1) ∈ h1 and s(i ) ∈ h2 . Then h2 contains all vertices adjacent
to s(i ) that have distance i +1 to v . In particular, s(i +1), gi r (i +1) ∈ h2 . Since G is locally 2-transitive
and ∣E(s(i ))∣ ≥ 3, there is h ∈ Ge satisfying h(gi r (i + 1)) = s(i + 1). Because G has the independence
property, we obtain the product decomposition Ge = FixG (h1 ) × FixG (h2 ) and can write h = (h1 , h2 )
with h1 ∈ FixG (h1 ) and h2 ∈ FixG (h2 ). Then h1 gi r (i + 1) = h2−1 hgi r (i + 1) = h2−1 s(i + 1) = s(i + 1).
Further, h1 v = v , since v ∈ V(h1 ). We put gi+1 ∶= h1 gi and finish the induction. Now the existence of
gn with gn v = v and gn x = gn r (n) = s(n) = y finishes the proof of the lemma.
Let us reformulate Lemma 6.8 in terms of Burger-Mozes groups.
Lemma 6.9 (Burger–Mozes [7, Section 3]). Let F ≤ Sn for n ≥ 3 be given. Then the following
statements are equivalent.
• U(F ) is boundary transitive,
• U(F ) is locally 2-transitive,
• F is 2-transitive.
Combining Theorem 6.7, Lemma 6.8 and Theorem A, we obtain the characterisation of vertex
transitive type I groups with the independence property, stated as Theorem B.
Proof of Theorem B. All statements of the theorem are obvious in case T is a line or n = 2.
Let T be a locally finite tree and G ≤c Aut(T ) a closed vertex transitive subgroup with Tits’
independence property. If G is locally 2-transitive, then G is boundary transitive by Lemma 6.8. So
Theorem 6.7 says that G is a type I group. If G is not locally 2-transitive, then T has at least one
vertex of degree 3. So T is not a line and it follows from vertex transitivity, minimality of G ↷ ∂T and
Proposition 2.5 that G is not amenable. So Theorem A applies to show that G is not a type I group.
It remains to prove the statement about Burger-Mozes groups. Since for every F ≤ Sn the closed
subgroup U(F )+ ≤ U(F ) has index 2, it suffices to characterise when U(F ) is a type I group. Now
U(F ) is vertex transitive and has Tits’ independence property. So the first part of the statement says
that U(F ) is a type I group if and only if it acts locally 2-transitively. Now Lemma 6.9 finishes the
proof of the theorem.
26
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
References
[1] O. E. Amann. Groups of tree-automorphisms and their unitary representations. PhD thesis, ETH
Zürich, 2003.
[2] C. Banks, M. Elder, and G. A. Willis. Simple groups of automorphisms of trees determined by
their actions on finite subtrees. J. Group Theory, 18:235–261, 2015.
[3] J. Bernstein. All reductive p-adic groups are of type I. Functional Anal. Appl., 8:91–93, 1974.
[4] J. N. Berstein and A. V. Zelevinsky. Representations of the group GL(n,F) where F is a nonArchimedean local field. Russian Math. Surveys, 31(3):1–68, 1976.
[5] N. P. Brown and N. Ozawa. C*-algebras and finite-dimensional approximations., volume 88 of
Graduate Studies in Mathematics. Providence, RI: American Mathematical Society, 2008.
[6] F. Bruhat and J. Tits. Groupes réductifs sur un corps local. Inst. Hautes Études Sci. Publ. Math.,
41:5–251, 1972.
[7] M. Burger and S. Mozes. Groups acting on trees: From local to global structure. Publ. Math.
Inst. Hautes Étud. Sci., 92:113–150, 2000.
[8] M. Burger and S. Mozes. Lattices in product of trees. Publ. Math. Inst. Hautes Étud. Sci.,
92:151–194, 2000.
[9] P.-E. Caprace and T. de Medts. Simple locally compact groups acting on trees and their germs
of automorphisms. Transform. Groups, 16(2):375–411, 2011.
[10] C. Ciobotaru. A note on type I groups acting on d-regular trees. arXiv:1506:02950, 2015.
[11] C. Ciobotaru. Infinitely generated Hecke algebras with infinite presentation. arXiv:1603.04599,
2016.
[12] L. Clozel. Spectral theory of automorphic forms. In P. Sarnak and F. Shahidi, editors, Automorphic forms and applications. Lecture notes from the IAS/Park City Mathematics Institute
held in Park City, UT, July 1-20, 2002, volume 12 of IAS/Park City Math. Ser., pages 43–93.
Providence, RI: American Mathematical Society and Princeton, NJ: Institute for Advanced Study.
[13] A. Connes. Classification of injective factors. Cases II1 , II∞ , IIIλ , λ ≠ 1. Ann. Math. (2),
74:73–115, 1976.
[14] A. Connes. On the cohomology of operator algebras. J. Funct. Anal., 28(2):248–253, 1978.
[15] A. Deitmar and S. Echterhoff. Principles of harmonic analysis. Universitext. Cham-HeidelbergNew York-Dordrecht-London: Springer-Verlag, second edition edition, 2014.
[16] J. Dixmier. Sur les représentations unitaires des groupes de Lie nilpotents. V. Bull. Soc. Math.
Fr., 87:65–79, 1959.
[17] K. J. Dykema. Free products of hyperfinite von Neumann algebras and free dimension. Duke
Math. J., 69(1):97–119, 1993.
27
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
[18] K. J. Dykema. Interpolated free group factors. Pac. J. Math., 163(1):123–135, 1994.
[19] P. Eymard. L’algèbre de Fourier d’un groupe localement compact. Bull. Soc. Math. Fr., 92:181–
236, 1964.
[20] J. Glimm. Type I C∗ -algebras. Ann. Math. (2), 73:572–612, 1961.
[21] U. Haagerup. On the dual weights for crossed products of von Neumann algebras II. Math.
Scand., 43:119–140, 1978.
[22] Harish-Chandra and G. van Dijk. Harmonic analysis on reductive p-adic groups, volume 162 of
Lecture Notes in Mathematics. 1970. Notes taken by by G. van Dijk on lectures of HarishChandra.
[23] C. Herz. Harmonic synthesis for subgroups. Ann. Inst. Fourier, 23(3):91–123, 1973.
[24] A. W. Knapp. Representation theory of semisimple groups., volume 36 of Princeton Mathematical
Series. Princeton, NJ: Princeton University Press, 1986.
[25] A. T.-M. Lau and A. L. Paterson. Inner amenable locally compact groups. Trans. Am. Math.
Soc., 325(1):155–169, 1991.
[26] J. R. McMullen. Extensions of positive-definite functions. Number 117. Providence, RI: American
Mathematical Society, 1972.
[27] C. C. Moore. Decomposition of unitary representations defined by discrete subgroups of nilpotent
groups. Ann. Math (2), 82(1):146–182.
[28] C. Nebbia. Amenabilitz and Kunze-Stein property for groups acting on a tree. Pac. J. Math.,
135(2), 1988.
[29] G. I. Olshanskii. Classification of irreducible representations of groups of automorphisms of
Bruhat-Tits trees. Funkcional. Anal. i Priložen., 11(1):26–34, 1977.
[30] G. I. Olshanskii. New “large” groups of type I. Akad. Nauk SSSR, Vsesoyuz. Inst. Nauchn. i
Tekhn. Informatsii, 16:31–52, 1980.
[31] F. Rădulescu. Random matrices, amalgamated free products and subfactors of the von Neumann
algebra of a free group, of noninteger index. Invent. Math., 115(2):347–389, 1994.
[32] S. Raum. C*-simplicity of locally compact Powers groups. Accepted for publication in J. Reine
Angew. Math., arXiv:1505.07793, 2015.
[33] S. Raum. Cocompact amenable closed subgroups: weakly inequivalent representations in the
left-regular representation. Accepted for publication in Int. Math. Res. Not., arXiv:1510.06215,
2015.
[34] S. Sakai. C*-algebras and W*-algebras, volume 60 of Ergebnisse der Mathematik und ihrer
Grenzgebiete. New York, Heidelberg: Springer-Verlag, 1971.
[35] J.-P. Serre. Trees. Springer Monographs in Mathematics. Berlin-Heidelberg-New York: SpringerVerlag, 2 edition, 2003.
28
Von Neumann algebras of groups acting on trees
by Cyril Houdayer and Sven Raum
[36] S. Smith. A product for permutation groups and topological groups. arXiv:1407.5697, 2015.
[37] M. Takesaki. Conditional expectations in von Neumann algebras. J. Funct. Ana., 9:306–321,
1972.
[38] M. Takesaki. Theory of operator algebras II. Berlin-Heidelberg-New York: Springer-Verlag, 2003.
[39] E. Thoma. Eine Charakterisierung diskreter Gruppen vom Typ I. Invent. Math., 6:190–196,
1968.
[40] J. Tits. Sur le groupe des automorphismes d’un arbre. In Essays on topology and related topics,
Mem. dédiés à Georges de Rham, pages 188–211. Berlin-Heidelberg-New York: Springer, 1970.
[41] K. Tzanev. Hecke C*-algebras and amenability. J. Oper. Theory, 50(1):169–178, 2003.
[42] Y. Ueda. Amalgamated free product over Cartan subalgebra. Pac. J. Math., 191(2):359–392,
1999.
[43] Y. Ueda. Factoriality, type classification and fullness for free product von Neumann algebras.
Adv. Math., 228:2647–2671, 2011.
[44] D. van Dantzig. Zur topologischen Algebra. III. Brouwersche und Cantorsche Gruppen. Compos.
Math., 3:408–426, 1936.
[45] D. V. Voiculescu, K. J. Dykema, and A. Nica. Free random variables. A noncommutative probability approach to free products with applications to random matrices, operator algebras and
harmonic analysis on free groups. CRM Monograph Series. 1. Providence, RI: American Mathematical Society, 1992.
[46] R. M. Weiss. The structure of affine buildings, volume 168 of Annals of Mathematics Studies.
Princeton, NJ: Princeton University Press, 2009.
Cyril Houdayer
Laboratoire de Mathématiques d’Orsay
Université Paris-Sud
CNRS
Université Paris-Saclay
F-91405 Orsay
Sven Raum
EPFL SB SMA
Station 8
CH–1015 Lausanne
Switzerland
[email protected]
[email protected]
29
| 4math.GR
|
Under consideration for publication in Theory and Practice of Logic Programming
1
SWI-Prolog and the Web
arXiv:0711.0917v1 [cs.PL] 6 Nov 2007
JAN WIELEMAKER
Human-Computer Studies laboratory
University of Amsterdam
Matrix I
Kruislaan 419
1098 VA, Amsterdam
The Netherlands
(e-mail: [email protected])
ZHISHENG HUANG, LOURENS VAN DER MEIJ
Computer Science Department
Vrije University Amsterdam
De Boelelaan 1081a
1081 HV, Amsterdam
The Netherlands
(e-mail: huang,[email protected])
submitted 28 April 2006; revised 23 August 2007; accepted 18 October 2007
Abstract
Prolog is an excellent tool for representing and manipulating data written in formal languages as well as natural language. Its safe semantics and automatic memory management
make it a prime candidate for programming robust Web services.
Where Prolog is commonly seen as a component in a Web application that is either
embedded or communicates using a proprietary protocol, we propose an architecture where
Prolog communicates to other components in a Web application using the standard HTTP
protocol. By avoiding embedding in external Web servers development and deployment
become much easier. To support this architecture, in addition to the transfer protocol, we
must also support parsing, representing and generating the key Web document types such
as HTML, XML and RDF.
This paper motivates the design decisions in the libraries and extensions to Prolog for
handling Web documents and protocols. The design has been guided by the requirement
to handle large documents efficiently. The described libraries support a wide range of Web
applications ranging from HTML and XML documents to Semantic Web RDF processing.
The benefits of using Prolog for Web related tasks is illustrated using three case studies.
KEYWORDS: Prolog, HTTP, HTML, XML, RDF, DOM, Semantic Web
1 Introduction
The Web is an exciting place offering new opportunities to artificial intelligence,
natural language processing and Logic Programming. Information extraction from
the Web, reasoning in Web applications and the Semantic Web are just a few
examples. We have deployed Prolog in Web related tasks over a long period. As
2
J. Wielemaker, Z. Huang and L. van der Meij
most of the development on SWI-Prolog takes place in the context of projects that
require new features, the system and its libraries provide extensive support for Web
programming.
There are two views on deploying Prolog for Web related tasks. In the most
commonly used view, Prolog acts as an embedded component in a general Web
processing environment. In this role it generally provides reasoning tasks such as
searching or configuration within constraints. Alternatively, Prolog itself can act as
a stand-alone HTTP server as also proposed by ECLiPSe (Leth et al. 1996). In this
view it is a component that can be part of any of the layers of the popular three-tier
architecture for Web applications. Components generally exchange XML if used as
part of the backend or middleware services and HTML if used in the presentation
layer.
The latter view is in our vision more attractive. Using HTTP and XML over
HTTP, the service is cleanly isolated using standard protocols rather than proprietary communication. Running as a stand-alone application, the attractive interactive development nature of Prolog can be maintained much more easily than
embedded in a C, C++, Java or C# application. Using HTTP, automatic testing of
the Prolog components can be done using any Web oriented test framework. HTTP
allows Prolog to be deployed in any part of the service architecture, including the
realisation of complete Web applications in one or more Prolog processes.
When deploying Prolog in a Web application using HTTP, we must not only
implement the HTTP transfer protocol, but also support parsing, representing and
generating the important document types used on the Web, especially HTML, XML
and RDF. Note that, being widely used open standards, supporting these document
types is also valuable outside the context of Web applications.
This paper gives an overview of the Web infrastructure we have realised. Given
the range of libraries and Prolog extensions that facilitate Web applications we
cannot describe them in detail. Details on the library interfaces can be found in
the manuals available from the SWI-Prolog Web site.1 Details on the implementation are available in the source distribution. The aim of this paper is to give an
overview of the required infrastructure to use Prolog for realizing Web applications
where we concentrate on scalability and performance. We describe our decisions for
representing Web documents in Prolog and outline the interfaces provided by our
libraries.
The benefits of using Prolog for Web related tasks are illustrated using three case
studies: 1) SeRQL, an RDF query language for meta data management, retrieval
and reasoning; 2) XDIG, an eXtended Description Logic interface, which provides
ontology management and reasoning by processing DIG XML documents and communicating to external DL reasoners; and 3) A faceted browser on Semantic Web
databases integrating meta-data from multiple collections of art-works. This case
study serves as a complete Semantic Web application serving the end-user.
This paper is organized as follows. Section 2 to section 4 describe reading, writing
1
http://www.swi-prolog.org
SWI-Prolog and the Web
hdocumenti
hcontenti
helementi
hattributei
hpii
hsdatai
hndatai
hcdatai, hnamei
hvaluei
hsvaluei
::=
::=
::=
::=
::=
::=
::=
::=
::=
::=
3
list-of hcontenti
helementi | hpii | hcdatai | hsdatai | hndatai
element(htagi, list-of hattributei, list-of hcontenti)
hnamei = hvaluei
pi(hatomi)
sdata(hatomi)
ndata(hatomi)
hatomi
hsvaluei | list-of hsvaluei
hatomi | hnumberi
Fig. 1. SGML/XML tree representation in Prolog. The notation list-of hxi describes a
Prolog list of terms of type hxi.
and representation of Web related documents. Section 5 describes our HTTP client
and server libraries. Section 6 describes extensions to the Prolog language that
facilitate use in Web applications. Section 7 to section 9 describe the case studies.
2 Parsing and representing XML and HTML documents
The core of the Web is formed by document standards and exchange protocols.
Here we describe tree-structured documents transferred as SGML or XML. HTML,
an SGML application, is the most commonly used document format on the Web.
HTML represents documents as a tree using a fixed set of elements (tags), where
the SGML DTD (Document Type Declaration) puts constraints on how elements
can be nested. Each node in the hierarchy has a name (the element-name), a set of
name-value pairs known as its attributes and content, a sequence of sub-elements
and text (data).
XML is a rationalisation of SGML using the same tree-model, but removing
many rarely used features as well as abbreviations that were introduced in SGML
to make the markup easier to type and read by humans. XML documents are used
to represent text using custom application-oriented tags as well as a serialization
format for arbitrary data exchange between computers. XHTML is HTML based
on XML rather than SGML.
The first SGML parser for SWI-Prolog was created by Anjo Anjewierden based
on the SP parser2 . A stable Prolog term-representation for SGML/XML trees plays
a similar role as the DOM (Document Object Model ) representation in use in the
object-oriented world. The term-structure we use is described in figure 1. Some
issues have been subject to debate.
• Representation of text by a Prolog atom is biased by the use of SWI-Prolog
which has no length-limit on atoms and atoms that can represent Unicode text
as motivated in section 6.2. At the same time SWI-Prolog stacks are limited to
128MB each. Using atoms only the structure of the tree is represented on the
2
http://www.jclark.com/sp/
4
J. Wielemaker, Z. Huang and L. van der Meij
stack, while the bulk of the data is stored on the unlimited heap. Using lists
of character codes is another possibility adopted by both PiLLoW (Gras and
Hermenegildo 2001) and ECLiPSe (Leth et al. 1996). Two observations make
lists less attractive: lists use two cells per character while practical experience
shows text is frequently processed as a unit only. For (HTML) text-documents
we profit from the compact representation of atoms. For XML documents
representing serialized data-structures we profit from frequent repetition of
the same value.
• Attribute values of multi-value attributes (e.g. NAMES) are returned as a Prolog
list. This implies the DTD must be available to get unambiguous results. With
SGML this is always true, but not with XML.
• Optionally attribute values of type NUMBER or NUMBERS are mapped to Prolog
numbers. In addition to the DTD issues mentioned above, this conversion also
suffers from possible loss of information. Leading zeros and different floating
point number notations used are lost after conversion. Prolog systems with
bounded arithmetic may also not be able to represent all values. Still, automatic conversion is useful in many applications, especially those involving
serialized data-structures.
• Attribute values are represented as Name=Value. Using Name(Value) is an
alternative. The Name=Value representation was chosen for its similarity to
the SGML notation and because it avoids the need for univ (=..) for processing argument-lists.
Implementation The SWI-Prolog SGML/XML parser is implemented as a C-library
that has been built from scratch to create a lightweight parser. Total source
is 11,835 lines. The parser provides two interfaces. Most natural to Prolog is
load structure(+Src, -DOM, +Options) which parses a Prolog stream into a term
as described above. Alternatively, sgml parse/2 provides an event-based parser
making call-backs on Prolog for the SGML events. The call-back mode can deal with
unbounded documents in streaming mode. It can be mixed with the term-creation
mode, where the handler for begin calls the parser to create a term-representation
for the content of the element. This feature is used to process long files with a repetitive record structure in limited memory. Section 4.1 describes how this is used to
process RDF documents.
Full documentation is available from http://www.swi-prolog.org/packages/
sgml2pl.html The SWI-Prolog SGML parser has been adopted by XSB Prolog.
3 Generating Web documents
There are many approaches to generating Web pages from programs in general and
Prolog in particular. We believe the preferred choice depends on various aspects.
• How much of the document is generated from dynamic data and how much is
static? Pages that are static except for a few strings are best generated from
a template using variable substitution. Pages consisting of a table generated
from dynamic data are best entirely generated from the program.
SWI-Prolog and the Web
5
• For program generated pages we can choose between direct printing and generating using a language-native syntax, for example format(’<b>bold</b>’)
or print_html(b(bold)). The second approach can guarantee well-formed
output, but the first requires the programmer to learn about format/3 only.
• Documents that contain a significant static part are best represented in the
markup language where special constructs insert program-generated parts.
A popular approach implemented by PHP3 and ASP4 is to add a reserved
element such as hscripti and use the SGML/XML programming instruction
written as <?...?>. The obvious name PSP (Prolog Server Pages) is in use by
various projects taking this approach.5 Another approach is PWP6 (Prolog
Well-formed Pages). It is based on the principle that the source is well-formed
XML and interacts with Prolog through additional attributes. Output is guaranteed to be well-formed XML. Our infrastructure does not yet include any
of these approaches.
• Page transformation is realised by parsing the original document into its
tree representation, managing the tree and writing a new document from the
tree. Managing the source-text directly is not reliable as due to character
encoding choice, entity usage and SGML abbreviations there are many different source-texts that represent the same tree. The load structure/3 predicate described in section 2 together with output primitives from the library
sgml write.pl provide this functionality. The XDIG case study described in
section 8 follows this approach.
3.1 Generating documents using DCG
The traditional method for creating Web documents is using print routines such
as write/1 or format/2. Although simple and easily explained to novices, the
approach has serious drawbacks from a software engineering point of view. In particular the user is responsible for HTML quoting, character encoding issues and
proper nesting of HTML elements. Automated validation is virtually impossible
using this approach.
Alternatively we can produce a DOM term as described in section 2 and use the
library sgml write.pl to create the HTML or XML document. Such documents
are guaranteed to use proper nesting of elements, escape sequences and character
encoding. The terms however are big, deeply nested and hard to read and write.
Prolog allows them to be built from skeletons containing variables. This approach is
taken by PiLLoW (section 3.2) to control the complexity. In our opinion, the result
is not optimal due to the unnatural order of statements as illustrated in figure 2.
PiLLoW has partly overcome this shortcoming by defining a large number of ‘utility
3
4
5
6
www.php.net
www.microsoft.com
http://www.prologonlinereference.org/psp.psp,
http://www.benjaminjohnston.com.au/template.prolog?t=psp,
http://www.ifcomputer.com/inap/inap2001/program/inap bartenstein.ps
http://www.cs.otago.ac.nz/staffpriv/ok/pwp.pl
6
J. Wielemaker, Z. Huang and L. van der Meij
...,
mkthumbnail(URL, Caption, ThumbNail),
output_html([ h1("Photo gallery"),
ThumbNail
]).
mkthumbnail(URL, Caption, Term) :Term = table([ tr(td([halign=center], img([src=URL],[]))),
tr(td([halign=center], Caption))
])
Fig. 2. Building PiLLoW terms
terms’ that are translated in a special way, as discussed in section 6.2 of (Gras and
Hermenegildo 2001).
We introduced a DCG rule html//1.7 This rule translates proper trees into a list
of high-level HTML/XML commands that are handed to html print/1 to realise
proper quoting, character encoding and layout. The intermediate format is of no
concern to the user and similar in structure to the PiLLoW representation without
using environments. Generated from the tree representation however, consistent
opening and closing of elements is guaranteed. In addition to variable substitution
which is provided by Prolog we allow calling rules. Rules are invoked by a term
\Rule embedded in the argument of html//1. Figure 3 illustrates our approach.
Note that any reusable part of the page generation can easily be translated into
a DCG rule and the difference between direct translation of terms to HTML and
rule-invocation is eminent.
In our current implementation rules are called using meta-calling from html//1.
Using term expansion/2 it is straightforward to move the rule invocation out
of the term, using variable substitution similar to PiLLoW. It is also possible to
recursively expand the generated tree and validate it to the HTML DTD at compiletime and even insert omitted tags at compile-time to generate valid XHMTL from
an incomplete specification. An overview of the argument to html//1 is given in
figure 4.
3.2 Comparison with PiLLoW
The PiLLoW library (Gras and Hermenegildo 2001) is a well established framework
for Web programming based on Prolog. PiLLoW defines html2terms/2, converting between an HTML string and a document represented as a Herbrand term.
There are fundamental differences between PiLLoW and the primitives described
here.
7
The notation hnamei//harityi refers to the grammar rule hnamei with the given harityi, and
consequently the predicate hnamei with arity harityi+2.
SWI-Prolog and the Web
7
affiliation_table :findall(Name-Aff, affiliation(Name, Aff), Pairs0),
keysort(Pairs0, Pairs),
reply_page(table([border(2),align(center)],
[ tr([th(’Name’), th(’Affiliation’)])
| \affiliations(Pairs)
])).
affiliations([]) --> [].
affiliations([H|T]) --> affiliation(H), affiliations(T).
affiliation(Name-Aff) --> html(tr(td(Name), td(Aff))).
% database
affiliation(wielemaker, uva).
affiliation(huang, vu).
affiliation(’van der meij’, vu).
% Page template
reply_page(Term) :format(’Content-type: text/html~n~n’),
phrase(html(Term), Tokens),
print_html(Tokens).
Fig. 3. Library html write.pl in action
hhtmli
hcontenti
hattributei
htagi, hentityi
hvaluei
hrulei
::=
::=
|
|
|
::=
::=
::=
::=
list-of hcontenti | hcontenti
hatomi
htagi(list-of hattributei, hhtmli)
htagi(hhtmli)
\hrulei
hnamei(hvaluei)
hatomi
hatomi | hnumberi
hcallablei
Fig. 4. The html//1 argument specification
• PiLLoW creates an HTML document from a Herbrand term that is passed
to html2terms/2. Complex terms are composed of partial terms passed as
Prolog variables. Frequent HTML constructs are supported using reserved
terms using dedicated processing. We use DCGs and the \Rule construct,
which makes it eminent which terms directly refer to HTML elements and
which function as a ‘macro’. In addition, the user can define applicationspecific reusable fragments in a uniform way.
• The PiLLoW parser does not create the SGML document tree. It does not
insert omitted tags, default attributes, etcetera. As a result, HTML docu-
8
J. Wielemaker, Z. Huang and L. van der Meij
[env(table, [], [tr$[], td$[], "Hello"])]
[element(table, [],
[ element(tbody, [],
[ element(tr, [],
[ element(td, [rowspan=’1’, colspan=’1’],
[’Hello’])])])])]
Fig. 5. Term representations for <table><tr><td>Hello</table> in PiLLoW (top) and
our parser (bottom). Our parser completes the tr and td environments, inserts the omitted
tbody element and inserts the defaults for the rowspan and colspan attributes
Special Issue
Journal
rdf:type
rdf:type
"Massimo Marchiori"
rdf:label
Massimo
Marchiori
dc:editor
Logic Programming
and the Web
TPLP
issue_in
Fig. 6. Sample RDF graph. Ellipses are vertices representing URIs. Quoted text is a
literal. Edges are labelled with URIs.
ments that differ only in omitted tags and whether or not default attributes
are included in the source produce different terms. In our approach the term
representation is equivalent, regardless of the input document. This is illustrated in figure 5. Having a canonical DOM representation greatly simplifies
processing parsed HTML documents.
4 RDF documents
Where the datamodel of both HTML and XML is a tree-structure with attributes, the datamodel of the Semantic Web (SW) RDF8 language consists of
{Subject, Predicate, Object} triples. Both Subject and Predicate are URI s.9 Object
is either a URI or a Literal. As the Object of one triple can be the Subject of another, a set of triples forms a graph, where each edge is labelled with a URI (the
Predicate) and each vertex is either a URI or a literal. Literals have no out-going
edges. Figure 6 illustrates this.
A number of languages are layered on top of the RDF triple model. RDFS provides a frame-based representation. The OWL-dialects10 provide three increasingly
8
9
10
http://www.w3.org/RDF/
URI: Uniform Resource Identifier is like a URL, but need not refer to an existing resource on
the Web.
http://www.w3.org/2004/OWL/
SWI-Prolog and the Web
hsubjecti, hpredicatei
hobjecti
hlit valuei
hURIi, htexti
hlangidi
::=
::=
|
::=
|
|
::=
::=
9
hURIi
hURIi
literal(hlit valuei)
htexti
lang(hlangidi, htexti)
type(hURIi, htexti)
hatomi
hatomi (ISO639)
Fig. 7. RDF types in Prolog.
complex Web ontology languages. SWRL11 is a proposal for a rule language. The
W3C standard for exchanging these triple models is an XML application known as
RDF/XML.
As there are multiple XML tree representations for the same triple-set, RDF
documents cannot be processed at the level of the XML-DOM as described in
section 2. A triple- or graph-based structure is the most natural choice for representating an RDF document in Prolog. First we must decide on the representation
of URIs and literals. As a URI is a string and the only operation defined on URIs
by SW languages is equivalence test, using a Prolog atom is an obvious choice. One
may consider using a term hnamespacei:hlocalnamei, but given that decomposing a
URI into its namespace and localname is only relevant during I/O we consider this
an inferior choice. The RDF library comes with a compile-time rewrite mechanism
based on goal expansion/2 that allows for writing resources in Prolog sourcetext
as hnsi:hlocali. Literals are expressed as literal(Value). The full type description is
in figure 7.
The typical SW use-scenario is to ‘harvest’ triples from multiple sources and
collect them in a database before reasoning with them. Prolog can represent data
as a Herbrand term on the stack or as predicates in the database. Given the relatively static nature of the RDF data as well as desired access from multiple threads,
using the Prolog database is the most obvious choice. Here we have two options.
One is the predicate rdf(Subject, Predicate, Object) using the argument types described above. The alternative is to map each RDF predicate on a Prolog predicate
Predicate(Subject, Object). We have chosen for rdf/3 because it supports queries
with uninstantiated predicates better and a single predicate is easier to manage
than an unbounded set of predicates with unknown names.
4.1 Input and output of RDF documents
The RDF/XML parser is realised as a Prolog library on top of the XML parser described in section 2. Similar to the XML parser it has two interfaces. The predicate
load rdf(+Src, -Triples, +Options) parses a document and returns a Prolog list of
11
http://www.w3.org/Submission/SWRL/
10
J. Wielemaker, Z. Huang and L. van der Meij
load_triples(File, Options) :process_rdf(File, assert_triples, Options).
assert_triples([], _).
assert_triples([rdf(S,P,O)|T], Src) :rdf_assert(S, P, O, Src),
assert_triples(T, Src).
Fig. 8. Loading triples using process rdf/3
<Painting rdf:about="...">
<dimension>
<Dimension width="45" height="50"/>
</dimension>
</Painting>
Fig. 9. Blank node to express the compound dimension property
rdf(S,P,O) triples. Note that despite harvesting to the database is the typical usecase scenario, the parser delivers a list of triples for maximal flexibility. The predicate process rdf(+Src, :Action, +Options) exploits the mixed call-back/convert
mode of the XML parser to process the RDF file one description (record) at a
time, calling Action with a list of triples extracted from the description. Figure 8
illustrates how this is used by the storage module to load unbounded files with
limited stack usage. Source location as hfilei:hlinei is passed to the Src argument of
assert triples/2.
In addition to named URIs, RDF resources can be blank-nodes. A blank-node
(short bnode) is an anonymous resource that is created from an in-lined description.
Figure 9 describes the dimensions of a painting as a compound instance of class
Dimension with width and height properties. The Dimension instance has no URI.
Our parser generates an identifier that starts with a double underscore, followed by
the source and a number. The double underscore is used to identify bnodes. Source
and number are needed to guarantee the bnode is unique.
The parser from XML to RDF triples covers the full RDF specification, including
Unicode handling, RDF datatypes and RDF language tags. The Prolog source is
1,788 lines. It processes approximately 9,000 triples per second on an AMD 1600+
based computer. Implementation details and evaluation of the parser are described
in (Wielemaker et al. 2003).12
We
have
two
libraries
for
writing
RDF/XML.
One,
rdf write xml(+Stream, +Triples), provides the inverse of load rdf/2, writing
12
The parser described there did not yet support RDF datatypes and language tags, nor Unicode.
SWI-Prolog and the Web
11
Table 1. Call-statistics on a real-world system
Index pattern
+
+
+
+
+
+
+
+
+
+
Calls
58
253,554
62
23,292,353
633,733
7,807,846
26,969,003
an XML document from a list of rdf(S,P,O) terms. The other, called rdf save/2
is part of the RDF storage module described in section 4.2 and writes a database
directly to a file or stream. The first (rdf write xml/2) is used to exchange
computed graphs to external programs using network communication, while the
second (rdf save/2) is used to save modified graphs back to file. The resulting
code duplication is unfortunate, but unavoidable. Creating a temporary graph in a
database requires potentially much memory, and harms concurrency, while graphs
fetched from the database into a list may not fit in the Prolog stacks and is also
considerably slower than a direct write.
4.2 Storage of RDF
Assuming the ‘harvesting’ use-case, we need to implement a predicate
rdf(?S,?P,?O). Indexing the database is crucial for good performance. Table 1 illustrates the calling pattern from a real-world application counting 4 million triples.
Also note that our data is described by figure 7. The RDF store was developed in
the context of projects which formulated the following requirements.
•
•
•
•
•
•
•
•
Upto at least 10 million triples on 32-bit hardware.
Fast graph traversal using any instantiation pattern.
Case-insensitive search on literals.
Prefix search on literals for completion in the User Interface.
Searching for words that appear in literals.
Multi-threaded access based on read/write locks.
Transaction management and persistent store.
Maintain source information, so we can update, save or remove data based
on its source.
• Fast load/save of current state.
Our first version of the database used the Prolog database with secondary tables to improve indexing. As requirements pushed us against the limits of what is
achievable in a 32-bit address-space we decided to implement the low level store
in C. Profiting from the known uniform structure of the data we realised about
12
J. Wielemaker, Z. Huang and L. van der Meij
two times more compact storage with better indexing than using a pure Prolog
approach. We took the following design decisions for the C-based storage module:
• The RDF predicates are represented as unique entities and organised according to the rdfs:subPropertyOf relation in multiple hierarchies. The root of
each hierarchy is used to compute the hash for the triple. If there is no unique
root due to a cycle an arbitrary predicate is assigned to be the root.
• Literals are kept in an AVL tree, sorted case-insensitive and case-preserving
(e.g. AaBb. . . ). Numeric literals preceed all non-numeric and are kept sorted
on their numeric value. Storing literals in a separate sorted table allows for
indexed search for prefixes and numeric values. It also allows for monitoring creation and destruction of literals to maintain derived tables such as
stemming or double methaphone (Philips 2000) based on rdf monitor/3
described below. The space overhead of maintaining the table is roughly cancelled by avoiding duplicates. Experience on real data ranges between -5%
and +10%.
• Resources are represented by Prolog atom-handles. The hash is computed
from the handle-value. Note that avoiding the translation between Prolog
atom and text avoids both duplication of data and table-lookup. We consider
this a crucial aspect.
• Each triple is represented by the atom-handle for the subject, predicatepointer, atom-handle or literal pointer for object, a pointer to the source,
a line number, a general bit-flag field and 6 ‘hash-next’ pointers covering all
indexing patterns except for +,+,+ and +,-,+. Queries using the pattern
+,-,+ are rare. Fully instantiated queries internally use the pattern +,+,-,
assuming few values on the same property. Considering experience with real
data we will probably add a +,+,+ index in the future. The un-indexed table is a simple linked list. The others are hash-tables that are automatically
resized if they become too populated.
The store itself does not allow for writes while there are active reads in progress.
If another thread is reading, the write operation will stall until all threads have
finished reading. If the thread itself has an open choicepoint a permission error
exception is raised. To arrive at meaningful update semantics we introduced transactions. The thread starting a transaction obtains a write-lock, initially allowing
readers to proceed. During the transaction all changes are recorded in a linked list
of actions. If the transaction is ready for commit, the thread denies access to new
readers and waits for all readers to vanish before updating the database. Transactions are realised by rdf transaction(:Goal). If Goal succeeds, its choicepoints
are discarded and the transaction is committed. If Goal fails or raises an exception
the transaction is discarded and rdf transaction/1 returns failure or exception.
Transactions can be nested. Nesting a transaction places a transaction-mark in the
list of actions of the current transaction. Committing implies removing this mark
from the list. Discarding removes all action cells following the mark as well as the
mark itself.
SWI-Prolog and the Web
13
% entailment interface % RDFS interface
% triples
?- rdf(mary, type, X). ?- rdfs_individual_of(mary, X).
mary type woman .
X = woman ;
X = woman ;
woman type Class .
X = human ;
woman subClassOf human . X = human ;
No
No
human type Class .
Fig. 10. Different interface styles for RDFS
It is possible to monitor the database using rdf monitor(:Goal, +Events).
Whenever one of the monitored events happens Goal is called. Modifying actions
inside a transaction are called during the commit. Modifications by the monitors
are collected in a new transaction which is committed immediately after completing the preceeding commit. Monitor events are assert, retract, update, new literal,
old literal, transaction begin/end and file-load. Goal is called in the modifying
thread. As this thread is holding the database write lock, all invocations of monitor
calls are fully serialized.
Although the 9,000 triples per second of the RDF/XML parser ranks it among
the fast parsers, loading 10 million triples takes nearly 20 minutes. For this reason
we developed a binary format. The format is described in (Wielemaker et al. 2003)
and loads approximately 10 times faster than RDF/XML, while using about the
same space. The format is independent from byte-order and word-length, supporting
both 32- and 64-bit hardware.
Persistency is achieved through the library rdf persistency.pl, which uses
rdf monitor/3 to maintain a set of files in a directory. Each source known to
the database is represented by two files, one file representing the initial state using the quick-load binary format and one file containing Prolog terms representing
changes, called the journal.
4.3 Reasoning with RDF documents
We have identified two approaches for reasoning on top of the plain RDF predicate
for more high-level languages such as RDFS or OWL. One approach is taken by the
SeRQL query system described in section 7. It is based on the observation that these
languages provide rules to deduce new triples from the set of known triples. The
API for high level languages is now simply the rdf/3 predicate, where rdf(S,P,O)
is true for any triple in the deductive closure of the original triple set under the
given language. The deductive closure can be realised using full forward reasoning,
deducing new triples until this is no longer possible or by a combination of backward
reasoning and forward reasoning. An alternative approach is to consider RDFS or
OWL at the conceptual level and introduce a set of predicates that are inspired
on this level. This approach is taken by our library rdfs.pl, defining predicates
such as rdfs individual of(?Resource, ?Class), rdfs subclass of(?Sub, ?Super).
Figure 10 illustrates the difference in these approaches.
14
J. Wielemaker, Z. Huang and L. van der Meij
4.4 Experience
The RDF infrastructure is used in two of the three case-studies described at the
end of this paper. RDF has a natural representation in Prolog, either as a list of
terms or as a pure predicate. Prolog non-determinism greatly simplifies querying
the database. Transitivity is easily expressed using recursion. However, as cycles in
SW graphs are allowed and frequent, such algorithms must be protected against
them. Cycle-detection complicates the code and harms performance. We plan to
investigate tabling (Ramakrishnan et al. 1995) to improve on this situation.
Although designed as RDF store for SW-based projects, the infrastructure is
also commonly used to create RDF documents from other sources as well as for
filtering and reorganizing RDF documents. In the e-culture project13 it has been
used to convert WordNet (Miller 1995) from its Prolog representation and the Getty
thesauri14 from XML (4GB data) into RDF.
5 Supporting HTTP
HTTP, or HyperText Transfer Protocol, is the key W3C standard protocol for
exchanging Web documents. All browsers and Web servers implement it. The initial
version of the protocol was very simple. The client request consists of a single line
of the format hactioni hpathi, the server replies with the requested document and
closes the connection. Version 1.1 of the protocol is more complicated, providing
additional name-value pairs in the request as well as the reply, features to request
status such as modification time, transfer partial documents, etcetera.
Adding HTTP support in Prolog, we must consider both the client- and serverside. In both cases our choice is between doing it in Prolog or re-using an existing
application or library by providing an interface for it. We compare our work with
PiLLoW (Cabeza and Hermenegildo 2003) and the ECLiPSe HTTP services (Leth
et al. 1996).
Given a basic TCP/IP socket library, writing an HTTP client is trivial (our client
is 258 lines of code). Both PiLLoW and ECLiPSe include a client written in Prolog.
More issues complicate the choice for a pure Prolog based server.
• The server is more complex, which implies there is more to gain by re-using
external code. Our core server library counts 1,784 lines.
• A single computer can only host one server at port 80 used by default for
public HTTP. Using an alternate port for middleware and storage tier components is no problem, but use as a public server often conflicts with firewall
or proxy settings. This can be solved using a proxy server such as the Apache
mod proxy 15 configured as reverse proxy.
• Servers by definition introduce security risks. Administrators are reluctant to
see non-proven software in the role of a public server. Using a proxy as above
also reduces this risk, especially if the proxy blocks malformed requests.
13
14
15
e-culture.multimedian.nl
http://www.getty.edu/research/conducting research/vocabularies/
http://httpd.apache.org/docs/1.3/mod/mod proxy.html
SWI-Prolog and the Web
application
http_open.pl
http_client.pl
15
inetd_http.pl
http_wrapper.pl
http_header.pl
xpce_http.pl
N
e
t
w
o
r
k
thread_http.pl
Fig. 11. Module dependencies of the HTTP library
Despite these observations, we consider, like the ECLiPSe team, a pure Prolog
based server worthwhile. As argued in section 6.1, many Prolog Web applications
profit from using state stored in the server. Large resources such as WordNet (Miller
1995) cause long startup times. In such cases the use of CGI (Common Gateway
Interface) is not appropriate as a new copy of the application is started for each
request. PiLLoW resolves this issue by using Active Modules, where a small CGI
application talks to a continuously running Prolog server using a private protocol.
Using a Prolog HTTP server and optionally a reverse proxy has the same benefits,
but based on a standard protocol, it is much more flexible.
Another approach is embedding Prolog in another server framework such as the
Java based Tomcat server. Although feasible, embedding non-Java based Prolog
systems in Java is complicated. Embedding through jni introduces platform and
Java version dependent problems. Connecting Prolog and Java concurrency models
and garbage collection is difficult and the resulting system is much harder to manage
by the user than a pure Prolog based application.
In the following sections we describe our HTTP client and server libraries. An
overall overview of the modules and their dependencies is given in figure 11.
5.1 HTTP client libraries
We support two clients. The first is a lightweight client that only supports the
HTTP GET method by means of http open(+URL, -Stream, +Options). Options
allows for setting a timeout or proxy as well as getting information from the replyheader such as the size of the document. The http open/3 predicate internally
handles HTTP 3xx (redirect) replies. Other non-ok replies are mapped to a Prolog
exception. After reading the document the user must close the returned streamhandle using the standard Prolog close/1 predicate. This predicate makes accessing
an HTTP resource as simple as accessing a local file. The second library, called
http client.pl, provides support for HTTP POST and a plugin interface that
allows for installing handlers for documents of specified MIME-types. It shares
http header.pl with the server libraries for DCG based creation and parsing of
HTTP headers. Currently provided plugins include http mime plugin.pl to handle
multipart MIME messages and http sgml plugin.pl for automatically parsing
16
J. Wielemaker, Z. Huang and L. van der Meij
?- use_module(library(’http/http_client’)).
?- use_module(library(’http/http_sgml_plugin’)).
?- http_get(’http://www.swi-prolog.org/’, DOM, []).
DOM = [element(html, [version=’-//W3C//DTD HTML 4.0 Transitional//EN’],
[element(head, [],
[element(title, [],
[’SWI-Prolog\’s Home’]), ...
Fig. 12. Fetching an HTML document
HTML, XML and SGML documents. Figure 12 shows the code for fetching a URL
and parsing the returned HTML document it into a Prolog term as described in
section 2.
Both the PiLLoW and ECLiPSe approach return the documents content as a
string. Our interface is stream-based (http open/3) or allows for plugin-based
processing of the stream (http get/3, http post/4). This interface avoids potentially large intermediate data-structures and allows for processing unbounded
documents.
5.2 The HTTP server library
Both to simplify re-use of application code and to make it possible to use the
server without committing to a large infrastructure we adopted the reply-strategy
of the CGI protocol, where the handler writes a page consisting of an HTTP header
followed by the document content. Figure 13 provides a simple example that returns
the request-data to the client. By importing thread http.pl we implicitly selected
the multi-threaded server model. Other models provided are inetd http, causing
the (Unix) inet daemon to start a server for each request and xpce http which uses
I/O multiplexing realising multiple clients without using Prolog threads. The logic
of handling a single HTTP request given a predicate realising the handler, an input
and output stream is implemented by http wrapper.
Replies other than “200 OK” are generated using a Prolog exception. Recognised
replies are defined by the predicate http reply(+Reply, +Stream, +Header). For
example to indicate that the user has no access to a page we must use the following
call.
throw(http_reply(forbidden(URL))).
Failure of the handler raises a “404 existence error” reply, while exceptions other
than the ones described above raise a “500 Server error” reply.
5.2.1 Form parameters
The library http parameters.pl defines http parameters(+Request, ?Parameters)
to fetch and type-check parameters transparently for both GET and POST re-
SWI-Prolog and the Web
17
:- use_module(
library(’http/thread_httpd’)).
start_server(Port) :http_server(reply, [port(Port)]).
reply(Request) :format(’Content-type: text/plain~n~n’),
writeln(Request).
Fig. 13. A simple HTTP server. The right window shows the client and the format of
the parsed request.
reply(Request) :http_parameters(Request,
[ title(Title, [optional(true)]),
name(Name,
[length >= 2]),
age(Age,
[integer])
]), ...
Fig. 14. Fetching HTTP form data
quests. Figure 14 illustrates the functionality. Parameter values are returned as
atoms. If large documents are transferred using a POST request the user may wish
to revert to http read data(+Request, -Data, +Options) underlying http get/3
to process arguments using plugins.
5.2.2 Session management
The library http session.pl provides session over the stateless HTTP protocol. It
does so by adding a cookie using a randomly generated code if no valid session id is
found in the current request. The interface to the user consists of a predicate to set
options (timeout, cookie-name and path) and a set of wrappers around assert/1
and retract/1, the most important of which are http session assert(+Data),
http session retract(?Data) and http session data(?Data). In the current version the data associated with sessions that have timed out is simply discarded.
Session-data does not survive the server.
Note that a session generally consists of a number of HTTP requests and replies.
Each request is scheduled over the available worker threads and requests belonging
to the same session are therefore normally not handled by the same thread. This
implies no session state can be stored in global variables or in the control-structure
of a thread. If such style of programming is wanted the user must create a thread
that represents the session and setup communication from the HTTP-worker thread
to the session thread. Figure 15 illustrates the idea.
18
J. Wielemaker, Z. Huang and L. van der Meij
reply(Request) :% HTTP worker
(
http_session_data(thread(Thread))
-> true
;
thread_create(session_loop([]), Thread, [detached(true)]),
http_session_assert(thread(Thread))
),
current_output(CGIOut),
thread_self(Me),
thread_send_message(Thread, handle(Request, Me, CGIOut)),
thread_get_message(_Done).
session_loop(State) :% Session thread
thread_get_message(handle(Request, Sender, CGIOut)),
next_state(Request, State, NewState, CGIOut).
thread_send_message(Sender, done).
Fig. 15. Managing a session in a thread. The reply/1 predicate is part of the HTTP
worker pool, while session loop/1 is executed in the thread handling the session. We
omitted error handling for readability of the example.
Table 2. HTTP performance executing a trivial query 10,000 times. Times are in
seconds. Localhost, dual AMD 1600+ running SuSE Linux 10.0
Connection
Close
Keep-Alive
Elapsed
Server CPU
Client CPU
20.84
16.23
11.70
8.69
7.48
6.73
5.2.3 Evaluation
The presented server infrastructure is currently used by many internal and external
projects. Coding a server is very similar to writing CGI handlers and running in the
interactive Prolog process is much easier to debug. As Prolog is capable of reloading
source files in the running system, handlers can be updated while the server is
running. Handlers running during the update are likely to die on an exception
though. We plan to resolve this issue by introducing read/write locks. The protocol
overhead of the multi-threaded server is illustrated in table 2.
6 Enabling extensions to the Prolog language
SWI-Prolog has been developed in the context of projects many of which caused the
development to focus on managing Web documents and protocols. In the previous
sections we have described our Web enabling libraries. In this section we describe
extensions to the ISO-Prolog standard (Deransart et al. 1996) we consider crucial
SWI-Prolog and the Web
19
for scalable and comfortable deployment of Prolog as an agent in a Web centred
world.
6.1 Multi-threading
Concurrency is necessary for applications for the following reasons:
• Network delays may cause communication of a single transaction to take very
long. It is not acceptable if such incidents block access for other clients. This
can be achieved using multiplexed I/O, multiple processes handling requests
in a pool or multiple threads in one or more processes handling requests in a
pool.
• CPU intensive services must be able to deploy multiple CPUs. This can be
achieved using multiple instances of the service and load-balancing or a single
server running on multi-processor hardware or a combination of the two.
As indicated, none of the requirements above require multi-threading support
in Prolog. Nevertheless, we added multi-threading (Wielemaker 2003) because it
resolves the problems mentioned above for medium-scale applications while greatly
simplifying deployment and debugging in a platform independent way. A multithreaded server also allows maintaining state for a specific session or even shared
between multiple sessions simply in the Prolog database. The advantages of this
are described in (Szeredi et al. 1996), using the or-parallel Aurora to serve multiple
clients. This is particularly interesting for accessing the RDF database described in
section 4.2.
6.2 Atoms and Unicode support
Unicode16 is a character encoding system that assigns unique integers (code-points)
to all characters of almost all scripts known in the world. In Unicode 4.0, the codepoints range from 1 to 0x10FFFF. Unicode can handle documents from different
scripts as well as documents that contain multiple scripts in a single uniform representation, an important feature in applications processing Web data. Traditional
HTML applications commonly insert special symbols through entities such as the
copyright ( c ) sign, Greek and mathematical symbols, etcetera. Using Unicode we
can represent all entity values as plain text. As illustrated in the famous Semantic
Web layer cake in figure 16, Unicode is at the heart of the Semantic Web.
HTML documents can be represented using Prolog strings because Prolog integers can represent all Unicode code-points. As we have claimed in section 2 however,
using Prolog strings is not the most obvious choice. XML attribute names and values can contain arbitrary Unicode characters, which requires the unnatural use of
strings for these as well. If we consider RDF, URIs can have arbitrary Unicode
characters17 and we want to represent URIs as atoms to exploit compact storage as
16
17
http://www.Unicode.org/
http://www.w3.org/TR/rdf-concepts/#section-Graph-URIref
20
J. Wielemaker, Z. Huang and L. van der Meij
Trust
E
n
c
r
y
p
t
i
o
n
Namespaces
Rules
S
i
g
Logic
n
Ontology
a
t
RDF Schema
u
RDF Model&Syntax
r
XML Query XML Schema e
Proof
XML
URI
Unicode
Fig. 16. The Semantic Web layer cake by Tim Burners Lee
well as fast equivalence testing. Without Unicode support in atoms we would have
to encode Unicode in the atom using escape sequences. All this patchwork can be
avoided if we demand the properties below for Prolog atoms.
• Atoms represent text in Unicode
• Atoms have no limit on their length
• The Prolog implementation allows for a large number of atoms, both to represent URIs and to represent text in HTML/XML documents. SWI-Prolog’s
limit is 225 (32 million).
• Continuously running servers cannot allow memory leaks and therefore processing dynamic data using atoms requires atom garbage collection.
7 Case study — A Semantic Web Query Language
In this case-study we describe the SWI-Prolog SeRQL implementation.18 SeRQL is
an RDF query language developed as part of the Sesame project19 (Broekstra et al.
2002). SeRQL uses HTTP as its access protocol. Sesame consists of an implementation of the server as a Java servlet and a Java client-library. By implementing
a compatible framework we made our Prolog based RDF storage and reasoning
engine available to Java clients. The Prolog SeRQL implementation uses all of the
described SWI-Prolog infrastructure and building it has contributed significantly
to the development of the infrastructure. Figure 17 lists the main components of
the server.
The entailment modules are plugins that implement the entailment approach to
RDF reasoning described in section 4.3. They implement rdf/3 as a pure predicate, adding implicit triples to the raw triples loaded from RDF/XML documents.
Figure 18 shows the somewhat simplified entailment module for RDF. The multifile
rule registers the module as entailment module for the SeRQL system. New modules can be loaded dynamically into the platform, providing support for other SW
languages or application-specific server-side reasoning. Prolog’s dynamic loading
and re-loading allows for updating such reasoning modules on the live server.
The SeRQL parser is a DCG-based parser translating a SeRQL query text into a
18
19
http://www.swi-prolog.org/packages/SeRQL
http://www.openrdf.org
SWI-Prolog and the Web
Network
Query Optimiser
XML/RDF
over
HTTP
HTML
over
HTTP
SeRQL
HTTP API
User Frontend
SeRQL parser
RDF I/O
RDF Entailment
21
Login & user
Management
RDFS Entailment
RDF-Library
HTTP Server Library
Fig. 17. Module dependencies of the SeRQL system. Arrows denote ‘imports from’
relations.
:- module(rdf_entailment, [rdf/3]).
rdf(S, P, O) :rdf_db:rdf(S, P, O).
rdf(S, rdf:type, rdf:’Property’) :rdf_db:rdf(_, S, _),
\+ rdf_db:rdf(S, rdf:type, rdf:’Property’).
rdf(S, rdf:type, rdfs:’Resource’) :rdf_db:rdf_subject(S),
\+ rdf_db:rdf(S, rdf:type, rdfs:’Resource’).
:- multifile serql:entailment/2.
serql:entailment(rdf, rdf_entailment).
Fig. 18. RDF entailment module
compound goal calling rdf/3 and predicates from the SeRQL runtime library which
provide comparison and functions built into the SeRQL language. The resulting
control-structure is passed to the query optimiser (Wielemaker 2005) which uses
statistics maintained by the RDF database to reorder the pure rdf/3 calls for
best performance. The optimiser uses a generate-and-evaluate approach to find
the optimal order. Considering the frequently long conjunctions of rdf/3 calls,
the conjunction is split into independent parts. Figure 19 illustrates this in a very
simple example. During abstract execution, information on instantiation and types
implied by the runtime library predicates is attached to the variables using dynamic
attributed variables. (Demoen 2002).
HTTP access consists of two parts. The human-centred portal consists of HTML
pages with forms to administer the server as well as view statistics, load and unload
documents and run SeRQL queries interactively presenting the result as an HTML
table. Dynamic pages are generated using the html write.pl library described in
section 3.1. Static pages are served from HTML files by the Prolog server. Ma-
22
J. Wielemaker, Z. Huang and L. van der Meij
...
rdf(Paper, author, Author),
rdf(Author, name, Name),
rdf(Author, affiliation, Affil),
...
Fig. 19. Split rdf conjunctions. After executing the first rdf/3 query Author is bound and
the two subsequent queries become independent. This is also true for other orderings, so
we only need to evaluate 3 alternatives instead of 3! (6).
chines use HTTP POST requests to provide query data and get a reply in XML or
RDF/XML.
The system knows about various RDF input and output formats. To reach modularity the kernel exchanges RDF graphs as lists of terms rdf(S,P,O) and resulttables as lists of terms using the functor row and arity equal to the number of
columns in the table. The system calls a multifile predicate using the format identifier and data to realise the requested format. The HTML output format uses
html write.pl. The RDF/XML format uses rdf write xml/2 described in section 4.1. Both rdf write xml/2 and the other XML output format use straight
calls format/3 to write the document, where quoting values is realised by quoting
primitives provided by the SGML/XML parser described in section 2. Using direct
writing instead of techniques described in section 3.1 avoids potentially large intermediate datastructures and is not very complicated given the very simple structure
of the documents.
7.1 Evaluation
The SeRQL server and the SWI-Prolog library development is too closely integrated
to use it as an evaluation of the functionality provided by the Web enabling libraries.
We compared our server to Sesame, written in Java. The source code of the Prolog
based server is 6,700 lines, compared to 86,000 for Sesame. As both systems have
very different coverage in functionality and can re-use libraries at different levels
it is hard to judge these figures. Both answer trivial queries in approximately 5ms
on a dual AMD 1600+ PC running Linux 2.6. On complex queries the two systems
perform very differently. Sesame’s forward reasoning makes it handle some RDFS
queries much faster. Sesame does not contain a query optimizer which cause orderdependent and sometimes very long response times on large conjunctions.
The power of LP where programs can be handled as data is exploited by parsing
the SeRQL query into a program, optimizing the program by manipulating it as
data, after which we can simply call it to answer the query. The non-deterministic
nature of rdf/3 allows for a trivial translation of the query to a non-deterministic
program that produces the answers on backtracking.
The server only depends on the standard SWI-Prolog distribution and there-
SWI-Prolog and the Web
23
fore runs unmodified on all systems supporting SWI-Prolog. It has been tested on
Windows, Linux and MacOS X.
All infrastructure described is used in the server. We use format/3, exploiting XML quoting primitives provided by the Prolog XML library to print highly
repetitive XML files such as the SeRQL result-table. Alternatively we could have
created the corresponding DOM term and call xml write/2 from the library
sgml write.pl.
8 Case study — XDIG
In section 7 we have discussed the case study how SWI-Prolog is used for a RDF
query system, i.e., a meta-data management and reasoning system. In this section
we describe a Prolog-powered system for ontology management and reasoning based
on Description Logics (DL). DL has greatly influenced the design of the W3C ontology language OWL. The DL community, called DIG (DL Implementation Group)
have developed a standard for accessing DL reasoning engines called the DIG description logic interface20 (Bechhofer et al. 2003), DIG interface for short. Many
DL reasoners like Racer (Haarslev and Möller 2001) and FACT (Horrocks 1999)
support the DIG interface, allowing for the construction of highly portable and
reusable DL components or extensions.
In this case study, we describe XDIG, an eXtended DIG Description Logic interface, which has been implemented on top of the SWI-Prolog Web libraries. The
DIG interface uses an XML-based messaging protocol on top of HTTP. Clients of
a DL reasoner communicate by means of HTTP POST requests. The body of the
request is an XML encoded message which corresponds to the DL concept language.
Where OWL is based on the triple model described in section 4, DIG statements
are grounded directly in XML. Figure 20 shows a DIG statement which defines the
concept MadCow as a cow which eats brains, part of sheep.
8.1 Architecture of XDIG
The XDIG libraries form a framework to build DL reasoners that have additional
reasoning capabilities. XDIG serves as a regular DL reasoner via its corresponding
DIG interface. An intermediate XDIG server can make systems independent from
application specific characteristics. A highly decoupled infrastructure significantly
improves the reusability and applicability of software components.
The general architecture of XDIG is shown in figure 21. It consists of the following
components:
XDIG Server The XDIG server deals with requests from ontology applications. It
supports our extended DIG interface, i.e., it not only supports standard DIG/DL
requests, like ’tell’ and ’ask’, but also additional processing features like changing system settings. The library dig server.pl implements the XDIG protocol on top of the Prolog HTTP server described in section 5.2. The predicate
20
http://dl.kr.org/dig/
24
J. Wielemaker, Z. Huang and L. van der Meij
<equalc>
<catom name=’mad+cow’/>
<and>
<catom name=’cow’/>
<some>
<ratom name=’eats’/>
<and>
<catom name=’brain’/>
<some>
<ratom name=’part+of’/>
<catom name=’sheep’/>
</some></and></some></and></equalc>
Fig. 20. a DIG statement on MadCow
Application/
GUI
XDIG
Server
DIG
Client
External
Reasoner
Main Control Component
Ontology
Repository
Fig. 21. Architecture of XDIG
dig server(+Request) is called from the HTTP server to process a client’s Request as illustrated in figure 13. XDIG server developers have to define the predicate my dig server processing(+Data, -Answer, +Options), where Data is
the parsed DIG XML requests and Answer is term answer(-Header, -Reply).
Reply is the XML-DOM term representing the answer to the query.
DIG Client XDIG is designed to rely on an external DL reasoner. It implements a regular DIG interface client and calls the external DL reasoner to access the standard DL reasoning capabilities. The predicate
dig post(+Data, -Reply, +Options) posts the data to the external DIG server.
The predicates are defined in terms of the predicate http post/4 and others in
the HTTP and XML libraries.
Main Control Component The library dig process.pl provides facilities to
analyse DIG statements such as finding concepts, individuals and roles, but also
decide of satisfiability of concepts and consistency. Some of this processing is
done by analysing the XML-DOM representation of DIG statements in the local
repository, while satisfiability and consistency checking is achieved by accessing
external DL reasoners through the DIG client module.
Ontology Repository The Ontology Repository serves as an internal knowledge
SWI-Prolog and the Web
25
direct_concept_relevant(element(catom, Atts, _), Concept) :memberchk(name=Concept, Atts).
direct_concept_relevant(element(_, _, Content), Concept) :direct_concept_relevant(Content, Concept).
direct_concept_relevant([H|T], Concept) :(
direct_concept_relevant(H, Concept)
;
direct_concept_relevant(T, Concept)
).
Fig. 22. direct concept relevant checks that a concept is referenced by a DIG statement
base (KB), which is used to store multiple ontologies locally. These ontology
statements are used for further processing when the reasoner receives an ‘ask’
request. The main control component usually selects parts from the ontologies
to post them to an external DL reasoner and obtain the corresponding answers.
This internal KB is also used to store system settings.
As DIG statements are XML based, XDIG stores statements in the local repository using the XML-DOM representation described in section 2. The tree model of
XDIG data has been proved to be convenient for DIG data management.
Figure 22 shows a piece of code from the library XDIG defining the predicate
direct concept relevant(+DOM, ?Concept) which checks if a set of Statements
is directly relevant to a Concept, namely the Concept appears in the body of a
statement in the list. The predicate direct concept relevant/2 has been used to
develop PION for reasoning with inconsistent ontologies, and DION for inconsistent
ontology debugging.
8.2 Application
XDIG has been used to develop several DL reasoning services. PION is a reasoning system that deals with inconsistent ontologies21 (Huang and Visser 2004;
Huang et al. 2005). MORE is a multi-version ontology reasoner22 (Huang and Stuckenschmidt 2005). DION is a debugger of inconsistent ontologies23 (Schlobach and
Huang 2005). With the support of an external DL reasoner like Racer, DION can
serve as an inconsistent ontology debugger using a bottom-up approach.
9 Case study — Faceted browser on Semantic Web database
integrating multiple collections
In this case study we describe a pilot for the STITCH-project24 whose main aim is
21
22
23
24
http://wasp.cs.vu.nl/sekt/pion
http://wasp.cs.vu.nl/sekt/more
http://wasp.cs.vu.nl/sekt/dion
http://stitch.cs.vu.nl
26
J. Wielemaker, Z. Huang and L. van der Meij
studying and finding solutions for the problem of integrating controlled vocabularies
such as thesauri and classification systems in the Cultural Heritage domain. The
pilot consists of the integration of two collections – the Medieval Illuminations of the
Dutch National Library (Koninklijke Bibliotheek) and the Masterpieces collection
from the Rijksmuseum – and development of a user interface for browsing the
merged collections. One requirement within the pilot is to use “standard Semantic
Web techniques” during all stages, so as to be able to evaluate their added value.
An explicit research goal was to evaluate existing “ontology mapping” tools.
The problem could be split into three main tasks:
• Gathering data, i.e. records of the collections and controlled vocabularies they
use, and transforming it into RDF.
• Establishing semantic links between the vocabularies using off-the-shelf ontology mapping tools.
• Building a prototype User Interface (UI) to access (search and browse) the
integrated collections and experiment with different ways to access them using
a Web server.
SWI-Prolog has been used in all three tasks; to illustrate the use of the SWI-Prolog
Web libraries in the pilot, in the next section we focus on their application in the
prototype UI because it is the largest subsystem using these libraries.
9.1 Multi-Faceted Browser
Multi-Faceted Browsing is a search and browse paradigm where a collection is
accessed by refining multiple (preferably) structured aspects – called facets – of
its elements. For the user interface and user interaction we have been influenced
by the approach of Flamenco (Hearst et al. 2002). The Multi-Faceted Browser is
implemented in SWI-Prolog. All data is stored in an RDF database, which can be
either an external SeRQL repository or an in-memory SWI-Prolog RDF database.
The system consists of three components, RDF-interaction, which deals with RDFdatabase storage and access, HTML-code generation, for the creation of Web pages
and the Web server component, implementing the HTTP server. They are discussed
in the following sections.
9.1.1 RDF-interaction
We first describe the content of the RDF database before explaining how to access
it. The RDF database contains:
• 750 records from the Rijksmuseum, and 1000 from the Koninklijke Bibliotheek.
• RDF representation of the hierarchically structured facets, we use SKOS25 , a
model dedicated to the represention of controlled vocabularies.
25
http://www.w3.org/2004/02/skos/
SWI-Prolog and the Web
27
SELECT Rec, RecTitle, RecThumb, CollSpec
FROM {SiteId} rdfs:label {"ARIATOPS-NONE"};
mfs:collection-spec {CollSpec} mfs:record-type {RT};
mfs:shorttitle-prop {TitleProp};
mfs:thumbnail-prop {ThumbProp},
{Rec} rdf:type {RT}; TitleProp {RecTitle};
ThumbProp {RecThumb},
{Rec} rdf:type {<http://www.telin.nl/rdf/topia#AnimalPieces>},
{Rec} rdf:type {<http://www.telin.nl/rdf/topia#Paintings>}
USING NAMESPACE skos = <http://www.w3.org/2004/02/skos/core#>,
mfs = <http://www.cs.vu.nl/STITCH/pp/mf-schema#><br>
Fig. 23. An example of a SeRQL query, which returns details of records matching two
facet values (AnimalPieces and Paintings)
• Mappings between SKOS Concept Schemes used in the different collections.
• Portal-specific information as “Site Configuration Objects”, identified by URIs
with properties defining what collections are part of the setup, what facets
are shown, and also values for the constant text in the Web page presentation and other User Interface configuration properties. Multiple such Site
Configuration Objects may be defined in a repository.
The in-memory RDF store contains, depending on the number of mappings and
structured vocabularies that are stored in the database, about 300,000 RDF triples.
The Sesame store contains more triples – 520,000 – as its RDFS-entailment implementation implies generation of derived triples (see section 7).
RDF database access Querying the RDF store for more complex results based on
URL query arguments consisted of three steps: 1) building SeRQL queries from
URL query arguments, 2) passing them on to the SeRQL-engine, gathering the
result rows and 3) finally post-processing the output, e.g. counting elements and
sorting them. Figure 23 shows an example of a generated SeRQL query. Finding matching records involves finding records annotated by the facet value or by
a value that is a hierarchical descendant of facet value. We implemented this by
interpreting records as instances of SKOS concepts and using the transitive and
reflexive properties of the rdfs:subClassOf property. This explains for example
{Rec} rdf:type {<http://www.telin.nl/rdf/topia#Paintings>} in figure 23.
The SeRQL-query interface contains timing and debugging facilities for single
queries; for flexibility it provides access to an external SeRQL server26 for which
we used Sesame27 , but also to the in-memory store of the SWI-Prolog SeRQL
implementation described in section 7.
26
27
We used the sesame client.pl library that provides an interface to external SeRQL servers,
packaged with the SWI-Prolog SeRQL library
http://www.openrdf.org
28
J. Wielemaker, Z. Huang and L. van der Meij
objectstr([],_O, _Cols,_Args) --> [].
objectstr([RowObjects|ObjectsList], Offset, Cols, Args) -->
{ Percentage is 100/Cols },
html(tr(valign(top),\objectstd(RowObjects, Offset, Percentage, Args))),
{ Offset1 is Offset + Cols },
objectstr(ObjectsList, Offset1, Cols, Args).
objectstd([], _, _, _) --> [].
objectstd([Url|RowObjects], Index, Percentage, Args) -->
{
..
construct_href_index(..., HRef),
missing_picture_txt(Url, MP)
},
html(td(width(Percentage),a([href(HRef)],img([src(Url),alt(MP)])))),
{ Index1 is Index + 1 },
objectstd(RowObjects, Index1, Percentage, Args).
Fig. 24. Part of the html code generation for displaying all the images of a query result
in an HTML table
9.1.2 HTML-code generation
We used the SWI-Prolog html write.pl library described in section 3.1 for our
HTML-code generation. There are three distinct kinds of Web pages the multifaceted browser generates, the portal access page, the refinement page and the single
collection-item page. The DCG approach to generating HTML code made it easy
to share HTML-code generating procedures such as common headers and HTML
code for refinement of choices. The HTML-code generation component contains
some 140 DCG rules (1200 lines of Prolog code of which 800 lines are DCG rules),
part of which are simple list-traversing rules such as the example of Figure 24.
9.1.3 Web Server
The Web server is implemented using the HTTP server library described in section 5.2. The Web server component itself is very small. It follows the skeleton code
described in Figure 13. In our case the reply/1 predicate extracts the URL root
and parameters from the URL. The Site Configuration Object, which is introduced
in section 9.1.1, is returned by the RDF-interaction component based on the URL
root. It is passed on to the HTML-code generation component which generates Web
content as shown in reply page/1 in Figure 3.
9.2 Evaluation
This case study shows that SWI-Prolog is effective for building applications in
the context of the Semantic Web. In a single month a fully functional prototype
portal has been created providing structured access to multiple collections. The
independence of any external libraries and the full support of all libraries on different
SWI-Prolog and the Web
29
platforms made it easy to develop and install in different operating systems. All
case study software has been tested to install and run transparently both on Linux
and on Microsoft Windows.
At the start of the pilot project we briefly evaluated existing environments for
creating multi-faceted browsing portals: We considered the software available from
the Flamenco Project (Hearst et al. 2002) and the OntoViews Semantic Portal Creation Tool (Mäkelä et al. 2004). The Flamenco software would require developing a
translation from RDF to the Flamenco data representation. OntoViews heavily uses
Semantic Web techniques, but the software was unnecessarily complex for our pilot,
requiring a number of external libraries. This together with our need for flexible
experiments with various setups made us decide to build our own prototype.
The prototype allowed us to easily experiment with and develop various interesting ways of providing users access to integrated heterogeneous collections (van
Gendt et al. 2006). The pilot project portal is accessible on line for your evaluation
at http://stitch.cs.vu.nl/demo.html.
10 Conclusion
We have presented an overview of the libraries and Prolog language extensions we
have implemented and which we provide to the LP community as Open Source
resources. As the presented libraries cover very different functionality we have compared our approach with related work throughout the document. We have demonstrated that Prolog, equipped with an HTTP server library, libraries for reading and
writing markup documents, multi-threading, Unicode support, unbounded atoms
and atom garbage collection, becomes a flexible component in a multi-tier server architecture. Automatic memory management and the absence of pointers and other
dangerous language constructs justify the use of Prolog in security sensitive environments. The middleware, typically dealing with the application logic is the most
obvious tier for exploiting Prolog. In our case studies we have seen Prolog active as
storage component (section 7), middleware (section 8) and in the presentation tier
(section 9).
Development in the near future is expected to concentrate on Semantic Web reasoning, such as the translation of SWRL rules to logic programs. Such translations
will benefit from tabling to reach at more predictable response-times and allow
for more declarative programs. We plan to add more support for OWL reasoning,
possibly supporting vital relations for ontology mapping such as owl:sameAs in
the low-level store. We also plan to add PSP/PWP-like (section 3) page-generation
facilities.
Acknowledgements
Recent development and preparing this paper was supported by the MultimediaN28
project funded through the BSIK programme of the Dutch Government. Older
28
www.multimedian.nl
30
J. Wielemaker, Z. Huang and L. van der Meij
projects that provided a major contribution to this research include the European
projects IMAT and GRASP and the Dutch project MIA. The XDIG case-study is
supported by FP6 Network of Excellence European project SEKT (IST-506826).
The faceted browser case-study is funded by CATCH, a program of the Dutch
Organization NWO. We acknowledge the SWI-Prolog community for feedback on
the presented facilities, in particular Richard O’Keefe for his extensive feedback on
SGML support.
References
Bechhofer, S., Möller, R., and Crowther, P. 2003. The DIG description logic
interface. In International Workshop on Description Logics (DL2003). Rome.
Broekstra, J., Kampman, A., and van Harmelen, F. 2002. Sesame: An architecture
for storing and querying rdf and rdf schema. In Proc. First International Semantic Web
Conference ISWC 2002, Sardinia, Italy. LNCS, vol. 2342. Springer-Verlag, 54–68.
Cabeza, D. and Hermenegildo, M. V. 2003. Distributed WWW programming using
(ciao-)prolog and the piLLoW library. CoRR cs.DC/0312031.
Demoen,
B.
2002.
Dynamic
attributes,
their
hProlog
implementation, and a first evaluation.
Report CW 350, Department of
Computer
Science,
K.U.Leuven,
Leuven,
Belgium.
oct.
URL
=
http://www.cs.kuleuven.ac.be/publicaties/rapporten/cw/CW350.abs.html.
Deransart, P., Ed-Dbali, A., and Cervoni, L. 1996. Prolog: The Standard. SpringerVerlag, New York.
Gras, D. C. and Hermenegildo, M. V. 2001. Distributed WWW programming using
(ciao-)prolog and the piLLoW library. TPLP 1, 3, 251–282.
Haarslev, V. and Möller, R. 2001. Description of the racer system and its applications. In Proceedings of the International Workshop on Description Logics (DL-2001).
Stanford, USA, 132–141.
Hearst, M., English, J., Sinha, R., Swearingen, K., and Yee, K. 2002. Finding the
flow in web site search. Communications of the ACM 45, 9 (September), 42–49.
Horrocks, I. 1999. Fact and ifact. In Proceedings of the International Workshop on
Description Logics (DL’99). 133–135.
Huang, Z. and Stuckenschmidt, H. 2005. Reasoning with multiversion ontologies:
a temporal logic approach. In Proceedings of the 2005 International Semantic Web
Conference (ISWC2005).
Huang, Z., van Harmelen, F., and ten Teije, A. 2005. Reasoning with inconsistent
ontologies. In Proceedings of the International Joint Conference on Artificial Intelligence
- IJCAI’05.
Huang, Z. and Visser, C. 2004. Extended DIG description logic interface support for
PROLOG. Deliverable D3.4.1.2, SEKT.
Leth, L., Bonnet, P., Bressan, S., and Thomsen, B. 1996. Towards ECLiPSe agents
on the internet. In Proceedings of the 1st Workshop on Logic Programming Tools for
INTERNET Applications.
Mäkelä, E., Hyvönen, E., Saarela, S., and Viljanen, K. 2004. Ontoviews - a tool for
creating semantic web portals. In International Semantic Web Conference, S. A. McIlraith, D. Plexousakis, and F. van Harmelen, Eds. Lecture Notes in Computer Science,
vol. 3298. Springer, 797–811.
Miller, G. 1995. WordNet: A lexical database for english. Comm. ACM 38, 11 (November).
SWI-Prolog and the Web
31
Philips, L. 2000. The double metaphone search algorithm. j-CCCUJ 18, 6 (June), ??–??
Ramakrishnan, I. V., Rao, P., Sagonas, K., Swift, T., and Warren, D. S. 1995.
Efficient tabling mechanisms for logic programs. In Proceedings of the 12th International
Conference on Logic Programming, L. Sterling, Ed. MIT Press, Cambridge, 697–714.
Schlobach, S. and Huang, Z. 2005.
Inconsistent ontology diagnosis: Framework and prototype.
Deliverable D3.6.1, SEKT.
URL =
http://www.cs.vu.nl/~huang/sekt/sekt361.pdf.
Szeredi, P., Molnár, K., and Scott, R. 1996. Serving Multiple HTML Clients
from a Prolog Application. In Proceedings of the 1st Workshop on Logic Programming Tools for INTERNET Applications. JICSLP”96, Bonn. Available from
http://clement.info.umoncton.ca/˜lpnet/lpnet9.html.
van Gendt, M., Isaac, A., van der Meij, L., and Schlobach, S. 2006. Semantic web
techniques for multiple views on heterogeneous collections: a case study.
Wielemaker, J. 2003. Native preemptive threads in SWI-Prolog. In Practical Aspects of
Declarative Languages, C. Palamidessi, Ed. Springer Verlag, Berlin, Germany, 331–345.
LNCS 2916.
Wielemaker, J. 2005. An optimised semantic web query language implementation in prolog. In ICLP 2005, M. Baggrielli and G. Gupta, Eds. Springer Verlag, Berlin, Germany,
128–142. LNCS 3668.
Wielemaker, J., Schreiber, G., and Wielinga, B. 2003. Prolog-based infrastructure
for RDF: performance and scalability. In The Semantic Web - Proceedings ISWC’03,
Sanibel Island, Florida, D. Fensel, K. Sycara, and J. Mylopoulos, Eds. Springer Verlag,
Berlin, Germany, 644–658. LNCS 2870.
| 6cs.PL
|
Faster exon assembly by sparse spliced alignment
Alexander Tiskin
arXiv:0707.3409v1 [cs.DS] 23 Jul 2007
Department of Computer Science, The University of Warwick,
Coventry CV4 7AL, United Kingdom
[email protected]
Abstract. Assembling a gene from candidate exons is an important
problem in computational biology. Among the most successful approaches
to this problem is spliced alignment, proposed by Gelfand et al., which
scores different candidate exon chains within a DNA sequence of length
m by comparing them to a known related gene sequence of length n,
m = Θ(n). Gelfand et al. gave an algorithm for spliced alignment running
in time O(n3 ). Kent et al. considered sparse spliced alignment, where the
number of candidate exons is O(n), and proposed an algorithm for this
problem running in time O(n2.5 ). We improve on this result, by proposing
an algorithm for sparse spliced alignment running in time O(n2.25 ). Our
approach is based on a new framework of quasi-local string comparison.
1
Introduction
Assembling a gene from candidate exons is an important problem in computational biology. Several alternative approaches to this problem have been developed over time. Among the most successful approaches is spliced alignment
[6],which scores different candidate exon chains within a DNA sequence by comparing them to a known related gene sequence. In this method, the two sequences
are modelled respectively by strings a, b of lengths m, n. We usually assume that
m = Θ(n). A subset of substrings in string a are marked as candidate exons.
The comparison between sequences is made by string alignment. Gelfand et al.
[6] give an algorithm for spliced alignment running in time O(n3 ).
In general, the number of candidate exons k may be as high as O(n2 ). The
method of sparse spliced alignment makes a realistic assumption that, prior to
the assembly, the set of candidate exons undergoes some filtering, after which
only a small fraction of candidate exons remains. Kent et al. [9] give an algorithm
for sparse spliced alignment that, in the special case k = O(n), runs in time
O(n2.5 ). For asymptotically higher values of k, the algorithm provides a smooth
transition in running time to the dense case k = O(n2 ), where its running time
is asymptotically equal to the general spliced alignment algorithm of [6].
In this paper, we improve on the results of [9], by proposing an algorithm
for sparse spliced alignment that, in the special case k = O(n), runs in time
O(n2.25 ). Like its predecessor, the algorithm also provides a smooth transition
in running time to the dense case. Our approach is based on a new framework of
quasi-local string comparison, that unifies the semi-local string comparison from
[12] and fully-local string comparison.
This paper is a sequel to paper [12]; we include most of its relevant material
here for completeness. However, we omit some definitions and proofs due to
space constraints, referring the reader to [12] for the details.
2
Semi-local longest common subsequences
We consider strings of characters from a fixed finite alphabet, denoting string
concatenation by juxtaposition. Given a string, we distinguish between its contiguous substrings, and not necessarily contiguous subsequences. Special cases of
a substring are a prefix and a suffix of a string. Given a string a, we denote
by a(k) and a(k) respectively its prefix and suffix of length k. For two strings
a = α1 α2 . . . αm and b = β1 β2 . . . βn of lengths m, n respectively, the longest
common subsequence (LCS) problem consists in computing the length of the
longest string that is a subsequence both of a and b. We will call this length the
LCS score of the strings.
We define a generalisation of the LCS problem, which we introduced in [12]
as the all semi-local LCS problem. It consists in computing the LCS scores on
substrings of a and b as follows:
• the all string-substring LCS problem: a against every substring of b;
• the all prefix-suffix LCS problem: every prefix of a against every suffix of b;
• symmetrically, the all substring-string LCS problem and the all suffix-prefix
LCS problem, defined as above but with the roles of a and b exchanged.
It turns out that by considering this combination of problems rather than each
problem separately, the algorithms can be greatly simplified.
A traditional distinction, especially in computational biology, is between
global (full string against full string) and local (all substrings against all substrings) comparison. Our problem lies in between, hence the term “semi-local”.
Many string comparison algorithms output either a single optimal comparison
score across all local comparisons, or a number of local comparison scores that
are “sufficiently close” to the globally optimal. In contrast with this approach,
we require to output all the locally optimal comparison scores.
In addition to standard integer indices . . . , −2, −1, 0, 1, 2, . . ., we use odd halfinteger indices . . . , − 25 , − 32 , − 12 , 12 , 32 , 52 , . . .. For two numbers i, j, we write i E j
if j − i ∈ {0, 1}, and i C j if j − i = 1. We denote
[i : j] = {i, i + 1, . . . , j − 1, j}
hi : ji = i + 12 , i + 32 , . . . , j − 23 , j −
1
2
To denote infinite intervals of integers and odd half-integers, we will use −∞ for
i and +∞ for j where appropriate. For both interval types [i : j] and hi : ji, we
call the difference j − i interval length.
We will make extensive use of finite and infinite matrices, with integer elements and integer or odd half-integer indices. A permutation matrix is a (0,1)matrix containing exactly one nonzero in every row and every column. An identity matrix is a permutation matrix I, such that I(i, j) = 1 if i = j, and
b a a b c a b c a b a c a
b
a
a
b
c
b
c
a
Fig. 1. An alignment dag and its implicit highest-score matrix
I(i, j) = 0 otherwise. Each of these definitions applies both to finite and infinite matrices.
A finite permutation matrix can be represented by its nonzeros’ index set.
When we deal with an infinite matrix, it will typically have a finite non-trivial
core, and will be trivial (e.g. equal to an infinite identity matrix) outside of
this core. An infinite permutation matrix with finite non-trivial core can be
represented by its core nonzeros’ index set.
Let D be an arbitrary numerical matrix with indices ranging over h0 : ni. Its
distribution matrix, with indices ranging over [0 : n], is defined by
X
d(i0 , j0 ) =
D(i, j)
i ∈ hi0 : ni, j ∈ h0 : j0 i
for all i0 , j0 ∈ [0 : n].
When matrix d is a distribution matrix of D, matrix D is called the density
matrix of d. The definitions of distribution and density matrices extend naturally to infinite matrices. We will only deal with distribution matrices where all
elements are defined and finite.
We will use the term permutation-distribution matrix as an abbreviation of
“distribution matrix of a permutation matrix”.
We refer the reader to [12] for the definition of alignment dag. In the context of
the alignment dag, a substring αi αi+1 . . . αj corresponds to the interval [i−1 : j];
we will make substantial use of this correspondence in Section 4.
We also refer the reader to [12] for the definitions of (extended) highest-score
matrix, and of its implicit representation. Figure 1 shows an alignment dag of two
strings, along with the nonzeros of its implicit highest-score matrix. In particular,
a nonzero (i, j), where i, j ∈ h0 : ni, is represented by a “seaweed” curve1 ,
originating between the nodes v0,i− 21 and v0,i+ 12 , and terminating between the
1
For the purposes of this illustration, the specific layout of the curves between their
endpoints is not important. However, notice that each pair of curves have at most
one crossing, and that the same property is true for highest-scoring paths in the
alignment dag.
nodes vm,j− 12 and vm,j+ 21 . The remaining curves, originating or terminating at
the sides of the dag, correspond to nonzeros (i, j), where either i 6∈ h0 : ni or
j 6∈ h0 : ni. For details, see [12].
Essentially, an extended highest-score matrix represents in a unified form the
solutions of the string-substring, substring-string, prefix-suffix and suffix-prefix
LCS problems. In particular, row 0 of this matrix contains the LCS scores of
string a against every prefix of string b. When considering such an array of n + 1
LCS scores on its own, we will call it highest-score vector for a against b. Every
highest-score vector will be represented explicitly by an integer array of size
n + 1 (as opposed to the implicit representation of the complete highest-score
matrix, which allows one to store all the rows compactly in a data structure of
size O(m + n)).
3
Fast highest-score matrix multiplication
Our algorithms are based on the framework for the all semi-local LCS problem
developed in [12], which refines the approach of [11,1].
A common pattern in the problems considered in this paper is partitioning
the alignment dag into alignment subdags. Without loss of generality, consider
a partitioning of an (M + m) × n alignment dag G into an M × n alignment dag
G1 and an m × n alignment dag G2 , where M ≥ m. The dags G1 , G2 share a
horizontal row of n nodes, which is simultaneously the bottom row of G1 and
the top row of G2 ; the dags also share the corresponding n − 1 horizontal edges.
We will say that dag G is the concatenation of dags G1 and G2 . Let A, B, C
denote the extended highest-score matrices defined respectively by dags G1 , G2 ,
G. In every recursive call our goal is, given matrices A, B, to compute matrix C
efficiently. We call this procedure highest-score matrix multiplication.
The implicit representation of matrices A, B, C consists of respectively M +n,
m + n, M + m + n non-trivial nonzeros.
The results of this paper are based on the following results from [12]; see the
original paper for proofs and discussion.
Definition 1. Let n ∈ N. Let A, B, C be arbitrary numerical matrices with
indices ranging over [0 : n]. The
(min, +)-product A B = C is defined by
C(i, k) = minj A(i, j) + B(j, k) , where i, j, k ∈ [0 : n].
Lemma 1 ([12]). Let DA , DB , DC be permutation matrices with indices ranging over h0 : ni, and let dA , dB , dC be their respective distribution matrices. Let
dA dB = dC . Given the set of nonzero elements’ index pairs in each of DA ,
DB , the
set of nonzero elements’ index pairs in DC can be computed in time
O n1.5 and memory O(n).
Lemma 2 ([12]). Let DA , DB , DC be permutation matrices with indices ranging over h−∞ : +∞i, such that
DA (i, j) = I(i, j)
for i, j ∈ h−∞ : 0i
DB (j, k) = I(j, k)
for j, k ∈ hn : +∞i
0
n
DA
DB
Fig. 2. An illustration of Lemma 2
Let dA , dB , dC be their respective distribution matrices. Let dA
have
dB = dC . We
DA (i, j) = DC (i, j)
for i ∈ h−∞ : +∞i, j ∈ hn : +∞i
(1)
DB (j, k) = DC (j, k)
for j ∈ h−∞ : 0i, k ∈ h−∞ : +∞i
(2)
Given the set of all n remaining nonzero elements’ index pairs in each of DA ,
DB , i.e. the set of all nonzero elements’ index pairs (i, j) in DA and (j, k) in DB
with i ∈ h0 : +∞i, j ∈ h0 : ni, k ∈ h−∞ : 0i, the set of all n remaining nonzero
elements’ index pairs in DC can be computed in time O n1.5 and memory O(n).
The lemma is illustrated by Figure 2. Three horizontal lines represent respectively the index ranges of i, j, k. The nonzeros in DA and DB are shown respectively by top-to-middle and middle-to-bottom “seaweed” curves. The nonzeros
in DC described by (1), (2) are shown by top-to-bottom thick “seaweed” curves.
The remaining nonzeros in DC are not shown; they are determined by application of Lemma 1 from nonzeros in DA and DB shown by top-to-middle and
middle-to-bottom thin “seaweed” curves.
Lemma 2 gives a method for multiplying infinite permutation-distribution
matrices, in the special case where both multiplicands have semi-infinite core.
We now consider the complementary special case, where one multiplicand’s core
is unbounded, and the other’s is finite.
Lemma 3. Let DA , DB , DC be permutation matrices with indices ranging over
h−∞ : +∞i, such that
DB (j, k) = I(j, k)
for j, k ∈ h−∞ : 0i ∪ hn : +∞i
Let dA , dB , dC be their respective distribution matrices. Let dA
have
DA (i, j) = DC (i, j)
dB = dC . We
for i ∈ h−∞ : +∞i, j ∈ h−∞ : 0i ∪ hn : +∞i
(3)
Given the set of all n remaining nonzero elements’ index pairs in each of DA ,
DB , i.e. the set of all nonzero elements’ index pairs (i, j) in DA and (j, k) in DB
with i ∈ h−∞ : +∞i, j, k ∈ h0 : ni, the set of all n remaining nonzero elements’
index pairs in DC can be computed in time O n1.5 and memory O(n).
0
n
DA
DB
Fig. 3. An illustration of Lemma 3
Proof. By Lemma 1; see Appendix.
t
u
The lemma is illustrated by Figure 3, using the same conventions as Figure 2.
Lemma 4. Consider the concatenation of alignment dags as described above,
with highest-score matrices A, B, C. Given the implicit representations of A,
B, the implicit representation of C can be computed in time O M + m0.5 n and
memory O(M + n).
Proof. By Lemma 3; see Appendix.
t
u
We will also need a separate efficient algorithm for obtaining highest-score
vectors instead of full highest-score matrices. This algorithm, which we call
highest-score matrix-vector multiplication, is complementary to the highest-score
matrix multiplication algorithm of Lemma 1. An equivalent procedure is given
(using different terminology and notation) in [10,5,9], based on techniques from
[8,3].
Lemma 5 ([10,5,9]). Let DA be a permutation matrix with indices ranging
over h0 : ni, and let dA be its distribution matrix. Let x, y be numerical (column)
vectors with indices ranging over h0 : ni. Let dA x = y. Given the set of nonzero
elements’ index pairs in DA , and the elements of x, the elements of y can be
computed in time O(n log n) and memory O(n).
4
Quasi-local string comparison
Consider an arbitrary set of substrings of string a. We call substrings in this set
prescribed substrings, and denote their number by k. Our aim is to compare the
LCS scores on substrings of a and b as follows:
• the quasi-local LCS problem: every prescribed substring of a against every
substring of b.
This problem includes as special cases the semi-local string comparison from
[12] and fully-local string comparison, as well as length-constrained local alignment from [2]. Note that the solution of the quasi-local LCS problem can be
represented in space O(kn) by giving the implicit highest-score matrix for each
prescribed substring of a against b. An individual quasi-local LCS score query
can be performed on this data structure in time O(log2 n) (or even O logloglogn n
with a higher multiplicative constant).
In the rest of this section, we propose an efficient algorithm for the quasilocal LCS problem. For simplicity, we first consider the case k = O(m). Intervals
corresponding to prescribed substrings of a will be called prescribed intervals.
Algorithm 1 (Quasi-local LCS).
Input: strings a, b of length m, n, respectively; a set of k = O(m) endpoint
index pairs for the prescribed substrings in a.
Output: implicit highest-score matrix for every prescribed substring of a against
full b.
Description. For simplicity, we assume that m is a power of 4. We call an
interval of the form [k · 2s : (k + 1) · 2s ], k, s ∈ Z, as well as the corresponding
substring of a, canonical. In particular, all individual characters of a are canonical
substrings. Every substring of a can be decomposed into a concatenation of
O(log m) canonical substrings.
In the following, by processing an interval we mean computing the implicit
highest-score matrix for the corresponding substring of a against b.
First phase. Canonical intervals are processed in a balanced binary tree, in order
of increasing length. Every interval of length 20 = 1 is canonical, and is processed
by a simple scan of string b. Every canonical interval of length 2s+1 is processed
as a concatenation of two already processed half-sized canonical intervals of
length 2s .
Second phase. We represent each prescribed interval [i, j] by an odd half-integer
prescribed point (i, j) ∈ h0 : mi2 (2 ). On the set of prescribed points, we build a
data structure allowing efficient orthogonal range counting queries. A classical
example of such a data structure is the range tree [4].
We then proceed by partitioning the square index pair range h0 : mi2 recursively into regular half-sized square blocks.
Consider an h × h block hi0 − h : i0 i × hj0 : j0 + hi. The computation is
organised so that when a recursive call is made on this block, either we have
i0 ≥ j0 , or the interval [i0 : j0 ] is already processed.
For the current block, we query the number of prescribed points it contains. If
this number is zero, no further computation on the block or recursive partitioning
is performed. Otherwise, we have j −i ∈ {−h, 0, h, 2h, . . .}. If j −i = −h, then the
intervals [i0 − h : j0 ], [i0 : j0 + h] have length 0, and the interval [i0 − h : j0 + h]
is canonical. If j − i ≥ 0, we process the intervals [i0 − h : j0 ], [i0 : j0 + h],
[i0 − h : j0 + h]. Each of these intervals can be processed by Lemma 4, appending
and/or prepending a canonical interval of length h to the already processed
2
The overall algorithm structure is essentially equivalent to building a onedimensional range tree [4] on the interval [0 : m], and then performing on this range
tree a batch query consisting of all the prescribed points. However, in contrast with
standard range trees, the cost of processing nodes in our algorithm is not uniform.
interval [i0 : j0 ]. We then perform further partitioning of the block, and call the
procedure recursively on each of the four subblocks.
The base of the recursion is h = 1. At this point, we process all 1 × 1 blocks
containing a prescribed point, which is equivalent to processing the original
prescribed intervals. The computation is completed.
Cost analysis.
First phase. The computation is dominated by the cost of the bottom level of
the computation tree, equal to m/2 · O(n) = O(mn).
Second phase. The recursion tree has maximum degree 4, height log m, and
O(m) leaves corresponding to the prescribed points.
Consider the top-to-middle levels of the recursion tree. In each level from
the top down to the middle level, the maximum number of nodes increases by a
factor of 4, and the maximum amount of computation work per node decreases
by a factor of 20.5 . Hence, the maximum amount of work per level increases in
geometric progression, and is dominated by the middle level log2 m .
Consider the middle-to-bottom levels of the recursion tree. Since the tree
has O(m) leaves, each level contains at most O(m) nodes. In each level from
the middle down to the bottom level, the maximum amount of computation
work per node still decreases by a factor of 20.5 . Hence, the maximum amount
of work per level decreases in geometric progression, and is again dominated by
the middle level log2 m .
Thus, the computational work in the whole recursion tree is dominated by the
maximum amount of work in the middle level log2 m . This level has at most O(m)
log m
nodes, each requiring at most O(m0.5 n)/20.5· 2 = O(m0.25 n) work. Therefore,
the overall computation cost of the recursion is at most O(m) · O(m0.25 n) =
O(m1.25 n).
t
u
The same algorithm can be applied in the case of general k, 1 ≤ k ≤ m
2 .
For 1 ≤ k ≤ m2/3 , the first phase dominates, so the overall computation cost is
O(mn). For m2/3 ≤ k ≤ m
2 , the second phase dominates. The dominant level in
the recursion tree will have k nodes, each requiring at most O(m0.5 n/k 0.25 ) work.
0.5
0.25
Therefore, the overall computation cost will be
) =
at most k · O(m2 n/k
m
0.5 0.75
O(m k
n). In the fully-local case k = 2 , the cost is O(m n); the same
result can be obtained by m independent runs of algorithms from [11,1], at the
same asymptotic cost.
5
Sparse spliced alignment
We now consider the problem of sparse spliced alignment. We keep the notation and terminology of the previous sections; in particular, candidate exons are
represented by prescribed substrings of string a. We say that prescribed substring a0 = αi0 . . . αj 0 precedes prescribed substring a00 = αi00 . . . αj 00 , if j 0 < i00 .
A chain of substrings is a chain in the partial order of substring precedence. We
identify every chain with the string obtained by concatenating all its constituent
substrings in the order of precedence.
Our sparse spliced alignment algorithm is based on the efficient method of
quasi-local string comparison developed in the previous section. This improves
the running time of the bottleneck procedure from [9]. The algorithm also uses
a generalisation of the standard network alignment method, equivalent to the
one used by [9]. For simplicity, we describe our algorithm for the special case of
unit-cost LCS score.
Algorithm 2 (Sparse spliced alignment).
Input: strings a, b of length m, n, respectively; a set of k = O(m) endpoint
index pairs for the prescribed substrings in a.
Output: the chain of prescribed substrings in a, giving the highest LCS score
against string b.
Description. The algorithm runs in two phases.
First phase. By running Algorithm 1, we compute the implicit highest-score
matrix for every prescribed substring of a against b.
Second phase. We represent the problem by a dag (directed acyclic graph) on
the set of nodes ui , where i ∈ [0 : m]. For each prescribed substring αi . . . αj , the
dag contains the edge ui−1 → uj . Overall, the dag contains k = O(m) edges.
The problem can now be solved by dynamic programming on the representing
dag as follows. Let s[i, j] denote the highest LCS score for a chain of prescribed
substrings in prefix string a(i) against prefix string b(j) . With each node vi , we
associate the integer vector s[i, ·]. The nodes are processed in increasing order
of their indices. For the node u0 , vector s[0, ·] is initialised by all zeros. For
a node uj , we consider every edge ui−1 → uj , and compute the highest-score
matrix-vector product between vector s[i − 1, ·] and the highest-score matrix
corresponding to prescribed string αi . . . αj by the algorithm of Lemma 5. Vector
s[j, ·] is now obtained by taking the elementwise maximum between vector s[j −
1, ·] and all the above highest-score matrix-vector products.
The solution score is given by the value s[m, n]. The solution chain of prescribed substrings can now be obtained by tracing the dynamic programming
sequence backwards from node um to node u0 .
Cost analysis.
First phase. Algorithm 1 runs in time O(m1.25 n).
Second phase. For each of the k = O(m) edges in the representing dag, the
algorithm of Lemma 5 runs in time O(n log n). Therefore, the total cost of this
phase is O(m) · O(n log n) = O(mn log n).
The overall cost of the algorithm is dominated by the cost of the first phase,
equal to O(m1.25 n).
t
u
In the case of general k, the analysis of the previous section can be applied to
obtain a smooth transition between the sparse and dense versions of the problem.
By a constant-factor blow-up of the alignment dag, our algorithms can be
extended from the LCS score to the more general edit score, where the insertion,
deletion and substitution costs are any constant rationals.
6
Conclusions
We have presented an improved algorithm for sparse spliced alignment, running
in time O(n2.25 ), and providing a smooth transition in the running time to
the dense case. A natural question is whether this running time can be further
improved.
Our algorithm is based on the previously developed framework of semi-local
string comparison by implicit highest-score matrix multiplication. The method
compares strings locally by the LCS score, or, more generally, by an edit score
where the insertion, deletion and substitution costs are any constant rationals. It
remains an open question whether this framework can be extended to arbitrary
real costs, or to sequence alignment with non-linear gap penalties.
7
Acknowledgement
The problem was introduced to the author by Michal Ziv-Ukelson.
References
1. C. E. R. Alves, E. N. Cáceres, and S. W. Song. An all-substrings common subsequence algorithm. Electronic Notes in Discrete Mathematics, 19:133–139, 2005.
2. A. N. Arslan and Ö. Eğecioğlu. Approximation algorithms for local alignment with
length constraints. International Journal of Foundations of Computer Science,
13(5):751–767, 2002.
3. G. Benson. A space efficient algorithm for finding the best nonoverlapping alignment score. Theoretical Computer Science, 145:357–369, 1995.
4. J. L. Bentley. Multidimensional divide-and-conquer. Communications of the ACM,
23(4):214–229, 1980.
5. M. Crochemore, G. M. Landau, B. Schieber, and M. Ziv-Ukelson. Re-use dynamic
programming for sequence alignment: An algorithmic toolkit. In String Algorithmics, volume 2 of Texts in Algorithmics. King’s College Publications, 2004.
6. M. S. Gelfand, A. A. Mironov, and P. A. Pevzner. Gene recognition via spliced
sequence alignment. Proceedings of the National Academy of Sciences of the USA,
93(17):9061–9066, 1996.
7. D. Gusfield. Algorithms on Strings, Trees, and Sequences: Computer Science and
Computational Biology. Cambridge University Press, 1997.
8. S. K. Kannan and E. W. Myers. An algorithm for locating non-overlapping regions
of maximum alignment score. SIAM Journal on Computing, 25(3):648–662, 1996.
9. C. Kent, G. M. Landau, and M. Ziv-Ukelson. On the complexity of sparse exon
assembly. Journal of Computational Biology, 13(5):1013–1027, 2006.
10. G. M. Landau and M. Ziv-Ukelson. On the common substring alignment problem.
Journal of Algorithms, 41(2):338–359, 2001.
11. J. P. Schmidt. All highest scoring paths in weighted grid graphs and their application to finding all approximate repeats in strings. SIAM Journal on Computing,
27(4):972–992, 1998.
12. A. Tiskin. Semi-local longest common subsequences in subquadratic time. Journal
of Discrete Algorithms. To appear, available from http://www.dcs.warwick.ac.
uk/~tiskin/pub.
A
Proof of Lemma 3
Proof (Lemma 3). It is straightforward to check equality (3), by (2) and
Definition 1. Informally, each nonzero of DC appearing in (3) is obtained as a
direct combination of a non-trivial nonzero of DA and a trivial nonzero of DB . All
remaining nonzeros of DA and DB are non-trivial, and determine collectively the
remaining nonzeros of DC . However, this time the direct one-to-one relationship
between nonzeros of DC and pairs of nonzeros of DA and DB need not hold.
Observe that all the nonzeros of DA appearing in (3) with j ∈ h−∞ : 0i
are dominated by each of the remaining nonzeros of DA . Furthermore, none
of the nonzeros of DA appearing in (3) with j ∈ hn : +∞i can be dominated
by any of the remaining nonzeros of DA . Hence, the nonzeros appearing in (3)
cannot affect the computation of the remaining nonzeros of DC . We can therefore
simplify the problem by eliminating all half-integer indices i, j, k that correspond
to nonzero index pairs (i, j) and (j, k) appearing in (3), and then renumbering
the remaining indices i, so that their new range becomes h0 : ni (which is already
the range of j, k after the elimination). More precisely, we define permutation
0
0
0
0
, with indices ranging over h0 : ni, as follows. Matrix DA
, DC
, DB
matrices DA
is obtained from DA by selecting all rows i with a nonzero DA (i, j), j ∈ h0 : ni,
and then selecting all columns that contain a nonzero in at least one (in fact,
0
is obtained from DB by selecting
exactly one) of the selected rows. Matrix DB
0
is obtained from DC
all rows j and columns k, where j, k ∈ h0 : ni. Matrix DC
by selecting all rows i with a nonzero DC (i, k), k ∈ h0 : ni, and then selecting
all columns that contain a nonzero in at least one (in fact, exactly one) of the
selected rows. We define d0A , d0B , d0C accordingly. The index order is preserved
by the above matrix transformation, so the dominance relation is not affected.
Both the matrix transformation and its inverse can be performed in time and
memory O(n).
0
0
,
, DB
It is easy to check that d0A d0B = d0C , iff dA dB = dC . Matrices DA
0
DC satisfy the conditions of Lemma 1. Therefore, given the set of nonzero index
0
0
0
pairs of DA
, DB
, the set of nonzero index pairs of DC
can be computed in time
1.5
O(n ) and memory O(n).
t
u
B
Proof of Lemma 4
Proof (Lemma 4). By Lemma 3, all but n non-trivial nonzeros of DC can
be obtained in time and memory O(M + n). We
now show how to obtain the
remaining non-trivial nonzeros in time O m0.5 n , instead of time O(n1.5 ) given
by Lemma 3.
The main idea is to decompose matrix dB into a (min, +)-product of permutationdistribution matrices with small core. The decomposition is described in terms of
density matrices, and proceeds recursively. In each recursive step, we define in0
00
finite permutation matrices DB
, DB
, that are obtained from the density matrix
DB as follows.
Recall that non-trivial nonzeros in DB belong to the index pair range h−m :
ni × h0 : m + ni. Intuitively, the idea is to split the range of each index into two
blocks:
h−m : ni = −m :
h0 : m + ni = 0 :
n
2
n
2
∪
∪
n
2 :n
n
2 :m
+n
Note that the splits are not uniform, and that among the resulting four index
pair blocks in DB , the block n2 : n × 0 : n2 cannot contain any nonzeros. We
process the remaining three index pair blocks individually, gradually introducing
0
00
nonzeros in matrices DB
, DB
until they become permutation matrices. Non0
00
trivial nonzeros in DB , DB will belong respectively to the index ranges −m :
n
n
n
n
2 × 0 : m + 2 and 2 : m + n × m + 2 : 2m + n .
First, we consider all nonzeros in DB with indices (j, k) ∈ −m : n2 × 0 : n2 .
0
at index pair (j, k). We
For every such nonzero, we introduce a nonzero in DB
also consider all nonzeros in DB with indices (j, k) ∈ n2 : n × n2 : m + n . For
00
at index pair (m + j, m + k).
every such nonzero, we introduce a nonzero in DB
Now consider all nonzeros in DB with indices (j, k) ∈ −m : n2 × n2 : m+n .
There are exactly m such nonzeros. Denote their index pairs by (j0 , k0 ), (j1 , k1 ),
. . . , (jm−1 , km−1 ), where j0 < j1 < · · · < jm−1 . For each nonzero with
index
n+1
0
at
index
pair
j
,
pair (jt , kt ), we introduce a nonzero in DB
+
t
, and a
t
2
00
nonzero in DB
at index pair n+1
+
t,
m
+
k
.
t
2
0
00
Finally, we introduce the trivial nonzeros in DB
, DB
at index pairs (j, k),
k − j = m, outside the above non-trivial ranges. The recursive step is completed.
00
0
. Let d∗B = d0B d00B ,
, DB
Let d0B , d00B be the distribution matrices of DB
∗
∗
00
0
∗
∗
and dC = dA dB = dA dB dB , and define DB , DC accordingly. By the
∗
construction of the decomposition of dB , matrices DB and DB
(as well as dB
∗
and dB ) are related by a simple shift: for all (i, k), i, k ∈ h−∞, +∞i, we have
∗
∗
are related by a
(j, k + m). Consequently, matrices DC and DC
DB (j, k) = DB
∗
(i, k + m).
similar shift: for all (i, k), i, k ∈ h−∞, +∞i, we have DC (i, k) = DC
The described decomposition process continues recursively, as long as n ≥ m.
The problem of computing matrix dC is thus reduced, up to an index shift, to
n/m instances of multiplying permutation-distribution matrices. In every instance, one of the multiplied matrices has core of size O(m). By Lemma 3, the
non-trivial part of every such multiplication can be performed in time O(m1.5 )
and memory O(m). The trivial parts of all these multiplications can be combined
into a single scan of the nonzero sets of DA , DB , and can therefore be performed
in time and memory O(M + n).
Hence, the whole computation can be performed
in time O M + (n/m) · m1.5 = O(M + m0.5 n) and memory O(M + n).
t
u
The decomposition of matrix DB in the proof of Lemma 4 is illustrated by
Figures 4, 5. The rectangle corresponding to DB is split into two half-sized rect0
00
angles, corresponding to DB
and DB
. Each of the new rectangles is completed
to a full-sized rectangle by trivial extension; then, the rectangles are arranged
vertically with a shift by m. The “seaweed” curves that do not cross the partition are preserved by the construction, up to a shift by m. The “seaweed”
−m
↓
0
↓
n
2
n
↓
↑
↑
n
↓
DB
↑
0
n
2
↑
m+n
Fig. 4. Proof of Lemma 4: the original matrix DB
−m
↓
n
2
0
↓
n
↓
↓
0
DB
00
DB
↑
m
↑
m+ n
2
↑
m+n
↑
2m+n
Fig. 5. Proof of Lemma 4: the decomposition of DB
curves that cross the partition are also preserved up to a shift by m, by passing
them through a parallelogram-shaped “buffer zone”. Note that this construction
0
makes the latter class of curves uncrossed in DB
, and preserves all their original
00
crossings in DB .
| 5cs.CE
|
Object-oriented programming:
some history, and challenges for the next fifty years
Andrew P. Black
arXiv:1303.0427v1 [cs.PL] 2 Mar 2013
Portland State University, Portland, Oregon, USA
Abstract
Object-oriented programming is inextricably linked to the pioneering work of
Ole-Johan Dahl and Kristen Nygaard on the design of the Simula language,
which started at the Norwegian Computing Centre in the Spring of 1961.
However, object-orientation, as we think of it today — fifty years later — is
the result of a complex interplay of ideas, constraints and people. Dahl and
Nygaard would certainly recognise it as their progeny, but might also be
amazed at how much it has grown up.
This article is based on a lecture given on 22nd August 2011, on the
occasion of the scientific opening of the Ole-Johan Dahl hus at the University
of Oslo. It looks at the foundational ideas from Simula that stand behind
object-orientation, how those ideas have evolved to become the dominant
programming paradigm, and what they have to offer as we approach the
challenges of the next fifty years of informatics.
Keywords: Ole-Johan Dahl, class, object, prefix, Simula, process,
cost-model,
2010 MSC: 01A60, 6803, 68N19, 68N19
Email address: [email protected] (Andrew P. Black)
URL: http://www.cs.pdx.edu/~black (Andrew P. Black)
Preprint submitted to Information and Computation
March 5, 2013
1. Introduction
On 22nd August 2011, a public event was scheduled to open both the 18th
International Symposium on Fundamentals of Computation Theory and the
Ole-Johan Dahl hus [1], the new building that is home to the University of
Oslo’s Department of Informatics, and which is shown in Figure 1. The morning session opened with an Introduction by Morten Dæhlen, followed by two
invited talks, one by myself, Andrew Black, and one by Jose Meseguer. These
talks were followed by a panel discussion on the future of object-orientation
and programming languages, chaired by Arne Maus, and comprising Andrew
Black, Yuri Gurevich, Eric Jul, Jose Meseguer, and Olaf Owe.
As it happened, none of
these events took place in
the Dahl hus, because the
beautiful lecture room that
had been scheduled for the
conference was put out of
commission, less than half
an hour before the start of
the session, by an electrical
fault: the scientific opening
of the Dahl hus was actually
conducted in the neighbouring Kristen Nygaard building. Thus, nine years after
their deaths, Dahl and Ny- Figure 1: The Ole-Johan Dahl hus. (Photograph c
gaard were still able to form the author)
a symbolic partnership to solve a pressing problem.
This article is based on the invited talk that I delivered at this event. It
is not a transcript; I have taken the opportunity to elaborate on some themes
and to précises others, to add references, and to tidy up some arguments that
seemed, in hindsight, a bit too ragged to set down in print.
2. The Birth of Simula
In American usage, the word “drafted” has many related meanings. It can
mean that you have been conscripted into military service, and it can mean
2
that you have been given a job that is necessary, but that no one else wants
to take on.
In 1948, Kristen Nygaard was drafted, in both of these senses. He
started his conscript service at the Norwegian Defence Research Establishment, where his assignment was to carry out calculations related to the construction of Norway’s first nuclear reactor [2]. Years later, Nygaard recalled
that he had no wish to be responsible for the first nuclear accident on the
continent of Europe1 .
After extensive work on a traditional numerical approach, Nygaard turned
to Monte Carlo simulation methods. He was made head of the “computing
office” at the Defence Establishment, and in 1952 turned full-time to operational research. He earned a Master of Science degree from the University
of Oslo in 1956, with a thesis on probability theory entitled “Theoretical Aspects of Monte Carlo Methods” [3]. In 1960, Nygaard moved to the Norwegian Computing Centre (NCC), a semi-governmental research institute that
had been established in 1958. His brief was to expand the NCC’s research
capabilities in computer science and operational research. He wrote “Many
of the civilian tasks turned out to present the same kind of methodological
problems [as his earlier military work]: the necessity of using simulation, the
need of concepts and a language for system description, lack of tools for generating simulation programs” [2]. In 1961, he started designing a simulation
language as a way of attacking those problems.
In January 1962, Nygaard wrote what has become a famous letter. It
was addressed to the French operational research specialist Charles Salzmann. Nygaard wrote: “The status of the Simulation Language (Monte
Carlo Compiler) is that I have rather clear ideas on how to describe queueing systems, and have developed concepts which I feel allow a reasonably
easy description of large classes of situations. I believe that these results
have some interest even isolated from the compiler, since the presently used
ways of describing such systems are not very satisfactory. . . . The work on
the compiler could not start before the language was fairly well developed,
but this stage seems now to have been reached. The expert programmer who
is interested in this part of the job will meet me tomorrow. He has been
rather optimistic during our previous meetings.”
1
David Ungar, private communication.
3
The “expert programmer” was of course Ole-Johan Dahl, shown in Figure 2, and now widely recognised as Norway’s foremost computer scientist.
Along with Nygaard, Dahl produced the initial ideas for object-oriented programming, which is now the dominant style of programming for commercial
and industrial applications. Dahl was made Commander of the Order of
Saint Olav by the King of Norway in 2000, and in 2001 Dahl and Nygaard
received the ACM Turing Award “for ideas fundamental to the emergence of
object-oriented programming, through their design of the programming languages Simula I and Simula 67.” In 2002, Dahl and Nygaard were awarded
the IEEE von Neumann medal. Dahl died on 29th June 2002.
There has been some confusion about the naming of the various Simula
languages. The first version of Simula, whose preliminary design was presented in May 1963 and whose compiler was completed in January 1965, was
a process description and simulation language, and was at the time named
just Simula. This language had neither classes nor inheritance. In their paper presented at the first History of Programming Languages Conference [2],
Nygaard and Dahl refer to this language as Simula I, and I shall do the
same here. Combining their experience with this language with some important new ideas, Dahl and Nygaard went on to design a general purpose
language, which was finished in 1967. This language was at the time named
Simula 67, although there is now a Simula ’87 standard, and this language
is now usually referred to simply as Simula; it was Simula 67 that explored
the ideas that were to become “fundamental to the emergence of objectoriented programming.” I will use the name Simula when referring to the
common thread of ideas that pervade both languages, and the forms with
post-scripted numerals when referring to a specific language.
In this article, I will try to identify the core concepts embodied in Dahl
and Nygaard’s early languages, and see how these concepts have evolved in
the fifty years that have passed since their invention. I will also hazard some
guesses as to how they will adapt to the future.
3. History in Context
In seeking guidance for the perilous process of predicting the future, I looked
at another event that also took place in 1961: the Centennial Celebration of
the Massachusetts Institute of Technology. Richard Feynman joined Sir John
Cockcroft (known for splitting the atom), Rudolf Peierls (who first conceptu4
alised a bomb based on U-235) and Chen Ning Yang (who received the 1957
Nobel Prize for his work on parity laws for elementary particles) to speak on
“The Future in the Physical Sciences.” While the other speakers contented
themselves with predicting 10, or perhaps 25, years ahead, Feynman, who
was not to receive his own Nobel Prize for another four years, decided to be
really safe by predicting 1000 years ahead. He said “I do not think that you
can read history without wondering what is the future of your own field, in a
wider sense. I do not think that you can predict the future of physics alone
[without] the context of the political and social world in which it lies.” Feynman argued that in 1000 years the discovery of fundamental physical laws
would have ended, but his argument was based on the social and political
context in which physics must operate, rather than on any intrinsic property
of physics itself [4].
If social and political context is ultimately what determines the future of science, we must start our exploration of the
Simula languages by looking at the context that gave them birth. Nygaard’s concern with modelling the phenomena associated with nuclear fission meant that Simula was designed as a process description
language as well as a programming language. “When Simula I was put to practical work it turned out that to a large extent it was used as a system description
language. A common attitude among its
simulation users seemed to be: sometimes
actual simulation runs on the computer provided useful information. The writing of the
Simula program was almost always useful,
Figure 2: Ole-Johan Dahl (photosince . . . it resulted in a better understand- graph courtesy of Department of Ining of the system”[2]. Notice that modelling formatics, University of Oslo).
means that the actions and interactions of
the objects created by the program model the actions and interactions of the
real-world objects that they are designed to simulate. It is not the Simula code that models the real-world system, but the objects created by that
5
code2 . The ability to see past the code to the objects that it creates seems
to have been key to the success of Simula’s designers.
Because of Nygaard’s concern with modelling nuclear phenomena, both
of the Simula languages followed Algol 60 in providing what was then called
“security”: the languages were designed to reduce the possibility of programming errors, and so that those errors that remained could be cheaply detected
at run time [5]. A combination of compile-time and run-time checks ensured
that a Simula program could never give rise to machine- or implementationdependent effects, so the behaviour of a program could be explained entirely
in terms of the semantics of the programming language in which it was written [2]. We would call this property “type-safety”.
This was quite different from assembly languages, and even some contemporary high-level languages such as NPL [6], where references were untyped,
and the consequences of mistaking a reference to a string for a reference to a
floating-point number could not be explained in terms of the language itself,
but could be understood only once one knew the representations of strings
and floating-point numbers chosen by the language implementer.
Simula addressed “security” by a combination of static and dynamic
checks. References were qualified to refer to objects of a particular class, so
the attributes of those objects — their variables and methods — were known
to the compiler. These attributes could be accessed by using a connection
statement (more commonly called an inspect statement), and in Simula 67
by means of the now-ubiquitous “dot” notation. If an attribute were present
in some subclasses of the statically-known class, but not in others, then
accessing it required a dynamic check, which could be requested by means
of one or more when clauses in the inspect statement [7]. This form of the
inspect statement, containing multiple when clauses, is thus very similar to a
type-case expression in a language like Haskell.
Dahl’s view of the contributions of Simula 67, as he explained in his rôle
as discussant at the first HOPL conference [8], was the generalisation and
liberalisation of the Algol 60 block to create the class. This gave Simula 67:
2
More precisely, the classes of a Simula 67 program abstract from its objects, and those
objects abstract from the real world. So the program can be regarded as a meta-model of
the real world. I return to this topic in Section 10.
6
1. record structures (class blocks with variable declarations but no statements);
2. procedural data abstraction (class blocks with variable and procedure
declarations, but liberated from the stack discipline so that instances
of a class could outlast their callers);
3. processes (class blocks that continue to execute after their callers have
resumed execution);
4. incremental abstraction, through prefixing one class block with the
name of another class block; and
5. modules, obtained by prefixing an ordinary, Algol-60–style, in-line block
with the name of a class block; this let the prefix block play the rôle of
what Dahl called a context object.
Nowadays, item 4 is called inheritance, or subclassing; I will use the terms
interchangeably. It remains one of the signature features of object-oriented
programming; I’ll return to inheritance in Section 5. In contrast, item 5 —
using classes as a packaging or modularity construct — has been forgotten
by mainstream languages. Dahl wrote: “This [item 5] turned out to be a
very interesting application, rather like collecting together sets of interrelated
concepts into a larger class, and then using that larger class as a prefix to a
block. The class would function as a kind of predefined context that would
enable the programmer to program in a certain style directed toward a certain
application area. Simulation, of course, was one of the ideas that we thought
of, and we included a standard class in the language for that particular
purpose” [8]. Bracha’s Newspeak language has revived this idea [9], and the
Grace language [10] does something very similar by using objects as modules.
This list of what could be done with a Simula class might lead one to
think that the real contribution of Simula 67 was the class construct, and
that classes, rather than the programming style that they enabled, were the
real contribution of Simula 67. However, I don’t believe that this is correct,
either technically or historically. Dahl himself wrote “I know that Simula
has been criticized for perhaps having put too many things into that single
basket of class. Maybe that is correct; I’m not sure myself. . . . It was great
fun to see how easily the block concept could be remodeled and used for all
these purposes. It is quite possible, however, that it would have been wiser to
introduce a few more specialized concepts, for instance, a “context” concept
7
for the contextlike classes” [8]. I interpret these remarks as Dahl saying that
the important contributions of Simula 67 were the five individual features,
and not the fact that they were all realised by a single piece of syntax. Using
identical syntax for features that users perceive as different may be a mistake.
4. The Origin of Simula’s Core Ideas
Dahl was inspired to create the Simula 67 class by visualising the runtime
representation of an Algol 60 program. To those with sufficient vision, objects
were already in existence inside every executing Algol program — they just
needed to be freed from the “stack discipline”. In 1972 Dahl and Hoare wrote
in Hierarchical Program Structures [11]:
In Algol 60, the rules of the language have been carefully designed to ensure that the lifetimes of block instances are nested,
in the sense that those instances that are latest activated are
the first to go out of existence. It is this feature that permits
an Algol 60 implementation to take advantage of a stack as a
method of dynamic storage allocation and relinquishment. But it
has the disadvantage that a program which creates a new block instance can never interact with it as an object which exists and has
attributes, since it has disappeared by the time the calling program regains control. Thus the calling program can observe only
the results of the actions of the procedures it calls. Consequently,
the operational aspects of a block are overemphasised; and algorithms (for example, matrix multiplication) are the only concepts
that can be modelled.
In Simula I, Dahl made two changes to the Algol 60 block: small changes,
but changes with far-reaching consequences. First, “a block instance is permitted to outlive its calling statement, and to remain in existence for as long
as the program needs to refer to it” [11]. Second, references to those block
instances are treated as data, which gives the program a way to refer to them
as independent objects. As a consequence of these changes, a more general
storage allocation mechanism than the stack is needed: a garbage collector
is required to reclaim those areas of storage occupied by objects that can no
longer be referenced by the running program. Dahl and Nygaard may not
have been the first to see the possibilities of generalising the Algol block,
8
but they were the first to realise that the extra complexity of a garbage collector was a small price to pay for the wide range of concepts that could
be expressed using blocks that outlive their calling creators. It was Dahl’s
creation of a storage management package “based on a two-dimensional free
area list” that made possible the Simula process [2]. This package seems to
have been completed by May 1963, the date of a preliminary presentation of
the Simula I Language. The Simula I compiler was completed in January
1965; a year later it was accepted by Univac, who had partially financed its
development — see Figure 3.
Simula 67’s class concept
was developed as a solution
to a number of shortcomings of Simula I, but particularly the cumbersome nature of the inspect mechanism for accessing the local
variables of a process, and
the need to share properties
between different simulation
models. Hoare had previously introduced the notion
of a class of record instances;
one of Hoare’s classes could Figure 3: Dahl and Nygaard during the develbe specialised into several opment of Simula, around 1962. The Univac
1107 was obtained by Nygaard for the NCC at
subclasses, but the class and a favourable price in exchange for a contract to
its subclasses had to be de- implement Simula I. (Photograph from the Virclared together, with the sub- tual Exhibition People Behind Informatics, http:
class definitions nested inside //cs-exhibitions.uni-klu.ac.at/)
the class definition [12]. Dahl probably became familiar with Hoare’s proposals when they both lectured (Hoare on Record Handling and Dahl on
Simulation Languages) at the NATO Summer school at Villard-de-Lans in
1966 [13]. Dahl and Nygaard realised that it would be useful if a class could
be declared separately from its subclasses, so that it could be specialised
separately for diverse purposes. However, Hoare’s nesting of subclass declarations within class declarations precluded such a separation. “The solution
came with the idea of class prefixing: using C as a prefix to another class,
the latter would be taken to be a subclass of C inheriting all properties of
9
C” [14]. The semantics of prefixing was defined as concatenation, but the
syntactic separation of class and subclass turned the class of Simula 67 into
a unit of reuse. This was so successful that the special-purpose simulation
facilities of Simula I were replaced in Simula 67 by a simulation class,
which was defined in Simula 67 itself. To make use of simulation facilities in
a class, all that a programmer then needed to do was to prefix the definition
of their class by the name of the class simulation.
An important part of prefixing was the notion of virtual, although Dahl
writes that it was “a last minute extension” [14]. In Simula 67, the binding
of attribute identifiers is static. This means that the syntax of the reference,
rather than the identity of the target object, determines which variable to
reference or which procedure to call. The binding can be made dynamic
by marking the attribute as virtual; this means that the choice of variable
or procedure depends only on the target object. Making static binding the
default was, I think, a concession to efficiency; with the benefit of hindsight,
it seems clear to me that it was a mistake. Smalltalk made all methods
virtual, and raised dynamic binding to the status of a fundamental part of
object-orientation. Even Java made dynamic binding the default, although in
Java the final keyword can be used to specify static binding and to prohibit
overriding.
5. Why Inheritance Matters
Inheritance, combined with dynamic binding, has become the sine qua non
of object-orientation; indeed, an influential 1987 paper by Wegner defined
“object-oriented” as “objects + classes + inheritance” [15]. Is this inevitable?
Can one have a viable object-oriented programming language without inheritance? I was one who objected to Wegner’s definition at the time: it seemed
more reasonable for the term “object-oriented” to imply only the use of objects. We had just published the first papers on Emerald [16, 17], a language
that we certainly thought of as object-oriented, even though it had no inheritance, so perhaps we were not unbiased observers. Wegner’s definition
excluded not only Emerald, but also any delegation-based language from
being object-oriented, even though such languages were of growing interest.
Wegner did admit that delegation, which he defined as “a mechanism that
allows objects to delegate responsibility for performing an operation or finding a value to one or more designated ‘ancestors’,” was actually the purer
feature.
10
Inheritance has stood the test of time as a useful feature, and I now
believe that some code-reuse mechanism like inheritance or delegation is a
very important part of any object-oriented programming system. Since 1989,
thanks to Cook, we have known that inheritance can be explained using fixpoints of generators of higher-order functions [18]. What this means is that,
in theory, functions parametrized by functions are “as good as” inheritance,
in the sense that they can enable equivalent reuse. Nevertheless, in practice
they are not as good, because to enable reuse, the programmer has to plan
ahead and make every part that could possibly change an explicit parameter.
Functional programmers call this process of explicit parametrization “abstraction”. It has two problems: life is uncertain, and most people think
better about the concrete than the abstract. Let’s examine these problems
briefly. We have all experienced trying to plan ahead for change. Inevitably,
when change comes, it is not the change we anticipated. We find that we have
paid the price of generality and forward planning, but still have to modify
the software to adapt to the change in requirements. Agile methodologies use
this: they tell us to eschew generality, and instead make the software ready
to embrace change by being changeable. That is, they tell us that rather
than trying to write software that is general enough to cope with all possible situations without need of modification, we should write software that
is specific to the current requirements, but easily modifiable to other possible requirements. Inheritance aids in this process by allowing us to write
software in terms of the concrete, specific actions that the application needs
now, and to abstract over them only when we know — because the change
request is in our hands — that the abstraction is needed.
Let’s explain how this works using an example. Suppose that in the original program the concrete actions are represented as methods in a class.3
An example might be a Text class with a method emphasise that italicises a
segment of the text. If the requirements change so that the visual appearance of emphasis should controlled by a preference, then a new class can be
created that inherits from Text and overrides the emphasise method with a
3
The methods of an object are the set of named operations that it can perform. They
execute in response to messages from itself or another object; each message contains the
name of a method, and the necessary arguments. A class can be thought of as a template
that describes the structure of a group of similarly-structured objects; in particular, it
defines their methods.
11
new method; this new method would use a preference object to determine
the visual appearance of the emphasised text. In doing so, the emphasise
method, previously a constant, has become a parameter. Notice that we
didn’t have to anticipate that this particular change would be requested, and
the original Text class didn’t have to make every possible locus of change a
parameter. Even if the process of emphasising text had not been broken out
into a separate method in the original code, it could have been extracted into
an emphasise method using a simple, behaviour-preserving refactoring before
any change to the functionality of the program was attempted.
As with so much of object-oriented programming, Dahl foresaw the value
of turning methods — which he called virtual procedures — into parameters.
Indeed, virtual procedures were originally conceived, in part, as a replacement for what he felt to be the excessive power of Algol 60’s call-by-name
parameters. In the discussion at the Lysebu conference following the presentation of the paper on Classes and Subclasses, Dahl is reported to have
remarked [7] “virtual quantities are in many ways similar to Algol’s name
parameters, but not quite as powerful. It turns out that there is no analogy
to Jensen’s device. This, I feel, is a good thing, because I hate to implement
Jensen’s device. It is awful.”
The second problem with abstraction derives from the way that we think.
Most of us can grasp new ideas most easily if we are first introduced to one
or two concrete instances, and are then shown the generalisation. To put
this another way: people learn best from examples. So we might first solve a
problem for n = 4, and then make the changes necessary for 4 to approach infinity. There are people who are able to first grasp abstractions directly, and
then apply them: such people tend to become mathematicians, and sometimes programming language designers. Languages designed by and for such
people tend to provide good support for creating abstractions. Languages for
the rest of us should support inheritance, which lets the programmer provide
a concrete instance first, and then generalise from it later.
To illustrate the power of inheritance to make complex abstractions easier
to understand, let’s look at a case study from the functional programming
literature. In Programming Erlang [19, Chapter 16], Armstrong introduces
the OTP (Open Telecom Platform) generic server. He comments: “This is
the most important section of the entire book, so read it once, read it twice,
read it 100 times — just make sure the message sinks in”. To make sure
that the message about the way that the OTP server works does sink in,
12
Armstrong presents us with
four little servers . . . each slightly different from the last. server1
runs some supplied code in a server, where it responds to remote
requests; server2 makes each remote request an atomic transaction; server3 adds hot code swapping, and server4 provides both
transactions and hot code swapping.
Each of these “four little servers” is self-contained: server4, for example,
makes no reference to any of the preceding three servers.
Why does Armstrong describe the OTP sever in this way, rather than just
presenting server4, which is his destination? Because he views server4 as too
complicated for the reader to understand in one go. Something as complex as
server4 needs to be introduced step-by-step. However, his language, lacking
inheritance (and higher-order functions) does not provide a way of capturing
this stepwise development.
In an effort to understand the OTP server, I coded it up in Smalltalk.
As George Forsythe is reputed to have said: “People have said you don’t
understand something until you’ve taught it in a class. The truth is you
don’t understand something until you’ve taught it to a computer — until
you’ve been able to program it” [20]. First I translated server1 into Smalltalk;
I called it BasicServer, and it had three methods and 21 lines of code. Then I
needed to test my code, so I wrote a name server plug-in for BasicServer, set
up unit tests, and made them pass. In the process, as Forsyth had predicted,
I gained a much clearer understanding of how Armstrong’s server1 worked.
Thus equipped, I was able to implement TransactionServer by subclassing
BasicServer, and HotSwapServer by subclassing TransactionServer, bringing me
to something that was equivalent to Armstrong’s server4 in two steps, each
of which added just one new concern. Then I refactored the code to increase
the commonality between TransactionServer and BasicServer, a commonality
that had been obscured by following Armstrong in re-writing the whole of
the main server loop to implement transactions. This refactoring added one
method and one line of code.
Once I was done, I discussed what I had learned with Phil Wadler. Wadler
has been thinking deeply, and writing, about functional programming since
the 1980s; amongst other influential articles he has authored “The Essence of
Functional Programming” [21] and “Comprehending Monads” [22]. His first
reaction was that the Erlang version was simpler because it could be described
13
in straight-line code with no need for inheritance. I pointed out that I could
refactor the Smalltalk version to remove the inheritance, simply by copying
down all of the methods from the superclasses into HotSwapServer, but that
doing so would be a bad idea. Why? Because the series of three classes, each
building on its superclass, explained how HotSwapServer worked in much the
same way that Armstrong explained it in Chapter 16 of Programming Erlang.
This was an “ah-ha moment” for Phil.
To summarise: most people understand complex ideas incrementally, by
starting with a simple concrete example, and then taking a series of generalisation steps. A program that uses inheritance can explain complex behaviour
incrementally, by starting with a simple class or object, and then generalising it in a series of inheritance steps. The power of inheritance is that it
enables us to organise our programs incrementally, that is, in a fashion that
corresponds to the way that most people think.
6. Object-Oriented Frameworks
In my view, one of the most significant contributions of Simula 67 was the
idea of the object-oriented framework. By the 1960s, subroutine libraries
were well-known. A subroutine, as its name suggests, is always subordinate
to the main program. The relationship is always “don’t call us, we’ll call
you”. The only exception to this is when a user-written subroutine p is
passed as an argument to a library subroutine q; in this case, q can indeed
call p. However, in the languages of the time, any such “callback” to p could
occur only during the lifetime of the original call from the main program
to q. An object-oriented framework allows subroutines, but also enables the
reverse relationship: the user’s program can provide one or more components
that are invoked by the framework, and the framework takes on the rôle of
“main program”.
This reversal of rôles enabled the special-purpose simulation constructs
of Simula I to be expressed as a framework within Simula 67. The programmer of a simulation wrote code that described the behaviour of the
individual objects in the system: the nuclear fuel rods, or the air masses, or
the products and customers, depending on the domain being simulated. The
simulation framework then called those user-defined objects. The framework
was in control, but users were able to populate it with objects that achieved
their diverse goals.
14
Of course, for this scheme to work, the user-provided objects called by
the framework must have methods for all of the requests made of them.
Providing these methods from scratch could be a lot of work, but the work
can be made much more manageable by having the user-provided objects
inherit from a framework object that already has methods for the majority
of these requests; all the user need do is override the inherited behaviour
selectively.
One of the most common instantiations of this idea is the user-interface
framework. In most object-oriented languages, objects that are to appear on
the computer’s screen are sent requests by the display framework. They must
respond to enquires such as those that ask for their bounds on the screen,
and to draw themselves within these bounds. This is normally achieved
by having displayable objects subclass a framework-provided class, such as
Squeak’s Morph or java.awt.Component.
Dahl and Nygaard realised that these ideas could be used to provide
the simulation features of Simula I. Simula 67 is a general-purpose language; its simulation features are provided by an object-oriented framework
called simulation. simulation is itself implemented using a more primitive
framework called simset.
7. From Simula to Smalltalk
Smalltalk-72 was an early version of Smalltalk, designed by Alan Kay and
used only within Xerox PARC. It is described in Kay’s paper on the Early
History of Smalltalk [23]. Kay was inspired by the simplicity of LISP and
the classes and objects of Simula 67, and took from Simula the ideas of
classes, objects and object references.
Smalltalk-72 refined and explored the idea of objects as little computers,
or, as Kay puts it: “a recursion on the notion of computer itself”. Objects
combined data with the operations on that data, in the same way that a
computer combines a memory to store data with an arithmetic and logic
unit to operate on the data.
Smalltalk-72 did not include inheritance, not because it was thought to be
unimportant, but because Simula’s single, static inheritance seemed too limiting; Kay believed that inheritance could be simulated within Smalltalk using its LISP-like flexibility. Smalltalk-72 also dropped some other ideas from
Simula, notably the idea that objects were also processes, and that classes,
15
used as prefixes to inline blocks, were also modules. Smalltalk-76 included
a Simula-like inheritance scheme that allowed superclasses to be changed incrementally; this was kept in Smalltalk-80. However, neither Smalltalk-76
nor Smalltalk-80 supported objects as processes or classes as modules; as a
result, what we may call the “North-American School” of object-orientation
views these features as relatively unimportant, and perhaps not really part
of what it means to support objects.
By the late 1980s, objects had attracted significant industrial interest.
Many different object-oriented languages had been developed, differing in the
weight that they gave to classes, objects, inheritance, and other features. In
1991, Alan Snyder of Hewlett Packard wrote an influential survey paper “The
Essence of Objects” [24], which was later published in IEEE Software [25].
Unlike Simula and Smalltalk, Snyder’s is a descriptive work, not a prescriptive one. He describes and classifies the features of contemporary objectoriented languages, including Smalltalk, C++, the Common Lisp Object System, and the Xerox Star, as well as a handful of other systems. In Snyder’s
view, the essential concepts were as follows:
• An object embodies an abstraction.
• Objects provide services.
• Clients issue requests for those services.
• Objects are encapsulated.
• Requests identify operations.
• Requests can identify objects.
• New Objects can be created.
• The same operation on distinct objects can have different implementations and observably different behaviour.
• Objects can be classified in terms of their services (interface hierarchy).
• Objects can share implementations.
– Objects can share a common implementation (multiple instances).
– Objects can share partial implementations (implementation inheritance or delegation).
16
Table 1: Evolution of the features of an “object”. Simula’s idea of procedural encapsulation became more complete, both over time and as objects migrated across the Atlantic.
However, other features of the Simula Class, such as Active Objects and Classes as Modules, have not persisted in mainstream object-oriented languages.
Feature
Simula 67
Smalltalk-80
Snyder (1991)
Procedural
Abstraction
Attributes
exposed
Attributes
encapsulated
Objects characterised
by services
Active Objects
Yes
No
“Associated Concept”
Dynamic object
creation
Yes
Yes
Yes
Classes
Yes
Yes
Shared
implementations
Inheritance
Class prefixing
Subclassing
“Shared partial
implementations”
Classes as
Modules
Yes
No
No
We see that objects as processes and classes as modules are not on this
list. Snyder does mention “Active Objects” as an “associated concept”, that
is, an idea “associated with the notion of objects, but not essential to it”.
The idea that classes can serve as modules does not appear at all.
The evolution of the features that make up object-orientation is summarised in Table 1. Encapsulation has become more important, and the
mechanisms for sharing implementation have become more diverse and refined. The idea that multiple objects can share an implementation, and that
new kinds of objects can be defined that share parts of existing objects, have
become recognised as important, although both kinds of sharing can be implemented either using Simula-like classes, or through other mechanisms
such as delegation. In contrast, active objects, and classes as modules, are
no longer regarded as key to object-orientation.
8. Objects as Abstractions
The word “abstraction” is used in Informatics in many different senses. As I
mentioned previously, it is used by functional programmers to mean “param17
eterisation”. In the context of objects, I am going to use the word abstraction
to capture the idea that what matters about an object is its protocol: the set
of messages that it understands, and the way that it behaves in response to
those messages. Nowadays, this is sometimes also referred to as the object’s
interface. The key idea is that when we use an object, we focus on how it
appears from the outside, and “abstract away from” its internal structure:
more simply, that the internal structure of an object is hidden from all other
objects. That’s why I said “the set of messages that it understands”, and
not “the set of methods that it implements”. In many languages they are
the same, but I wanted to emphasise the external rather than the internal
view.
Abstraction, in this sense, is one of the key ideas behind objects, but
it does not appear explicitly until Snyder’s 1991 survey. Simula doesn’t
mention abstraction specifically; it speaks instead of modelling, which captures the importance of interacting with the object, but not the information
hiding aspect. Dahl remarks [14] “According to a comment in [the Simula user’s manual, 1965] it was a pity that the variable attributes of a Car
process could not be hidden away in a subblock”, but this was not possible
in Simula without also hiding the procedures that defined the Car’s behaviour — and exposing those procedures was the whole reason for defining
the Car. (Around 1972, Simula addressed this problem by adding the protected and hidden keywords to control the visibility of a name declared in a
class [26].)
Smalltalk-80 approached abstraction by protecting the instance variables
that contain the data representing an object: these variables are accessible
only by the object itself. The Smalltalk-80 programming environment also
provided a way for the programmer to say that a method should be private,
but that was merely a convention, and not enforced by the system. Nevertheless, Dan Ingalls’ sweeping 1981 Byte article “Design Principles behind
Smalltalk” [27] refers to abstraction only indirectly. Ingalls singles out Classification, not abstraction, as the key idea embodied in the Smalltalk class.
Ingalls writes:
Classification is the objectification of nessness. In other words,
when a human sees a chair, the experience is taken both literally
as “that very thing” and abstractly as “that chair-like thing”.
Such abstraction results from the marvellous ability of the mind
to merge “similar” experience, and this abstraction manifests it18
self as another object in the mind, the Platonic chair or chairness.
The fact remains that Smalltalk classes do not classify objects on the basis of
their behaviour, but according to their representation. It seems that the idea
of separating the internal (concrete) and external (abstract) view of an object
was yet to mature, although it can be glimpsed elsewhere in Ingalls’ article.
After Classification, his next principle is “Polymorphism: A program should
specify only the behaviour of objects [that it uses], not their representation”.
Also, in the “Future Work” section, Ingalls remarks that
message protocols have not been formalised. The organisation
provides for protocols, but it is currently only a matter of style
for protocols to be consistent from one class to another. This
can be remedied easily by providing proper protocol objects that
can be consistently shared. This will then allow formal typing of
variables by protocol without losing the advantages of polymorphism.
Borning and Ingalls did indeed attempt just that [28], but it proved not to
be as easy to get the details right as Ingalls had predicted. Still, Smalltalk as
defined in 1980 did succeed in abstracting away from representation details:
it never assumes that two objects that exhibit the same protocol must also
have the same representation.
The related concept of data abstraction can be traced back to the 1970s.
Landmarks in its evolution are of course Hoare’s “Proof of Correctness of
Data Representations” [29], Parnas’ “On the Criteria to be used in Decomposing Systems into Modules” [30], both published in 1972, and the CLU
programming language, which provided explicit conversion between the concrete data representation seen by the implementation and the abstract view
seen by the client [31]. However, none of these papers pursues the idea that
multiple representations of a single abstraction could co-exist. They assume
that abstraction is either a personal discipline backed by training the programmer, or a linguistic discipline backed by types. The latter view — that
abstraction comes from types — has been heavily marketed by the proponents
of ML and Haskell. It is based on the idea that types capture representation
as well as interface, but that by hiding the actual type from code that lies
outside the abstraction boundary, the programmer can be prohibited from
taking advantage of the representation. The mechanism used for this hiding
is the notion of an existentially quantified type.
19
Objects follow a different route to abstraction, which can be traced back
to Alonzo Church’s work on representing numbers in the lambda-calculus [32].
Church represented the number 2 in the pure lambda-calculus by the function
that did something twice; the number 3 by the function that did something
three times, and so on. In 1975 John Reynolds named this route “procedural
abstraction” and showed how it could be used to support data abstraction in
a programming language, as well as contrasting it with the approach based
on existential types [33]. Procedural abstraction is characterised by using
a computational mechanism — a representation of a function as executable
code — rather than type discipline or self-discipline — to enforce abstraction
boundaries. Reynolds writes:
Procedural data structures provide a decentralised form of data
abstraction. Each part of the program which creates procedural
data will specify its own form of representation, independently
of the representations used elsewhere for the same kind of data,
and will provide versions of the primitive operations (the components of the procedural data item) suitable for this representation.
There need be no part of the program, corresponding to a type
definition, in which all forms of representation for the same kind
of data are known.
Unfortunately, Reynolds’ paper is not widely known, even though it has
been reprinted in two collections of notable papers [34, 35]; many practitioners do not understand the fundamental distinction between type abstraction
and procedural abstraction. This distinction is important because there is a
trade-off between the costs and benefits of the two approaches. The benefit of
procedural abstraction — the approach used by objects — is that it provides
secure data abstraction without types, and allows multiple implementations
of the same abstraction to co-exist; this supports change, and helps to keep
software soft. The cost is that it is harder to program “binary operations” —
those that work on two or more objects — with maximal efficiency.
William Cook made another attempt to explain this distinction in a 2009
OOPSLA Essay [36]. Cook coined the term “autognostic”, meaning “selfknowing”, for what Reynolds called “decentralisation”. Autognosis means
that an object can have detailed knowledge only of itself: all other objects
are encapsulated. Cook remarks: “The converse is quite useful: any programming model that allows inspection of the representation of more than
20
one abstraction at a time is not object-oriented.”
To see the benefits of procedural abstraction, consider a selection of objects implementing numbers. The objects 1 and 2 might include a representation as a machine integer, and methods +, −, etc., that eventually use
hardware arithmetic. However, the object representing 267 might use a representation comprising three radix 232 digits, and the methods for + and −
would then perform multi-digit arithmetic. This scheme, augmented with
operations to convert machine integers to the multi-digit representation, is
essentially how Smalltalk implements numbers. Notice that new representations of number objects can be added at any time, without having to change
existing code, so long as the new objects provide the necessary methods.
Because the representation of an object can never be observed by any other
object4 , there is no need to protect it with an existential type.
The idea of procedural data abstraction was certainly in Simula I —
indeed, I believe that it was the genesis of Simula I. As Dahl and Hoare
observe in “Hierarchical Program Structures” [11], the key concept is already
present in Algol 60: an Algol 60 block is a single mechanism that can contain
both procedures and data. The limitation of Algol is that blocks are constrained not to outlive their callers. Dahl saw this, and “set blocks free” by
devising a dynamic storage allocation scheme to replace Algol 60’s stack [2].
However, procedural data structures were not the only kind of data aggregate supported by Simula 67. As I mentioned in Section 3, Dahl realised
that Simula’s class construct could be used to generate both records (unprotected, or protected by type abstraction) and objects (protected by procedural abstraction), and saw that as a strength of the mechanism; some later
languages, notably C++, also support this dualism. Records were made possible by Simula I’s inspect statement, and later by Simula 67’s more flexible
“dot” notation; these mechanisms were introduced specifically to expose the
variables that would otherwise have been local to an object. In contrast,
Smalltalk made the instance variables of an object inaccessible to any other
object. If a Smalltalk object wanted to expose an instance variable, it could
provide methods that permitted the reading and writing of the variable, but
the client would nevertheless see only methods, never variables.
4
To take advantage of the hardware arithmetic unit, that piece of hardware must naturally be allowed to “see” the representation of small integers. Nevertheless, the representation of a object remains hidden from all other objects in the program.
21
The modern view of an object is that only the behaviour of the object
matters: whether that behaviour is ultimately supported by methods or
variables is one of the hidden implementation details of the object. The
text-book illustration of this consists of two Point objects, one encapsulating
a representation using cartesian coordinates, and the other encapsulating
a representation using polar coordinates. Both objects can offer methods
to retrieve, and to change, the x, y, ρ and θ of the points. Some of these
methods will involve direct access to single instance variables, while other will
involve trigonometric calculations and access to multiple instance variables.
Nevertheless, clients of the two objects will be able to treat them identically.
It seems to me that Dahl himself eventually came to believe that procedural abstraction was the fundamental contribution of the Simula languages.
In 2002, Dahl wrote [14]:
The most important new concept of Simula 67 is surely the idea
of data structures with associated operators . . . called objects.
There is an important difference, except in trivial cases, between
• the inside view of an object, understood in terms of local
variables, possibly initialising operations establishing an invariant, and implemented procedures operating on the variables maintaining the invariant, and
• the outside view, as presented by the remotely accessible
procedures, including some generating mechanism, dealing
with more “abstract” entities.
There is one more consequence of the difference between type abstraction
and procedural abstraction. Type abstraction makes a fundamental distinction between the collection of representation variables — the value that is
protected by the existential type — and the functions that operate on it.
Procedural abstraction makes no such distinction: it exposes only methods,
with the consequence that there may be no representation variables at all!
Perhaps the clearest example of this is the implementation of the Booleans
in Smalltalk, which is outlined in Figure 4. The objects true and false implement methods &, |, not, ifTrue:ifFalse:, etc. (The colon in a Smalltalk
method name introduces an argument, in the same way that parentheses are
used to indicate arguments in other languages, and ↑ means “return from
this method with the following answer”.) For example, the & method of true
22
class Boolean
==> aBlock
↑self not or: [aBlock value]
eqv: aBoolean
↑self == aBoolean
class True
class False
& aBoolean
↑aBoolean
& aBoolean
↑self
| aBoolean
↑self
| aBooolean
↑aBoolean
not
not
↑true
↑false
ifTrue: trueBlock ifFalse: falseBlock
↑falseBlock value
ifTrue: trueBlock ifFalse: falseBlock
↑trueBlock value
Figure 4: An extract from the Boolean classes in Smalltalk. Boolean is an abstract class
with no instances; true and false are the only instances of the classes True and False, which
inherit from Boolean. There is no “hidden state”; the methods themselves are sufficient
to define the expected behaviour of the Booleans.
answers its argument, while the | method always answers self, i.e., true. The
ifTrue:ifFalse: methods are a bit more complicated, because the arguments
must be evaluated conditionally. The ifTrue:ifFalse: method of false ignores
its first argument, evaluates its second (by sending it the message value), and
returns the result, while the ifTrue:ifFalse: method of true does the reverse.
9. Active Objects
Active objects seem to be one idea from Simula that has been lost to the
object-oriented community. Activity was an important part of Simula; after all, the original purpose of the language was to simulate activities from
the real world. Simula’s “quasi-parallelism” was a sweet-spot in 1961: it
allowed programmers to think about concurrency while ignoring synchronisation. Because another task could execute only when explicitly resumed,
programmers could be confident that their data would not change “out from
23
under them” at unexpected times.
Hewitt’s Actor model [37] built on this idea, as did Emerald [16], in
which every object could potentially contain a process. Other languages, like
Erlang, have made the process the main focus of the language rather than
the object [38]. However, active objects have disappeared from “mainstream”
object-oriented languages. Since Smalltalk, processes and objects have been
independent concepts.
Why are Smalltalk objects passive? I don’t know. Perhaps Kay and
Ingalls had a philosophical objection to combining what they saw as separate
ideas. Perhaps the realities of programming on the Alto set limits as to
what was possible at that time. Perhaps they wanted real processes, not
coroutines; this requires explicit synchronisation, and thus a significantly
different language design. Smalltalk does indeed have processes, rather than
coroutines. However, these processes are realised as separate objects, rather
than being a part of every object.
10. Classes and Objects
It’s interesting to revisit Wegner’s idea, discussed in Section 5, that objectorientation is characterised by the existence of classes. As we have seen, most
of the concepts that we recognise as making up object-orientation came from
the class concept of Simula. Nevertheless, as both Dahl and Nygaard would
be quick to point out, it’s the dynamic objects, not the classes, that form
the system model, and modelling was the raison d’être of Simula. Classes
were interesting only as a way of creating the dynamic system model, which
comprised interacting objects. Classes are indeed a most convenient tool if
you want hundreds of similar objects. But what if you want just one object?
In that case they impose a conceptual overhead. The fact is that classes are
meta, and meta isn’t always better!
Classes are meta, relative to objects, because a class describes the implementation and behaviour of those objects that we call instances of the
class, as well as providing a factory (procedure or method) that realises that
description by actually making an instance. If you need just one or two
objects, it’s simpler and clearer to describe those objects in the program
directly, rather than to describe a factory (that is, a class) to make them,
and then use it once or twice. This is the idea behind Self, Emerald, NewtonScript and JavaScript, which take the object, rather than the class, as
24
their fundamental concept. Which is not to say that classes are not useful:
just that, at least in their rôle of object factories, they can be represented as
objects.
In class-based languages, classes typically play many rôles. Borning, in
an early paper contrasting class-based and object-based languages [39], lists
eight rôles for Smalltalk classes. The primary rôle of a class is acting as a
factory that creates new objects; some of the other rôles served by Smalltalk
classes include describing the representation of those objects, defining their
message protocol, serving as repositories for the methods that implement
that protocol, providing a taxonomy for classifying objects, and serving as
the unit of inheritance. Nevertheless, as we saw when discussing Ingalls’
view of classification (Section 8 above), the taxonomy provided by classes,
being based on implementation rather than behaviour, is fundamentally at
odds with the autognosis that other researchers, such as Cook, believe to be
fundamental to object-orientation.
If the only thing that we can know about an object is its behaviour, the
most useful way of classifying objects must be according to their behaviour.
This brings us back to types. Recall that types are not needed to provide
abstraction in object-oriented systems. Nor are they needed to provide what
Hoare called “security”: that comes from procedural encapsulation. The true
rôle of types is to provide a behavioural taxonomy: types provide a way of
classifying objects so that programmers can better understand what rôles a
parameter is expected to play, or what can be requested of an object returned
as an answer from some “foreign” code.
Why is such a taxonomy useful? To add redundancy. A type annotation
is an assertion about the value of a variable, no different in concept from
asserting that a collection is empty or that an object is not nil. Redundancy
is in general a good thing: it provides additional information for readers, and
it means that more errors — at least those errors that manifest themselves as
an inconsistency between the type annotation and the code — can be detected
sooner. Nevertheless, I am about to argue that types can be harmful.
On the surface, this is a ridiculous argument. If types add redundancy,
and redundancy is good, how can types be harmful? The problem is not that
adding type annotations will mess up your program, but that adding types
to a language can, unless one is very careful, mess up the language design.
I have been an advocate of types for many years. The Emerald type system [17, 40] was one of the ways in which the Emerald project attempted
25
to improve on Smalltalk. Our original perception was that types would enable us to generate more efficient code. However, as we implemented the
language, we realised that behavioural types, that is, types that constrain
the interface of an object, but not its implementation, did not help us to
generate more efficient code. Emerald was efficient, but that was because
we inferred implementation information, for example, that no reference to
a particular object ever escaped an enclosing object.5 I spent many years
working on type-checking and type systems, but have only recently really
understood the sage advice that I was given by Butler Lampson in the late
1980s: stay away from types — they are just too difficult.
The difficulty springs from two sources. One is Gödel’s second incompleteness theorem, which tells us that there will be true facts about any
(sufficiently rich) formal system that cannot be proved, no matter what system of proof we adopt. In this case, the formal system is our programming
language, and the system of proof is that language’s type system. The other
source is our very natural desire to know that our program won’t go wrong,
or at least that it won’t go wrong in certain circumscribed ways. This has led
most statically-typed languages to adopt an interpretation of type-checking
that I am going to refer to as “The Wilson interpretation”. The name honours Harold Wilson, prime minister of the UK 1964–70 and 1974–76. Wilson
fostered what has been called “the Nanny State”, a social and political apparatus that said the government will look after you: if it is even remotely
possible that something will go wrong, we won’t even let you try. The Wilson interpretation of type-checking embraces two tenets: first, that the type
system must be complete, that is, every type assertion that can be made in
it must be provably true or false, and second, that every part of the program
must be type-checked before it can be executed. The consequence is that if
it is even remotely possible that something may go wrong when executing
your program, then the language implementation won’t even let you try to
run it — assuming that the “something” that might go wrong is a thing over
which the type system believes it has control.
5
In one place we did get efficiency from types, but that was because we cheated. Emerald made it impossible to re-implement integers: a user-defined object would never be
recognised as having type Integer, even though it provided all of the necessary operations.
This was cheating because Emerald types are supposed to constrain interface, not implementation. The consequence was that we could indeed infer representation information
from the Integer type, and directly invoke machine operations to perform arithmetic.
26
The Wilson interpretation is not the only interpretation of type-checking.
At the other extreme is what I will call the “Bush interpretation”, honouring
George W. Bush, president of the USA 2001–2009, who abolished many government regulations and weakened the enforcement of those that remained.
The Bush interpretation of type checking is that type information is advisory. The program should be allowed to do what you, the programmer,
said that it should do. The type-checker won’t stop it; if you mess up, the
PDIC — the Program Debugger and Interactive Checker — will bail you out.
The Bush interpretation amounts to not having static type-checking at all:
the programmer is allowed to write type annotations in the program, but
they won’t be checked until runtime.
There is a third interpretation of type-checking that lies between these
extremes. This might be called the “Proceed with caution” approach. If the
type-checker has been unable to prove that there are no type errors in your
program, you are given a warning. “It may work; it may give you a run-time
error. Good night, and good luck.” I’m going to name this interpretation in
honour of Edward R. Murrow, 1908–1965, an American broadcaster, whose
signature line that was. Like Murrow, I intend that phrase to be taken
positively.
I’m referring to Wilson, Murrow, and Bush as interpretations of typechecking, because I’m taking the view that the semantics of the language,
in the absence of errors, is independent of its type system. In other words,
the language has an untyped semantics, and the rôle of the type system is
to control how and when errors are reported, not to determine the meaning of the code. There is an alternative view in which the meaning of a
program depends on the types of its terms; programs that don’t type-check
have no semantics. In this view, Wilson, Murrow, and Bush define different
languages, connected by a superset relation. Since, as we have seen, types
are not necessary to define object-based abstraction, I feel that the untyped
view is simpler and more appropriate, but the comparison of these views is
a topic worthy of its own essay, and I’m not going to discuss it further here.
So, under all three interpretations, an error-free program has the same
meaning. Under Wilson, we have conventional static typing: an erroneous
program will result in a static error, and won’t be permitted to run. Because of Gödel’s theorem, this means that some error-free programs won’t
be permitted to run either, no matter how complex the type system becomes.
Under Bush, we have conventional dynamic typing: all checks will be
27
performed at runtime — even those that are guaranteed to fail. Enthusiasts
for dynamic typing point out that a counter-example is often more useful
than a type-error message. The disadvantage is that the programmer will
not receive any compile-time warnings. Under the Murrow interpretation, the
programmer will get a mix of compile-time warnings and run-time checks. A
Murrow warning may say that a certain construct is provably a type error,
or it may merely say that the type-checker has been unable to prove that
it is type safe. The difference between these two warnings is dramatic, but
Wilson treats them as if they were the same, conflating a deficiency in the
type system with a programming error.
Let me say here that I’m for Murrow! I believe that the Murrow interpretation of types is the most useful, not only for programmers, but also for
language designers. Wilson’s “Nanny Statism” is an invitation to mess up
your language design. The temptation for the language designer is to exclude
any construct that can’t be statically checked, regardless of how useful it may
be to the practicing programmer.
I believe that Simula was for Murrow too. Recall that, according to
Nygaard, the core ideas of SIMULA were first modelling, and second, security. Modelling meant that the actions and interactions of the objects that
the program created represented the actions and interactions of the corresponding real-world objects. Security meant that the behaviour of a program
could be understood and explained entirely in terms of the semantics of the
programming language in which it is written. Modelling came first! Simula
did not compromise its modelling ability to achieve security; it compromised
its run-time performance, by incorporating explicit checks when a construct
necessary for modelling was not statically provable to be safe.
I feel that many current language designers are suffering from a Wilson
obsession. This obsession has resulted in type systems of overwhelming complexity, and languages that are larger, less regular, and less expressive than
they need to be. The fallacy is the assumption that, because type checking
is good, more type-checking is necessarily better.
Let’s consider an example of how the insistence on static type-checking
can mess up a language design. I’m currently engaged, with Kim Bruce,
James Noble and their students, in the design of a language for teaching
programming. The language is called Grace, both to honour Rear Admiral
Grace Hopper, and because we hope that it will enable the Graceful expression of programs. Grace is an object-oriented language, and contains an
28
inheritance facility, inspired by early proposals by Antero Taivalsaari [41].
Here is an example:
class aDictionary . ofSize ( initialSize ) {
inherits aHashtable. ofSize ( initialSize )
method findIndex (predicate) is override { . . . }
method at (key) put (value) is public { . . . }
...
}
In Grace, classes are nothing more than objects that have a single method
that, when executed, creates a new object. This example declares a class object named aDictionary, which has an object-creation method with the name
ofSize. Executing aDictionary.ofSize(10) will create and return a new object
that inherits (that is, contains copies of) all of the methods and instance
variables of the object aHashtable.ofSize(10), except that the given method
for findIndex() will override that inherited from the Hashtable object, and the
given method for at()put() will be added.
Checks that one might expect to take place on such a definition would
include that aHashtable actually does have a method ofSize, and that the
object answered by that method does have a findIndex() method and does
not have an at()put() method. There is no difficulty constructing a static
type system to perform such checks so long as aHashtable is a globally known
class. However, suppose that I want to let the client choose the actual object
that is extended, perhaps because prefix trees are better than hash tables
for some applications. In other words, suppose that I want to make the
super-object a parameter:
class aDictionary .basedOn(superObj) {
inherits superObj
method findIndex (predicate) is override { . . . }
method at (key) put (value) is public { . . . }
...
}
Although the same implementation mechanisms will still work, constructing a static type system to perform the checking is no longer simple. What
arguments may be substituted for superObj? The requirement that superObj
have a (possibly protected) method findIndex (because of the override annotation) and that it not have a method at()put() (because of the absence of an
override annotation) cannot be captured by the usual notion of type, which
29
holds only the information necessary to use an object, not the information
necessary to inherit from it. One solution to the problem of type-checking
this definition in a modular way is to invent a new notion of “heir types”,
that is, types that capture the information needed to check inheritance, such
as the presence of a method that is either public or protected, or the absence
of a method.
Another possible solution is to make classes a new sort of entity, different
from objects, and to give them their own, non-Turing-complete sublanguage,
including class-specific function, parameter and choice mechanisms. A third
possibility is to ban parametric super-objects. How then does one use inheritance, that is, how does the programmer gain access to a superclass to
extend? The classical solution is to refer to classes by global variables, and to
require that the superclass be globally known. This is a premature commitment that runs in opposition to the late-binding that otherwise characterises
object-orientation; its effect is to reduce reusability. Some languages alleviate this problem by introducing yet another feature, known as open classes,
which allows the programmer to add methods to, or override methods in, an
existing globally-known class.
Virtual classes, as found in beta [42], are yet another approach to the
problem of parametrizing inheritance. A virtual class can be overridden in a
subclass by a “more specialised” class. Virtual classes don’t by themselves
solve our problem: they need to be accompanied by a set of type rules that
specify when it is legal to override a virtual class. The difficulty is that one
set of rules — conventional subtyping — is appropriate for clients of the class,
and another set of rules — heir typing — is appropriate for inheritors. beta
chooses to use co-variant method subtyping for virtual classes, that is, to
allow the arguments of methods to be specialised along with their results.
This means that virtual classes cannot be guaranteed to be safe by a modular
static analysis. Nevertheless, virtual classes are useful for modelling real
systems; beta follows Simula in ranking modelling above static checking.
Regardless of the path chosen, the resulting language is both larger and
less-expressive than the simple parametric scheme sketched above. Indeed,
in the scheme suggested by my original example, a new object could extend
any existing object, and classes did not need to be “special” in any way; they
could be ordinary objects that did not require any dedicated mechanisms for
their creation, categorisation or use.
Another example of this sort of problem is collections that are parametrized
30
by types. If types are desirable, it certainly seems reasonable to give programmers the opportunity to parameterize their collection objects, so that a
library implementor can offer Bags of Employees as well as Bags of Numbers.
The obvious solution is to represent types as objects, and use the normal
method and parameter mechanisms to parameterize collections — which is
exactly what we did in Emerald [43]. Unfortunately, the consequence of this
is that type checking is no longer decidable [44]. The reaction of the Wilsonian faction to this is to recoil in shock and horror. After all, if the mission
of types is to prevent the execution of any program that might possibly go
wrong, and the very act of deciding whether the program will go wrong might
not terminate, what is the language implementor to do?
A language that takes the Murrow interpretation of type can react to the
possibility that type-checking is statically undecidable more pragmatically.
The language implementation will inevitably use run-time type checks, to
deal with situations in which the program cannot be guaranteed to be free
of type errors statically. If a type assertion cannot be proved true or false
after a reasonable amount of effort, it suffices to insert a run-time check.
An alternative way of parameterising objects by types is to invent a new
parameter passing mechanism for types, with new syntax and semantics,
but with restricted expressivity to ensure decidability. Nevertheless, because
of Gödel’s second incompleteness theorem, some programs will still be untypeable. The consequence is that the language becomes larger, while at the
same time less expressive.
11. The Future of Objects
Now I’m going to turn to some speculations about the programming languages of the future. I could follow Feynman, and predict, with a certain
confidence, that in 1000 years object-oriented programming will no longer exist as we know it. This is either because our civilisation will have collapsed,
or because it has not; in the latter case humans will no longer be engaged in
any activity resembling programming, which will instead be something that
computers do for themselves.
I think that it is probably more useful, and certainly more within my
capabilities, to look 10 or 20 years ahead. The changes in computing that
will challenge our ability to write programs over that time span are already
clear: the trend from multicore towards manycore, the need to control energy
31
consumption, the growth of mobility and “computing in the cloud”, the
demand for increased reliability and failure resilience, and the emergence of
distributed software development teams.
11.1. Multicore and Manycore
Although the exact number of “cores”, that is, processing units, that will be
present on the computer chips of 2021–31 is unknown, we can predict that
with current technology trends we will be able to provide at least thousands,
and perhaps hundreds of thousands. It also seems clear that manycore chips
will be much more heterogeneous than the two- and four-processor multicore
chips of today, exactly because not all algorithms can take advantage of
thousands of small, slow but power-efficient cores, and will need larger, more
power-hungry cores to perform acceptably. The degree of parallelism that will
be available on commodity chips will thus depend both on the imperatives
of electronic design and on whether we can solve the problem of writing
programs that use many small cores effectively.
What do objects have to offer us in the era of manycore? The answer
seems obvious. Processes that interact through messages, which are the
essence of Simula’s view of computation, are exactly what populate a manycore chip. The similarity becomes even clearer if we recall Kay’s vision of
objects as little computers. Objects also offer us heterogeneity. There are
many different classes of objects interacting in a typical program, and there
is no reason that these different classes of objects need be coded in the same
programming language, or compiled to the same instruction set. This is because the “encapsulation boundary” around each object means that no other
object need know about its implementation. For example, objects that perform matrix arithmetic to calculate an image might be compiled to run on
a general-purpose graphical processing unit, while other parts of the same
application might be compiled to more conventional architectures.
11.2. Controlling Energy Consumption
How does one write an energy-efficient program? How, indeed, can one compare two programs for efficiency? A common approach is to use a cost model :
a way of thinking about computation that is closely-enough aligned with the
real costs of execution that it can help us to reason about performance. Objects can provide us with a cost model for manycore computing.
32
Most current computing models date from the 1950s and treat computation as the expensive resource — which was the case when those models were
developed, and computation made use of expensive and power-hungry valve
or discrete-transistor logic. In contrast, data movement is regarded as free,
since it was implemented by copper wires, which were inexpensive compared
to valves or transistors, and consumed little energy. As a consequence, these
models lead us to think, for example, of moving an operation out of a loop
and caching the result in memory as an “optimisation”. The reality today
is very different: computation is essentially free, because it happens “in the
cracks” between data fetch and data store; on a clocked processor, there is
little difference in energy consumption between performing an arithmetic operation and performing a no-op. In contrast, data movement is expensive: it
takes both time and energy to propagate signals along wires. So controlling
the cost of computation on a manycore chip requires one to exercise control of
data movement. Today’s programming languages are unable to even express
the thing that needs to be carefully controlled: data movement.
I would like to propose an object-oriented cost model for manycore computing. To Simula’s concurrent object model we need add only three things.
The first addition is to treat objects as enjoying true parallelism, rather than
Simula’s quasi-parallelism. The second and third additions are to attribute
to each object a size and a spatial location. Size is important because an
object needs to fit into the cache memory available at its location. “Size”
includes not just the data inside the object, but also the code that makes up
its method suite, since both must be cached.
With this model, local operations on an object can be regarded as free,
since they require only local computations on local data. Optimising an
object means reducing its size until it fits into the available cache, or, if this
is not possible, partitioning the object into two or more smaller objects. In
contrast to local operations, requesting that a method be evaluated in another
object may be costly, since this may require communicating with an object at
a remote location. The cost of such a message is proportional to the product
of the amount of data in it and the distance to the receiver; we can imagine
measuring cost in byte-nanometers. The cost of a whole computation can
be estimated as proportional to the number of messages sent and the cost
of each message and reply. We can reduce this cost by relocating objects so
that frequent communication partners are close together, and by recomputing
answers locally rather than requesting them from a remote location. Such a
33
cost model does not, of course, capture all of the costs of a real computation,
but it seems to me that it will provide a better guide for optimisation than
models from the 1950s.
11.3. Mobility and the Cloud
The world of mobile computing is a world in which communication is transient, failure is common, and replication and caching are the main techniques
used to improve performance and availability. The best models for accessing
objects in such an environment seem to me to be those used for distributed
version control, as realised in systems like subversion and git [45, 46]. In
such systems, all objects are immutable; updating an object creates a new
version. Objects can be referred to in two ways: by version identifier and by
name. A version identifier is typically a long numeric string and refers to (a
copy of) a particular immutable object. In contrast, a name is designed for
human consumption, but the binding of a name to a version object changes
over time: when a name is resolved, it is by default bound to a recent version
of the object that it describes.
Erlang uses a similar naming mechanism to realise its fail-over facility.
Erlang messages can be sent either to a process identifier or to a process name.
A particular process identifier will always refer to the same process; sending a
message to a process identifier will fail if the process is no longer running. In
contrast, a process name indirects through the name server, so messages sent
to a process name will be delivered to the most recent process to register that
name. When a process crashes and its monitor creates a replacement process,
the replacement will usually register itself with the name sever under the same
name that was used by the crashed process. This ensures that messages sent
to the process name will still be delivered, although successive messages may
be delivered to different processes [19]. Perhaps a similar mechanism should
be part of every object-oriented system? This would mean that we would
be able to reference an object either by a descriptor, such as “Most recent
version of the Oslo talk”, or by a unique identifier, such as Object16x45d023f.
The former would resolve to a replica of a recent version; the latter would
resolve to a specific object.
11.4. Reliability
Total failures are easy for the programmer to handle, because there is nothing
that the programmer can do! However, they can be disastrous for the users of
34
a computer system, who are prevented from doing their work. As distributed
systems have become ubiquitous, total failures have become rare, and partial
failures much more common. This trend will accelerate as the size of the
features on a chip decreases, and massively manycore chips proliferate. The
reliability of individual transistors on such chips will be much lower than we
are accustomed to today: this is a consequence of the physics of doping. By
2014, if there are 100 billion transistors on a chip, 20 percent of them will not
work when it is fabricated, and another 10 percent will fail over the life of the
chip6 . This means two things: that processor design will have to change quite
radically, and that single chips will have reliability characteristics similar to
today’s Internet, where computers may fail at any time.
What techniques are available for dealing with partial failures? It is often
possible to mask a failure using replication in time or space, but this is not
always wise. Resending lost messages is not a good idea if they contain
out-of date information; it may be better to send a new message containing
fresh information. Similarly, it may be more cost-effective to recompute lost
data than to retrieve it from a remote replica. So automatic masking of
failure by the programming model is not a good idea. Instead, it seems to
me that a variety of facilities for dealing with failures must be visible in
the programming model, so that programmers can choose whether or not to
use them. What, then, should be the unit of failure in an object-oriented
model of computation? Is it the object, or is there some other, larger, unit?
Whatever unit we choose, it must “leak” failure, in the sense that its clients
must be able to see it fail, and then take some action appropriate to the
context.
11.5. Distributed Development Teams
With the ubiquity of computing equipment and the globalisation of industry,
distributed software development teams have become common. You may
wonder what this trend has to do with objects. The answer is packaging:
collaborating in loosely-knit teams demands better tools for packaging code
and sharing it between the parts of a distributed team. Modules, by which I
mean the unit of code sharing, are typically parametrized by other modules.
6
These predictions are from Shekhar Borkar, an Intel fellow and Director of the Microprocessor Technology lab. They were presented by Bob Colwell at FCRC, San Diego,
2007.
35
To take full advantage of the power of objects, module parameters should be
bound at the time that a module is used, not when it is written. In other
words, module parameters should be “late bound”. This means that there
is no need for a global namespace in the programming language itself; URLs
or versioned objects provide a perfectly adequate mechanism for referring to
the arguments of modules.
12. The Value of Dynamism
Before I conclude, I’m going to offer some speculations on the value of dynamism. These speculations are motivated by the fact that in designing
Simula, Dahl recognised that the runtime structures of Algol 60 already
contained the mechanisms that were necessary for simulation, and by Nygaard and Dahl’s understanding that it is the runtime behaviour of a simulation program that models the real world, not the program’s text. I see
a similar recognition of the value of dynamism in Agile software development, a methodology in which a program is developed in small increments,
in close consultation with the customer. The idea is that a primitive version
of the code runs at the end of the first week, and new functionality is added
every week, under the guidance of the customer, who determines which functions have the greatest value. Extensive test suites make sure that the old
functionality continues to work while new functionality is added.
If you have never tried such a methodology, you may wonder how it
could possibly work! After all, isn’t it important to design the program?
The answer is yes: design is important. In fact, design is so important,
that Agile practitioners don’t do it only at the start of the software creation
process, when they know nothing about the program. Instead, they design
every day: the program is continuously re-designed as the programmers learn
from the code, and from the behaviour of the running system.
What I’ve learned from watching this process is that the program’s runtime behaviour is a powerful teaching tool. This is obvious if you view
programs as existing to control computers, but perhaps less obvious if you
view programs as static descriptions of a system, like a set of mathematical
equations. As Dahl and Nygaard discovered, it is a program’s behaviour, not
the program itself, that models the real-world system. The program’s text
is a meta-description of the program behaviour, and it is not always easy to
infer the behaviour from the meta-description.
36
As a teacher of object-oriented programming, I know that I have succeeded when students anthropomorphise their objects, that is, when they
turn to their partners and speak of one object asking another object to do
something. I have found that this happens more often, and more quickly,
when I teach with Smalltalk than when I teach with Java: Smalltalk programmers tend to talk about objects, while Java programmers tend to talk
about classes. I suspect that this is because Smalltalk is the more dynamic
language: the language and the programming environment are designed to
help programmers interact with objects, as well as with code. Indeed, I am
tempted to define a “Dynamic Programming Language” as one designed to
help the programmer learn from the run-time behaviour of the program.
13. Summary
Fifty years ago, Dahl and Nygaard had some profound insights about the nature of both computation and human understanding. “Modelling the world”
is not only a powerful technique for designing and re-designing a computer
program: it is also one of the most effective ways yet found of communicating
that design to other humans.
What are the major concepts of object-orientation? The answer depends
on the social and political context. Dahl listed five, including the use of
objects as record structures and as procedurally-encapsulated data abstractions, and the use of inheritance for incremental definition. Some of what he
though were key ideas, such as active objects and objects as modules, have
been neglected over the last 30 years, but may yet be revived as the context
in which we program becomes increasingly distributed and heterogeneous,
in terms of both execution platform and programming team. It certainly
seems to me that, after 50 years, there are still ideas in Simula that can be
mined to solve twenty-first century problems. There may not be any programming a thousand years from now, but I’m willing to wager that some
form of Dahl’s ideas will still be familiar to programmers in fifty years, when
Simula celebrates its centenary.
Acknowledgements
I thank Stein Krogdahl, Olaf Owe and the committee of FCT11 for honouring me with the invitation to speak at the scientific opening of the OleJohan Dahl hus, as well as supplying me with information about Dahl and
37
Simula. I also thank Olaf for convincing me to turn my lecture into this
article. William Cook was an encouraging and through critic; Ivan Sutherland helped me understand where the costs lie in modern processors, and
provided useful feedback on a draft of my manuscript. David Ungar shared
many insights with me and helped to improve both my lecture and this article. The anonymous referees were especially helpful in pointing out errors
of fact and passages in need of clarification.
References
[1] FCT 2011, Scientific opening of the Ole-Johan Dahl building, 2011.
Web page last visited 2 March 2013. http://fct11.ifi.uio.no/index.php?
n=General.SocialEvents.
[2] K. Nygaard, O.-J. Dahl, The development of the SIMULA languages,
in: R. L. Wexelblat (Ed.), History of programming languages I, ACM,
New York, NY, USA, 1981, pp. 439–480.
[3] University of Oslo, Department of Informatics, Kristen Nygaard —
Education and Career, 2011. Web page last visited 25 December 2011.
http://www.mn.uio.no/ifi/english/about/kristen-nygaard/career/.
[4] R. Feynman, Speech at Centenial Celebration of Massachusetts Institute
of Technology (version C), 1961. The Feynman Archives at California
Institute of Technology.
[5] C. Hoare, Hints on Programming Language Design, Memo AIM-224,
Stanford Artificial Intelligence Laboratory, 1973. Invited address, 1st
POPL conference.
[6] G. Radin, The early history and characteristics of PL/I, SIGPLAN
Notices 13 (1978) 227–241.
[7] O.-J. Dahl, K. Nygaard, Class and subclass declarations, in: J. N.
Buxton (Ed.), Simulation Programming Languages, Proceedings from
the IFIP working conference in Oslo, North Holland, Amsterdam, 1967,
pp. 158–174.
[8] O.-J. Dahl, Transcript of discussant’s remarks, in: R. L. Wexelblat
(Ed.), History of programming languages I, ACM, New York, NY, USA,
1981, pp. 488–490.
38
[9] G. Bracha, P. von der Ahé, V. Bykov, Y. Kashai, W. Maddox, E. Miranda, Modules as objects in newspeak, in: T. D’Hondt (Ed.), ECOOP,
volume 6183 of Lecture Notes in Computer Science, Springer, 2010, pp.
405–428.
[10] A. P. Black, K. B. Bruce, M. Homer, J. Noble, Grace: the absence of
(inessential) difficulty, in: Onward! ’12: Proceedings 12th SIGPLAN
Symp. on New Ideas in Programming and Reflections on Software, ACM,
New York, NY, 2012, pp. 85–98.
[11] O.-J. Dahl, C. Hoare, Hierarchical program structures, in: Structured
Programming, Academic Press, 1972, pp. 175–220.
[12] C. A. R. Hoare, Record handling, ALGOL Bull. 21 (1965) 39–69.
[13] F. Genuys (Ed.), Programming Languages: NATO Advanced Study
Institute, Academic Press, London, New York, 1968.
[14] O.-J. Dahl, The roots of object-oriented programming: the Simula language, in: M. Broy, E. Denert (Eds.), Software Pioneers: Contributions
to Software Engineering, Springer-Verlag, Berlin, Heidelberg, 2002, pp.
79–90.
[15] P. Wegner, Dimensions of object-based language design, in: N. Meyrowitz (Ed.), Proceedings Second ACM Conference on Object-Oriented
Programming Systems, Languages and Applications, ACM Press, Orlando, Florida, 1987, pp. 168–182.
[16] A. P. Black, N. Hutchinson, E. Jul, H. Levy, Object structure in the
Emerald system, in: Proceedings First ACM Conference on ObjectOriented Programming Systems, Languages and Applications, ACM
Press, Portland, Oregon, 1986, pp. 78–86.
[17] A. P. Black, N. Hutchinson, E. Jul, H. M. Levy, L. Carter, Distribution
and abstract types in Emerald, IEEE Trans. Software Engineering SE-13
(1987) 65–76.
[18] W. R. Cook, A Denotational Semantics of Inheritance, Ph.D. thesis,
Brown University, Department of Computer Science, 1989.
39
[19] J. Armstrong, Programming Erlang: Software for a concurrent world,
Pragmatic Bookshelf, 2007.
[20] B. Archer, Programming quotations, 2011. Web page last visited 22
January 2012. http://www.bobarcher.org/software/programming quotes.
html.
[21] P. Wadler, The essence of functional programming, in: Conference
Record of the Nineteenth ACM Symposium on Principles of Programming Languages, ACM Press, Albuquerque, NM, 1992, pp. 1–14.
[22] P. Wadler, Comprehending monads, Mathematical Structures in Computer Science 2 (1992) 461–493. Originally published in ACM Conference
on Lisp and Functional Programming, June 1990.
[23] A. C. Kay, The early history of Smalltalk, in: The second ACM SIGPLAN conference on History of programming languages, HOPL-II, ACM
Press, New York, NY, USA, 1993, pp. 511–598.
[24] A. Snyder, The Essence of Objects: Common Concepts and Terminology, Technical Report HPL-91-50, Hewlett Packard Laboratories, 1991.
[25] A. Snyder, The essence of objects: Concepts and terms, IEEE Software
10 (1993) 31–42.
[26] S. Krogdahl, Concepts and terminology in the Simula programming
language — an introduction for new readers of Simula literature, 2010.
http://folk.uio.no/simula67/Archive/concepts.pdf.
[27] D. H. Ingalls, Design principles behind Smalltalk, Byte 6 (1981) 286–
298.
[28] A. H. Borning, D. H. H. Ingalls, A type declaration and inference system
for Smalltalk, in: Conference Record of the Ninth ACM Symposium on
Principles of Programming Languages, ACM Press, Albuquerque, NM,
USA, 1982, pp. 133–141.
[29] C. Hoare, Proof of correctness of data representations, Acta Informatica
1 (1972) 271–281.
[30] D. L. Parnas, On the criteria to be used in decomposing systems into
modules, Comm. ACM 15 (1972) 1053–1058.
40
[31] B. Liskov, A. Snyder, R. Atkinson, C. Schaffert, Abstraction mechanisms in CLU, Comm. ACM 20 (1977) 564–576.
[32] A. Church, The Calculi of Lambda-Conversion, volume 6 of Annals of
Mathematical Studies, Princeton University Press, 1941.
[33] J. C. Reynolds, User-defined types and procedural data structures as
complementary approaches to data abstraction, in: S. A. Schuman
(Ed.), Conference on New Directions in Algorithmic Languages, IFIP
Working Group 2.1 on Algol, INRIA, Munich, Germany, 1975, pp. 157–
168.
[34] J. C. Reynolds, User-defined types and procedural data structures as
complementary approaches to data abstraction, in: D. Gries (Ed.),
Programming Methodology: A Collection of Articles by members of
IFIP WG2.3, Texts and Monographs in Computer Science, Springer
Verlag, New York, Heidelberg, Berlin, 1978, pp. 309–317.
[35] J. C. Reynolds, User-defined types and procedural data structures as
complementary approaches to data abstraction, in: C. A. Gunter, J. C.
Mitchell (Eds.), Theoretical Aspects of Object-Oriented Programming:
Types, Semantics, and Language Design, MIT Press, Cambridge, MA,
1994, pp. 13–23.
[36] W. R. Cook, On understanding data abstraction, revisited, in: S. Arora,
G. T. Leavens (Eds.), Proceedings 24th ACM Conference on ObjectOriented Programming, Systems, Languages, and Applications, ACM,
2009, pp. 557–572.
[37] C. Hewitt, P. Bishop, R. Steiger, A universal modular actor formalism for artificial intelligence, in: N. J. Nilsson (Ed.), IJCAI, William
Kaufmann, Stanford University, California, 1973, pp. 235–245.
[38] J. Armstrong, Concurrency oriented programming in Erlang, in:
Frühjahrsfachgespräch 2003 (FG 2003), German Unix Users Group,
Köln, Germany, 2003.
[39] A. Borning, Classes versus prototypes in object-oriented languages, in:
FJCC, IEEE Computer Society, 1986, pp. 36–40.
41
[40] A. P. Black, N. Hutchinson, Typechecking Polymorphism in Emerald,
Technical Report Technical Report CRL 91/1 (Revised), Digital Equipment Corporation Cambridge Research Laboratory, 1991.
[41] A. Taivalsaari, Delegation versus concatenation, or cloning is inheritance
too, SIGPLAN OOPS Mess. 6 (1995) 20–49.
[42] O. L. Madsen, An overview of BETA, in: J. Knudsen, O. Madsen,
B. Magnusson (Eds.), Object-Oriented Environments, Prentice Hall, Englewood Cliffs, NJ, 1993, pp. 99–118.
[43] A. P. Black, N. C. Hutchinson, E. Jul, H. M. Levy, The development
of the Emerald programming language, in: B. G. Ryder, B. Hailpern
(Eds.), HOPL III: Proceedings Third ACM SIGPLAN conference on
History of Programming Languages, ACM Press, San Diego, CA, 2007,
pp. 11–1–11–51.
[44] A. R. Meyer, M. B. Reinhold, ‘Type’ is not a Type: Preliminary report,
in: Conference Record of the Thirteenth ACM Symposium on Principles
of Programming Languages, ACM Press, St Petersburg Beach, FL, USA,
1986, pp. 287–295.
[45] M. Mason, Pragmatic Version Control Using Subversion, Pragmatic
Bookshelf, Pragmatic Programmers, 2005.
[46] C. Duan, Understanding Git: Repositories, 2012. http://www.eecs.
harvard.edu/∼cduan/technical/git/git-1.shtml. Accessed on 3 Aug 2012.
42
| 6cs.PL
|
1
Modeling and Design of Millimeter-Wave Networks
for Highway Vehicular Communication
arXiv:1706.00298v5 [cs.IT] 15 Aug 2017
Andrea Tassi, Malcolm Egan, Robert J. Piechocki and Andrew Nix
Abstract—Connected and autonomous vehicles will play a
pivotal role in future Intelligent Transportation Systems (ITSs)
and smart cities, in general. High-speed and low-latency wireless
communication links will allow municipalities to warn vehicles
against safety hazards, as well as support cloud-driving solutions to drastically reduce traffic jams and air pollution. To
achieve these goals, vehicles need to be equipped with a wide
range of sensors generating and exchanging high rate data
streams. Recently, millimeter wave (mmWave) techniques have
been introduced as a means of fulfilling such high data rate
requirements. In this paper, we model a highway communication
network and characterize its fundamental link budget metrics.
In particular, we specifically consider a network where vehicles
are served by mmWave Base Stations (BSs) deployed alongside
the road. To evaluate our highway network, we develop a new
theoretical model that accounts for a typical scenario where heavy
vehicles (such as buses and lorries) in slow lanes obstruct Lineof-Sight (LOS) paths of vehicles in fast lanes and, hence, act
as blockages. Using tools from stochastic geometry, we derive
approximations for the Signal-to-Interference-plus-Noise Ratio
(SINR) outage probability, as well as the probability that a user
achieves a target communication rate (rate coverage probability).
Our analysis provides new design insights for mmWave highway
communication networks. In considered highway scenarios, we
show that reducing the horizontal beamwidth from 90◦ to 30◦
determines a minimal reduction in the SINR outage probability
(namely, 4 · 10−2 at maximum). Also, unlike bi-dimensional
mmWave cellular networks, for small BS densities (namely, one
BS every 500 m) it is still possible to achieve an SINR outage
probability smaller than 0.2.
Index Terms—Vehicular communications, millimeter-wave networks, performance modeling, stochastic geometry.
I. I NTRODUCTION
By 2020, fifty billion devices will have connectivity capabilities [1]. Among these, ten million vehicles equipped
with on-board communication systems and with a variety
of autonomous capabilities will be progressively rolled out.
According to the National Highway Traffic Safety Administration (U.S. Department of Transportation) and the European
Commission’s Connected-Intelligent Transportation System
(C-ITS) initiative [2], [3], connectivity will allow vehicles
to engage with future ITS services, such as See-Through,
Automated Overtake, High-Density Platooning, etc [4].
This work is partially supported by the VENTURER Project and
FLOURISH Project, which are supported by Innovate UK under Grant
Numbers 102202 and 102582, respectively.
A. Tassi, R. J. Piechocki and A. Nix are with the Department of Electrical
and Electronic Engineering, University of Bristol, UK (e-mail: {a.tassi,
r.j.piechocki, andy.nix}@bristol.ac.uk).
M. Egan is with the CITI Laboratory of the Institut National de Recherche
en Informatique et en Automatique (INRIA), Université de Lyon, and
Institut National de Sciences Apliquées (INSA) de Lyon, FR (e-mail:
[email protected]).
As identified by the European Commission’s C-ITS initiative, the number of sensors mounted on each vehicle
has increased. A typical sensor setup is expected to range
from ultra-sound proximity sensors to more sophisticated
camcorders and ‘Light Detection And Ranging’ (LiDAR)
systems [4]. Currently, the number of on-board sensors are
around 100 units and this number is expected to double by
2020 [5]. Ideally, the higher the number of on-board sensors,
the “smarter” the vehicle. However, this holds true only if
vehicles are able to exchange the locally sensed data [6]. For
instance, multiple LiDAR-equipped vehicles may approach
a road hazard and share their real-time LiDAR data with
incoming vehicles by means of the road-side infrastructure.
This allows the approaching vehicles to compensate for their
lack of sensor data (blind-spot removal) and, for instance,
help smart cruise-control systems make decisions. As such,
there are strong constraints on LiDAR data delivery, which
can be generated at rates up to 100 Mbps. More generally,
semi-autonomous and fully autonomous vehicles will require
high rate and low latency communication links to support
the applications envisaged by the 5G Infrastructure Public
Private Partnership’s (5G-PPP). These applications include the
See-Through use case (maximum latency equal to 50 ms),
which enables vehicles to share live video feeds of their
onboard cameras to following vehicles. Other applications
such as Automated Overtake and High-Density Platooning are
also expected to require communication latencies smaller than
10 ms [4, Table 1].
Recently, communication systems operating in the
millimeter-wave (mmWave) range of the wireless spectrum
have been proposed as a means of overcoming the rate
and latency limitations of existing technologies [7], [8].
In fact, currently commercialized mmWave systems can
already ensure up to 7 Gbps and latencies smaller than
10 ms [9]. Table I summarizes the general performance
metrics of mmWave systems and compares them with the
main technologies adopted to enable infrastructure-to-vehicle
communications. Traditionally, ITSs rely on Dedicated
Short-Range Communication (DSRC) standards, such as
IEEE 802.11p/DSRC and ITS-G5/DSRC [10]–[13]. Even
though these technologies operate in a licensed band and
ensure low communication latencies, their maximum realistic
data rate hardly exceeds 6 Mbps [10]. As such, several
papers [14], [15] suggest the adoption of 3GPP’s Long Term
Evolution-Advanced (LTE-A) [16], [17], which can guarantee
higher communication rates. Nevertheless, the maximum
supported data rate is limited to 100 Mbps and end-to-end
latencies cannot go below 100 ms [6]. As a result, both
2
TABLE I
R ADIO ACCESS SOLUTIONS FOR VEHICULAR COMMUNICATIONS .
IEEE 802.11p/DSRC,
ITS-G5/DSRC [13]
LTE-A [15]
mmWave
Systems [9]
Frequency
Band
5.85 GHz 5.925 GHz
Spanning
multiple bands in
450 MHz 4.99 GHz
28 GHz,
38 GHz,
60 GHz bands
and E-band
Channel
Bandwidth
10 MHz
Up to 100 MHz
100 MHz2.16 GHz
Bit Rate
3 Mbps-27 Mbps
Up to 1 Gbps
Up to 7 Gbps
Latency
≤ 10 ms
100 ms-200 ms
≤ 10 ms
Mobility
Support
≤ 130 km h
−1
≤ 350 km h
−1
≤ 100 km h−1
DSRC and LTE-A cannot always meet the communication
constraints dictated by delay and bandwidth sensitive services
that will be offered by future ITSs [4, Table 1].
In mmWave systems, both Base Stations (BSs) and users
are equipped with large antenna arrays to achieve high array
gains via beamforming techniques [18]. As mmWave systems
operate in the portion of the spectrum between 30 GHz and
300 GHz [9], mmWave links are highly sensitive to blockages.
In particular, line-of-sight (LOS) communications are characterized by path loss exponents that tend to be smaller than
2.8, while non-line-of-sight (NLOS) path loss exponents are
at least equal to 3.8 [19]. Due to of the the difficulty of beam
alignment, commercial mmWave solutions cannot support user
speeds greater than 100 km h−1 . Despite this, research to cope
with mobile users is gaining momentum [20]. For instance,
in the UK, the main railway stakeholders are already trialing
mmWave systems with enhanced beam searching techniques
to provide broadband wireless connectivity onboard moving
trains [21]. In addition, multiple research initiatives already
regard mmWave systems as suitable to deploy 5G cellular
networks [22]–[24].
In this paper, we consider a typical road-side infrastructure for ITSs [25]. In particular, the infrastructure-to-vehicle
communications required by ITS services are handled by a
dedicated network of BSs placed on dedicated antenna masts
and or other street furniture, typically on both sides of the
road [26], [27]. We deal with a highway system where vehicles
receive high data rate streams transmitted by mmWave BSs, although we do not consider scenarios where there is no roadside
deployment of BSs where vehicle-to-vehicle communication
technologies may provide a more effective solution. A key
feature of our highway system is that vehicles with different
sizes are likely to drive along the same set of lanes. In a lefthand traffic system, any slow vehicle (such as double decker
buses or lorries) typically travels in the outermost lanes of
the highway, while the other vehicles tend to drive along the
innermost lanes. If a larger vehicle drives between a user and
its serving BS, the BS is no longer in LOS. In other words,
large vehicles may act as communication blockages.
We develop a new framework to analyze and design
mmWave communication systems. The original contributions
of this paper are summarized as follows:
• We propose the first theoretical model to characterize
the link budget requirements of mmWave networks pro-
viding downlink connectivity to highway vehicles where
communications are impaired by large vehicles acting as
communication blockages. Specifically, we offer design
insights that take into account how the BS and blockage
densities impact the user achievable data rate.
• We show that the performance of mmWave highway
networks can be well approximated by our theoretical
model, which assumes that both BS and blockage positions are governed by multiple time-independent monodimensional PPPs. Traditional vehicular models assume
that BSs are equally spaced [28] – thus making them incapable of describing irregular BS deployments. In addition,
the impact of blockages is either not considered [29] or
the blockage positions are deterministically known [30],
which makes the latter kind of models suitable to be
included in large-scale network simulators but also makes
them analytically intractable. On the other hand, the
proposed model is analytically tractable and allows us
to predict user performance in scenarios characterized
by different BS densities, traffic intensities, antenna gain
and directivity without assuming the BS and blockage
positions being known in advance.
• Our numerical validation demonstrates that the proposed
theoretical model is accurate and provides the following
design insights: (i) a smaller antenna beamwidth does not
necessarily reduce the Signal-to-Interference-plus-Noise
Ratio (SINR) outage probability, and (ii) a reduced SINR
outage probability in highway mmWave networks can be
achieved even by low-density BS deployments, for a fixed
probability threshold.
The remainder of the paper is organized as follows. Section II discusses the related work on mmWave and vehicular
communication systems. Section III presents our mmWave
communication system providing downlink coverage in highway mmWave networks. We evaluate the network performance
in terms of the SINR outage and rate coverage probabilities,
which are derived in Section IV. Section V validates our
theoretical model. In Section VI, we conclude and outline
avenues of future research.
II. R ELATED W ORK
As summarized in Table II, over the past few years,
mmWave systems have been proposed as a viable alternative to traditional wireless local area networks [9] or as a
wireless backhauling technology for BSs of the same cellular
network [31], [32]. Furthermore, mmWave technology has
also been considered for deploying dense cellular networks
characterized by high data rates [23], [34], [35]. With regards
to the vehicular communication domain, J. Choi et al. [7]
pioneered the application of the mmWave technology to partially or completely enable ITS communications. A mmWave
approach to ITS communications is also being supported by
the European Commission [3].
As both the BS deployment and vehicle locations differ
over both time and in different highway regions, any highway
network model must account for these variations. In this
setting, stochastic geometry provides a means of characterizing
3
TABLE II
R ELATED W ORKS ON MM WAVE S YSTEMS AND V EHICULAR C OMMUNICATIONS .
Ref.
Radio Access
Technology
Network Topology
[7]
mmWave
Vehicle-to-vehicle,
Vehicle-to-infrastructure
[23]
mmWave
Dense cellular network
[9]
mmWave
Dense cellular system
[31]
mmWave
Network backhauling
[32]
mmWave
Cellular network with
self-backhauling
[33]
mmWave
Co-operative cellular network
[34]
mmWave
Cellular network with
self-backhauling
[35]
mmWave
Multi-tier cellular network
[28]
DSRC
Vehicle-to-vehicle,
Vehicle-to-infrastructure
[29]
DSRC
Vehicle-to-vehicle,
Vehicle-to-infrastructure
[30]
DSRC
Vehicle-to-vehicle
[36]
DSRC
Vehicle-to-vehicle
Channel and Path Loss
Models
Mobility
Based on ray-tracing
Vehicles moving on urban roads
Nakagami small-scale fading;
BGG path loss model
Based on measurements
Constant small-scale fading
(i.e., the square norm of the
small-scale fading
contribution is equal to 1)
Constant small-scale fading;
BGG path loss model
Nakagami for the signal,
Rayleigh for the interference
contribution; BGG path loss
model
Based on measurements
Constant small-scale fading;
Probabilistic path loss model
Coverage-based (i.e., no
packet errors from nodes
within the radio range)
Rice small-scale fading
Obstacle-based channel and
path loss models
Rayleigh small-scale fading
the performance of the system by modeling BS locations via a
spatial process, such as the Poisson Point Process (PPP) [37].
Generally, PPP models for wireless networks are now a wellestablished methodology [8], [37]–[41]; however, there are
challenges in translating standard results into the context
of mmWave networks for road-side deployments due to the
presence of NLOS links resulting from blockages [23]. In
particular, the presence of blockages has only been addressed
in the context of mmWave cellular networks in urban and
suburban environments that are substantially different to a
highway deployment [23]. In particular, in mmWave cellular
networks: (i) the positions of BSs follow a bi-dimensional
PPP, and (ii) the positions of blockages are governed by
a stationary and isotropic process. Even though this is a
commonly accepted assumption for bi-dimensional cellular
networks [37], this is not satisfied by highway scenarios, where
both blockages and BS distributions are clearly not invariant
to rotations or translations. With regards to Table II, the path
loss contribution of blockages has either been modeled by
means of the Boolean Germ Grain (BGG) principle (i.e.,
only the BSs within a target distance are in LOS) or in a
probabilistic fashion (i.e., a BS is in LOS/NLOS with a given
probability). To the best of our knowledge, no models for
road-side mmWave BS deployment accounting for vehicular
blockages have been proposed to date.
Given the simplicity of their topology and their high level
of automation, highway scenarios have been well investigated
in the literature [28]–[30]. In particular, [28] addresses the
issue of optimizing the density of fixed transmitting nodes
placed at the side of the road, with the objective of maximizing
the stability of reactive routing strategies for Vehicular AdHoc Networks (VANETs) based on the IEEE 802.11p/DSRC
stack. Similar performance investigations are conducted in [29]
where a performance framework jointly combining physical
Communication
Blockages
Not analytically
investigated
Static blockages
Buildings
Static blockages
Buildings
Static blockages
None
Static blockages
Buildings
Static blockages
Buildings
Static blockages
Indoor objects
Static blockages
Buildings
Vehicles moving on urban roads
None
Vehicles moving on urban roads
None
Vehicles moving on a highway
Vehicles
Vehicles moving on a highway
None
and Media Access Control (MAC) layer quality metrics is
devised. In contrast to [28] and [29], [30] addresses the issue
of blockage-effects caused by large surrounding vehicles; once
more, [30] strictly deals with IEEE 802.11p/DSRC communication systems. The proposal in [28]–[30] is not applicable for
mmWave highway networks as the propagation conditions of
a mmWave communication system are not comparable with
those characterizing a system operating between 5.855 GHz
and 5.925 GHz. Another fundamental difference with between
mmWave and IEEE 802.11p/DSRC networks is the lack of
support for antenna arrays capable of beamforming as the
IEEE 802.11p/DSRC stack is restricted to omnidirectional or
non-steering sectorial antennas.
Highway networks have also been studied using stochastic
geometry. In particular, M. J. Farooq et al. [36] propose a
model for highway vehicular communications that relies on
the physical and MAC layers of an IEEE 802.11p/DSRC
or ITS-G5/DSRC system. In particular, the key differences
between this paper and [36] are: (i) the focus on multihop LOS inter-vehicle communications and routing strategies
while our paper deals with one-hop infrastructure-to-vehicle
coverage issues, and (ii) the adoption of devices with no
beamforming capabilities while beamforming is a key aspect
of our mmWave system.
III. S YSTEM M ODEL AND P ROPOSED BS-S TANDARD
U SER A SSOCIATION S CHEME
We consider a system model where mmWave BSs provide
network coverage over a section of a highway, illustrated in
Fig. 1. The goal of our performance model is to characterize
the coverage probability of a user surrounded by several
moving blockages (i.e., other vehicles) that may prevent a
target user to be in LOS with the serving BS. Without loss
4
y
NLOS BS
w
LOS and serving BS
obstacle lane 2
ψ
τ
standard user (theory model)
(o)
O
xℓ,i
AL
*
No
w
λBS
obstacle lane 1
ǫ(U)
user lane 1
speed
user lane high
2 x
i lane 2
standard user (numerical validation)
x
)
r
ǫi
(r
AL
)
(r
TABLE III
C OMMONLY USED NOTATION .
obstacle lane 1
ǫj
obstacle lane 2
Closest NLOS BS
Fig. 1. Considered highway system model, composed of No = 2 obstacle
lanes in each traffic direction.
of generality, we consider the scenario where vehicles drive
on the left-hand side of the road1 . For clarity, Table III
summarizes the symbols commonly used in the paper. In order
to gain insight into the behavior of the model, we make the
following set of assumptions.
Assumption 3.1 (Road Layout): We assume that the whole
road section is constrained within two infinitely long parallel
lines, the upper and bottom sides of the considered road
section. Vehicles flow along multiple parallel lanes in only two
possible directions: West-to-East (for the upper-most lanes)
and East-to-West (for the lowermost lanes). Each lane has the
same width w. For each direction, there are No obstacle lanes
and one user lane closer to the innermost part of the road. The
closer a lane is to the horizontal symmetry axis of the road
section, the more the average speed is likely to increase – thus,
the large/tall vehicles are assumed to drive along obstacle lanes
most of the time. Vehicles move along the horizontal symmetry
axis of each lane. We use a coordinate system centered on
a point on the line separating the directions of traffic. The
upper side of the road intercepts the y-axis of our system of
coordinates at the point (0, w(No + 1)), while the bottom side
intercepts at (0, −w(No + 1)).
In the following sections, we will focus on characterizing
the performance of the downlink phase of a mmWave cellular
network providing connectivity to the vehicles flowing in
the high speed lanes of the considered model, which is
challenging. In fact, communication links targeting users in
the high speed lanes are impacted by the largest number of
communication blockages (namely, large vehicles) flowing on
on the outer road lanes. In addition, we adopt the standard
assumption of the BSs being distributed according to a PPP.
Assumption 3.2 (BS Distribution): Let ΦBS = {xi }bi=1
be the one-dimensional PPP, with density λBS of the
x-components of the BS locations on the road. We assume
that BSs are located along with the upper and bottom sides
of the road section. In particular, the i-th BS lies on the
upper or bottom sides with a probability equal to q = 0.5.
In other words, the y-axis coordinate of the i-th BS is defined
as yi = w(−2Bq +1)(No +1), where Bq is a Bernoulli random
variable with parameter q.
By Assumption 3.2 and from the independent thinning
theorem of PPP [42, Theorem 2.36], it follows that the x-axis
1 The proposed theoretical framework also applies to road systems where
drivers are required to drive on the right-hand side of the road.
λO,ℓ
τ
pL , pN
λL , λN
ℓ(ri )
AL
AN
PL , PN
GT X , GRX
gT X , gRX
LIS,E ,E1 (s)
LI,E1 (s)
PT (θ)
RC (κ)
Number of obstacle lanes per driving direction
Width of a road lane
Density of the PPP ΦBS of x-components of the BS locations
on the road
Density of the PPP ΦO,ℓ of x-components of the blockages on
the ℓ-th obstacle lane
Footprint segment of each blockage
Approximated probabilities of a BS being in LOS or NLOS with
respect to the standard user, respectively
Densities of the PPPs of the x-components of LOS and NLOS
BSs, respectively
Path loss component associated with the i-th BS
Assuming the standard user connects to a NLOS BS at a distance
r, it follows that there are no LOS BSs at a distance less or equal
to AL (r)
Assuming the standard user connects to a LOS BS at a distance
r, it follows that there are no NLOS BSs at a distance less or
equal to AN (r)
Probabilities that the standard user connects to a LOS or a NLOS
BS, respectively
Maximum transmit and receive antenna gains, respectively
Minimum transmit and receive antenna gains, respectively
Laplace transform of the interference component determined by
BSs placed on the upper (S = U) or bottom side (S = B) of
the road that are in LOS (E = L) or in NLOS (E = N) to
the standard user, conditioned on the serving BS being in LOS
(E1 = L) or NLOS (E1 = N)
Laplace transform of the interference I, given that the standard
user connects to a LOS BS (E1 = L) or NLOS BS (E1 = N)
SINR outage probability as a function of SINR threshold θ
Rate coverage probability as a function of target rate κ
coordinates of the BSs at the upper and bottom sides of the
road form two independent PPPs with density 0.5 · λBS .
Assumption 3.3 (Blockage Distribution): We assume that the
ℓ-th obstacle lane on a traffic direction and the coordinates
(o) (o)
(o)
(xℓ,i , yℓ,i ) of blockage i, xℓ,i belongs to a one-dimensional
PPP ΦO,ℓ with density λo,ℓ , for ℓ ∈ {1, . . . , No } [36]. The
(o)
term yℓ,i is equal to wℓ, or −wℓ, depending on whether we refer to the West-to-East or East-to-West direction, respectively.
We assume that the density of the blockages of lane ℓ in each
traffic direction is the same. Each blockage point is associated
with a segment of length τ , centered on the position of the
blockage itself and placed onto the horizontal symmetry axis
of the lane (hereafter referred to as the “footprint segment”).
Obstacles can be partially overlapped and the blockage widths
and heights are not part of our modeling. The presence of large
vehicles in the user lanes is regarded as sporadic hence, it is
ignored.
From Assumption 3.3, given a driving direction, we observe
that the blockage density of each obstacle lane can be different.
This means that our model has the flexibility to cope with
different traffic levels per obstacle lane; namely, the larger
the traffic density, the larger the traffic intensity. In a real (or
simulated) scenario, the obstacle density λo,ℓ of a road section
is function of the mobility model, which in turn depends on the
vehicle speed, maximum acceleration/deceleration, etc. Form
a logic point of view, in our theoretical model, we observe
that at the beginning of a time step, process ΦO,ℓ is sampled
and a new blockage position is extracted, for ℓ = 1, . . . , No .
In Section V, we will show that the considered PPP-based
mobility model provides a tight approximation of the investigated network performance, in the case of blockages moving
according to a Krauss car-following mobility model [43].
Our primary goal consists of characterizing the SINR outage
5
and rate coverage probability of users located on the user lanes,
as these are the most challenging to serve due to the fact
that vehicles in the other lanes can behave as blockages. For
the sake of tractability, our theoretical model tractable will
consider the service of a standard user placed at the origin
O = (0, 0) of the axis.
A. BS-Standard User Association and Antenna Model
Since vehicles in the slow lanes can block a direct link
between the standard user and each BS, it is necessary to
distinguish between BSs that are in LOS with the standard
user and those that are in NLOS. BS i is said to be in
LOS if the footprint of any blockage does not intersect with
the ideal segment connecting the standard user and BS i.
The probability that BS i is in LOS is denoted by pi,L . We
assume that the blockages are of length τ , illustrated in Fig. 1.
In the case that the ideal segment connecting BS i to the
standard user intersects with one or more footprint segments,
BS i is in NLOS (this occurs with probability pi,N ) and the
relation pi,N = 1 − pi,L holds. For generality, we also assume
that signals from NLOS BSs are not necessarily completely
attenuated by the blockages located in the far-field of the
antenna systems. This can happen when the main lobe of
the antenna is only partially blocked and because of signal
diffraction [44], [45]. By Assumption 3.3, we observe that the
probability pi,E for E ∈ {L, N} of BS i being in LOS (E = L)
or NLOS (E = N) depends on the distance from O. This is
due to the fact that the further the BS is from the user, the
further away the center of an obstacle footprint segment needs
to be to avoid a blockage.
Consider Assumption 3.3 and the set of points where the
segment connecting O with BS i intersects the symmetry axis
of each obstacle lane. We approximate pi,L with the probability
pL that no blockages are present within a distance of τ /2 on
either side of the ray connecting the user to BS i. Hence, our
approximation is independent of the distance from BS i to O
and pi,L can be approximated independently of i as follows:
pL ∼
=
No
Y
e−λo,ℓ τ ,
(1)
ℓ=1
while pi,N is approximated as pN = 1 − pL . Observe also that
the term e−λO,ℓ τ is the void probability of a one-dimensional
PPP of density λo,ℓ [42].
Using the approximation in (1) and invoking the independent thinning theorem of PPP, it follows that the PPP of the
LOS BSs ΦL ⊆ ΦBS and of the NLOS BS ΦN ⊆ ΦBS are
independent and with density λL = pL λBS and λN = pN λBS ,
respectively. In addition, the relation ΦL ∩ ΦNp= ∅ holds.
Consider the i-th BS at a distance ri = x2i + yi2 from
the standard user. The indicator function 1i,L is equal to one
if BS i is in LOS with respect to the standard user, and
zero otherwise. The path loss component ℓ(ri ) impairing the
signal transmitted by BS i and received by the standard user
is defined as follows:
ℓ(ri ) = 1i,L CL ri−αL + (1 − 1i,L )CN ri−αN
(2)
where αL and αN are the path loss exponents, while CL
and CN are the path loss intercept factors in the LOS and
NLOS cases, respectively. Terms CL and CN can either be
the result of measurements of being analytically derived by a
free space path loss model; the intercept factors are essential
to capture the path loss component at a target transmitterreceiver distance, which for practical mmWave system is equal
to 1 m [46]. From Assumption 3.1, we remark that relation
ri ≥ w(No + 1) holds. Hence, for typical values of road
lane widths, path loss intercept factors and exponents, relation
1
α
1
α
w(No + 1) ≥ max{CL L , CN N } holds as well. This ensures
that both CL ri−αL and CN ri−αN are less than or equal to 1.
Assumption 3.4 (BS-Standard User Association): In our
system model, the standard user has perfect channel state
information and always connects to the BS with index i∗ ,
which is characterized by the minimum path loss component,
i.e., i∗ = arg max {ℓ(ri )}.
i=1,...,b
We assume that the standard user connects
to a1 NLOS BS at
1
αL
α
a distance r. Since w(No + 1) ≥ max{CL , CN N }, it follows
that there are no LOS BSs at a distance less than or equal to
AL (r), defined as:
(
− 1 )
CN −αN αL
AL (r) = max w(No + 1),
.
(3)
r
CL
We observe that AL (r) is the distance from O for which the
path loss component associated with a LOS BS is equal to
the path loss component associated with a NLOS BS at a
distance r. In a similar way, we observe that if the standard
user connects to a LOS BS at a distance r from O, it follows
that there are no NLOS BSs at a distance smaller than or equal
to AN (r), defined as:
(
− 1 )
CL −αL αN
.
(4)
AN (r) = max w(No + 1),
r
CN
We observe that definitions (3) and (4) prevent AL (r) and
AN (r) to be smaller than the distance w(No + 1) between O
and a side of the road.
The standard user will always connect to one BS at a
time, which is either the closest LOS or the closest NLOS
BS. This choice is made by the standard user according
to Assumption 3.4. In particular, is closest LOS BS is at
a distance greater than AL (r), the BS associated with the
smallest path loss component is the closest NLOS BS, which
is at a distance r to the standard user. Those facts allow us to
prove the following lemma.
Lemma 3.1: Let dL and dN be the random variables expressing the distance to the closest LOS and NLOS BSs from the
perspective of the standard user, respectively. The Probability
Density Function (PDF) of dL can be expressed as:
fL (r) =
2λL r −2λL b(r)
e
,
b(r)
(5)
while the PDF of dN can be expressed as
fN (r) =
2λN r −2λN b(r)
e
,
b(r)
(6)
p
where b(r) = r2 − w2 (No + 1)2 .
Proof: Considering the LOS case, the proof directly
follows from the expression of the PDF of the distance of the
6
closest point to the origin of the axis in a one-dimensional PPP
with density λL , which is fL (t) = 2λL e−2λL t [42, Eq. (2.12)].
By applying the change of variable t ← b(r) we obtain (5).
With similar reasoning, it is also possible to prove (6).
Using Lemma 3.1, (3) and (4), the following lemma holds.
Lemma 3.2: The standard user connects to a NLOS BS with
probability
Z ∞
PN =
fN (r)e−2λL b(AL (r)) dr.
(7)
w(No +1)
On the other hand, the standard user connects to a LOS BS
with probability
Z ∞
PL =
fL (r)e−2λN b(AN (r)) dr = 1 − PN . (8)
w(No +1)
Proof: Consider the event that the standard user connects
to a NLOS BS, which is at a distance r from O. This event
is equivalent to have all the LOS BSs at a distance greater
than or equal to AL (r). From (5), it follows that P[dL ≥
AL (r) and dN = r] is equal to e−2λL b(AL (r)) . Then if we
marginalize P[dL ≥ AL (r) and dN = r] with respect to r, we
obtain (7). The same reasoning applies to the proof of (8).
The gain of the signal received by the standard user depends
on the antenna pattern and beam steering performed by the
BS and the user. Each BS and the standard user are equipped
with antenna arrays capable of performing directional beamforming. To capture this feature, we follow [23] and use
a sectored approximation to the array pattern. We detail
the sectored approximation for our highway model in the
following assumption.
Assumption 3.5 (Antenna Pattern): The antenna pattern
consists of a main lobe with beamwidth ψ and a side lobe
that covers the remainder of the antenna pattern. We assume
that the gain of the main lobe is GT X and the gain of the
sidelobe is gT X . Similarly, the antenna pattern of the standard
user also consists of main lobe with beamwidth ψ and gain
GRX , and a side lobe with gain gRX .
The antenna of each BS and the user can be steered as
follows.
Assumption 3.6 (BS Beam Steering): Let ǫi be the angle
between the upper (bottom) side of the road and the antenna
boresight of BS i (see Fig. 1). We assume that ǫi takes values
in G = [ ψ2 , 2π − ψ2 ]. As such, the main lobe of each BS is
always entirely directed towards the road portion constrained
by the upper and bottom side. If the standard user connects
to BS i, the BS steers its antenna beam towards the standard
user. On the other hand, if the standard user is not connected
to BS i, we assume that ǫi takes a value that is uniformly
distributed in G.
Assumption 3.7 (Standard User Beam Steering): The angle
ǫ(U) between the positive x-axis and the boresight of the user
beam is selected to maximize the gain of the received signal
from the serving BS. We assume that ǫ(U) ∈ [ ψ2 , π − ψ2 ] or
ǫ(U) ∈ [π + ψ2 , 2π − ψ2 ] if the user is served by a BS on the
upper side or the bottom side of the road respectively. This
assumption ensures that interfering BSs on the opposite side
of the road are always received by a sidelobe, with gain gRX .
We also assume that the standard user directs its antenna beam
towards the serving BS, which is then received with gain GRX .
IV. SINR O UTAGE AND R ATE C OVERAGE
C HARACTERIZATION
For the sake of simplifying the notation and without loss of
generality, we assume that the BS with index 1 is the BS that
the standard user is connected to, while BSs 2, . . . , b define the
set of the interfering BSs. We define the SINR at the location
of the standard user as follows:
b
SINRO =
X
|h1 |2 ∆1 ℓ(r1 )
|hj |2 ∆j ℓ(rj ).(9)
, where I =
σ+I
j=2
Terms hi and ∆i are the small-scale fading component and
the overall transmit/receive antenna gain associated with BS
i, respectively, for i = 1, . . . , b. The term I is the total
interference contribution determined by all the BSs except the
one connected to the standard user, i.e., the total interference
determined by BSs 2, . . . , b. From Assumption 3.6 and 3.7, it
follows that ∆1 is equal to GTX GRX . Finally, σ represents
the thermal noise power normalized with respect to the transmission power Pt .
As acknowledged in [7] there is a lack of extensive
measurements for vehicular mmWave networks, as well as
widely accepted channel models. Therefore, it is necessary
to adopt conservative assumptions on signal propagation.
As summarized in Table II, several channel models have
been proposed in the literature. Typically publicly available
system-level mmWave simulators [47] adopt channel models
entirely [23] or partially [33] based on the Nakagami model,
which are more refined alternatives to the widely adopted
models dictating constant small-scale contributions [31], [32],
[35]. In particular, we adopt the same channel model in [33],
which is based on the following observations: (i) because
of the beamforming capabilities and the sectorial antenna
pattern, the signal is impaired only by a limited number of
scatterers, and (ii) the interfering transmissions cluster with
many scatterers and reach the standard user. Furthermore,
the considered sectorial antennal model at the transmitter
and receiver sides (see Assumption 3.5) significantly reduces
the angular spread of the incoming signals – thus reducing the Doppler spread. Moreover, the incoming signals are
concentrated in one direction. Hence, it is likely that there
is a non-zero bias in the Doppler spectrum, which can be
compensated by the automatic frequency control loop at the
receiver side [48]. For these reasons, the Doppler effect has
been assumed to be mitigated.
Assumption 4.1 (Channel Model): The channel between the
serving BS and standard user is described by a Nakagami
channel model with parameter m, and hence, |h1 |2 follows
a gamma distribution (with shape parameter m and rate
equal to 1). On the other hand, to capture the clustering of
interfering transmissions the channels between the standard
user and each interfering BS are modelled as independent
Rayleigh channels – thus |h2 |2 , . . . , |hb |2 are independently
and identically distributed as an exponential distribution with
mean equal to 1.
7
TABLE IV
VALUES OF (a, b, ∆) FOR DIFFERENT < |x1 |, S1 , E1 , S, E >.
Configuration of
< S1 , E1 , S, E >
< U, L, U, L >
< U, L, U, N >
< U, L, B, L >
< U, L, B, N >
< U, N, U, L >
< U, N, U, N >
< U, N, B, L >
< U, N, B, N >
Cases where
S1 = B, S = B
Cases where
S1 = B, S = U
Conditions on |x1 |
Enumeration of elements
(a, b, ∆) ∈ C|x1 |,S1 ,E1 ,S,E
(|x1 |, K, gTX GRX ),
(K, +∞, gTX gRX ),
(|x1 |, +∞, gTX gRX )
(|x1 |, K, gTX GRX ),
(K, +∞, gTX gRX ),
For any |x1 |
(|x1 |, |J|, gTX GRX ),
such that J ≤ 0
(|J|, +∞, gTX gRX )
(xN (r1 ), J, gTX gRX ),
(xN (r1 ), +∞, gTX gRX ),
For any |x1 |
(J, K, gTX GRX ),
such that J > 0
(K, +∞, gTX gRX )
Refer to the case
< U, L, U, L > (J ≤ 0)
For any |x1 |
and replace |x1 |
such that J ≤ 0
with xN (r1 )
(|x1 |, +∞, gTX gRX ),
For any |x1 |
(|x1 |, +∞, gTX gRX ),
Refer to the case < U, L, B, L > and
replace |x1 | with xN (r1 )
Refer to the case
For any |x1 |
< U, L, B, L > and
such that xL (r1 ) > K
replace |x1 | with xL (r1 )
Refer to the case
For any |x1 |
< U, L, U, L > and
such that xL (r1 ) ≤ K
replace |x1 | with xL (r1 )
Refer to the case < U, L, U, L >
Refer to the case < U, L, B, L > and
replace x1 with xL (r1 )
Refer to the case < U, L, B, L >
Refer to the correspondent cases
where S1 = U and S = U
Refer to the correspondent cases
where S1 = U and S = B
For any |x1 |
such that J > 0
A. Analytical Characterization of I
In order to provide an analytical characterization of the
interference power at O, it is convenient to split the term I into
four different components: (i) IU,L and IU,N representing the
interference power associated with LOS and NLOS BSs placed
on the upper side of the road whose positions are defined by
the PPPs ΦU,L and ΦU,N , respectively, and, (ii) IB,L and IB,N
the interference power generated by LOS and NLOS BSs on
the bottom side of the road placed at the location given by the
PPPs ΦB,L and
P ΦB,N . Overall, the total interference power is
In addition, the relations
given by I =S S∈{U,B},E∈{L,N} IS,E .S
ΦL = ΦU,L ΦB,L and ΦN = ΦU,N ΦB,N hold.
In the following result, we derive an approximation for the
Laplace transform LI (s) of I.
Theorem 4.1: Let S1 = U and S1 = B represent the cases
where the standard user connects to a BS on the upper or the
bottom side of the road, respectively. In addition, let E1 = L
and E1 = N signify the cases where the standard user connects
to a LOS or NLOS BS, respectively. The Laplace transform
LIS,E ,E1 (s) of IS,E , conditioned on E1 , for S ∈ {U, B} and
E ∈ {L, N}, can be approximated as follows:
q
Y
(10)
LIS,E ,E1 (s; a, b, ∆),
LIS,E ,E1 (s) ∼
=
S1 ∈{U,B},
(a,b,∆)∈C|x1 |,S1 ,E1 ,S,E
where LIS,E ,E1 (s; a, b, ∆) is defined as in (30). We define
the x-axis coordinates J = w(No + 1)/[tan(ǫ(U) + ψ/2)]
and K = w(No + 1)/[tan(ǫ(U) − ψ/2)] of the points where
the two rays defining the standard user beam intersect with
a side of the road, where ǫ(U) = ∓ tan−1 [w(No + 1)/x1 ],
for S1 = U or B, respectively
(see Fig. 8). Furthermore,
p
2
2
2 and
let us define
x
(r
)
=
(A
(r
L
1
L
1 )) − w (No + 1)
p
2
2
2
xN (r1 ) = (AN (r1 )) − w (No + 1) . Different combinations of parameters < |x1 |, S1 , E1 , S, E > determine different
sequences C|x1 |,S1 ,E1 ,S,E of parameter configurations (a, b, ∆),
as defined in Table IV.
Proof: See Appendix A.
Example 4.1: Consider the scenario where the standard user
connects to a LOS BS, i.e., E1 = L, and relation J > 0
holds. We evaluate the Laplace transform of the interference
associated with the BSs located on the upper side of the road
(S = U) that are in LOS with respect to the standard user
(E = L). The sequence C|x1 |,U,L,U,L is given by the first row
of Table IV, while C|x1 |,B,L,U,L consists of the same elements
of sequence C|x1 |,U,L,B,L (last row of Table IV). As a result,
LIS,E ,E1 (s) can be approximated as follows:
h
LIS,E ,E1 (s) ∼
= LIS,E ,E1 (s; |x1 |, K, gTX GRX )
· LIS,E ,E1 (s; K, +∞, gTX gRX )
· LIS,E ,E1 (s; |x1 |, +∞, gTX gRX )
i1/2
· LIS,E ,E1 (s; xN (r1 ), +∞, gTX gRX ). (11)
From Theorem 4.1 and the fact that I is defined as a
sum of statistically independent interference components, the
following corollary holds.
Corollary 4.1: The Laplace transform of I, for E1 = {L, N},
can be approximated as follows:
Y
(12)
LIS,E ,E1 (s)
LI,E1 (s) ∼
=
S∈{U,B},E∈{L,N}
Example 4.2: Consider the scenario where E1 = L, and
relation J > 0 holds. Using Corollary 4.1, LIS,E ,E1 (s) can be
approximated as follows:
LI,E1 (s) ∼
= LIS,E ,E1 (s; |x1 |, K, gTX GRX )
· LIS,E ,E1 (s; xN (r1 ), J, gTX gRX )
· LIS,E ,E1 (s; J, K, gTX GRX )
2
· LIS,E ,E1 (s; K, +∞, gTX gRX )
3
· LIS,E ,E1 (s; |x1 |, +∞, gTX gRX )
· LIS,E ,E1 (s; xN (r1 ), +∞, gTX gRX )
3
(13)
B. SINR Outage and Rate Coverage Probability Framework
The general framework for evaluating the SINR outage
probability is given in the following result.
Theorem 4.2: Let
√2 2
2
FL (t) = e−2λL t −w (No +1)
(14)
and
FN (t) = e−2λN
√2
t −w 2 (No +1)2
(15)
be the probability of a LOS or NLOS BS not being at a
distance smaller than t from O, respectively. We regard PT (θ)
to be the SINR outage probability with respect to a threshold
8
θ, i.e., the probability that SINRO is smaller than a threshold
θ. PT (θ) can be expressed as follows:
PCL (θ)
z
}|
{
PT (θ) = PL − P[SINRO > θ ∧ std. user served in LOS]
+ PN − P[SINRO > θ ∧ std. user served in NLOS] (16)
|
{z
}
PCN (θ)
where
m−1
X
Z +∞
α
m
− vσθ(m−k) r L
PCL (θ) = −
(−1)
e ∆ 1 CL 1
k
w(No +1)
k=0
E1 =L
αL
vθr1 (m − k)
· LI,E1
fL (r1 )FN (AN (r1 )) dr1 (17)
∆1 CL
m−k
and
m−1
X
Z +∞
α
m
− vσθ(m−k) r N
PCN (θ) = −
(−1)
e ∆ 1 CN 1
k
w(No +1)
k=0
E1 =N
αN
vθr1 (m − k)
fN (r1 )FL (AL (r1 )) dr1 ,(18)
· LI,E1
∆1 CN
m−k
represent the probability of the standard user not experiencing SINR outage while connected to a LOS or NLOS BS,
respectively.
Proof: The result (16) follows immediately once PCL (θ)
and PCN (θ) as known. In particular, the following relation
holds (for E1 = L):
"
#
|h1 |2 ∆1 ℓ(r1 )
PCL (θ) = P
> θ ∧ std. user served in LOS
σ+I
Z +∞
αL m
(i)
−v (σ+I)θ
∆ 1 CL r1
∼
1− 1−e
= EI
w(No +1)
· fL (r1 )FN (AN (r1 )) dr1
(19)
where v = m(m!)−1/m [23, Lemma 6] and (i) arise from
|h1 |2 being distributed as a gamma random variable. In addition, FN (AN (r1 )) is defined as the probability of a NLOS
BS not being at a distance smaller than AN (r1 ) to O, i.e., the
probability that the standard user is not connected to a NLOS
BS. The expression of FN (t), as in (15), immediately follows
from the simplification of the following relation:
Z t
FN (t) = 1 −
fN (r) dr.
(20)
w(No +1)
From the binomial theorem, we swap the integral and the
expectation with respect to I and invoke Corollary 4.1 to
obtain (17). By following the same reasoning, it is also
possible to derive expressions for PCN (θ) and FL (t).
Remark 4.1: As the value of αN increases it is less likely that
the standard user connects to a NLOS BS. Hence, from (4),
AN is likely to be equal to w(No + 1). As a result, the
exponential term in (8) approaches one and, hence, PL can
be approximated as follows:
Z ∞
Z ∞
fL (r) dr =
2λL e−2λL t = 1.
(21)
PL ∼
=
w(No +1)
0
Using this approximation, it follows that PN ∼
= 0 holds. In
addition, since PCN is always less than or equal to PN , the
relation PCN ∼
= w(No + 1), the
= 0 holds as well. If AN ∼
relation FN (AN (r1 )) ∼
= 1 holds. For these reasons, PT (θ)
can be approximated as follows:
Z +∞
m−1
X
vσθ(m−k) αL
m
−
r
(−1)m−k
e ∆ 1 CL 1
PT (θ) ∼
=1+
k
w(No +1)
k=0
vθr1αL (m − k)
fL (r1 ) dr1 .
(22)
· LI,L
∆1 CL
In addition, should signals from NLOS BSs be entirely attenuated by blockages, PN would be equal to 0 and (22) would
hold as well.
Remark 4.2: Should the BSs be deployed only along the
horizontal line separating the two driving directions, the value
of pL be equal to 1 and, hence, PL = 1. Under this circumstances (22) would hold with minimal changes. For instance,
let us focus on the East-to-West driving direction, assume that
the standard user be located in the middle of the No + 1 lanes,
and that the origin of the coordinate system overlaps with
the standard user position. In this case, an interfering BS i is
associated with ǫi , which takes values uniformly distributed in
[0, 2π). Let us regard with U the upper-most edge of the Eastto-West driving direction, E1 is equal to L and hence, relation
LI,L (s) ∼
= LIU,L ,L (s) holds. This allows us to approximate
PT as in (22), where term w(No + 1) has to be replaced with
w(No + 1)/2.
From [41, Theorem 1] and by using Theorem 4.2, it is now
possible to express the rate coverage probability RC (κ), i.e.,
the probability that the standard user experiences a rate that
is greater than or equal to κ. In particular, the rate coverage
probability is given by:
RC (κ) = P[rate of std. user ≥ κ]
= 1 − PT (2κ/W − 1),
(23)
where W is the system bandwidth.
V. N UMERICAL R ESULTS
A. Simulation Framework
In order to validate the proposed theoretical model, we
developed a novel MATLAB simulation framework capable
of estimating the SINR outage and rate coverage probabilities
by means of the Monte Carlo approach. Both our simulator
and the implementation of the proposed theoretical framework
are available online [52].
We remark that Assumption 3.1 models the highway as
infinitely long, which is not possible in a simulation. However, as noted in [53], [54], the radius R of the simulated
system (i.e., the length of the simulated road section 2R)
can be related to the simulation accuracy error ε, as in [53,
Eq. (3.5)]. In the case of a one-dimensional PPP, the radius
− 1
is related to the simulation error by R ≥ ε αL −1 . We
superimpose a normal approximation of the binomial proportionh confidence
interval [55] topour simulation
i results defined
p
as p̂ − z p̂(1 − p̂)/n; p̂ + z p̂(1 − p̂)/n , where p̂ is the
simulated probability value, n is the number of Monte Carlo
iterations and z is the (1−0.5·e)-th quantile, for 0 ≤ e ≤ 1. In
particular, z is set equal to 0.99, defining a confidence intervals
9
1
TABLE V
M AIN SIMULATION PARAMETERS .
Blockage
dimensions
No
{λo,1 , λo,2 }
λBS
Carrier frequency f
CL , CN
αL
αN
m
φ
GTX
GRX
gTX , gRX
W
Pt
Thermal noise
power (i.e, σ · Pt )
0.96
Value
13 h (for Figs. 3-7), 55 h (for Fig. 2)
20 km, 100 km
3.7 m, as per [49]
2 · 10−2
Blockages and the standard user move according to a
Krauss car-following mobility model [43]; maximum
acceleration and deceleration equal to 5.3 m/s2 [50],
maximum vehicle speed equal to 96 km h−1
(blockages) and 112 km h−1 (standard user).
The dimensions of a double decker bus, i.e., length τ
equal to 11.2 m and width equal to 2.52 m [51]
1, 2
{1 · 10−2 , 2 · 10−2 }, i.e., one blockage every
{100 m, 50 m}
From 2 · 10−4 to 1 · 10−2 , with a step of 2 · 10−4
28 GHz
−20 log10 (4πf /c), which is the free space path loss in
dB at a distance of 1 m and c is the speed of light [46]
2.8
{4, 5.76}, as per [19]
3, as per [33]
{30◦ , 90◦ }
{10 dB, 20 dB}
10 dB
−10 dB
100 MHz
27 dBm
10 log10 (k · T · W · 103 ) dBm, where k is the
Boltzmann constant and the temperature
T = 290 K [18]
of 98% and n is equal to 2 · 105 (for Fig. 2) or 5 · 104 (for
Figs. 3-7).
We simulated scenarios where the standard user drives on
the lower-most user lane along with multiple other vehicles;
in particular, we considered a number of vehicles equal to
⌊2Rλu ⌋, where λu is the vehicle density driving on each user
lane. In addition, a number of blockages equal to ⌊2Rλo,i ⌋ are
placed at random on each obstacle lane i, for i = 1, . . . , No .
During each simulated scenario, both the vehicles driving on
the user lanes and blockages move according to the Krauss
car-following mobility model [43] and their maximum speed
is set equal to 70 mph (i.e., 112 km h−1 ) and 60 mph (i.e.,
96 km h−1 ), respectively as dictated by the current British
speed limits2 . In order to keep the density of the simulated
blockages constant and hence allow a fair validation of the
proposed theoretical framework, we adopted the Krauss carfollowing mobility model with the wrap-around policy. In
particular, when a vehicle reaches the end of the simulated
road section, it re-enters at the beginning.
BSs are positioned uniformly at random at both sides of
the road. The simulator estimates the SINR outage probability
PT and rate coverage probability RC by averaging over the
total simulation time and across a number of BS random
locations and steering angle configurations (of the interfering
BSs); number allowing to the simulated average performance
metrics to converge to a stable value. We remark that the
adoption of highly directional antennas significantly reduces
the angular spread of the incoming signals. As such, in the
2 We refers to the Highway Code (https://www.gov.uk/speed-limits) valid
for England, Scotland and Wales.
0.94
PL
Parameter
Simulated time
Length of the
simulated road
section (i.e., 2R)
w
λu
Mobility model
0.98
0.92
0.9
Avg. Blockage Duration
0.88
No
No
No
No
0.86
0.84
2
= 1,
= 1,
= 2,
= 2,
20
Simulation
Theory
Simulation
Theory
40
60
80
pL
pN
No = 1
0.78
0.22
0.19 s
2.5 s
No = 2
0.69
0.31
0.23 s
2.68 s
100
λBS · 104
120
140
Upper Side
160
Bottom Side
180
200
Fig. 2. Probability PL that the standard users connects to a LOS BS as a
function of λBS , for No = {1, 2} and αN = 4.
simulated scenarios, we assume the standard user is equipped
with an automatic frequency control loop compensating for the
Doppler effect [48]. In addition, the simulated channel follows
Assumption 4.1.
With regards to Table V, we consider No = {1, 2} obstacle
lanes per driving direction. For No = 2, we assume different
traffic intensities by setting densities {λo,1 , λo,2 } as per row
six of Table V. Furthermore, we consider a typical highway
lane width w [49].
In Section III-A, we approximated the probabilities pL and
pN for a BS of being in LOS or NLOS with respect to the
standard user, respectively. It is worth noting that approximation (1) has been invoked only in the derivation of the proposed
theoretical model. In contrast, in the simulated scenarios a BS
is in NLOS only if the ideal segment connecting the standard
user and the BS intersects with one or more vehicles in the
obstacle lanes and not just with their footprint segments.
Communications between the standard user and the serving
BS are impaired only by large vehicles (namely, trucks,
double-decker buses, etc.) driving on the obstacle lanes.
Specifically, we consider blockages with length (τ ) and width
of a double-decker bus [51]. Without loss of generality, both
the proposed theoretical framework and our simulations consider bi-dimensional scenarios. Although it is always possible
to deploy BSs having an antenna height sufficient to prevent
the vehicles in the obstacle lanes to behave as blockages, it
is not always feasible in practice. For instance, in a 4-lane
road section (No = 2) with w = 3.7 m where the standard
user drives in the middle of the lower-most user lane and the
user antenna height is 1.5 m, the BS antenna height should be
greater than 12.5 m to allow a blockage-free scenario, which
is at least twice as much the antenna height in a typical
LTE-A urban deployment [56]. Therefore, we assume that the
BS antenna height is 5 m, which means that vehicles in the
obstacle lanes always behave as blockages. For this reason, we
do not further consider the height of the vehicles in our study.
All the remaining simulation parameters are summarized in
Table V.
B. Theoretical Model Assessment
In order to numerically study our mmWave highway network and assess the accuracy of our theoretical model, we
first consider αN = 4 and a road section with a length
10
ψ = 30 , GTX
ψ = 90◦ , GTX
ψ = 30◦ , GTX
ψ = 90◦ , GTX
Simulation
Theory
◦
0.7
0.6
PT (θ)
0.5
= 10dB
= 10dB
= 20dB
= 20dB
0.4
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
05
0.8
0.7
0.6
0.5
10
15
20
0.3
0.2
0.2
0.1
0.1
5
15
θ (dB)
25
35
0
−5
45
= 10dB
= 10dB
= 20dB
= 20dB
0.4
0.3
0
−5
ψ = 30 , GTX
ψ = 90◦ , GTX
ψ = 30◦ , GTX
ψ = 90◦ , GTX
Simulation
Theory
◦
PT (θ)
0.8
5
(a) λBS = 10−2 , x-ISD = 100 m
ψ = 30 , GTX
ψ = 90◦ , GTX
ψ = 30◦ , GTX
ψ = 90◦ , GTX
Simulation
Theory
◦
0.7
0.6
PT (θ)
0.5
= 10dB
= 10dB
= 20dB
= 20dB
0.4
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
05
0.8
0.7
0.6
0.5
15
20
0.3
0.2
0.2
0.1
0.1
5
15
θ (dB)
25
15
35
45
(b) λBS = 4 · 10−3 , x-ISD = 250 m
= 10dB
= 10dB
= 20dB
= 20dB
0.4
0.3
0
−5
ψ = 30 , GTX
ψ = 90◦ , GTX
ψ = 30◦ , GTX
ψ = 90◦ , GTX
Simulation
Theory
◦
10
10
θ (dB)
15
25
20
35
45
(a) λBS = 10−2 , x-ISD = 100 m
PT (θ)
0.8
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
05
0
−5
5
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
05
10
15
θ (dB)
15
25
20
35
45
(b) λBS = 4 · 10−3 , x-ISD = 250 m
Fig. 3. SINR outage probability PT as a function of the threshold θ, for
No = 1, αN = 4, ψ = {30◦ , 90◦ } and GTX = {10 dB, 20 dB}.
Fig. 4. SINR outage probability PT as a function of the threshold θ, for
No = 2, αN = 4, ψ = {30◦ , 90◦ } and GTX = {10 dB, 20 dB}.
2R = 100 km, which ensures a simulation accuracy error
of at least 10−7.2 . In addition, the adoption of a relatively
small but realistic value of αN makes more likely for the
standard user to a connect to an NLOS BS [23] and hence,
allows us to effectively validate the proposed LOS/NLOS user
association model (see Lemma 3.2). Considering the density
λBS of process ΦBS , we ideally project the BSs onto the xaxis and we define their projected mean Inter-Site-Distance
(x-ISD) as 1/λBS .
Let us consider a Fig. 2 shows the probability of the standard
user connecting to a LOS BS as a function of λBS for one and
two obstacle lanes on each side of the road. The equivalent
x-ISD spans between 5 km (λBS = 2 ·10−4) and 50 m (λBS =
2 · 10−2 ). In particular, as typically happens, we observe that
PL is significantly greater than PN . Specifically, if No = 1
then, for λBS = 4 · 10−3 , the simulated value of PL is equal
to 0.94. When No increases to 2, the simulated value of PL
reduces to 0.92, for λBS = 4 · 10−3 .
Fig. 2 also compares our approximated theoretical expression of PL , as in (8), with the simulated one. We note
that (8) overestimates PL , and, hence, (7) underestimates PN .
However, we observe that: (i) for λBS ∈ [2 · 10−4 , 10−2 ], the
overestimation error is smaller than 0.03), and (ii) for dense
scenarios (λBS > 10−2 ), it never exceeds 0.01. Generally, we
observe that the proposed theoretical model follows the trend
of the simulated values. From Fig. 2, we also conclude that
PL may have a non-trivial minimum. In our scenarios, this is
particularly evident when No = 2.
Remark 5.1: As we move from sparse to dense scenarios,
it becomes more likely for a NLOS BS to be closer to the
standard user; thus PL decreases. However, this reasoning
holds up to a certain value of density. In fact, at some point,
the BS density becomes so high that it becomes increasingly
unlikely not to have a LOS BS that is close enough to serve the
standard user. This phenomenon may determine a non-trivial
minimum in PL .
The table superimposed to Fig. 2 lists the (simulated) values
of pL , pN and the average duration of a blockage event
impairing transmissions from BSs on the upper and bottom
side of the road. In particular, we observe that a blockage
event can occur with a probability greater than 0.22 and can
last up to 2.68 s3 .
Fig. 3 shows the effect of the SINR threshold θ on the outage probability PT (θ), for No = 1, several antenna beamwidth
ψ and a range of BS transmit antenna gains GTX . Here, the
vehicular receive antenna gain is set to GRX = 10 dB. In
Fig. 3a, the x-ISD is fixed at 100 m. It should be noted that
the proposed theoretical model, as in Theorem 4.2, not only
follows the trend of the simulated values of PT (θ) but also
it is a tight upper-bound for our simulations for the majority
of the values of θ. In addition, the deviation between theory
3 The standard user drives in the East-to-West direction. Hence, the East-toWest blockages have an (average) relative speed equal to 16 km h−1 (namely,
112 km h−1 − 96 km h−1 ). For blockages with a length equal to 11.2 m a
blockage event is expected to last about 2.5 s, which is close to the result of
our simulations. The same reasoning applies to West-to-East blockages.
11
0.06
0.05
No
No
No
No
No
No
No
No
0.04
0.16
0.03
PT (θ)
0.02
0.01
0.12
0
50
60
70
80
= 1,
= 1,
= 2,
= 2,
= 1,
= 1,
= 2,
= 2,
θ
θ
θ
θ
θ
θ
θ
θ
= 5dB, Simulation
= 5dB, Theory
= 5dB, Simulation
= 5dB, Theory
= 15dB, Simulation
= 15dB, Theory
= 15dB, Simulation
= 15dB, Theory
90
0.08
0.04
0
10
30
50
70
λBS · 104
90
110
130
150
Fig. 5. SINR outage probability PT as a function of the BS density λBS ,
for θ = {5 dB, 15 dB} dB, No = {1, 2}, αN = 4, ψ = 30◦ and
GTX = 20 dB.
1
0.95
0.9
RC (κ)
and simulation is negligible when θ ∈ [−5 dB, 15 dB] or
θ ∈ [−5 dB, 10 dB], for GTX = 10 dB or GTX = 20 dB,
respectively. On the other hand, that deviation gradually increases as θ becomes larger. Nevertheless, the maximum Mean
Squared Error (MSE) between simulation and theory is smaller
than 3.2 · 10−3 . Overall, we observe the following facts:
◦
◦
• Changing the beamwidth from 30 to 90 alters the SINR
outage probability only by a maximum of 4 · 10−2 . This
can be intuitively explained by noting that the serving
BS is likely to be close to the vertical symmetry axis of
our system model. From Assumption 3.7, the standard
user aligns its beam towards the serving BS. As such,
the values of J and K (see Theorem 4.1) do not largely
change on passing from ψ = 30◦ to ψ = 90◦ . Thus,
for the interference component to become substantial, the
value of ψ should be quite large.
• Overall, we observe that when the beamwidth increases,
so does PT . Intuitively, that is because the standard user
is likely to receive a large interference contribution via
the main antenna lobe.
• Increasing the value of the maximum transmit antenna
gain (from 10 dB to 20 dB) results in a reduction of the
SINR outage probability that, for large values of θ, can be
greater than 0.25. This clearly suggests that the increment
on the interfering power is always smaller than or equal
to the correspondent increment on the signal power. This
is mainly because of the directivity of the considered
antenna model and the disposition of the BSs.
Fig. 3b refers to the same scenarios as in Fig. 3a except for
the x-ISD that is equal to 250 m. In general, we observe that
the comments to Fig. 3a still hold. Furthermore, the impact of
the value of ψ on PT becomes negligible. Intuitively, this can
be explained by noting that the number of interfering BSs that
are going to be received by the standard user at the maximum
antenna gain decreases as λBS decreases. However, as the BS
density decreases (the BSs are more sparsely deployed), it
becomes more likely (up to a certain extent) that the number
of interfering BSs remains the same, even for a beamwidth
equal to 90◦ .
Fig. 4 refers to the same scenarios as Fig. 3 with two
obstacle lanes on each side of the road (No = 2). In addition
to the discussion for Fig. 3, we note the following:
• For the smallest value of the antenna transmit gain
(GTX = 10 dB), both the simulated and the proposed
theoretical model produce values of PT that are negligibly greater that those when No = 1.
• For x-ISD = 100 m and GTX = 20 dB, the SINR
outage is slightly greater that the correspondent case as
in Fig. 3a. In particular, for θ ≥ 25 dB, we observe an
increment in the simulated PT bigger than 9 · 10−2 .
• As soon as we refer to a sparser network scenario,
x-ISD = 250 m, the conclusions drawn for Fig. 3b also
apply for Fig. 4b. Hence, the impact of ψ on PT vanishes.
From Fig. 3 and Fig. 4, we already observed that the
proposed theoretical model, as in Theorem 4.2, follows well
the trend of the corresponding simulated values, and it is
characterized by an error that is negligible for the most
0.2
0.85
0.8
GTX
GTX
GTX
GTX
0.75
0.7
0
= 10dB,
= 10dB,
= 20dB,
= 20dB,
100
Simulation
Theory
Simulation
Theory
200
300
400
500
κ (Mbps)
600
700
800
900
Fig. 6. Rate coverage probability RC as a function of the threshold κ, for
αN = 4, ψ = 30◦ , GTX = {10 dB, 20 dB}, λBS = 4 · 10−3 , No = 2.
important values of θ (e.g., θ ≤ 20 dB). These facts are
further confirmed by Fig. 5, which shows the value of PT
as a function of λBS , for θ = 5 dB or 15 dB, and ψ = 30◦ . In
particular, as also shown in Fig. 3 and Fig. 4, as θ increases the
deviation between the simulations and the theoretical model
increases. However, the MSE between theory and simulation
never exceeds 5 · 10−3 in Figs. 4a and 4b. Furthermore, Fig. 5
allows us to expand what was already observed for Fig. 3 and
Fig. 4:
• As expected, PT increases as No increases. However,
when No passes from 1 to 2, PT increases no more
than 1 · 10−2 . Hence, we conclude that the network is
particularly resilient to the number of obstacle lanes.
• The impact of λBS on the value of PT more evident for
sparse scenarios – λBS ≤ 3 · 10−3 and λBS ≤ 5 · 10−3 ,
for θ = 5 dB and θ = 15 dB, respectively. Otherwise, the
impact of λBS is reasonably small, if compared to what
happens in a typical bi-dimensional mmWave cellular
network [23]. This can be justified by the same reasoning
provided for Fig. 3a.
• As the value of λBS increases, the interference component
progressively becomes dominant again and hence, PT is
expected to increase. In Fig. 5, this can be appreciated
for No = 2 and θ = 15 dB.
Let us consider again Fig. 5. In the considered scenarios, it is
possible to achieve a value of PT smaller than 0.2 for values
of λBS ∼
= 2.2 · 10−3 .
Fig. 6 shows the rate coverage probability as a function
of the rate threshold κ, for ψ = 30◦ , λBS = 4 · 10−3 and
No = 2. From (23), we remark that the expression of RC
12
0.06
0.2
0.05
No
No
No
No
No
No
No
No
0.04
0.16
0.03
PT (θ)
0.02
0.01
0.12
0
50
60
70
80
= 1,
= 1,
= 2,
= 2,
= 1,
= 1,
= 2,
= 2,
θ
θ
θ
θ
θ
θ
θ
θ
= 5dB, Simulation
= 5dB, Theory
= 5dB, Simulation
= 5dB, Theory
= 15dB, Simulation
= 15dB, Theory
= 15dB, Simulation
= 15dB, Theory
•
90
0.08
•
0.04
0
10
30
50
70
λBS · 104
90
110
130
150
Fig. 7. SINR outage probability PT as a function of the BS density λBS ,
for θ = {5 dB, 15 dB} dB, No = {1, 2}, αN = 5.76, ψ = 30◦ and
GTX = 20 dB. Simulation results obtained for 2R = 20 km.
directly follows from PT . For this reason, we observe that the
greater the gain GTX , the higher the value of RC . Finally, we
observe that the MSE between simulations and the proposed
theoretical approximation is smaller than 5.8 · 10−3 .
For completeness, our model was validated by considering
αN = 5.76 and a significantly shorter highway section, namely
2R = 20 km. We observe that the considered value of αN is
among the highest NLOS path loss exponent that has ever
been measured in an outdoor performance investigation [23].
In particular, Fig. 7 compares simulation and theoretical results
for the same transmission parameters and road layout as in
Fig. 5. The bigger NLOS path loss exponent determines bigger
PT values than the correspondent cases reported in Fig. 5 (the
absolute difference is bigger than 1.1 · 10−2 ). Nevertheless,
what observed for Fig. 5 applies also for Fig. 7. In particular,
we conclude that the proposed theoretical model remains valid
for shorter road sections.
VI. C ONCLUSIONS AND F UTURE D EVELOPMENTS
This paper has addressed the issue of characterizing the
downlink performance of a mmWave network deployed along
a highway section. In particular, we proposed a novel theoretical framework for characterizing the SINR outage probability
and rate coverage probability of a user surrounded by large
vehicles sharing the other highway lanes. Our model treated
large vehicles as blockages, and hence, they impact on the
developed LOS/NLOS model. One of the prominent features
of our system model is that BSs are systematically placed at
the side of the road, and large vehicles are assumed to drive
along parallel lanes. Hence, unlike a typical stochastic geometry system, we assumed that both BS and blockage positions
are governed by multiple independent mono-dimensional PPPs
that are not independent of translations and rotations. This
modeling choice allowed the proposed theoretical framework
to model different road layouts.
We compared the proposed theoretical framework with
simulation results, for a number of scenarios. In particular,
we observed that the proposed theoretical framework can
efficiently describe the network performance, in terms of
SINR outage and rate coverage probability. Furthermore, we
observed the following fundamental properties:
◦
◦
• Reducing the antenna beamwidth from 90 to 30 does
not necessarily have a disruptive impact on the SINR
outage probability, and hence, on the rate coverage probability.
In contrast with bi-dimensional mmWave cellular networks, the network performance is not largely impacted
by values of BS density ranging from moderately sparse
to dense deployments.
Overall, for a fixed SINR threshold, a reduced SINR
outage probability can be achieved for moderately sparse
network deployments.
A PPENDIX A
P ROOF OF T HEOREM 4.1
For E1 = {L, N}, the Laplace transform of IS,E can be
expressed as:
Y
2
LIS,E ,E1 (s) = EΦS,E
Eh E∆ e−s|hj | ∆j ℓ(rj ) (24)
j∈ΦS,E
(i)
− E∆ Eh
= exp
Z
+∞
w(No +1)
2rqλE
(1 − e−sh∆CE r
dr
·p
r2 − w2 (No + 1)2
!
−αE
)
(25)
where EΦS,E represents the expectation with respect to the
distance of each BS in ΦS,E from O. Similarly, operators
E∆ and Eh signify the expectation with respect to the overall
antenna gain and the small-scale fading gain associated with
the transmissions of each BS, respectively. For the sake of
compactness, from (i) onward we refer to |h|2 simply as
h. We observe that equality (i) arises from the definition
of a probability generating functional (pgfl) of a PPP [42,
Definition 4.3] and the mapping theorem applied to ΦS,E [42,
Theorem 2.34]. In addition, the pgfl allows us to drop the
relation to a specific BS j in the terms expressing a distance
of a BS to O, its channel and antenna gains. For this reason,
in the integrand function, we simply refer to terms r, h and
∆.
Let a and b be two real numbers greater than or equal to
w(No + 1) and such that a ≤ b. With regards to (25), we
condition with respect to a specific value of h and ∆, and we
approximate the following term4 :
Z b
−αE
2rqλE
dr
(1 − e−sh∆CE r
)p
2
r − w2 (No + 1)2
a
Z √b2 −w2 (No +1)2
2
2
2 −αE /2
(i)
= √
2qλE dt
1 − e−sh∆CE (t +w (No +1) )
2
2
2
a −w (No +1)
(ii)
∼
=
Z
a
(iii)
b
−αE
1 − e−sh∆CE t
2qλE dt
= −2qλE
Z
b−αE
a−αE
(26)
−1
−αE
(1 − e−sh∆CE x )α−1
E x
−1
dx
Θ(h,∆)
}|
{
i −αE
h
−1 b
(iv)
= 2qλE (1 − e−sh∆CE x )x−αE
x=a−αE
Z b−αE
− 1
−2qλE
sh∆CE x αE e−sh∆CE x dx,
−α
a E
{z
}
|
z
Λ(h,∆)
4 For
clarity, we define [f (x)]bx=a = f (b) − f (a).
(27)
13
Term Eh [Λ(h, ∆)] can be found as follows:
Z b−αE
− 1
Eh [Λ(h, ∆)] = −2qλE
x αE
−αE
Z ∞ a
·
sh∆CE e−(s∆CE x+1)h dh dx
(29)
0
b−αE
1
−
dx
= −2qλE
x
s∆CE x + 1
a−αE
− α1
Z −(s∆CE b−αE +1)−1
1
E
1
(i)
− −1
= −2qλE (s∆CE ) αE
dt
t
−(s∆CE a−αE +1)−1
Z
(ii)
= −2qλE (s∆CE )
·2 F̃1
1
αE
"
t(−t
− α1
E
∂
∂x
1
−1 − αE
)
1
1 1
,
+ 1;
+ 2; −t
αE αE
αE
Γ
1
+1
αE
#−(s∆CE b−αE +1)−1
y
= exp
− Eh [Θ(h, ∆)] + Eh [Λ(h, ∆)]
a=0,b=+∞
,
S1 = U, E1 = L, S = U and E = L - we divide this case
into the following subcases:
– If the value of |x1 | is such that J > 0 - we observe
that there are no LOS BSs at a distance smaller than
|x1 |. Hence, it follows that
|x1 | ≤ r ≤ K
g = GRX if
(31)
p = RX
g = gRX
if
K ≤ r ≤ +∞
|x1 | ≤ r ≤ +∞
or
(32)
p = RX
p = LX
– If J ≤ 0 - by following the same reasoning as before,
in addition to the case as in (31), it follows that
K ≤ r ≤ +∞
|J| ≤ r ≤ +∞
g = gRX if
or
(33)
p = RX
p = LX
(30)
!
.
5 We observe that z is always a real number, which allows us to significantly
reduce the complexity of the whole numerical integration process [57].
x
With regards to (24), we observe that after conditioning on
the standard user being connected to a BS at a distance r1 from
O, then the receive antenna gain g of the interfering BSs j is
determined by the parameter list < |x1 |, p, r, S1 , E1 , S, E >,
where: (i) |x1 | is the absolute value of the x-axis coordinate
of BS 1, (ii) p captures the fact that the interfering BS is at
a location on the positive (right-hand side of y-axis, RX) or
negative side (left-hand side of y-axis, LX) of the x-axis, and
(iii) r is the distance of the interfering BS to O.
Let us consider the x-axis coordinates J and K of the points
where the two rays defining the antenna beam of the standard
user intersect the side of the road, as shown in Fig. 8. The
receive antenna gain of the interfering BS also depends on: (i)
the fact the standard user connects to a BS on the upper/bottom
side of the road (S1 ∈ {U, B}), that can be in LOS/NLOS
(E1 ∈ {L, N}) with respect to the standard user, (ii) the
values of S and E, and (iii) the specific configuration of the
values of |x1 |, J and K. By invoking
the same approximation
p
r2 − w2 (No + 1)2 and the
as in (26), we say that r ∼
=
following parameters determine the receiver gain:
•
Let us focus on the transmit antenna gain of the j-th
interfering BS, which has a PDF that depends on the distance
rj and the orientation of the beam ǫi . To take into account the
exact formulation of the BS transmit antenna gain would make
the performance model intractable. As such, we instead make
the approximation that the transmit antenna gain is always
equal to gTX .
K
O x1
Fig. 8. Case where the standard user is served by a BS from the upper side
of the road.
t=−(s∆CE a−αE +1)−1
a=0,b=+∞
ǫ(U)
J
where (i) arises from the change of variable t ← − s∆C1E x+1 .
P∞ {a}k {b}k zk
Let us signify with 2 F̃1 (a, b; c; z) =
k=0
{c}k
k! the
5
Gauss hypergeometric function . We observe that the integral as in equality (i) is closely related to that as in [58,
Eq. (3.228.3)], and after some manipulations we have equality (ii).
p it follows that
p From the approximation in (26),
b2 − w2 (No + 1)2 .
a∼
=
= a2 − w2 (No + 1)2 and b ∼
Hence, we observe that LIS,E ,E1 (s), conditioned on the gain
∆ (see (9)), can be expressed as follows:
LIS,E ,E1 (s) ∼
= LIS,E ,E1 (s; a, b, ∆)
r1
w(No + 1)
where
p (i) arises from the change of variable
t ← r2 − w2 (No + 1)2 , while (ii) assumes that w(No + 1)
is equal to 0 (see Section V-B for the validation of the
proposed theoretical framework). Equality (iii) arises by
applying the changes of variable y ← tαE and then x ← y −1 .
In addition, in (iv), we resort to an integration by parts.
With regards to (27), we keep the conditioning to ∆ and
calculate the expectation of Θ(h, ∆) and Λ(h, ∆), with respect
to h. From Assumption 4.1, it should be noted that we refer to
a Rayleigh channel model, and, hence, the following relation
holds:
b−αE
−1
1
. (28)
Eh [Θ(h, ∆)] = 2qλE x−αE 1 −
s∆CE x + 1 x=a−αE
g = GRX
•
if
|x1 | ≤ r ≤ |J|
p = LX
(34)
S1 = U, E1 = L, S = U and E = N - we apply the
same reasoning as before by bearing in mind that it is
impossible for a NLOS BS to be at a distance that is
smaller than AN (r1 ) to O. Equivalently, it is impossible
for a NLOS BS to be associated
with a x-axis coordinate
p
smaller than xN (r1 ) = (AN (r1 ))2 − w2 (No + 1)2 . In
particular, for J ≤ 0, the value of g can be derived as
in (31) and (33)-(34), where term |x1 | is replaced by
14
xN (r1 ). On the other hand, for J > 0, the value of g can
be expressed as follows:
g = gRX
if
or
xN (r1 ) ≤ r ≤ J or K ≤ r ≤ +∞
p = RX
xN (r1 ) ≤ r ≤ +∞
(35)
p = LX
∼
= exp −
(ii)
g = GRX
•
(36)
S1 = U, E1 = L, S = B and E = L - from
Assumption 3.7, we observe that g is always equal to
gRX . In addition, we note that it is not possible to have
a LOS BS at a distance smaller than r1 . Hence, we have
only two possible configurations:
g = gRX
•
if
if
|x1 | ≤ r ≤ +∞
|x1 | ≤ r ≤ +∞
or
(37)
p = RX
p = LX
S1 = U, E1 = L, S = B and E = N - similarly to the
previous case, we observe that g is equal to gRX and it
is not possible to have a NLOS BS at a distance smaller
than xN (r1 ). Hence, we have the following cases:
xN (r1 ) ≤ r ≤ +∞
g = gRX if
p = LX
(38)
• With regards the remaining parameter combinations
where S1 = U, E1 = N, we observe the following cases:
xN (r1 ) ≤ r ≤ +∞
or
p = RX
– S
p = U, E = L - we define xL (r1 ) =
(AL (r1 ))2 − w2 (No + 1)2 . If xL (r1 ) > K, refer
to (37) and replace |x1 | with xL (r1 ). Otherwise, refer
to (31)-(34) and replace |x1 | with xL (r1 ).
– S = U, E = N - refer to (31)-(34).
– S = B, E = L - refer to (37) and replace |x1 | with
xL (r1 ).
– S = B, E = N - refer to (37).
•
By following the above approach, it is possible to derive all the remaining configurations. In particular, the
characterization of g, for a parameter configuration where
S1 = B and S = B (S1 = B and S = U) follows exactly
the same rule of the corespondent parameter list, where
S1 = U and S = U (S1 = U and S = B).
The aforementioned parameter configurations are also summarized in Table IV. With regards to parameter p, we observe
that the probability P[p] of p being equal to DX or RX is 0.5.
Consider (30), all the elements are in place to explicitly
calculate the expectation with respect to ∆. In particular, it
follows that LIS,E (s) can be expressed as:
(i)
LIS,E,E1 (s) ∼
= exp −E∆ Eh [Θ(h, ∆)+Λ(h, ∆)]
a=0,b=+∞
X
P[p] Eh [Θ(h, ∆) + Λ(h, ∆)]
S1 ∈{U,B}
(a,b,∆)∈C|x1 |,S1 ,E1 ,S,E
=
Y
1
exp − Eh [Θ(h, ∆)]
2
S1 ∈{U,B},
(a,b,∆)∈C|x1 |,S1 ,E1 ,S,E
=
J ≤r≤K
p = RX
Y
S1 ∈{U,B},
(a,b,∆)∈C|x1 |,S1 ,E1 ,S,E
+ Eh [Λ(h, ∆)]
a,b,∆
q
LIS,E ,E1 (s; a, b, ∆),
a,b,∆
a,b,∆
(39)
where (i) is (30). From the previous discussion, for a given
|x1 |, ∆ can either be equal to gTX gRX or gTX GRX . In particular, the value of ∆ is determined by the list of parameters
< |x1 |, S1 , E1 , S, E >, where terms a and b are the minimum
and maximum distance r to an interfering BS, respectively.
We define sequence C|x1 |,S1 ,E1 ,S,E . This sequence consists
of all the possible parameter configurations (a, b, ∆). For
instance, if S1 = U, E1 = L, S = U, E = L and J > 0,
sequence C|x1 |,S1 ,E1 ,S,E consists of: (|x1 |, K, gTX GRX ),
(K, +∞, gTX gRX ) and (|x1 |, +∞, gTX gRX ). We note that
each element of a sequence C|x1 |,S1 ,E1 ,S,E occurs with
probability P[p]. Furthermore, we observe that like those
in (37) and (38), sequence C|x1 |,S1 ,E1 ,S,E lists twice the
same parameter configuration. Given these reasons, the term
E∆ (Eh [Θ(h, ∆) + Λ(h, ∆)] a=0,b=+∞ ) can be approximated
as in (ii). After some manipulations of (ii), we get to (10),
which concludes the proof.
R EFERENCES
[1] D.
Evans,
“The
Internet
of
Things,”
Cisco
IBSG,
Tech.
Rep.,
Apr.
2011.
[Online].
Available:
https://www.cisco.com/c/dam/en us/about/ac79/docs/innov/IoT IBSG 0411FINAL.p
[2] “Preliminary Statement of Policy Concerning Automated Vehicles,”
U.S. Department of Transportation, National Highway Traffic Safety
Administration (NHTSA), Tech. Rep., 2013. [Online]. Available:
http://www.nhtsa.gov/staticfiles/rulemaking/pdf/Automated Vehicles Policy.pdf
[3] “C-ITS
Platform,”
European
Commission,
Tech.
Rep.,
Jan.
2016.
[Online].
Available:
http://ec.europa.eu/transport/themes/its/doc/c-its-platform-final-report-january-2016.p
[4] “5G-PPP White Paper on Automotive Vertical Sector,” 5G Infrastructure
Public Private Partnership, Tech. Rep., Oct. 2015. [Online]. Available:
https://5g-ppp.eu/wp-content/uploads/2014/02/5G-PPP-White-Paper-on-Automotive[5] N. Lu, N. Cheng, N. Zhang, X. Shen, and J. W. Mark, “Connected
Vehicles: Solutions and Challenges,” IEEE Internet Things J., vol. 1,
no. 4, pp. 289–299, Aug. 2014.
[6] E. Uhlemann, “Connected-Vehicles Applications Are Emerging,” IEEE
Veh. Technol. Mag., vol. 11, no. 1, pp. 25–96, Mar. 2016.
[7] J. Choi, V. Va, N. Gonzalez-Prelcic, R. Daniels, C. R. Bhat, and R. W.
Heath, “Millimeter-Wave Vehicular Communication to Support Massive
Automotive Sensing,” IEEE Commun. Mag., vol. 54, no. 12, pp. 160–
167, Dec. 2016.
[8] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K.
Soong, and J. C. Zhang, “What Will 5G Be?” IEEE J. Sel. Areas
Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[9] A. Ghosh, T. A. Thomas, M. C. Cudak, R. Ratasuk, P. Moorut,
F. W. Vook, T. S. Rappaport, G. R. MacCartney, S. Sun, and S. Nie,
“Millimeter-Wave Enhanced Local Area Systems: A High-Data-Rate
Approach for Future Wireless Networks,” IEEE J. Sel. Areas Commun.,
vol. 32, no. 6, pp. 1152–1163, Jun. 2014.
[10] J. B. Kenney, “Dedicated Short-Range Communications (DSRC) Standards in the United States,” Proceedings of the IEEE, vol. 99, no. 7, pp.
1162–1182, Jul. 2011.
15
[11] G. Karagiannis, O. Altintas, E. Ekici, G. Heijenk, B. Jarupan, K. Lin, [33] D. Maamari, N. Devroye, and D. Tuninetti, “Coverage in mmWave Celand T. Weil, “Vehicular Networking: A Survey and Tutorial on Relular Networks With Base Station Co-Operation,” IEEE Trans. Wireless
quirements, Architectures, Challenges, Standards and Solutions,” IEEE
Commun., vol. 15, no. 4, pp. 2981–2994, Apr. 2016.
Commun. Surveys Tuts., vol. 13, no. 4, pp. 584–616, Fourth 2011.
[34] R. Baldemair, T. Irnich, K. Balachandran, E. Dahlman, G. Mildh,
Y. Selén, S. Parkvall, M. Meyer, and A. Osseiran, “Ultra-Dense Net[12] E. Uhlemann, “Introducing Connected Vehicles,” IEEE Veh. Technol.
works in Millimeter-Wave Frequencies,” IEEE Commun. Mag., vol. 53,
Mag., vol. 10, no. 1, pp. 23–31, Mar. 2015.
no. 1, pp. 202–208, Jan. 2015.
[13] E. G. Strom, “On Medium Access and Physical Layer Standards for
Cooperative Intelligent Transport Systems in Europe,” Proceedings of [35] M. D. Renzo, “Stochastic Geometry Modeling and Analysis of MultiTier Millimeter Wave Cellular Networks,” IEEE Trans. Wireless Comthe IEEE, vol. 99, no. 7, pp. 1183–1188, Jul. 2011.
mun., vol. 14, no. 9, pp. 5038–5057, Sep. 2015.
[14] M. Xie, Y. Shang, Z. Yang, Y. Jing, and H. Zhou, “A Novel MBSFN
Scheme for Vehicle-to-Vehicle Safety Communication Based on LTE [36] M. J. Farooq, H. ElSawy, and M. S. Alouini, “A Stochastic Geometry
Model for Multi-Hop Highway Vehicular Communication,” IEEE Trans.
Network,” in Proc. of IEEE VTC-Fall 2015, Boston, Massachusetts,
Wireless Commun., vol. 15, no. 3, pp. 2276–2291, Mar. 2016.
USA, Sep. 2015.
[15] G. Araniti, C. Campolo, M. Condoluci, A. Iera, and A. Molinaro, “LTE [37] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and
M. Franceschetti, “Stochastic Geometry and Random Graphs for
for Vehicular Networking: A Survey,” IEEE Commun. Mag., vol. 51,
the Analysis and Design of Wireless Networks,” IEEE J. Sel. Areas
no. 5, pp. 148–157, May 2013.
Commun., vol. 27, no. 7, pp. 1029–1046, Sep. 2009.
[16] A. Tassi, M. Egan, R. J. Piechocki, and A. Nix, “Wireless Vehicular
[38] S. P. Weber, X. Yang, J. G. Andrews, and G. de Veciana, “Transmission
Networks in Emergencies: A Single Frequency Network Approach,” in
Capacity of Wireless Ad-Hoc Networks with Outage Constraints,” IEEE
Proc. of SigTelCom 2017, Da Nang, Vietnam, VN, Jan 2017.
Trans. Inf. Theory, vol. 51, no. 12, pp. 4091–4102, Dec. 2005.
[17] A. Tassi, I. Chatzigeorgiou, and D. E. Lucani, “Analysis and Optimiza[39] C.-H. Lee, C.-Y. Shih, and Y.-S. Chen, “Stochastic Geometry Based
tion of Sparse Random Linear Network Coding for Reliable Multicast
Models for Modeling Cellular Networks in Urban,” Springer Wireless
Services,” IEEE Trans. Commun., vol. 64, no. 1, pp. 285–299, Jan 2016.
Networks, vol. 19, no. 6, pp. 1063–1072, 2013.
[18] T. Rappaport, R. Heath, R. Daniels, and J. Murdock, Millimeter Wave
[40] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic Geometry for
Wireless Communications, ser. Communication engineering and emergModeling, Analysis, and Design of Multi-Tier and Cognitive Cellular
ing technologies. Prentice Hall, 2014.
Wireless Networks: A Survey,” IEEE Commun. Surveys Tuts., vol. 15,
[19] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao,
no. 3, pp. 996–1019, Third 2013.
and J. I. Tamir, “Broadband Millimeter-Wave Propagation Measurements
[41] S. Singh, H. S. Dhillon, and J. G. Andrews, “Offloading in Heterogeand Models Using Adaptive-Beam Antennas for Outdoor Urban Cellular
neous Networks: Modeling, Analysis, and Design Insights,” IEEE Trans.
Communications,” IEEE Trans. Antennas Propag., vol. 61, no. 4, pp.
Wireless Commun., vol. 12, no. 5, pp. 2484–2497, May 2013.
1850–1859, Apr. 2013.
[42] M. Haenggi, Stochastic Geometry for Wireless Networks, ser. Stochastic
[20] H. Shokri-Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, and
Geometry for Wireless Networks. Cambridge University Press, 2013.
M. Zorzi, “Millimeter Wave Cellular Networks: A MAC Layer Per[43] V. Kanagaraj, G. Asaithambi, C. N. Kumar, K. K. Srinivasan, and
spective,” IEEE Trans. Commun., vol. 63, no. 10, pp. 3437–3458, Oct.
R. Sivanandan, “Evaluation of Different Vehicle Following Models
2015.
Under Mixed Traffic Conditions,” Procedia - Social and Behavioral
[21] “Leading
train
operators
granted
innovation
funding,”
Sciences, vol. 104, pp. 390–401, 2013.
RSSB,
Tech.
Rep.,
Jul.
2015.
[Online].
Available:
[44] A. Loch, A. Asadi, G. H. Sim, J. Widmer, and M. Hollick, “mm-Wave
https://www.rssb.co.uk/Library/about-rssb/2015-07-press-release-leading-train-operators-granted-innovation-funding.pdf
on Wheels: Practical 60 GHz Vehicular Communication Without Beam
Training,” in Proc. of COMSNETS 2017, Bangalore, India, IN, Jan. 2017.
[22] J. Gozalves, “Fifth-Generation Technologies Trials,” IEEE Veh. Technol.
[45] I. Rodriguez, H. C. Nguyen, T. B. Sorensen, J. Elling, J. A. Holm,
Mag., vol. 11, no. 2, pp. 5–13, Jun. 2016.
P. Mogensen, and B. Vejlgaard, “Analysis of 38 GHz mmWave Propa[23] T. Bai and R. W. Heath Jr., “Coverage and Rate Analysis for Millimetergation Characteristics of Urban Scenarios,” in Proc. of 21th European
Wave Cellular Networks,” IEEE Trans. Wireless Commun., vol. 14, no. 2,
Wireless Conference, Budapest, Hungary, HU, May 2015.
pp. 1100–1114, Feb. 2015.
[24] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. [46] S. Sun, G. R. MacCartney, and T. S. Rappaport, “Millimeter-Wave
Distance-dependent Large-scale Propagation Measurements and Path
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter Wave
Loss Models for Outdoor and Indoor 5G systems,” in Proc. of EuCAP
Mobile Communications for 5G Cellular: It Will Work!” IEEE Access,
2016, Davos, Switzerland, CH, Apr. 2016.
vol. 1, pp. 335–349, 2013.
[47] “ns3-mmwave.”
[Online].
Available:
[25] S. Hasan, N. Siddique, and S. Chakraborty, Intelligent Transport Syshttps://github.com/nyuwireless/ns3-mmwave
tems: 802.11-based Roadside-to-Vehicle Communications.
Springer
[48] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadNew York, 2012.
band systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101–107, Jun.
[26] C. Lochert, B. Scheuermann, C. Wewetzer, A. Luebke, and M. Mauve,
2011.
“Data Aggregation and Roadside Unit Placement for a VANET Traffic
[49] P. Ioannou, Automated Highway Systems. Springer US, 2013.
Information System,” in Proc. of the fifth ACM international workshop
[50] D. Fambro, K. Fitzpatrick, and R. Koppa, Determination of Stopping
on VehiculAr Inter-NETworking, San Francisco, California, USA, 2008.
Sight Distances, ser. Determination of stopping sight distances. National
[27] O. Trullols, M. Fiore, C. Casetti, C. Chiasserini, and J. B. Ordinas,
Academy Press, 1997.
“Planning roadside infrastructure for information dissemination in intel[51] “Routemaster
Spec
Sheet,”
Wrightbus,
ligent transportation systems,” ELSEVIER Computer Communications,
Tech.
Rep.,
2014.
[Online].
Available:
vol. 33, no. 4, pp. 432 – 442, 2010.
http://www.wrightsgroup.com/datasheets/Routemaster%20spec%20sheet.pdf
[28] A. B. Reis, S. Sargento, F. Neves, and O. Tonguz, “Deploying Roadside
[52] “Proposed simulation framework and theoretical model.” [Online].
Units in Sparse Vehicular Networks: What Really Works and What Does
Available: https://github.com/andreatassi/V2X-MMWAVE
Not,” IEEE Trans. Veh. Technol., vol. 63, no. 6, pp. 2794–2806, Jul.
[53] M. Haenggi and R. K. Ganti, “Interference in Large Wireless Networks,”
2014.
Foundations and Trends in Networking, Now Publishers, vol. 3, no. 2,
[29] K. Rostamzadeh and S. Gopalakrishnan, “Analysis of Message Delivery
pp. 127–248, 2008.
Delay in Vehicular Networks,” IEEE Trans. Veh. Technol., vol. 64,
[54] S. Weber and M. Kam, “Computational Complexity of Outage Probno. 10, pp. 4770–4779, Oct. 2015.
ability Simulations in Mobile Ad-Hoc Networks,” in Proc. of the
[30] N. Akhtar, S. C. Ergen, and O. Ozkasap, “Vehicle Mobility and Commu39th Annual Conference on Information Sciences and Systems (CISS),
nication Channel Models for Realistic and Efficient Highway VANET
Baltimore, Maryland, USA, Mar. 2005.
Simulation,” IEEE Trans. Veh. Technol., vol. 64, no. 1, pp. 248–262,
[55] T. Ryan, Modern Engineering Statistics. Wiley, 2007.
Jan. 2015.
[56] “3GPP TR 36.873,” 3GPP, Tech. Rep., 2013.
[31] G. Zhang, T. Q. S. Quek, M. Kountouris, A. Huang, and H. Shan,
[57] R. K. S. Hankin, “Numerical evaluation of the Gauss hypergeometric
“Fundamentals of Heterogeneous Backhaul Design - Analysis and
function with the hypergeo package,” The R Journal, vol. 7, no. 2, pp.
Optimization,” IEEE Trans. Commun., vol. 64, no. 2, pp. 876–889, Feb.
81–88, Dec. 2015.
2016.
[58] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products,
[32] S. Singh, M. N. Kulkarni, A. Ghosh, and J. G. Andrews, “Tractable
ser. Table of Integrals, Series, and Products Series. Elsevier Science,
Model for Rate in Self-Backhauled Millimeter Wave Cellular Networks,”
2007.
IEEE J. Sel. Areas Commun., vol. 33, no. 10, pp. 2196–2211, Oct. 2015.
| 7cs.IT
|
arXiv:1703.03868v2 [cs.AI] 23 May 2017
Front-to-End Bidirectional Heuristic Search with Near-Optimal Node Expansions
Jingwei Chen
Dept. of Comp. Sci.
University of Denver
USA
Robert C. Holte
Comp. Sci. Dept.
University of Alberta
Canada
Sandra Zilles
Comp. Sci. Dept.
University of Regina
Canada
[email protected]
[email protected]
[email protected]
Abstract
It is well-known that any admissible unidirectional
heuristic search algorithm must expand all states
whose f -value is smaller than the optimal solution
cost when using a consistent heuristic. Such states
are called “surely expanded” (s.e.). A recent study
characterized s.e. pairs of states for bidirectional
search with consistent heuristics: if a pair of states
is s.e. then at least one of the two states must be
expanded. This paper derives a lower bound, VC,
on the minimum number of expansions required
to cover all s.e. pairs, and present a new admissible front-to-end bidirectional heuristic search algorithm, Near-Optimal Bidirectional Search (NBS),
that is guaranteed to do no more than 2VC expansions. We further prove that no admissible frontto-end algorithm has a worst case better than 2VC.
Experimental results show that NBS competes with
or outperforms existing bidirectional search algorithms, and often outperforms A* as well.
1
Introduction
One method of formally assessing the efficiency of a heuristic search algorithm is to establish upper and lower bounds
on the number of nodes it expands on any given problem
instance I. Such bounds can then be compared to a theoretical minimum that a competing heuristic search algorithm
would have to expand on I. In this context, one is interested
in finding sufficient conditions for node expansion, i.e., conditions describing nodes that must provably be expanded by
any competing algorithm. In a unidirectional search an algorithm must expand every node whose f -value is less than the
optimal solution cost; this condition establishes the optimality of A* [Dechter and Pearl, 1985].
Sufficient conditions for node expansion have recently
been developed for front-to-end bidirectional heuristic
search [Eckerle et al., 2017], but no existing front-to-end bidirectional search algorithm is provably optimal. In this paper,
we use these sufficient conditions to derive a simple graphtheoretic characterization of nodes that must provably be expanded on a problem instance I by any admissible front-toend bidirectional search algorithm given a consistent heuristic. In particular, the set of nodes expanded must correspond
Nathan R. Sturtevant
Dept. of Comp. Sci.
University of Denver
USA
[email protected]
to a vertex cover of a specific graph derived from I. We
then adapt a known vertex cover algorithm [Papadimitriou
and Steiglitz, 1982] into a new admissible front-to-end bidirectional search algorithm, NBS (Near-Optimal Bidirectional
Search), and prove that NBS never expands more than twice
the number of nodes contained in a minimum vertex cover.
Hence, the number of nodes expanded by NBS is provably
within a factor of two of optimal.
We further establish that no admissible bidirectional frontto-end algorithm can be better than NBS in the worst case.
In that sense, we formally verify that NBS is near-optimal
in the general case and optimal in the worst case. In an experimental study on a set of standard benchmark problems,
NBS either competes with or outperforms existing bidirectional search algorithms, and it often outperforms the unidirectional algorithm A*, especially when the heuristic is weak
or the problem instance is hard.
2
Related Work
Bidirectional search has a long history, beginning with bidirectional brute force search [Nicholson, 1966], and proceeding to heuristic search algorithms such as BHPA [Pohl, 1971].
Other notable algorithms include BS* [Kwa, 1989], which
avoids re-expanding states in both directions, and MM [Holte
et al., 2016], which ensures that the search frontiers meet in
the middle. Along with these algorithms there have been explanations for the poor performance of bidirectional heuristic
search, including that the frontiers miss [Nilsson, 1982] or
that the frontiers meet early, and a long time is spent proving the optimal solution [Kaindl and Kainz, 1997]. Recent
work has refined this, showing that with strong heuristics the
frontiers meet later [Barker and Korf, 2015].
3
Terminology and Notation
We use the same notation and terminology as [Eckerle et al.,
2017]. A state space G is a finite directed graph whose vertices are states and whose edges are pairs of states.1 Each
edge (u, v) has a cost c(u, v) ≥ 0. A forward path in G
is a finite sequence U = (U0 , . . . , Un ) of states in G where
(Ui , Ui+1 ) is an edge in G for 0 ≤ i < n. We say that forward path U contains edge (u, v) if Ui = u and Ui+1 = v
1
If G has multiple edges from state u to state v, we ignore all but
the cheapest of them.
for some i. Likewise, a backward path is a finite sequence
V = (V0 , . . . , Vm ) of states where (Vi , Vi+1 ) is a “reverse”
edge, i.e. (Vi+1 , Vi ) is an edge in G for 0 ≤ i < m. Backward path V contains reverse edge (u, v) if Vi = u and
Vi+1 = v for some i. The reverse of path V = (V0 , . . . , Vm )
is V −1 = (Vm , . . . , V0 ). The cost of a reverse edge equals the
cost of the corresponding original edge. A path pair (U, V )
has a forward path (U ) as its first component and a backward
path (V ) as its second component.
If U is a path (forward or backward), |U | is the number of
edges in U , c(U ) is the cost of U (the sum of the costs of all
the edges in U ), and Ui is the ith state in U (0 ≤ i ≤ |U |).
U|U | is the last state in path U , which we also denote end(U ).
λF = (start) and λB = (goal) are the empty forward and
backward paths from start and goal, respectively. Note that
end(λF ) = start while end(λB ) = goal. Both λF and λB
have a cost of 0. Forward (backward, resp.) path U is optimal
if there is no cheaper forward (backward, resp.) path from U0
to end(U ). d(u, v) is the distance from state u to state v, i.e.,
the cost of the cheapest forward path from u to v. If there is
no forward path from u to v then d(u, v) = ∞. Given two
states in G, start and goal, a solution path is a forward path
from start to goal. C ∗ = d(start, goal) is the cost of the
cheapest solution path.
A heuristic maps an individual state in G to a non-negative
real number or to ∞. Heuristic hF is forward admissible iff
hF (u) ≤ d(u, goal) for all u in G and is forward consistent
iff hF (u) ≤ d(u, u0 )+hF (u0 ) for all u and u0 in G. Heuristic
hB is backward admissible iff hB (v) ≤ d(start, v) for all
v in G and is backward consistent iff hB (v) ≤ d(v 0 , v) +
hB (v 0 ) for all v and v 0 in G. For any forward path U with
U0 = start define fF (U ) = c(U )+hF (end(U )), and for any
backward path V with V0 = goal define fB (V ) = c(V ) +
hB (end(V )).
A problem instance is defined by specifying two front-toend heuristics, hF and hB , and a state space G represented
implicitly by a 5-tuple (start, goal, c, expandF , expandB )
consisting of a start state (start), a goal state (goal), an
edge cost function (c), a successor function (expandF ), and
a predecessor function (expandB ). The input to expandF
is a forward path U . Its output is a sequence (U 1 , . . . , U n ),
where each U k is a forward path consisting of U followed by
one additional state (end(U k )) such that (end(U ), end(U k ))
is an edge in G. There is one U k for every state s such
that (end(U ), s) is an edge in G. Likewise, the input to
expandB is a backward path V and its output is a sequence
(V 1 , . . . , V m ), where each V k is a backward path consisting
of V followed by one additional state (end(V k )) such that
(end(V k ), end(V )) is an edge in G. There is one V k for
every state s such that (s, end(V )) is an edge in G.
Although the expand functions operate on paths, it is sometimes convenient to talk about states being expanded. We say
state u has been expanded if one of the expand functions has
been applied to a path U for which end(U ) = u. Finally,
we say that a state pair (u, v) has been expanded if either u
has been expanded in the forward direction or v has been expanded in the backward direction (we do not require both).
A problem instance is solvable if there is a forward path
in G from start to goal. IAD is the set of solvable prob-
lem instances in which hF is forward admissible and hB is
backward admissible. ICON is the subset of IAD in which
hF is forward consistent and hB is backward consistent. A
search algorithm is admissible iff it is guaranteed to return an
optimal solution for any problem instance in IAD .
We only consider DXBB [Eckerle et al., 2017] algorithms.
These are deterministic algorithms that proceed by expanding states that have previously been generated and have only
black-box access to the expand, heuristic, and cost functions.
4
Sufficient Conditions for Node Expansion
This paper builds on recent theoretical work [Eckerle et al.,
2017] defining sufficient conditions for state expansion for
bidirectional DXBB search algorithms. While A* with a consistent heuristic necessarily expands all states with fF (s) <
C ∗ , in bidirectional search there is no single state that is necessarily expanded, as the search can proceed forward or backwards, avoiding the need to expand any single state. However,
given a path pair (U, V ), that meets the following conditions,
one of the paths’ end-states must necessarily be expanded.
Theorem 1. [Eckerle et al., 2017] Let I = (G, hF , hB ) ∈
ICON have an optimal solution cost of C ∗ . If U is an optimal
forward path and V is an optimal backward path such that
U0 = start, V0 = goal, and:
max{fF (U ), fB (V ), c(U ) + c(V )} < C ∗ ,
then, in solving problem instance I, any admissible DXBB
bidirectional front-to-end search algorithm must expand the
state pair (end(U ), end(V )).
The c(U ) + c(V ) condition was not used by BS*, is degenerate in A*, but is used by MM. When the heuristic is weak,
this is an important condition for early termination.
We now use these conditions for state-pair expansion to
define a bipartite graph.
Definition 1. For path pair (U, V ) define
lb(U, V ) = max{fF (U ), fB (V ), c(U ) + c(V )} .
When hF is forward admissible and hB is backward admissible, lb(U, V ) is a lower bound on the cost of a solution
path of the form U ZV −1 , where Z is a forward path from
end(U ) to end(V ).
Definition 2. The Must-Expand Graph GMX (I) of problem
instance I = (G, hF , hB ) ∈ ICON is an undirected, unweighted bipartite graph defined as follows. For each state
u ∈ G, there are two vertices in GMX (I), the left vertex
uF and right vertex uB . For each pair of states u, v ∈ G,
there is an edge in GMX (I) between uF and vB if and only
if there exist an optimal forward path U with U0 = start
and end(U ) = u and an optimal backward path V with
V0 = goal and end(V ) = v such that lb(U, V ) < C ∗ . Thus,
there is an edge in GMX (I) between uF and vB if and only
if Theorem 1 requires the state pair (u, v) to be expanded.
We illustrate this in Figures 1 and 2. Figure 1 shows a
problem instance I = (G, hF , hB ) ∈ ICON . In this example a is the start state, f is the goal, and C ∗ = 3. Figure 2
shows GMX (I), where d refers to the cost of the shortest
path to each state and f refers to the f -cost of that path. By
construction, the edges in GMX (I) exactly correspond to the
state pairs that must be expanded according to Theorem 1,
and therefore any vertex cover for GMX (I) will, by definition, represent a set of expansions that covers all the required
state pairs. For example, one possible vertex cover includes
exactly the vertices in the left side with at least one edge–
{aF , cF , dF , eF }. This represents expanding all the required
state pairs in the forward direction. This requires four expansions and is not optimal because the required state pairs can
be covered with just three expansions: a and c in the forward
direction and f in the backward direction. This corresponds
to a minimum vertex cover of GMX (I) : {aF , cF , fB }.
Theorem 2. Let I ∈ ICON . Let A be an admissible DXBB
bidirectional front-to-end search algorithm, and SF (resp.
SB ) be the set of states expanded by A on input I in the
forward (resp. backward) direction. Together, SF and SB
correspond to a vertex cover for GMX (I). In particular,
|SF |+|SB | is lower-bounded by the size of the smallest vertex
cover for GMX (I).
Proof. Let (uF , vB ) be an edge in GMX (I). Then Theorem 1
requires the state pair (u, v) to be expanded by A on input
I, i.e., A must expand u in the forward direction or v in the
backward direction. Thus the set of states expanded by A on
input I corresponds to a vertex cover of GMX (I).
We will show below (Theorem 5) that this lower bound
cannot be attained by any admissible DXBB bidirectional
front-to-end search algorithm. However, we devise an admissible algorithm, NBS, which efficiently finds a near-optimal
vertex cover and thus is near-optimal in terms of necessary
node expansions.
The claim of near-optimality is only with respect to
the state pairs that must be expanded according to Theo-
start
2
b
2
a
1
hF = 1
hB = 0
5
c
hF = 0
hB = 0
d
1
hF = 0
hB = 2
2
e
f
goal
1
hF = 0
hB = 1
NBS: A Near-Optimal Front-to-End
Bidirectional Search Algorithm
While a vertex cover can be computed efficiently on a bipartite graph, in practice, building GMX (I) is more expensive
than solving I. Instead, we adapt a greedy algorithm for vertex cover [Papadimitriou and Steiglitz, 1982] to achieve nearoptimal performance. The greedy algorithm selects a pair of
states that are not part of the vertex cover subset selected so
far and are connected by an edge in the graph. It then adds
both states to the vertex cover. We introduce a new algorithm,
NBS, which uses the same approach to achieve near-optimal
node expansions while finding the shortest path.
The pseudocode for NBS is shown in Algorithms 1 and 2.
NBS considers all pairs for which lb is smallest (line 9).
Among these pairs, it first chooses the pairs (U, V ) with
smallest cost c(U ) (line 12) and then, among those, the ones
with smallest cost c(V ) (line 15). NBS picks an arbitrary pair
(U, V ) from the remaining candidates and expands both U
and V (lines 16/17). Breaking ties in this way is necessary
to guarantee that NBS never expands a suboptimal path when
its heuristics are consistent; other tie-breaking rules can be
used.An efficient data structure for implementing this path
pair selection is described in Section 7.
The pseudocode for backwards expansion is not shown, as
it is analogous to forward expansion. We have proofs that, for
all problem instances in IAD , NBS returns C ∗ . These are not
included here because of space limitations, and because they
are very similar to the corresponding proof for MM.
6
hF = 4
hB = 1
hF = 2
hB = 0
rem 1, it does not take into account state pairs of the form
(end(U ), end(V )) when lb(U, V ) = C ∗ . In principle, NBS
could expand many such state pairs while some other algorithm does not. We investigate this further in our experiments.
Bounded Suboptimality in State Expansions
Theorem 1 identifies the set of state pairs that must be expanded by any admissible DXBB front-to-end bidirectional
algorithm. We refer to these as surely expanded (s.e.) path
pairs. The theorem does not stipulate which state in each
s.e. pair must be expanded; an algorithm is free to make that
choice in any manner. Different choices can lead to vastly
Algorithm 1 NBS
Figure 1: A sample problem instance.
left
right
d:0, f:2
aF
fB
d:0, f:2
d:2, f:6
bF
eB
d:1, f:2
d:1, f:2
cF
dB
d:2, f:2
d:2, f:2
dF
cB
d:2, f:2
d:2, f:2
eF
bB
d:5, f:6
d:3, f:3
fF
aB
d:3, f:3
Figure 2: The Must-Expand Graph for Figure 1, where C ∗ =3.
1: C ← ∞
2: OpenF ← {λF }; OpenB ← {λB }
3: ClosedF ← ∅ ; ClosedB ← ∅
4: while OpenF 6= ∅ and OpenB 6= ∅ do
5:
P airs ← OpenF × OpenB
6:
lbmin ← min{lb(X, Y ) | (X, Y ) ∈ P airs}
7:
if lbmin ≥ C then return C
8:
end if
9:
minset ← {(X, Y ) ∈ P airs | lb(X, Y ) = lbmin}
10:
U set ← {X | ∃Y (X, Y ) ∈ minset}
11:
U min ← min{c(X) | X ∈ U set}
12:
Choose any U ∈ U set such that c(U ) = U min
13:
V set ← {Y | (U, Y ) ∈ minset}
14:
V min ← min{c(Y ) | Y ∈ V set}
15:
Choose any V ∈ V set such that c(V ) = V min
16:
Forward-Expand(U )
17:
Backward-Expand(V )
18: end while
19: return C
Algorithm 2 NBS: Forward-Expand(U )
1: Move U from OpenF to ClosedF
2: for each W ∈ expandF (U ) do
3:
if ∃Y ∈ OpenB with end(Y ) = end(W ) then
4:
C = min(C, c(W ) + c(Y ))
5:
end if
6:
if ∃X ∈ OpenF ∪ ClosedF with end(X) = end(W ) then
7:
if c(X) ≤ c(W ) then
8:
Continue for loop // discard W
9:
else
10:
remove X from OpenF /ClosedF
11:
end if
12:
Add W to OpenF
13:
end if
14: end for
different numbers of expansions. Given that VC is the size of
a minimum vertex cover for GMX , we have shown above that
at least VC expansions are required. In this section we prove
that on the subset of consistent problem instances NBS never
expands more than 2VC states to cover all the s.e. pairs, and
that for every DXBB front-to-end bidirectional algorithm A
there exists a problem instance in ICON on which A expands
at least 2VC states to cover all the s.e. pairs. That means that
the suboptimality of NBS is bounded by a factor of two, and
that no competing algorithm can do better in the worst case.
Theorem 3. Let I ∈ ICON , let GMX (I) be the Must-Expand
Graph, and let VC(I) be the size of the smallest vertex cover
of GMX . Then NBS does no more than 2VC(I) state expansions on GMX (I) to cover its s.e. pairs.
∗
Proof. If (u, v) is a s.e. pair then lb(U, V ) < C for every
optimal forward path U from start to u and every optimal
backward path from goal to v. NBS will select exactly one
such (U, V ) pair for expansion and expand both end(U ) = u
and end(V ) = v. A minimum vertex cover for I might require only one of them to be expanded, so for each expansion
required by a minimum vertex cover, NBS might do two.
Theorems 2 and 3 yield the following result.
Corollary 4. Let I ∈ ICON and let A be any admissible
front-to-end DXBB bidirectional algorithm. Then NBS makes
no more than twice the number of state expansions on input I
than A does in covering I’s s.e. pairs.
For any algorithm A, let us use the term worst-case expansion ratio of A to refer to the ratio maxI∈ICON #A(I)
#(I) ,
where #A(I) is the number of states A expands in covering
the s.e. pairs in instance I, and #(I) is the smallest number
of states any DXBB front-to-end bidirectional search algorithm expands in covering the s.e. pairs in I. By definition,
#(I) ≥ V C(I), so we can rephrase Corollary 4 as follows:
NBS’s worst-case expansion ratio is at most 2.
(1)
We now demonstrate that NBS is optimal in the sense that
no admissible DXBB front-to-end bidirectional search algorithm has a worst-case expansion ratio smaller than 2.
Theorem 5. Let A be any admissible DXBB front-to-end
bidirectional search algorithm. Then there exists a problem
instance I and a DXBB front-to-end bidirectional search algorithm B such that A expands at least twice as many states
in solving I as B expands in solving I.
Proof. Consider the two problem instances I1 and I2 in Figure 3. In these instances hF (n) = hB (n) = 0 for all n, s
is the start and g is the goal. Assume A is given either one
of these instances as input. Since A is DXBB and cannot
initially distinguish I1 from I2 , it must initially behave the
same on both instances. Hence, on both instances, A will initially either expand s in the forward direction, expand g in the
backward direction, or expand both s and g.
If A first expands s in the forward direction, consider I =
I1 . On instance I, the algorithm A has to expand a second
state (either g in the backward direction or t in the forward
direction) in order to be able to terminate with the optimal
solution path (s, g). By comparison, an algorithm B that first
expands g in the backward direction will terminate with the
optimal solution after just a single state expansion. Here we
assume that B terminates when there are no pairs satisfying
the sufficient condition for node expansion.
If A first expands g in the backward direction, one can argue completely symmetrically, with I = I2 and B being an
algorithm that first expands s in the forward direction.
If A begins by expanding both s and g, as NBS does, then
on both these instances it will have expanded two states when
only one expansion was required.
7
Efficient Selection of Paths for Expansion
Algorithm 1 assumes that NBS can efficiently compute lbmin
and select the best path pair (U, V ) for expansion. In this
section we provide a new Open list that can do this efficiently.
The data structure works by maintaining a lower bound on
lbmin, Clb . NBS initalizes Clb to 0 prior to its first iteration
and each time a new path pair is needed, Clb is raised until it
reaches lbmin and a new path pair is found.
This data structure is illustrated in Figure 4. The data structure is composed of two priority queues for each direction.
The first priority queue is waitingF . It contains paths with
fF ≥ Clb sorted from low to high fF . The second priority
queue is readyF . It contains paths with fF ≤ Clb sorted from
low to high c. Analogous queues are maintained in the backward direction. When paths are added to Open, they are first
added to waitingF and waitingB . After processing, paths are
removed from the front of readyF and readyB . The current
value of Clb is the minimum of the f -costs at the front of
waitingF and waitingB and the sum of the c-costs at the front
of readyF and readyB .
Pseudocode for the data structure is in Algorithm 3. Where
forward or backwards queues are designated with a D, operations must be performed twice, once in each direction. We
use the notation readyF .c to indicate the smallest c-cost on
ready in the forward direction. At the end of the procedure
I1
1
s
t
3
I2
4
4
g
s
t
3
1
g
Figure 3: Two problem instances with C ∗ = 3 differing only
in the costs of the edges (s, t) and (t, g).
OPENB (WaitingB)
Low-to-high f
Low-to-high f
Clb
OPENF (ReadyF)
Low-to-high c
OPENB (ReadyB)
+
Low-to-high c
Next Pair
Figure 4: The open list data structure.
15k
10k
5k
0
Algorithm 3 NBS pseudocode for selecting the best pair from
Open list. Clb is set to 0 when the search begins.
1: procedure P REPARE B EST
2:
while min f in waitingD < Clb do
3:
move best node from waitingD to readyD
4:
end while
5:
while true do
6:
if readyD ∪ waitingD empty then return false
7:
end if
8:
if readyF .c + readyB .c ≤ Clb then return true
9:
end if
10:
if waitingD .f ≤ Clb then
11:
move best node from waitingD to readyD
12:
else
13:
Clb = min(waitingF .f , waitingB .f , readyF .c+readyB .c)
14:
end if
15:
end while
16: end procedure
the paths on readyF and readyB with the smallest individual
c-costs together form the pair to be expanded.
The procedure works as follows. First, paths with f -cost
lower than Clb must immediately be moved to ready (line 2).
If readyD and waitingD are jointly empty in either direction,
the procedure is halted and the search will terminate (line 6).
If the best paths in ready have c(U ) + c(V ) ≤ Clb , the procedure completes; these paths will be expanded next (line 8).
If the readyD queue is empty in either direction, any paths
with f = Clb can be moved to ready (line 10). While we
could, in theory, move all such paths to ready in one step,
doing so incrementally allows us to break ties on equal f towards higher c first, which slightly improves performance. If
there are no paths with f ≤ Clb in waiting and in ready with
c(U ) + c(V ) ≤ Clb , then the Clb estimate is too low, and Clb
must be increased (line 13).
We illustrate this on an artificial example from Figure 5.2
2
This example also illustrates why we cannot just sort by minimum fF or c when performing expansions.
c: 8
h: 1
A
D
c: 8
h: 1
c: 6
h: 6
B
E
c: 6
h: 6
c: 3
h: 10
C
F
c: 3
h: 10
Lower Bound
AD
16
AE
14
AF
13
BD
14
BE
12
BF
13
CD
13
CE
13
CF
13
Figure 5: Sample state space and priorities of state pairs
Necessary Node Expansions (brc203d)
Best Alternate
y=x
y=2x
y=x/2
20k
Necessary Expansions by NBS
OPENF (WaitingF)
0
10k
20k
Necessary Expansions by Best Alternate
Figure 6: A comparison between necessary expansions by
NBS (y-axis) and the minimum of the expansions with f <
C ∗ by MMe, BS* and A* (x-axis) on each problem instance.
To begin, Clb ← 0 and we assume that (A, B, C) are on
waitingF and (D, E, F ) are on waitingB . First, Clb is set to 9
(line 13). Then, A and D can be added to ready because they
have lowest f -cost, and Clb = 9. However, c(A) + c(D) =
16 > Clb = 9, so we cannot expand A and D. Instead, we
increase Clb to 12 and then add B and E to ready. Now,
because c(B) + c(E) = 12 ≤ Clb we can expand B and E.
We can prove that the amortized runtime over a sequence
of basic operations of our data structure (including insertion
and removal operations) is O(log(n)), where n is the size of
waiting ∪ ready.
8
Experimental Results
There are two primary purposes of the experimental results.
First, we want to validate that our implementation matches
the theoretical claims about NBS. Second, we want to compare the overall performance of NBS to existing algorithms.
This comparison includes total node expansions, necessary
expansions, and time. We compare A*, BS*3 , MMe (a variant of MM) [Sharon et al., 2016], NBS, and MM0 (bidirectional brute-force search). We also looked at IDA*, but it was
not competitive in most domains due to cycles.
In Table 1 we present results on problems from four
different domains, including grid-based pathfinding problems [Sturtevant, 2012] (‘brc’ maps from Dragon Age: Origins (DAO)), random 4-peg Tower of Hanoi (TOH) problems, random pancake puzzles, and the standard 15 puzzle
instances [Korf, 1985]. The canonical goal state is used for all
puzzle problems.4 On each of these domains we use standard
heuristics of different strength. The octile, GAP [Helmert,
2010], and Manhattan Distance heuristics can be easily computed at runtime for any two states. The additive pattern
databases used for Towers of Hanoi are computed for each
3
BS* is not an admissible algorithm (it will not optimally solve
problems with an inconsistent heuristic) so the theory in this paper
does not fully apply to BS*.
4
On the 15-puzzle and TOH it is more efficient to search backwards because of the lower branching factor, but we search forward
to the standard goal states.
Table 1: Average state expansions for unidirectional (A*) and bidirectional search across domains.
Domain
Grids
Grids
4 Peg TOH
4 Peg TOH
16 Pancake
16 Pancake
16 Pancake
15 puzzle
Instances
DAO
Mazes
50 random
50 random
50 random
50 random
50 random
[Korf,1985]
Heuristic
Octile
Octile
12+2 PDB
10+4 PDB
GAP
GAP-2
GAP-3
MD
Strength
+
−
++
−
+++
−
−−
+
A*
9,646
64,002
1,437,644
19,340,099
125
1,254,082
unsolvable
15,549,689
Table 2: Average running time and expansions per second for
unidirectional (A*) and bidirectional search across domains.
Average Running Time (in seconds)
Domain
h
A*
BS* MMe
DAO
Octile
0.005
0.006 0.015
Mazes
Octile
0.035
0.022 0.060
TOH4
12+2
3.23
2.44
4.17
TOH4
10+4
52.08
23.06 30.64
Pancake
GAP
0.00
0.00
0.00
Pancake
GAP-2
14.16
4.91
5.25
Pancake
GAP-3
N/A
212.33 72.13
15 puzzle
MD
47.68
29.59 41.38
Expansion Rate (×103 nodes per seconds)
DAO
Octile
1,896
1,912
851
Mazes
Octile
2,225
2,366
848
TOH4
12+2
444
453
418
TOH4
10+4
371
376
375
Pancake
GAP
156
564
564
Pancake
GAP-2
89
193
109
Pancake
GAP-3
N/A
137
98
15 puzzle
MD
326
406
318
instance. We selected the size of problems to ensure that as
many algorithms as possible could solve them in RAM.
In grid maps we varied the difficulty of the problem by
changing the map type/topology. In TOH and the pancake
puzzle we varied the strength of the heuristic. The GAP-k
heuristic is the same as the GAP heuristic, except that gaps
involving the first k pancakes are not added to the heuristic.
The approximate heuristic strength on a problem is indicated
by a + or −. The general trend is that with very strong heuristics, A* has the best performance. As heuristics get weaker,
or the problems get harder, the bidirectional approaches improve relative to A*. NBS is never far from the best algorithm,
and on some problems, such as TOH, it has significantly better performance than all previous approaches. Runtime and
node expansions/second are found in Table 2. NBS is 30%
slower than A* on the DAO problems, but competitive on
other problems. NBS is slower than BS*, but this is often
compensated for by performing fewer node expansions.
In Table 3 we look at the percentage of total nodes on
closed compared to total expansions with f = C ∗ by each
algorithm in each domain. For the majority of domain and
heuristic combinations there are very few expansions with
f = C ∗ . The exception is the pancake puzzle with the GAP
MMe
13,013
51,074
1,741,480
11,499,867
283
578,283
7,100,998
13,162,312
NBS
12,085
34,474
1,420,554
6,283,143
335
625,900
6,682,497
12,851,889
MM0
17,634
51,075
12,644,722
12,644,722
unsolvable
unsolvable
unsolvable
unsolvable
Table 3: Percent of expansions with (f -cost = C ∗ ) for each
algorithm/domains.
Domain
DAO
Mazes
TOH4
TOH4
Pancake
Pancake
Pancake
15 puzzle
NBS
0.007
0.019
3.54
16.60
0.00
5.23
77.17
37.67
1,662
2,290
401
379
153
120
87
338
BS*
11,501
42,164
1,106,189
8,679,443
339
947,545
29,040,138
12,001,024
Heuristic
Octile
Octile
12+2 PDB
10+4 PDB
GAP
GAP-2
GAP-3
MD
A*
1.3%
0.0%
0.0%
0.0%
60.7%
0.2%
N/A
5.5%
BS*
0.6%
0.0%
0.0%
0.0%
81.6%
1.6%
-0.6%
0.5%
MMe
0.7%
0.0%
0.0%
0.0%
76.2%
6.0%
5.7%
0.6%
NBS
1.2%
0.0%
0.0%
0.0%
81.0%
0.0%
0.0%
0.3%
heuristic. On random instances this heuristic is often perfect,
so all states expanded have f = C ∗ . This is why NBS does
more than twice the number of expansions as A* on these
problems—these expansions are not accounted for in our theoretical analysis. BS* puts nodes on closed that it does not
expand, which is why it has a negative percentage.
Figure 6 shows a scatter plot of necessary node expansions
on 1171 instances from the brc203d grid map where NBS has
slightly worse performance than A* and BS*, but better performance than MM. Each point represents one problem instance, and plots NBS necessary expansions against the minimum of the expansions with f < C ∗ by A*, BS* and MMe.
The y = 2x line represents the theoretical maximum expansion ratio (Theorem 4), which is never crossed. NBS often
does much better than this theoretical worst case, and on approximately 30 instances is more than 2x better than all alternate algorithms, as they do not have similar 2x expansion
bounds.
9
Conclusion
This paper presents the first front-to-end heuristic search algorithm that is near-optimal in necessary node expansions. It
thus addresses questions dating back to Pohl’s work on the
applicability of bidirectional heuristic search [Pohl, 1971].
When the problems are hard or the heuristic is not strong,
NBS provides a compelling alternative to A*.
10
Acknowledgements
Financial support for this research was in part provided by
Canada’s Natural Sciences and Engineering Research Council (NSERC). This material is based upon work supported by
the National Science Foundation under Grant No. 1551406.
References
[Barker and Korf, 2015] Joseph Kelly Barker and Richard E.
Korf. Limitations of front-to-end bidirectional heuristic
search. In Proc. 29th AAAI Conference on Artificial Intelligence, pages 1086–1092, 2015.
[Dechter and Pearl, 1985] Rina Dechter and Judea Pearl.
Generalized best-first search strategies and the optimality
of A*. J. ACM, 32(3):505–536, 1985.
[Eckerle et al., 2017] Jürgen Eckerle, Jingwei Chen, Nathan
Sturtevant, Sandra Zilles, and Robert Holte. Sufficient
conditions for node expansion in bidirectional heuristic
search. In International Conference on Automated Planning and Scheduling (ICAPS), 2017.
[Helmert, 2010] Malte Helmert. Landmark heuristics for
the pancake problem. In Symposium on Combinatorial
Search, pages 109–110, 2010.
[Holte et al., 2016] Robert C. Holte, Ariel Felner, Guni
Sharon, and Nathan R. Sturtevant. Bidirectional search
that is guaranteed to meet in the middle. In AAAI Conference on Artificial Intelligence, pages 3411–3417, 2016.
[Kaindl and Kainz, 1997] Hermann Kaindl and Gerhard
Kainz. Bidirectional heuristic search reconsidered. J. Artificial Intelligence Resesearch (JAIR), 7:283–317, 1997.
[Korf, 1985] Richard E. Korf.
Depth-first iterativedeepening: An optimal admissible tree search. Artificial
Intelligence, 27(1):97–109, 1985.
[Kwa, 1989] James B. H. Kwa. BS*: An admissible bidirectional staged heuristic search algorithm. Artificial Intelligence, 38(1):95–109, 1989.
[Nicholson, 1966] T. A. J. Nicholson. Finding the shortest
route between two points in a network. The Computer
Journal, 9(3):275–280, 1966.
[Nilsson, 1982] Nils J. Nilsson. Principles of Artificial Intelligence. Springer, 1982.
[Papadimitriou and Steiglitz, 1982] Christos H Papadimitriou and Kenneth Steiglitz. Combinatorial optimization:
algorithms and complexity. Courier Corporation, 1982.
[Pohl, 1971] Ira Pohl. Bi-directional search. Machine Intelligence, 6:127–140, 1971.
[Sharon et al., 2016] Guni Sharon, Robert C. Holte, Ariel
Felner, and Nathan R. Sturtevant. Extended abstract:
An improved priority function for bidirectional heuristic search. Symposium on Combinatorial Search (SoCS),
pages 139–140, 2016.
[Sturtevant, 2012] Nathan R. Sturtevant. Benchmarks for
grid-based pathfinding. Transactions on Computational
Intelligence and AI in Games, 4(2):144–148, 2012.
| 2cs.AI
|
arXiv:math/9810022v1 [math.NA] 5 Oct 1998
Active Libraries: Rethinking the roles of compilers
and libraries
Todd L. Veldhuizen and Dennis Gannon∗
Abstract
We describe Active Libraries, which take an active role in compilation.
Unlike traditional libraries which are passive collections of functions
and objects, Active Libraries may generate components, specialize algorithms, optimize code, configure and tune themselves for a target
machine, and describe themselves to tools (such as profilers and debuggers) in an intelligible way. Several such libraries are described, as
are implementation technologies.
1
Introduction
This paper attempts to document a trend toward libraries which take an
active role in generating code and interacting with programming tools. We
call these Active Libraries. They solve the problem of how to provide efficient domain-specific abstractions (Section 1.1): active libraries are able
to define abstractions, and also control how they are optimized. Several
existing libraries for arrays, parallel physics, linear algebra and Fast Fourier
Transforms fit the description of active libraries (Section 2). These libraries
take novel approaches to generating optimized code (Section 3). To implement active libraries, programming systems and tools which open up the
development environment are needed (Section 4).
1.1
Why are active libraries needed?
To produce readable, maintainable scientific computing codes, we need abstractions. Every subdomain in scientific computing has its own requirements: interval arithmetic, tensors, polynomials, automatic differentiation,
∗
Indiana University Computer Science Department. [email protected]
1
2
1
INTRODUCTION
sparse arrays, spinors, meshes, and so on. How should these abstractions be
provided? The alternatives are:
Extend mainstream languages. In the past, syntax and efficiency
concerns have encouraged building abstractions into languages and compilers. Fortran 95, for example, has built-in complex numbers and numerical
arrays, and many intrinsic functions for common numerical operations. Efforts are underway to extend the Java language for scientific computing [18],
and similar efforts in the past have extended C [25]. However, loading up
these languages with features for the scientific market can meet with only
limited success. The costs of compiler development are quite high, and scientific computing is a comparatively small segment of the market. Even
Fortran, whose bread and butter is scientific computing, is showing signs
of hitting economic limits to its size. A controversial 60-page proposal to
add interval arithmetic to Fortran 2000 was eventually discarded after much
debate. The committee had to balance the limited demand for interval
arithmetic against the large implementation costs for vendors. Although
we may succeed in getting mainstream languages to incorporate basic scientific features such as numeric arrays and complex numbers, the likelihood
of these languages providing specialized features such as interval arithmetic
and sparse arrays is small.
There are other disadvantages to building abstractions into mainstream
languages: feature turnaround is very slow, since extensions require championing a proposal through years of standards committee meetings. Experimental features must be prematurely standardized, before experience can
produce a consensus on the “right way” to implement them.
Domain-Specific Languages (DSLs). Dozens of DSLs for scientific
computing have been produced to handle sparse arrays, automatic differentiation, interval arithmetic, adaptive mesh refinement, and so on. DSLs
are often implemented as preprocessors for mainstream languages such as C,
Fortran, C++, or Java. The growing availability of compiler construction
tools has encouraged a proliferation of DSLs. Although DSLs are an attractive alternative, many people would prefer to work in mainstream languages
for the wealth of tools support and libraries available. DSLs frequently
have problems with portability and long-term support, since they tend to
be research projects. It is also impossible to use multiple DSLs at once;
for example, although separate Fortran-based DSLs are available for both
sparse arrays and interval arithmetic, you cannot use the features of both
DSLs in the same source file.
Object-oriented language features. An alternative to building abstractions into languages is to provide language features which allow library
3
developers to construct their own abstractions. In C++ and Fortran 90/95,
we are seeing libraries for many applications (sparse arrays, interval arithmetic, data-parallel arrays) which were previously solved by domain-specific
languages. Unfortunately, such libraries are hard to optimize. Compilers
have difficulty because they lack semantic knowledge of the abstractions:
instead of seeing array operations, they see loops and pointers. Libraries
also tend to have layers of abstraction and side effects which confound optimizations. The heroic optimizers needed to overcome these problems may
never appear, because the economics of the scientific market may not support their development. It is doubtful that the optimization problems admit
a general-purpose solution, since every problem domain has its own tricks
and peculiarities.
What we really need are language features which allow library developers to define their own abstractions, and also to specify how these
abstractions are optimized. We call this solution Active Libraries. Active
Libraries combine the benefits of built-in language abstractions (nice syntax and efficient code) with those of library-level abstractions (adaptability,
quick feature turnaround, cheap to implement).
2
Examples of Active Libraries
In defining Active Libraries, we are not proposing a new concept, but rather
trying to summarize what many people are already trying to do. In the
following sections, we highlight existing software packages which illustrate
the characteristics of Active Libraries.
2.1
Blitz++
The Blitz++ library [32] provides generic array objects for C++ similar to
those in Fortran 90, but with many additional features. In the past, C++ array libraries have been 3-10 times slower than Fortran, due to the temporary
arrays which result from overloaded operators. Blitz++ solves this problem
using the expression templates technique [30] to generate custom evaluation
kernels for array expressions. The library performs many loop transformations (tiling, reordering, collapsing, unit stride optimizations, etc.) which
have until now been the responsibility of optimizing compilers. Blitz++ also
generates different code depending on the target architecture.
For operations on small vectors and matrices, Blitz++ uses the template
metaprogram technique [31] to generate specialized algorithms. This avoids
4
2 EXAMPLES OF ACTIVE LIBRARIES
the performance penalty often associated with small objects by completely
unrolling loops and inlining code.
2.2
POOMA
POOMA is a C++ library for parallel physics which uses many of the
same techniques as Blitz++. Users write simple array expressions, such
as “A=B+C”, which trigger the generation of data-parallel implementation
routines using threads and message passing [21]. POOMA uses template
techniques to generate components (such as Fields) using a variety of types,
geometries, addressing schemes, data distribution and communication parameters.
2.3
Matrix Template Library
The Matrix Template Library (MTL) [28] is a C++ library which extends
the ideas of STL [24] to linear algebra. MTL handles both sparse and dense
matrices. For dense matrices, MTL uses template metaprograms to generate
tiled algorithms. Tiling is a crucial technique for obtaining top performance
from cache-based memory systems; MTL uses template metaprograms to
tile on both the register and cache level. For register tiling, it uses template metaprograms to completely unroll loops. MTL provides generic,
high-performance algorithms which are competitive with vendor-supplied
kernels.
2.4
Generative Matrix Computation Library
The Generative Matrix Computation Library (GMCL) [16] provides heavily
parameterized matrix classes. Users can specify the element type, whether
the matrix is dense or sparse, the storage format (including several sparse
formats), dynamic or static memory allocation, error checking, and several other parameters. The GMCL uses template metaprograms to examine
the parameters and determine any interactions between them (for example,
sparse matrices cannot use static memory allocation); it then instantiates a
matrix class with the desired characteristics. The implementation is roughly
7500 lines of C++ code, yet covers more than 1840 different kinds of matrices. Despite this flexibility, the authors report performance on par with
manually generated code.
2.5
FFTW: The Fastest Fourier Transform in the West
2.5
5
FFTW: The Fastest Fourier Transform in the West
FFTW [14] is a C library for Fast Fourier Transforms. It generates a collection of small “codelets,” each of which is a small step in the transform process. At installation, FFTW evaluates the performance of each codelet for
the target architecture. At run-time, the codelets are dynamically stitched
together to perform FFTs. FFTW records the ordering of these codelets
using bytecode, which is generated and interpreted at run time. Based on
extensive benchmarking, the authors report superior performance over all
commonly used FFT packages.
2.6
PhiPAC, ATLAS
Obtaining top performance for matrix operations requires substantial expertise and hand-tuning. PhiPAC [4] provides a methodology to achieve
near-peak performance automatically. It uses parameterized code generators, whose parameters are related to machine-specific tuning. PhiPAC
searches the parameter space to find the best implementation for a given
target architecture. On dense matrix-matrix multiplication, PhiPAC performs better then vendor-supplied kernels on many platforms. The ATLAS
package [33] provides similar capabilities.
3
Optimization models
Active Libraries make it possible to approach code optimization in several
new and exciting ways. In describing some of the new approaches to optimization being explored, it is useful to distinguish between two flavours of
optimization, which we call low-level and high-level:
• Low-level optimizations can be applied without knowing what the
code is supposed to do (copy propagation, dead code elimination, instruction scheduling, loop pipelining).
• High-level optimizations which require some understanding of the
operation being performed. Examples: tiling for stencils; fusing loops
over sparse arrays; iteration-space tiling.
3.1
Transformational optimization
Traditional approaches to optimization are transformational. In transformational optimization, low-level code is transformed into an equivalent (but
6
3 OPTIMIZATION MODELS
hopefully faster) implementation. Typical transformational optimizations
are constant propagation, dead code elimination, and loop transformations.
One of the difficulties with transformational optimization is that the
optimizer lacks an understanding of the intent of the code. This makes it
harder to apply high-level optimizations. For example, rather than seeing
an array stencil operation, a transformational optimizer will just see loops,
pointers, and variables. To apply interesting optimizations, the optimizer
must recognize that this low-level code represents a stencil operation (or
more accurately, that it possesses the sort of data dependencies which benefit
from tiling). More generally, the optimizer must infer the intent of the code
to apply higher-level optimizations. To this end, sophisticated optimizers
employ algebraic reasoning, pattern recognition and matching techniques.
However, such optimizers can still only apply radical optimizations in simple
situations, and with limited success.
Another problem with transformational optimizers is the lack of extensibility. If the optimizer doesn’t recognize what your code is trying to do, you
are probably out of luck. Optimizations for dense arrays are fairly reliable,
because their use is very common; if you are working with sparse arrays
or interval arithmetic, the likelihood of achieving optimal performance is
smaller.
3.2
Generative optimization
Many environments which provide higher-level abstractions (for example,
arrays in Fortran 90) can use generative optimization. Code written using
higher-level abstractions can be regarded as a specification of what operation
needs to be performed; an efficient implementation is then generated to fulfill
the “specification”. This approach to optimization can be much simpler
than transformational optimization, since the full semantics are available to
whatever generates the optimized code. Generative optimization can make
it simpler to apply radical, high-level optimizations.
While generative optimization can produce good code for individual operations, it tends to miss optimizations which depend on context. For example, one problem encountered in libraries which use expression templates
(e.g. Blitz++, POOMA) is that while individual array statements can be optimized well, opportunities for between-statement optimizations are missed.
For example, in the array statements A=B+C; D=B-C; there is a substantial
gain if the two expressions are evaluated simultaneously, since they share the
same operands. Attempts to solve this problem have focused on increasing
granularity, so that code is generated for basic blocks of array statements,
3.3
Explorative optimization
7
rather than individual statements.
3.3
Explorative optimization
For library developers, finding a near-optimal implementation of a routine
is very difficult. Modern architectures can behave unpredictably due to
pipeline and cache effects. Between the library writer and the hardware
lies the compiler, a black box which transforms one’s code in sometimes
mysterious ways. Aside from some basic guidelines relating to cache reuse,
performance tuning usually requires randomly adjusting code and measuring
the result.
Explorative optimization gives up on the notion that performance is
somehow predictable. The basic approach is to examine an algorithm and
identify parameters which might affect performance (for example loop structures, tile sizes, and unrolling factors). One then writes a parameterized
code generator which produces variants of the basic algorithm. The parameter space is then explored point by point, and the performance of each
variant is measured until the best implementation is found. This approach
was pioneered by FFTW (Section 2.5) and PhiPAC (Section 2.6). On dense
matrix-matrix multiplication, PhiPAC performs better than vendor-supplied
kernels on many platforms.
Explorative optimization is expensive in time, but is worth it if a near
peak-performance kernel is required. So far, there is no way to automatically construct variants of a given piece of code; one must write the code
generators manually.
3.4
Compositional optimization
Compositional optimization is useful when a problem can be decomposed
into a sequence of calls to well-tuned kernels. The FFTW package (Section 2.5) uses this approach to decompose FFTs into a sequence of calls to
high-speed kernels which were found using explorative optimization. FFTW
records and interprets the sequence of calls using a bytecode. Compositional
optimization has also been used in the context of compilers, to generate
high-performance communication code [29].
For compositional optimization to be effective, the problem must be
decomposable into coarse chunks, so that the overhead of composition is
negligible.
8
4
4 TECHNOLOGIES FOR ACTIVE LIBRARIES
Technologies for Active Libraries
In the following sections, we describe technologies relevant to Active Libraries. Some of these are already being used, for example generic programming and C++ templates; others are promising technologies still being
explored.
4.1
Component generation
Traditional libraries define concrete components. We use the term components loosely here, to mean an algorithm, class, or collection of these things.
By concrete, we mean that the behavior is fixed: the component operates
on a fixed kind of data (for example, arrays of double), using a specific data
structure, and handles errors in a fixed way. Concrete components might
allow for some flexibility through configuration variables, but the code of
the component is unchanging. Concrete components require a trade-off between flexibility and efficiency: if the library developer wants to provide a
customizable component, this must be done through callbacks and runtime
checking of configuration variables, which are often inefficient. To solve this
problem, we need ways to generate customized components on demand. We
contrast two approaches: Generic Programming and Generative Programming.
4.1.1
Generic Programming
The aim of Generic Programming can be summarized as “reuse through
parameterization”. Generic components have parameters which customize
their behaviour. When a generic component is instantiated using a particular
choice of parameters, a concrete component is generated. This allows library
developers to create components which are very customizable, yet retain the
efficiency of statically configured code. Probably the greatest achievement of
the Standard Template Library in C++ [24] was to separate algorithms from
the data structures on which they operate, allowing them to be combined
in a mostly orthogonal way. In C++, template parameters can effectively
be types, data structures, or even pieces of code and algorithms. Other
languages which support generic programming are Ada (via its generics
mechanism), and Fortran 2000 (albeit in a limited way).
There are two main benefits of generic programming: (1) Library developers have to develop and maintain less code; (2) Application developers
find it easier to find components which match their needs. There are some
4.2
C++ Templates
9
limitations of generic programming as currently supported: (1) Aggressive
parameterization is quite ugly; providing more than 2-3 template parameters introduces serious usability problems. This can be alleviated somewhat
by named template parameters.1 (2) Generic programming may result in
code bloat, due to multiple versions of generic components. (3) Perhaps
most importantly, generic programming limits code generation to substituting concrete types for generic type parameters, and welding together preexisting fragments of code in fixed patterns. It does not allow generation
of completely new code, nor does it allow computations to be performed at
compile time.2
4.1.2
Generative Programming
Generative Programming [9, 8] is a broader term which encompasses generic
programming, code generation, code analysis and transformation, and compiletime computation. In general, it refers to systems which generate customized
components to fulfill specified requirements.
A central idea in Generative Programming (and also in Aspect-Oriented
Programming [1]) is separation of concerns: the notion that important issues
should be dealt with one at a time. In current languages, there are many
aspects such as error handling, data distribution, and synchronization which
cannot be dealt with in a localized way. Instead, these aspects are scattered
throughout the code. One of the goals of Generative Programming is to
separate these aspects into distinct pieces of code. These pieces of code are
combined to produce a needed component. Doing so often requires more
than cutting and pasting, and this is where the need for code generation,
analysis, and transformation arises. Active Libraries can be viewed as a
vehicle for implementing the goals of Generative Programming [9].
4.2
C++ Templates
In addition to enabling generic programming, the template mechanism of
C++ unwittingly introduced powerful code generation mechanisms. Nested
templates allow data structures to be created and manipulated at compile time, by encoding them as types. This is the basis of the expression
templates technique [30], which creates parse trees of array expressions at
1
Named template parameters can be listed in any order, with missing parameters
assuming default values. Thought not directly supported in C++, the technique can be
faked (see [16]).
2
We regard techniques such as template metaprograms and expression templates to lie
outside the domain of generic programming.
10
4 TECHNOLOGIES FOR ACTIVE LIBRARIES
compile time, and uses these parse trees to generate customized kernels.
Expression templates also provides a crude facility similar to the lambda
operator of functional languages, which may be used to replace callbacks.
Template metaprograms [31] use the template instantiation mechanism to
perform computations at compile-time, and generate specialized algorithms
by selectively inlining code as they are executed. Although these techniques
are powerful, the accidental nature of their presence has resulted in a clumsy
syntax. Nonetheless, they provide a reasonable way to implement Active Libraries in C++, and several libraries based on these techniques (Blitz++,
POOMA, MTL) are being distributed. Recently, a package which simplifies
the construction of template metaprograms has been made available [7].
4.3
Extensible compilation, Reflection, and Metalevel Processing
In metalevel processing systems, library writers are given the ability to directly manipulate language constructs. They can analyze and transform
syntax trees, and generate new source code at compile time. The MPC++
metalevel architecture system [17] provides this capability for the C++ language. MPC++ even allows library developers to extend the syntax of the
language in certain ways (for example, adding new keywords). Other examples of metalevel processing systems are Xroma [9], Open C++ [6], and
Magik [11]. A potential disadvantage of metalevel processing systems is the
complexity of code which one must write: modern languages have complicated syntax trees, and so code which manipulates these trees tends to be
complex as well.
4.4
Run-Time Code Generation (RTCG)
RTCG systems allow libraries to generate customized code at run-time. This
makes it possible to perform optimizations which depend on information not
available until run-time, for example, the structure of a sparse matrix or the
number of processors in a parallel application. Examples of such systems
which generate native code are ‘C (Tick-C) [12, 26], and Fabius [23]. Speeds
as high as 6 cycles per generated instruction have been achieved. Recently,
this technology has been extended to C++ [15].
4.5
Partial Evaluation
Code generation is an essential part of active libraries. Over the past two
decades, researchers in the field of Partial Evaluation have developed an
4.6
Multilevel Languages
11
extensive theory and literature of code generation [19]. In its simplest form,
a partial evaluator analyzes a program and separates its data into a static
portion (values known at compile-time) and a dynamic portion (values not
known until run-time). It then evaluates as much of the program as possible
(using the static values) and outputs a specialized residual program. For
example, a partial evaluator could take a dot-product routine, and produce
a specialized version for a particular vector length. These techniques have
been applied to scientific codes with promising results [2, 3, 22].
However, this just provides a taste of the field; partial evaluation has
evolved into a comprehensive toolbox containing both theories and practical software. One of the most important theoretical contributions was
that the concept of generating extensions [13] unifies a very wide category
of apparently different program generators. Using partial evaluation, concrete components which check configuration variables at run-time can be
transformed into component generators (or generating extensions in the terminology of the field [5, 10, 20]) which produce customized components
(eliminating the run-time checking of configuration variables). Automatic
tools for turning a general component into a component generator now exist
for C and Scheme [20]. However, such tools are not yet available for C++.
4.6
Multilevel Languages
Another important contribution of Partial Evaluation is the concept of twolevel (or more generally, multi-level) languages. Two-level languages contain
static constructs (which are evaluated at compile-time) and dynamic code
(which is compiled and later evaluated at run-time). Two-level languages
provide a simpler notation for writing code generators (compared to systems
which generate source code or intermediate representations). For example,
consider a (fictional) two-level language based on C++, in which static variables and control flow are annotated with the @ symbol. A code generator
for dot products of length N would be written as:
double dot(double* a, double* b, int@ N) {
double sum = 0;
for@ (int@ i=0; i < N; ++i)
sum += a[i] * b[i];
return sum;
}
In this example, the @ symbol indicates that N and i are compile-time variables. The for@ loop is evaluated at compile time, and the residual code
12
5 CONCLUSIONS
is equivalent to N statements of the form sum+=a[i]*b[i], with i replaced
by appropriate integer literals. This dot-product generator is dramatically
simpler than an equivalent implementation using template metaprograms
or metalevel processing. Two-level languages have been used to generate
customized run-time library code for parallel compilers [29].
4.7
Extensible Programming Tools
Libraries typically have two (or more) layers: there are user-level classes and
functions, and behind them are one or more implementation layers. Using
tools such as debuggers and profilers with large libraries is troublesome, since
the tools make no distinction between user-level and implementation code.
For example, a user who invokes a debugger on an array class library will be
confronted with many irrelevant (and undocumented) private data members,
when all they wanted to see was the array data. Using a profiler with such
a library will expose many implementation routines, instead of indicating
which array expressions were responsible for a slow program. This problem
is compounded when template libraries are used, with their long symbol
names and many instances.
The solution may be extensible tools, which provide hooks for libraries
to define customized support for debugging, profiling, etc. An example
of such interaction is Blitz++ and the Tau [27] profiling package. Tau is
unique in that it allows libraries to instrument themselves. This allows
Blitz++ to hide implementation routines, and only expose user-level routines. Time spent in the library internals is correctly attributed to the
responsible user-level routine. Blitz++ describes array evaluation kernels
to Tau using pretty-printing. When users profile applications, they do not
see incomprehensible expression template types; rather, they see expressions
such as “A=B+C+D”.
If debuggers provided similar hooks, users could debug scientific codes
by looking at visualizations and animations of the array data. The general
goal of such tool/library interactions is to provide a user-oriented view of
libraries, rather than an implementation-oriented view.
5
Conclusions
Active Libraries are able to define domain-specific abstractions, and also
control how these abstractions are optimized. This may involve compiletime computations, code generation, and even code analysis and transformation. It is no coincidence that all of the existing examples are libraries
13
for scientific computing: this is a field which requires many abstractions,
and also demands high performance. Active libraries may be the best way
to construct and deliver efficient, configurable abstractions.
6
Acknowledgements
This work was supported in part by NSF grants CDA-9601632 and CCR9527130. Portions of this paper grew out of joint work with David Vandevoorde (who coined the term active library), Krzysztof Czarnecki, Ulrich
Eisenecker, and Robert Glück [9].
References
[1] Aspect-Oriented
Programming
http://www.parc.xerox.com/aop/.
Project
home
page.
[2] A. Berlin and R. Surati, Partial evaluation for scientific computing:
the super computer experience, tech. rep., Artificial Intelligence Laboratory, Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts, May 1994.
[3] A. Berlin and D. Weise, Compiling scientific code using partial evaluation, Computer, 23 (1990), pp. 25–37.
[4] J. Bilmes, K. Asanović, C. whye Chin, and J. Demmel, Optimizing
matrix multiply using PHiPAC: a Portable, High-Performance, ANSI
C coding methodology, in Proceedings of International Conference on
Supercomputing, Vienna, Austria, July 1997.
[5] D. Björner, A. P. Ershov, and N. D. Jones, eds., Partial Evaluation and Mixed Computation, North-Holland, Amsterdam, 1988.
[6] S. Chiba, A Metaobject Protocol for C++, in OOPSLA’95, 1995,
pp. 285–299.
[7] K.
Czarnecki
and
U.
Eisenecker,
tures
for
template
metaprogramming.
online.de/home/Ulrich.Eisenecker/meta.htm.
[8]
Meta-control
struchttp://home.t-
, Generative Programming: Methods, Techniques and Applications,
Addison-Wesley Longman, 1999. (to appear).
14
REFERENCES
[9] K. Czarnecki, U. Eisenecker, R. Glück, D. Vandevoorde, and T. L. Veldhuizen, Generative Programming and Active Libraries, in Proceedings
of the 1998 Dagstuhl-Seminar on Generic Programming, vol. TBA of
Lecture Notes in Computer Science, 1998. (in review).
[10] O. Danvy, R. Glück, and P. Thiemann, eds., Partial Evaluation,
vol. 1110 of Lecture Notes in Computer Science, Springer-Verlag, 1996.
[11] D. R. Engler, Incorporating application semantics and control into
compilation, in USENIX Conference on Domain-Specific Languages
(DSL’97), October 15–17, 1997.
[12] D. R. Engler, W. C. Hsieh, and M. F. Kaashoek, ‘C: A language for
high-level, efficient, and machine-independent dynamic code generation,
in POPL’96, 1996, pp. 131–144.
[13] A. P. Ershov, On the essence of compilation, in Formal Description
of Programming Concepts, E. J. Neuhold, ed., North-Holland, 1978,
pp. 391–420.
[14] M. Frigo and S. G. Johnson, FFTW: The Fastest Fourier Transform
in the West. MIT LCS Technical Report No. MIT-LCS-TR-728, 1997.
[15] N. Fujinami, Automatic run-time code generation in C++, in ISCOPE’97, vol. TBA of Lecture Notes in Computer Science, 1997.
[16] Generative
Matrix
Computation
Library
http://nero.prakinf.tu-ilmenau.de/ czarn/gmcl.
home
page.
[17] Y. Ishikawa, A. Hori, M. Sato, M. Matsuda, J. Nolte, H. Tezuka,
H. Konaka, M. Maeda, and K. Kubota, Design and implementation of
metalevel architecture in C++ – MPC++ approach, in Reflection’96,
1996.
[18] Java Grande Forum home page. http://www.javagrande.org.
[19] N. D. Jones, An introduction to partial evaluation, ACM Computing
Surveys, 28 (1996), pp. 480–503.
[20] N. D. Jones, C. K. Gomard, and P. Sestoft, Partial Evaluation and Automatic Program Generation, Prentice Hall International, International
Series in Computer Science, June 1993.
REFERENCES
15
[21] S. Karmesin, J. Crotinger, J. Cummings, S. Haney, W. Humphrey,
J. Reynders, S. Smith, and T. Williams, Array design and expression
evaluation in POOMA II, in ISCOPE’98, vol. 1505, Springer-Verlag,
1998. Lecture Notes in Computer Science.
[22] P. Kleinrubatscher, A. Kriegshaber, R. Zöchling, and R. Glück, Fortran
program specialization, SIGPLAN Notices, 30 (1995), pp. 61–70.
[23] M. Leone and P. Lee, Lightweight run-time code generation, in Proceedings of the 1994 ACM SIGPLAN Workshop on Partial Evaluation
and Semantics-Based Program Manipulation, Technical Report 94/9,
Department of Computer Science, University of Melbourne, June 1994,
pp. 97–106.
[24] D. R. Musser and A. A. Stepanov, Algorithm-oriented generic libraries,
Software: Practice and Experience, 24 (1994), pp. 632–642.
[25] P. J. Plauger, Numerical C extensions group, C Users Journal, 11
(1993).
[26] M. Poletto, D. R. Engler, and M. F. Kaashoek, tcc: A system for fast,
flexible, and high-level dynamic code generation, in PLDI’97, 1996.
[27] S. Shende, A. D. Malony, J. Cuny, K. Lindlan, P. Beckman, and
S. Karmesin, Portable profiling and tracing for parallel, scientific applications using C++, in Proceedings of SPDT’98: ACM SIGMETRICS
Symposium on Parallel and Distributed Tools, August, 1998, pp. 134–
145.
[28] J. G. Siek and A. Lumsdaine, A rational approach to portable high performance: The basic linear algebra instruction set (BLAIS) and the
fixed algorithm size template (FAST) library, in Parallel Object Oriented Scientific Computing, ECOOP, 1998.
[29] J. Stichnoth and T. Gross, Code composition as an implementation language for compilers, in USENIX Conference on Domain-Specific Languages, 1997.
[30] T. L. Veldhuizen, Expression templates, C++ Report, 7 (1995), pp. 26–
31. Reprinted in C++ Gems, ed. Stanley Lippman.
[31]
, Using C++ template metaprograms, C++ Report, 7 (1995),
pp. 36–43. Reprinted in C++ Gems, ed. Stanley Lippman.
16
[32]
REFERENCES
, Arrays in Blitz++, in ISCOPE’98, vol. 1505 of Lecture Notes in
Computer Science, 1998.
[33] R. C. Whaley and J. J. Dongarra, Automatically Tuned Linear Algebra
Software, in Supercomputing’98, 1998.
| 6cs.PL
|
On decoding procedures of intertwining codes
Shyambhu Mukherjee
arXiv:1801.02010v1 [cs.IT] 6 Jan 2018
SMU Department, Indian Statistical Institute
Bangalore, Karnataka, India.
Joydeb Pal
Department of Mathematics
National Institute of Technology Durgapur
Burdwan, India.
Satya Bagchi
Department of Mathematics
National Institute of Technology Durgapur
Burdwan, India.
Abstract
One of the main weakness of the family of centralizer codes is that its length is
always n2 . Thus we have taken a new matrix equation code called intertwining code.
Specialty of this code is the length of it, which is of the form nk. We establish two
decoding methods which can be fitted to intertwining codes as well as for any linear
codes. We also show an inclusion of linear codes into a special class of intertwining
codes.
Keywords: Linear codes, Centralizer codes, Syndrome decoding.
MSC: 94B05, 94B35, 15A24.
Email addresses: [email protected] (Shyambhu Mukherjee),
[email protected] (Joydeb Pal), [email protected] (Satya Bagchi)
Preprint submitted to Elsevier
January 9, 2018
1. Introduction
A code of length n2 is obtained by taking centralizer of a matrix from the vector
space Fqn×n . As a consequence, it cannot reach to most of the sizes. Whereas a code
of length n · k is formed by taking the solutions of matrix equation for some matrices
A ∈ Fqn×n and C ∈ Fqk×k over Fq . Here we have taken one such matrix equation of
the form AB = BC where the matrix A ∈ Fqn×n and the matrix C ∈ Fqk×k . Thus
the set of solutions R(A, C) = {B ∈ Fqn×k |AB = BC} gives a code of length n · k by
a similar construction in [1] [2] [4]. This code is named intertwining code [3]. This
code can extend the use of better decoding ability of GTC codes [4] into a vast class
of linear codes.
Finding efficient error correcting procedure for a linear code is a challenging problem.
If we look upon centralizer codes and twisted centralizer codes, we see that there was
a very nice method to detect and correct single error using syndrome. This technique
cannot provide an easy task for correcting more than single errors. Thus we present
two algorithms to do better decoding procedure. Our algorithms work for the family
of intertwining codes as well as for any linear codes.
In this paper, we explore few properties on intertwining codes. Then we show that
there exists a intertwining code for which a certain linear code is a subcode of it.
We find a way to effectively find an upper bound on the minimum distance of the
code along with proving the existence of a certain weight codeword based upon the
matrices A and C. At last we show a possible way to find the matrix pair (A, C) of
intertwining code related to a linear code in total computational perspective.
Throughout this paper we denote Fq as a finite field with q elements and Fqn×k as the
set of all matrices of order n × k over Fq . We take two matrices A and C from the
vector spaces Fqn×n and Fqk×k respectively and also O denotes the null matrix.
Definition 1.1. For any matrices A ∈ Fqn×n and C ∈ Fqk×k , the set R(A, C) = {B ∈
Fqn×k |AB = BC} is called intertwining code [3].
Clearly, the set R(A, C) is a linear subspace of the vector space Fqn×n and hence it is
a linear code. The formation of the code is very similar to [1] [2] [4]. We develop few
basic results on intertwining codes as follows.
1. If A ∈ R(A, C) then C should be an n × n matrix. Now let A belongs to the
intertwining code then A2 = AC. If A is idempotent then A(C − In ) = O and
2
if A2 = 0 then C belongs to the matrix wise null space of A. But these are not
possibly the whole solution of the equation A2 = AC.
2. If A and C are invertible matrices of order n and k respectively, then R(A, C)
is isomorphic to R(A−1 , C −1 ).
3. Let X ∈ Fqn×n and Y ∈ Fqk×k are two invertible matrices. Then R(A, C) is
isomorphic to conjugate code R(XAX −1 , Y CY −1 ).
4. Let A is such a matrix that c is not an eigenvalue and C = cIk . Then the matrix
equation AB = BC possesses a trivial solution, i.e., R(A, C) = {0}.
2. Analysis on weight distribution
Theorem 2.1. Let J ∈ F2n×k be the matrix with all entries are 1. If J ∈ R(A, C)
then the weight distribution of the code R(A, C) is symmetric, i.e., Ai = Ank−i for
all i. In addition, if the matrix A is of row sum equal to 0 and C is of column sum
equal to 0 then J ∈ R(A, C).
Proof. Let B1 ∈ R(A, C) with weight i where 0 ≤ i ≤ nk. If J ∈ R(A, C) then
J + B1 ∈ R(A, C), since R(A, C) is a linear code. Now weight of J + B1 is nk − i.
So, whenever a codeword of weight i appears, simultaneously there exists a codeword
with weight nk − i. Thus we can say weight distribution is symmetric, i.e., number
of codewords with weight i = number of codewords with weight nk − i, where i is a
positive integer with 0 ≤ i ≤ nk, i.e., Ai = Ank−i .
For the second part, observe that AJ = [row-1-sum row-2-sum . . . row-n-sum]T ·
[1 1 1 1 . . . 1] and similarly observe that JC = [1 1 1 . . . 1]T · [col-1-sum col-2-sum
. . . col-k-sum]. Hence it is clear that if the matrices A has row sum equal to 0 and C
has column sum equal to 0, then J satisfies the equation AJ − JC = O since AJ = O
and JC = O.
2.1. Existence of a certain weight codeword
Consider the matrices A and C for which R(A, C) has been constructed. Let us
assume that none of them are invertible. Therefore, there will be dependent columns
and rows. If dA is the minimum number of columns which are dependent in A and
dC is the minimum number of rows which are dependent in C, then at least one
P
codeword in R(A, C) of weight dA · dC exists. In this case, we have
ci Ai = 0 with
ci 6= 0, ∀i as the relation is taken to be of minimum number of columns. Then we
3
have a column vector of those ci ’s. Similarly, we get a row vector from C. Then we
take the multiplication of these vectors as n × 1 into 1 × k matrix multiplication. So
this vector will give us an element of TO = {B ∈ Fqn×k |AB = BC = O}.
As the codeword of weight dA · dC exists, by construction this weight dA · dC is less
than or equal to the (rA + 1) · (rC + 1) hence the minimum weight is also less than or
equal to (rA + 1) · (rC + 1). We list the result in the above discussion as follows.
Theorem 2.2. For an intertwining code generated by A and C, there exists at least
one codeword of weight dA · dC and therefore the minimum distance d of R(A, C) is
less than or equal to dA · dC , i. e., d(R(A, C)) ≤ dA · dC .
Proof. The product code Ker A⊗Ker C T belongs to intertwining code. Now according
to definition of dA there exists dA number of columns dependent such that no set of
lesser cardinality is dependent. Therefore a column vector v of weight dA exists such
that Av = 0. Similarly we find a row vector w of weight dC with wC = 0 therefore
vw exists in the code with the rest following.
3. Decoding process
Encoding procedure for intertwining codes are similar to the encoding procedures
mentioned in [1] [2] [4]. To check whether a message is erroneous, it is required to
define syndrome.
Definition 3.1. Let A be a square matrix of order n and C be a square matrix of
order k. Then the syndrome of an element B ∈ Fqn×k with respect to the intertwining
code R(A, C) is defined as SA,C (B) = AB − BC.
If a word is erroneous then SA,C (B) = AB − BC 6= O. It is an easiest technique to
check whether a codeword belongs to the code or not. Here we propose two algorithms
to correct errors in an intertwining code.
3.1. Algorithm 1:
As we do not know how to find the minimum weight element algorithmically hence
we can not keep it inside our decoding process as it needs a lot of time. That’s
why we modify our procedure a bit for the Step 3. As the vector space Fqn×k can
be broken down into intertwining code and its cosets so we will calculate minimum
4
weight element first for all the cosets separately. Now we input list of coset leaders
as a table inside our decoding system.
Step 1. Receiver received a word B ′ from the channel.
Step 2. Calculate the syndrome SA,C (B ′ ) = AB ′ − B ′ C. If SA,C (B ′ ) = O then the
transmitted codeword is B ′ and goto Step 5. Otherwise, goto Step 3.
Step 3. Find this syndromes corresponding least weight error matrix E, already stored
in the table. The matrix E is the error pattern for the word B ′ .
Step 4. Since, both B ′ and the E belong to the same coset B ′ + R(A, C), then B ′ − E
is the transmitted codeword of the code R(A, C).
Step 5. End.
Caution! This table may look like syndrome look-up table and it will work similarly,
but here we must remember following differences.
1. This syndrome is different from the standard syndrome which is obtained by
multiplying the codeword of length n with a (n − k) × n parity-check matrix,
resulting into n − k length i.e. 7 − 4 = 3 length for hamming code. Here we will
get a whole matrix of length n2 , same length as our codewords, as the minimum
weight codeword of a coset.
2. This table is not the syndrome look-up table for the intertwining code. A
different one can be constructed but the construction will be more complex if
we try to do so.
3.2. Algorithm 2:
In the previous algorithm, we have to store whole of a table which occupy a good
amount of memory. So, we provide a better algorithm which does not take the
memory that much and achieves the least weight error matrix or error pattern by this
algorithm.
For this algorithm, partition the codewords of the code R(A, C) by its weight distribution. Let R(A, C) = ∪ki=1 Ak , where Aj = {c ∈ R(A, C) : wt(c) = aj }. We easily
see that Ai ∩ Aj = φ, ∀ i 6= j. Now weights available are a0 , a1 , . . . , ak . Now, the
algorithm is presented below.
Step 1. Receiver received a word B ′ and start to calculate its syndrome SA,C (B ′ ) =
AB ′ − B ′ C. To reduce complexity for computing the syndrome, we calculate
bitwise syndrome. Whenever a nonzero bit appears in the syndrome, we stop
5
the computation of the syndrome and goto Step 2. Otherwise if SA,C (B ′ ) = 0
then the transmitted word is B ′ and goto Step 5.
Step 2. Find b = wt(B ′ ), weight of B ′ . Now evaluate the intervals in two cases. If
n ≥ b + ai , then we take the interval [|b − ai |, b + ai ], otherwise take the interval
[|b − ai |, 2n − b − ai ].
⌋
Step 3. Delete those intervals whose lower bounds are greater than t, where t = ⌊ d−1
2
and d is the minimum distance of the code R(A, C).
Step 4. Arrange remaining intervals in ascending order of the lower bounds of intervals.
We store ordering of index. Take the first interval then take corresponding
weight. Let am be the corresponding weight. Now find S = {A + B ′ : wt(A +
B ′ ) ≤ wt(A′ + B ′ ) ∀ A′ ∈ Am }. Select A such that wt(A + B ′ ) is minimum
for A ∈ Am . If not unique, choose one A randomly. Let E = A + B ′ . Find
weight of E and then delete intervals with lower bound greater than or equal
to wt(E). Go to next partition. In same procedure find the least weight word
E ′ . Compare with the previous least weight word E. Choose minimum weight
among these and save it to E. Continuing this process for further partitions
we will get a matrix E, which is the error pattern. Therefore the transmitted
codeword was B ′ − E.
Step 5. End.
Note: In Step 4 of the above algorithm, error pattern matrix E is always unique
because if there are E1 and E2 such that both B ′ − E1 and B ′ − E2 have weight less
than t then distance between E1 and E2 will be less than 2t < d which is impossible
for two codewords.
Analysis: In our algorithm we will store the code sorted according to weights. So we
will have to store 2k codewords at our worse, where k is the dimension of the code.
This is less than 2n−k if k <
less than
n
,
2
n
,
2
where n is the length. So for a code of dimension
our algorithm takes less memory. Now we can view the weight wise
sorted codewords as non-linear codes of constant weight. So we can store each sorted
partition using the standard representation of a non-linear code using kernel of it and
it’s coset representatives as discussed in [5]. Thus this will take shorter memory even.
4. Can any linear code be represented as a subcode of intertwining code?
Let us consider an l-dimensional subspace of the nk-dimensional vector space over
Fq . This nk dimensional vector space can be represented as the vector space of
6
matrices Fqn×k . Can we find a pair of matrices A ∈ Fqn×n and C ∈ Fqk×k to construct
an intertwining code R(A, C) which contains the l dimensional linear code? This
problem can be formulated as follows.
Given a linear code C over Fq of length nk and dimension l. So, the linear code C
has a generator matrix G of order l × nk. Can we represent the linear code C as a
subcode of an intertwining code R(A, C)?
4.1. Forming the equations and existence of solutions
We represent each row of the generating matrix as an n×k order matrix. So, there are
l linearly independent matrices B1 , B2 , . . . , Bl corresponding to the generator matrix
of the known linear code C. Our aim is to find two such non-zero matrices A and C
which satisfy the equation ABi = Bi C for each Bi , i = 1, 2, . . . , l. To find A and C,
let us consider the entries of A and C are variables. Then we get n2 + k 2 variables.
For each matrix Bi , there are nk equations and there will be a total nkl equations
satisfying these variables. Here we use the mapping ¯ : Fqn×n → Fqn
2 ×1
with B 7→ B̄,
where the matrix B̄ is formed by concatenating columns of B. Now the equation
ABi = Bi C can be written as
" #
" #
h
i Ā
Ā
= O,
= O ⇒ Di
ABi − Bi C = O ⇒ In ⊗ BiT | −Bi ⊗ Ik
C̄
C̄
i
h
iT
h
T
where Di = In ⊗ Bi | − Bi ⊗ Ik . Let D = D1 D2 · · · Dl . Then the above
system of l equations is written as
D
" #
Ā
C̄
= O.
(1)
Here Di is coming from each Bi . Now the final solution is the solution of the equation
(1). The matrix D is of order nkl×(n2 +k 2 ). Thus the existence of non-trivial solution
of above equation is reached if n2 + k 2 ≥ nkl. This is a sufficient condition.
Now we see
block In ⊗BiT .
h that each Di consists
i ofh two blocks, i.e.,
i hBi ⊗Ik and another
i
So Di = In ⊗ BiT | − Bi ⊗ Ik = In ⊗ BiT | O + O | − Bi ⊗ Ik = A1 + A2 .
Now rank(Di ) ≤ rank(A1 ) + rank(A2 ). Clearly, rank(A1 ) = n · rank(BiT ) = n ·
rank(Bi ) and rank(A2 ) = k · rank(Bi ). Now, we get rank(Di ) ≤ (n + k) · rank(Bi ).
Pl
So, rank(D) ≤
i=1 (n + k) · rank(Bi ). Now the final solution space will have
P
2
2
dimension ≥ n + k − li=1 (n + k) · rank(Bi ).
7
5. Conclusion
Centralizer codes, twisted centralizer codes and generalized twisted centralizer codes
have length n2 which is a reason that it cannot fit to most of the famous linear codes.
But intertwining codes can reach most of the linear codes because it is of length nk.
So, we have taken intertwining codes and try to make a correspondence between it
and existing linear codes. We have found an upper bound on minimum distance and
proposed two decoding algorithms which take less storage memory.
Acknowledgements
The author Joydeb Pal is thankful to DST-INSPIRE for financial support to pursue
his research work.
References
[1] A. Alahmadi, S. Glasby, C. E. Praeger, P. Solé, B. Yildiz, Twisted centralizer
codes, Linear Algebra and its Applications, 524 (2017) 235 – 249.
[2] A. Alahmadi, S. Alamoudi, S. Karadeniz, B. Yildiz, C. E. Praeger, P. Solé, Centraliser codes, Linear Algebra and its Applications, 463 (2014) 68 – 77.
[3] S. Glasby, C. E. Praeger, On the minimum distance of intertwining codes,
arXiv:1711.04104, 2017.
[4] J. Pal, P. K. Maurya, S. Mukherjee, S. Bagchi, Generalized twisted centralizer
codes, arXiv:1709.01825, 2017.
[5] M. Villanueva, F. Zeng, J. Pujol, Efficient representation of binary nonlinear
codes: constructions and minimum distance computation, Designs, Codes and
Cryptography, 76 (2015) 3 – 21.
8
| 7cs.IT
|
Structured Memory for Neural Turing Machines
arXiv:1510.03931v3 [cs.AI] 25 Oct 2015
Wei Zhang ∗ Yang Yu Bowen Zhou
IBM Watson
zhangwei,yu,[email protected]
Abstract
Neural Turing Machines (NTM) [2] contain memory component that simulates
“working memroy” in the brain to store and retrieve information to ease simple
algorithms learning. So far, only linearly organized memory is proposed, and
during experiments, we observed that the model does not always converge, and
overfits easily when handling certain tasks. We think memory component is key
to some faulty behaviors of NTM, and better organization of memory component
could help fight those problems. In this paper, we propose several different structures of memory for NTM, and we proved in experiments that two of our proposed
structured-memory NTMs could lead to better convergence, in term of speed and
prediction accuracy on copy task and associative recall task as in [2].
1
Introduction
Memory components for Neural Networks has been recently introduced in Neural Turing Machines
(NTMs) [2], Memory Networks [1], and Dynamic Memory Networks [3]. The purposes of those
memory components are similar, which is to simulate “working memory” in the brain to store temporary information through time to be used by attention module for reading and writing. In this
paper, our focus is on NTMs. They are attractive in that memory could be randomly accessed by
controller through blurry “erase” and “write” operations, which well aligns with Turing Machine
operations. When memory size is small, nice convergence of the NTM model could be frequently
observed, although not always. But when memory size is large, the model struggles to convergence,
and sometimes the test loss jumps drastically in a large range, which is a sign of overfitting, as was
observed in our experiments. We think that the dynamics of memory contents takes a crucial role
in controlling model convergence speed and quality. Thus, we proposed and experimented different memory structures to explore if specific memory structure could lead to more stable memory
contents, or say, perform “memory smoothing” so that the memory content generated after read and
write operations does not deviate too far from the “expected” memory content, so that overfitting
of parameters could be in turn alleviated. We proposed three different architectures, NTM1, NTM2
and NTM3, which are explained in detail in section 2, and experiments are shown in section 3.
2
Architectures
We show the NTM original architecture introduced by Graves et al. [2] and three variants NTM1,
NTM2 and NTM3 in Figure 1. Details of each model is explained in a moment. We introduce a
notion of Memory visibility: We call memory “Controlled” if it is modified by controller outputs
through write heads directly, or “Hidden” if not.
NTM1 Compared to NTM, NTM1 has an additional hidden memory Mh , which is not controlled by
controller module, but is connected to the controlled memory Mc . The hidden memory accumulates
the content in the controlled memory, so that a type of memory smoothing is performed to prevent
∗
Alternative email for Wei Zhang: [email protected]
1
Figure 1: NTM and NTM variants that use Long short-term memory [5] as controllers. Note that
every module in those modules are updated recurrently through time using their previous states.
NTM1 contains only one write head, and one of the memory Mh is not written directly by controller,
which is different from M2 in NTM2 that is written both by M1 and write head simultaneously.
NTM3 is different from NTM2 in that NTM3 write heads takes inputs from two layers. NTM3
could also be expanded to multiple layers as well.
memory from deviating from the “expected content”. Specifically, the memory content for hidden
memory Mh (t) of time t is generated as:
write : Mc (t) = h Mc (t − 1), w(t − 1), c(t)
update : Mh (t) = aMh (t − 1) + bMc (t)
read : r(t) = wr (t)Mh (t)
Mc is updated by t − 1 time write head to generate t time controlled memory output, which in turn
is used to update hidden memory Mh , and then the new hidden memory is used to generate read
head at time t. w(t) is write weights, c(t) is the controller output of time t, and h() is the function
that updates controlled memory and write weights that implement “erase and add” operations. wr
is read weights. a and b are scalar mixture weights, which could be further extended to tensors. We
use scalars in this work. The read head is reading from Mh (t) instead of M(t) as is did in [2].
NTM2 The second architecture NTM2 is similar to NTM1 in that two memory blocks are used, and
they are connected hierarchically. However, the difference from NTM1 is that hidden memory in
NTM2 is no longer hidden from the controller, but connected to controller outputs through another
write head. So two memory blocks are all controlled memories connecting to two different write
heads. The upper level memory is denoted as M1 , and the lower level one as M2 . M1 is modified
solely by L1 write head, but M2 is modified by both M1 contents and L2 write head. The logic is:
write : M1 (t), w1 (t) = h M1 (t − 1), w1 (t − 1), c(t)
f 2 (t), w2 (t) = h M2 (t − 1), w2 (t − 1), c(t)
M
f 2 (t) + bM1 (t)
update : M2 (t) = aM
read : r(t) = wr (t)M2 (t)
Additionally, M1 and M2 are all generated by the same head function with same controller input,
but with different write weights. Single read head r is used, which means the output is only read
from the lowest layer memory (M2 in this case). Note that the architecture could be easily expanded
into multiple layers by increasing number of heads and memory blocks.
NTM3 The third architecture is significantly different from the previous two. First, the controller
in the model have multiple layers. Second, each layer output is connecting to a memory by write
heads. Different layers of memory contains different level of transformation of the input. In the
case where multi-layer LSTM is used as NTM controller, output from each layer goes through the
non-linear transformation of write heads, and writes to each individual memory. Then the deeper
2
Figure 2: comparison of convergence speed and quality on copy task and associative recall task. Four
figures to the left are all from copy task, and four on the right are for associative recall task. For
all the figures, X-axis is the number of iterations (scaled by sampling once for every 25 iterations),
and for copy task, Y-axis shows the binary cross entropy loss per item (sequence of length 8), and in
recall task, Y-axis shows log of binary cross entropy loss. Memory is 128 vectors of length 20, for
every memory block shown in Figure 1.
level memory is updated by upper level memory and write head jointly.
write : M1 (t), wL1 (t) = h M1 (t − 1), wL1 (t − 1), cL1 (t)
f 2 (t), wL2 (t) = h M2 (t − 1), wL2 (t − 1), cL2 (t)
M
f 2 (t) + bM1 (t)
update : M2 (t) = aM
read : r(t) = wr (t)M2 (t)
In so doing, the memory blocks receives write operations not only from the final layer LSTM output, but also from the intermediate layer outputs as well. The purpose is to smooth the final layer
memoroy with intermediate LSTM outputs. We can see in a moment how those three performs on
copy and associative recall task.
3
Experiments
The experiments is to show the convergence speed and quality of those three variants, compared to
the NTM setting. In theory, the convergence should happen for every run of every model, which
is not the case in practice. Randomness of parameter initialization might be the reason1 . But if
model convergences, the number of iterations used were quite stable across runs. Thus, to make
the evaluation reliable, we repeat experiments at least 5 times for each model and choose the most
frequent circumstance for showing.
The tasks we choose for model testing are copy task and associative recall task as in [2]. Associative
recall task requires the model to target an item in sequence, and later on shift to the immediate next
item to give as an answer. Copy task requires the model to consecutively output items in original
input order, which requires long-term location based addressing capability. In each training iteration,
we randomly generated each item as a binary vector of length 8. The number of items is random as
well. In recall task, the items, and number of items to be generated are all chosen at random. All
those randomness is to guarantee that every training iteration uses different training example, so that
overfit to specific data set is no longer an issue, to guarantee that algorithms could be learned. This
is critical to our further analysis.For all the models, the training criteria is binary cross entropy, and
we use RMSProp [4] for optimization, with learning rate 1E − 4, momentum 0.9, and decay 0.95.
In Figure 2 we show our results on copy and recall respectively. Copy tasks are four graphs to the
left. X-axis shows the number of iteration used (sampled every 25 iterations), and Y-axis is the
1
We applied the technique used by https://github.com/kaishengtai/torch-ntm to initialize each memory slot
to define “write/read order” prior
3
binary cross entropy loss of each iteration. The shown result in each graph is representative among
5 runs, among which the selected curve is observed most frequently. We can see that NTM1 and
NTM2 shows faster convergence and less outliers above the convergence range (closely around 0
near X-axis) than NTM or NTM3. The outliers (extremely high loss) means that the prediction on
the item in that iteration is significantly incorrect, which is a sign of overfitting or underfitting. In
NTM1 and NTM2 runs we see very few outliers, but NTM shows more outliers when using two
read and write heads (the curve with sharp values), or does not converge as fast when using 1 read
and 1 write head (flat curve). This pattern is consistently observed over 5 runs as well. We can
also read out from Figure “NTM-copy” that using 1 read and 1 write head converges significantly
slower than 2w/r heads. The series of experiments on copy task confirms our assumption that the
introduction of additional memory in NTM1 and NTM2 do help stabilize the memory component,
which in turn leads to better tuning of the parameters. However, NTM3 does not show nice curves,
and produce outlier almost every 1000 iterations or so. This means that the memory that is attached
to the intermediate layers of LSTM introduces more noise, which is unexpected. Although less
stable, NTM does converge about 500 iterations faster than NTM1 and NTM2, but we observed the
opposite in associative recall task.
For associative recall, we can see that outliers are produced much more frequently when loss significantly reduces, and we rarely observe convergence of original NTM or NTM3. In “NTM3-recall”
graph, we show a non-converged run since it happens frequently for NTM3. But NTM1 and NTM2
converges more frequently across runs. And NTM2 shows much faster convergence, roughly with
37,000 iterations, compared to 50,000 for NTM or NTM1. Moreover, NTM1 and NTM2 generate less outliers than NTM. This means that, the prediction accuracy for NTM1 and NTM2 will be
higher than NTM, since more training examples are correctly predicted than NTM. And, NTM1 and
NTM2 will have higher probability to be a non-overfitting model than NTM. Unfortunately, NTM3
model does not converge as frequent as NTM1 and NTM2, which aligns with the observation in
copy task about the difference of those models.
4
Conclusion
This paper discussed three new structured memory architectures for Neural Turing Machines, and
showed that organizing memory blocks in a proper hierarchical manner could alleviate overfitting
and sometimes increase predictive accuracy compared to NTM.
5
Future work
In the future we would also try NTM1 NTM2 on data sets that require more complex reasoning. We
tested NTM on a synthetic QA data set proposed in [1], and observed the instability in convergence.
We would try NTM1 and NTM2 on the same data set.
References
[1] Jason Weston, Sumit Chopra, & Antoine Bordes. Memory Networks. International Conference
on Representation Learning, 2015
[2] Alex Graves, Greg Wayne, & Ivo Danihelka.
arXiv:1410.5401. 2014
Neural Turing Machines.
arXiv preprint
[3] Ankit Kumar, Ozan Irsoy, Jonathan Su, James Bradbury, Robert English, Brian Pierce, Peter
Ondruska, Ishaan Gulrajani, & Richard Socher. Ask Me Anything: Dynamic Memory Networks for
Natural Language Processing. arXiv preprint arXiv:1506.07285. 2015
[4] Tijmen Tieleman, & Geoffrey Hinton. ”Lecture 6.5-rmsprop: Divide the gradient by a running
average of its recent magnitude.” COURSERA: Neural Networks for Machine Learning 4. 2012
[5] Sepp Hochreiter, & Jrgen Schmidhuber. Long short-term memory. Neural computation, no. 8:
1735-1780. 1997
4
| 9cs.NE
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:cs/0401022v1 [cs.PL] 26 Jan 2004
Enhanced Sharing Analysis Techniques:
A Comprehensive Evaluation
ROBERTO BAGNARA, ENEA ZAFFANELLA∗
Department of Mathematics, University of Parma, Italy
(e-mail: {bagnara,zaffanella}@cs.unipr.it)
PATRICIA M. HILL†
School of Computing, University of Leeds, Leeds, U.K.
(e-mail: [email protected])
Abstract
Sharing, an abstract domain developed by D. Jacobs and A. Langen for the analysis of
logic programs, derives useful aliasing information. It is well-known that a commonly
used core of techniques, such as the integration of Sharing with freeness and linearity
information, can significantly improve the precision of the analysis. However, a number
of other proposals for refined domain combinations have been circulating for years. One
feature that is common to these proposals is that they do not seem to have undergone
a thorough experimental evaluation even with respect to the expected precision gains. In
this paper we experimentally evaluate: helping Sharing with the definitely ground variables
found using Pos, the domain of positive Boolean formulas; the incorporation of explicit
structural information; a full implementation of the reduced product of Sharing and Pos;
the issue of reordering the bindings in the computation of the abstract mgu; an original
proposal for the addition of a new mode recording the set of variables that are deemed to
be ground or free; a refined way of using linearity to improve the analysis; the recovery
of hidden information in the combination of Sharing with freeness information. Finally,
we discuss the issue of whether tracking compoundness allows the computation of more
sharing information.
KEYWORDS: Abstract Interpretation; Logic Programming; Sharing Analysis; Experimental Evaluation.
1 Introduction
In the execution of a logic program, two variables are aliased or share at some
program point if they are bound to terms that have a common variable. Conversely,
two variables are independent if they are bound to terms that have no variables in
common. Thus by providing information about possible variable aliasing, we also
provide information about definite variable independence. In logic programming,
∗ The work of the first and second authors has been partly supported by MURST projects “Certificazione automatica di programmi mediante interpretazione astratta” and “Interpretazione
astratta, sistemi di tipo e analisi control-flow.”
† This work was partly supported by EPSRC under grant GR/M05645.
2
R. Bagnara, E. Zaffanella, and P. M. Hill
a knowledge of the possible aliasing (and hence definite independence) between
variables has some important applications.
Information about variable aliasing is essential for the efficient exploitation of
AND-parallelism (Bueno et al. 1994; Bueno et al. 1999; Chang et al. 1985; Hermenegildo and Greene 1990;
Hermenegildo and Rossi 1995; Jacobs and Langen 1992; Muthukumar and Hermenegildo 1992).
Informally, two atoms in a goal are executed in parallel if, by a mixture of compiletime and run-time checks, it can be guaranteed that they do not share any variable.
This implies the absence of binding conflicts at run-time: it will never happen that
the processes associated to the two atoms try to bind the same variable.
Another significant application is occurs-check reduction (Crnogorac et al. 1996;
Søndergaard 1986). It is well-known that many implemented logic programming
languages (e.g., almost all Prolog systems) omit the occurs-check from the unification procedure. Occurs-check reduction amounts to identifying the unifications
where such an omission is safe, and, for this purpose, information on the possible
aliasing of program variables is crucial.
Aliasing information can also be used indirectly in the computation of other
interesting program properties. For instance, the precision with which freeness information can be computed depends, in a critical way, on the precision with which
aliasing can be tracked (Bruynooghe et al. 1994a; Codish et al. 1993; Filé 1994;
King and Soper 1994; Langen 1990; Muthukumar and Hermenegildo 1991).
In addition to these well-known applications, a recent line of research has shown
that aliasing information can be exploited in Inductive Logic Programming (ILP).
Several optimizations have been proposed for speeding up the refinement of inductively defined predicates in ILP systems (Blockeel et al. 2000; Santos Costa et al. 2000).
It has been observed that the applicability of some of these optimizations, formulated in terms of syntactic conditions on the considered predicate, could be recast
as tests on variable aliasing (Blockeel et al. 2000, Appendix D).
Sharing, a domain due to D. Jacobs and A. Langen (Jacobs and Langen 1989;
Jacobs and Langen 1992; Langen 1990), is based on the concept of set-sharing.
An element of the Sharing domain, which is a set of sharing-groups (i.e., a set
of sets of variables), represents information on groundness,1 groundness dependencies, possible aliasing, and more complex sharing-dependencies among the variables that are involved in the execution of a logic program (Bagnara et al. 1997;
Bagnara et al. 2002; Bueno et al. 1994; Bueno et al. 1999).
Even though Sharing is quite precise, it is well-known that more precision is attainable by combining it with other domains. Nowadays, nobody would seriously
consider performing sharing analysis without exploiting the combination of aliasing
information with groundness and linearity information. As a consequence, expressions such as ‘sharing information’, ‘sharing domain’ and ‘sharing analysis’ usually
capture groundness, aliasing, linearity and quite often also freeness. Notice that
this idiom is nothing more than a historical accident: as we will see in the sequel,
1
A variable is ground if it is bound to a term containing no variables, it is compound if it is
bound to a non-variable term, it is free if it is not compound, it is linear if it is bound to a
term that does not contain multiple occurrences of a variable.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
3
compoundness and other kinds of structural information could also be included in
the collective term ‘sharing information’.
As argued informally by H. Søndergaard (Søndergaard 1986), linearity information can be suitably exploited to improve the accuracy of a sharing analysis.
This observation has been formally applied in (Codish et al. 1991) to the specification of the abstract mgu operator for ASub, a sharing domain based on the
concept of pair-sharing (i.e., aliasing and linearity information is encoded by a set
of pairs of variables). A similar integration with linearity for the domain Sharing
was proposed by Langen in his PhD thesis (Langen 1990). The synergy attainable
from the integration between aliasing and freeness information was pointed out
by K. Muthukumar and M. Hermenegildo (Muthukumar and Hermenegildo 1992).
Building on these works, W. Hans and S. Winkler (Hans and Winkler 1992) proposed a combined integration of freeness and linearity information with sharing,
but small variations (such as the one we will present as the starting point for our
work) have been developed by M. Bruynooghe et al. (Bruynooghe and Codish 1993;
Bruynooghe et al. 1994a).
There have been a number of other proposals for more refined combinations
which have the potential for improving the precision of the sharing analysis over
and above that obtainable using the classical combinations of Sharing with linearity
and freeness. These include the implementation of more powerful abstract semantic operators (since it is well-known that the commonly used ones are sub-optimal)
and/or the integration with other domains. Not one of these proposals seem to have
undergone a thorough experimental evaluation, even with respect to the expected
precision gains. The goal of this paper is to systematically study these enhancements and provide a uniform theoretical presentation together with an extensive
experimental evaluation that will give a strong indication of their impact on the
accuracy of the sharing information.
Our investigation is primarily from the point of view of precision. Reasonable
efficiency is also clearly of interest but this has to be secondary to the question
as to whether precision is significantly improved: only if this is established, should
better implementations be researched. One of the investigated enhancements is
the integration of explicit structural information in the sharing analysis and an
important contribution of this paper is that it shows both the feasibility and the
positive impact of this combination.
Note that, regardless of its practicality, any feasible sharing analysis technique
that offers good precision may be valuable. While inefficiency may prevent its adoption in production analyzers, it can help in assessing the precision of the more
competitive techniques.
The present paper, which is an improved and extended version of (Bagnara et al. 2000),
is structured as follows. In Section 2, we define some notation and recall the definitions of the domain Sharing and its standard integration with freeness and linearity
information denoted as SFL. In Section 3, we briefly describe the China analyzer,
the benchmark suite and the methodology we follow in the experimental evaluations. In each of the next seven sections, we describe and experimentally evaluate
different enhancements and precision optimizations for the domain SFL. Section 4
4
R. Bagnara, E. Zaffanella, and P. M. Hill
considers a simple combination of Pos with SFL; Section 5 investigates the effect of
including explicit structural information by means of the Pattern(·) construction;
Section 6 discusses possible heuristics for reordering the bindings so as to maximize
the precision of SFL; Section 7 studies the implementation of a more precise combination between Pos and SFL; Section 8 describes a new mode ‘ground or free’ to
be included in SFL; Section 9 and Section 10 study the possibility of improving the
exploitation of the linearity and freeness information already encoded in SFL. In
Section 11 we discuss (without an experimental evaluation) whether compoundness
information can be useful for precision gains. Section 12 concludes with some final
remarks.
2 Preliminaries
For any set S, ℘(S) denotes the powerset of S. For ease of presentation, we assume
there is a finite set of variables of interest denoted by VI . If t is a syntactic object
then vars(t) and mvars(t) denote the set and the multiset of variables in t, respectively. If a occurs more than once in a multiset M we write a A M . We let Terms
denote the set of first-order terms over VI . Bind denotes the set of equations of
the form x = t where x ∈ VI and t ∈ Terms is distinct from x. Note that we
do not impose the occurs-check condition x ∈
/ vars(t), since we target the analysis
of Prolog and CLP systems possibly omitting this check. The following simplification of the standard definitions for the Sharing domain (Cortesi and Filé 1999;
Hill et al. 1998; Jacobs and Langen 1992) assumes that the set of variables of interest is always given by VI .2
Definition 1
(The set-sharing domain SH .) The set SH is defined by
def
SH = ℘(SG),
where the set of sharing-groups SG is given by
def
SG = ℘(VI ) \ {∅}.
SH is ordered by subset inclusion. Thus the lub and glb of the domain are set union
and intersection, respectively.
Definition 2
(Abstract operations over SH .) The abstract existential quantification on SH
causes an element of SH to “forget everything” about a subset of the variables of
interest. It is encoded by the binary function aexists : SH × ℘(VI ) → SH such that,
2
Note that, during the analysis process, the set of variables of interest may expand (when solving
the body of a clause) and contract (when abstract descriptions are projected onto the variables
occurring in the head of a clause). However, at any given time the set of variables of interest is
fixed. By consistently denoting this set by VI , we simplify the presentation, since we can omit
the set of variables of interest to which an abstract description refers.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
5
for each sh ∈ SH and V ∈ ℘(VI ),
def
aexists(sh, V ) = S \ V S ∈ sh, S \ V 6= ∅ ∪ {x} x ∈ V .
For each sh ∈ SH and each V ∈ ℘(VI ), the extraction of the relevant component
of sh with respect to V is given by the function rel : ℘(VI ) × SH → SH defined as
def
rel(V, sh) = { S ∈ sh | S ∩ V 6= ∅ }.
For each sh ∈ SH and each V ∈ ℘(VI ), the function rel : ℘(VI ) × SH → SH
gives the irrelevant component of sh with respect to V . It is defined as
def
rel(V, sh) = sh \ rel(V, sh).
The function (·)⋆ : SH → SH , also called star-union, is given, for each sh ∈ SH ,
by
n
[
⋆ def
Ti .
sh = S ∈ SG ∃n ≥ 1 . ∃T1 , . . . , Tn ∈ sh . S =
i=1
For each sh 1 , sh 2 ∈ SH , the function bin : SH × SH → SH , called binary union,
is given by
def
bin(sh 1 , sh 2 ) = { S1 ∪ S2 | S1 ∈ sh 1 , S2 ∈ sh 2 }.
We also use the self-bin-union function sbin : SH → SH , which is given, for each
sh ∈ SH , by
def
sbin(sh) = bin(sh, sh).
The function amgu: SH × Bind → SH captures the effect of a binding on an
element of SH . Assume (x = t) ∈ Bind , sh ∈ SH , Vx = {x}, Vt = vars(t), and
Vxt = Vx ∪ Vt . Then
def
(1)
amgu(sh, x = t) = rel(Vxt , sh) ∪ bin rel(Vx , sh)⋆ , rel(Vt , sh)⋆ .
We now briefly recall the standard integration of set-sharing with freeness and
linearity information. These properties are each represented by a set of variables,
namely those variables that are bound to terms that definitely enjoy the given
property. These sets are partially ordered by reverse subset inclusion so that the
lub and glb operators are given by set intersection and union, respectively.
Definition 3
def
def
(The domain SFL.) Let F = ℘(VI ) and L = ℘(VI ) be partially ordered by
reverse subset inclusion. The domain SFL is defined by the Cartesian product
def
SFL = SH × F × L
ordered by the component-wise extension of the orderings defined on the three
subdomains.
A complete definition would explicitly deal with the set of variables of interest VI .
We could even define an equivalence relation on SFL identifying the bottom element
def
⊥ = h∅, VI , VI i with all the elements corresponding to an impossible concrete
6
R. Bagnara, E. Zaffanella, and P. M. Hill
computation state: for example, elements hsh, f, li ∈ SFL such that f * vars(sh)
(because a free variable does share with itself) or VI \ vars(sh) * l (because variables that cannot share are also linear). Note however that these and other similar
spurious elements rarely occur in practice and cannot compromise the correctness
of the results.
In a bottom-up abstract interpretation framework, such as the one we focus on,
abstract unification is the only critical operation. Besides unification, the analysis
depends on the ‘merge-over-all-paths’ operator, corresponding to the lub of the
domain, and the abstract projection operator, which can be defined in terms of an
abstract existential quantification operator.
Definition 4
(Abstract operations over SFL.) The abstract existential quantification on SFL
is encoded by the binary function aexists: SFL × ℘(VI ) → SFL such that, for each
d = hsh, f, li ∈ SFL and V ∈ ℘(VI ),
def
aexists(d , V ) = aexists(sh, V ), f ∪ V, l ∪ V .
For each d = hsh, f, li ∈ SFL, we define the following predicates. The predicate
ind d : Terms × Terms → Bool expresses definite independence of terms. Two terms
s, t ∈ Terms are independent in d if and only if ind d (s, t) holds, where
def
ind d (s, t) =
rel vars(s), sh ∩ rel vars(t), sh = ∅ .
A term t ∈ Terms is free in d if and only if the predicate free d : Terms → Bool
holds for t, that is,
def
free d (t) = ∃x ∈ VI . x = t ∧ x ∈ f .
A term t ∈ Terms is linear in d if and only if lin d (t), where lin d : Terms → Bool
is given by
def
lin d (t) = vars(t) ⊆ l
∧ ∀x, y ∈ vars(t) : x = y ∨ ind d (x, y)
∧ ∀x ∈ vars(t) : x A mvars(t) ⇒ x ∈
/ vars(sh) .
The function amgu : SFL × Bind → SFL captures the effects of a binding on an
element of SFL. Let (x = t) ∈ Bind and d = hsh, f, li ∈ SFL. Let also Vx = {x},
Vt = vars(t), Vxt = Vx ∪ Vt , Rx = rel(Vx , sh) and Rt = rel(Vt , sh). Then
def
amgu d , x = t = hsh ′ , f ′ , l′ i,
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
7
where
def
sh ′ = rel(Vxt , sh) ∪ bin Sx , St ;
(
Rx , if free d (x) ∨ free d (t) ∨ lin d (t) ∧ ind d (x, t) ;
def
Sx =
Rx⋆ , otherwise;
(
Rt , if free d (x) ∨ free d (t) ∨ lin d (x) ∧ ind d (x, t) ;
def
St =
Rt⋆ , otherwise;
f,
if free d (x) ∧ free d (t);
f \ vars(R ),
if free d (x);
x
def
f′ =
f \ vars(Rt ),
if free d (t);
f \ vars(Rx ∪ Rt ), otherwise;
def
l′ = VI \ vars(sh ′ ) ∪ f ′ ∪ l′′ ;
l \ vars(Rx ) ∩ vars(Rt ) , if lin d (x) ∧ lin d (t);
l \ vars(R ),
if lin d (x);
x
def
l′′ =
l
\
vars(R
),
if
lin d (t);
t
l \ vars(Rx ∪ Rt ),
otherwise.
This specification of the abstract unification operator is equivalent (modulo the
lack of the explicit structural information provided by abstract equation systems) to
that given in (Bruynooghe et al. 1994a), provided x ∈
/ vars(t). Indeed, as done in all
the previous papers on the subject, in (Bruynooghe et al. 1994a) it is assumed that
the analyzed language does perform the occurs-check. As a consequence, whenever
considering a definitely cyclic binding, that is a binding x = t such that x ∈ vars(t),
the abstract operator can detect the definite failure of the concrete computation
and thus return the bottom element of the domain. Such an improvement would not
be safe in our case, since we also consider languages possibly omitting the occurscheck. However, when dealing with definitely cyclic bindings, the specification given
by the previous definition can still be refined as follows.
Definition 5
(Improvement for definitely cyclic bindings.) Consider the specification of the
abstract operations over SFL given in Definition 4. Then, whenever x ∈ vars(t), the
computation of the new sharing component sh ′ can be replaced by the following.3
def
sh ′ = rel(Vxt , sh) ∪ bin Sx , CS t ,
3
Note that, in this special case, it also holds that free d (t) = false and ind d (x, t) = (Rx = ∅).
8
R. Bagnara, E. Zaffanella, and P. M. Hill
where
def
CS t =
(
CR t ,
if free d (x);
CR ⋆t ,
otherwise;
def
CR t = rel vars(t) \ {x}, sh .
This enhancement, already implemented in the China analyzer, is the rewording
of a similar one proposed in (Bagnara 1997a) for the domain Pos in the context
of groundness analysis. Its net effect is to recover some groundness and sharing
dependencies that are unnecessarily lost when using the standard operators.
The domain SH captures set-sharing. However, the property we wish to detect
is pair-sharing and, for this, it has been shown in (Bagnara et al. 2002) that SH
includes unwanted redundancy. The same paper introduces an upper-closure operdef
ator ρ on SH and the domain PSD = ρ(SH ), which is the weakest abstraction
of SH that is as precise as SH as far as tracking groundness and pair-sharing is
concerned.4 A notable advantage of PSD is that we can replace the star-union operation in the definition of the amgu by self-bin-union without loss of precision. In
particular, in (Bagnara et al. 2002) it is shown that
amgu(sh, x = t) =ρ rel(Vxt , sh) ∪ bin sbin rel(Vx , sh) , sbin rel(Vt , sh) , (2)
where the notation sh 1 =ρ sh 2 means ρ(sh 1 ) = ρ(sh 2 ).
It is important to observe that the complexity of the amgu operator on SH (1)
is exponential in the
number of sharing-groups of sh. In contrast, the operator on
PSD (2) is O |sh|4 . Moreover, checking whether a fixpoint
has been reached by
testing sh 1 =ρ sh 2 has complexity O |sh 1 |3 + |sh 2 |3 . Practically speaking, very
often this makes the difference between thrashing and termination of the analysis
in reasonable time.
The above observations on SH and PSD can be generalized to apply to the
def
domain combinations SFL and SFL2 = PSD × F × L. In particular, SFL2 achieves
the same precision as SFL for groundness, pair-sharing, freeness and linearity and
the complexity of the corresponding abstract unification operator is polynomial.
For this reason, all the experimental work in this paper, with the exception of part
of the one described in Section 7, has been conducted using the SFL2 domain.
3 Experimental Evaluation
Since the main purpose of this paper is to provide an experimental measure of the
precision gains that might be achieved by enhancing a standard sharing analysis
with several new techniques we found in the literature, it is clear that the implementation of the various domain combinations was a major part of the work. However,
so as to adapt these assorted proposals into a uniform framework and provide a fair
4
The name PSD, which stands for Pair-Sharing Dependencies, was introduced
in (Zaffanella et al. 1999). All previous papers, including (Bagnara et al. 2002), denoted
this domain by SH ρ .
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
9
comparison of their results, a large amount of underlying conceptual work was also
required. For instance, almost all of the proposed enhancements were designed for
systems that perform the occurs-check and some of them were developed for rather
different abstract domains: besides changing the representation of the domain elements, such a situation usually requires a reconsideration of the specification of the
abstract operators.
All the experiments have been conducted using the China analyzer (Bagnara 1997a)
on a GNU/Linux PC system equipped with an AMD Athlon clocked at 700 MHz
and 256 MB of RAM. China is a data-flow analyzer for CLP(HN ) languages (i.e.,
ISO Prolog, CLP(R), clp(FD) and so forth), HN being an extended Herbrand system where the values of a numeric domain N can occur as leaves of the terms.
China, which is written in C++, performs bottom-up analysis deriving information
on both call-patterns and success-patterns by means of program transformations
and optimized fixpoint computation techniques. An abstract description is computed for the call- and success-patterns for each predicate defined in the program
using a sophisticated chaotic iteration strategy proposed in (Bourdoncle 1993a;
Bourdoncle 1993b).5
A major point of the experimental evaluation is given by the test-suite, which
is probably the largest one ever reported in the literature on data-flow analysis of
(constraint) logic programs. The suite comprises all the programs we have access to
(i.e., everything we could find by systematically dredging the Internet): more than
330 programs, 24 MB of code, 800 K lines. Besides classical benchmarks, several real
programs of respectable size are included, the largest one containing 10063 clauses in
45658 lines of code. The suite also comprises a few synthetic benchmarks, which are
artificial programs explicitly constructed to stress the capabilities of the analyzer
and of its abstract domains with respect to precision and/or efficiency.
Because of the exponential complexity of the base domain SFL, a data-flow analysis that includes this domain will only be practical if it incorporates widening
operators (Zaffanella et al. 1999).6 However, since almost none of the investigated
combinations come with specialized widening operators, for a fair assessment of
the precision improvements we decided to disable all the widenings available in
our SFL implementation. As a consequence, there are a few benchmarks for which
the analysis does not terminate in reasonable time or absorbs memory beyond acceptable limits, so that a precision comparison is not possible. Note however that
the motivations behind this choice go beyond the simple observation that widening
operators affect the precision of the analysis: the problem is also that, if we use
the widenings defined and tuned for our implementation of the domain SFL, the
results would be biased. In fact, the definition of a good widening for an analysis
domain normally depends on both the representation and the implementation of
the domain. In other words, different implementations even of the same domain
5
6
China uses the recursive fixpoint iteration strategy on the weak topological ordering defined by
partitioning of the call graph into strongly-connected subcomponents (Bourdoncle 1993b).
Note that we use the term ‘widening operator’ in its broadest sense: any mechanism whereby,
in the course of the analysis, an abstract description is substituted by one that is less precise.
10
R. Bagnara, E. Zaffanella, and P. M. Hill
will require different tunings of the widening operators (or even, possibly, brand
new widenings). This means that adopting the same widening operators for all the
domain combinations would weaken, if not invalidate, any conclusions regarding
the relative benefits of the investigated enhancements. On the other hand, the definition of a new specialized widening operator for each one of the considered domain
combinations, besides being a formidable task, would also be wasted effort as the
number of benchmark programs for which termination cannot be obtained within
reasonable time is really small.
For space reasons, the experimental results are only summarized here. The interested reader can find more information (including a description of the constantly
growing benchmark suite and detailed results for each benchmark) at the URI
http://www.cs.unipr.it/China/. Indeed, given the high number of benchmark
programs and the many domain combinations considered,7 even finding a concise,
meaningful and practical way to summarize the results has been a non-trivial task.
For each benchmark, precision is measured by counting the number of independent pairs (the corresponding columns are labeled ‘I’ in the tables) as well as the
numbers of definitely ground (labeled ‘G’), free (‘F’) and linear (‘L’) variables detected by each abstract domain. The results obtained for different analyses are
compared by computing the relative precision improvements or degradations on
each of these quantities and expressing them using percentages. The “overall” (‘O’)
precision improvement for the benchmark is also computed as the maximum improvement on all the measured quantities.8 The benchmark suite is then partitioned
into several precision equivalence classes: the cardinalities of these classes are expressed again using percentages. For example, when looking at the precision results
reported in Table 1 for goal-dependent analysis, the value 2.3 that can be found at
the intersection of the row labeled ‘0 < p ≤ 2’ with the column labeled ‘G’ is to
be read as follows: “for 2.3 percent of the benchmarks the increase in the number
of ground variables is less than or equal to 2 percent.” The precision class labeled
‘unknown’ identifies those benchmarks for which a precision comparison was not
possible, because one or both of the analyses was timed-out (for all comparisons,
the time-out threshold is 600 seconds). In summary, a precision table gives an approximation of the distribution of the programs in the benchmark suite with respect
to the obtained precision gains.
For a rough estimate of the efficiency of the different analyses, for each comparison
we provide two tables that summarize the times taken by the fixpoint computations. It should be stressed that these by no means provide a faithful account of the
7
8
We compute the results of 40 different variations of the static analysis, which are then used
to perform 36 comparisons. The results are computed over 332 programs for goal-independent
analyses and over 221 programs for goal-dependent analyses. This difference in the number of
benchmarks considered comes from the fact that many programs either are not provided with
a set of entry goals or use constructs such as call(G) where G is a term whose principal functor
is not known. In these cases the analyzer recognizes that goal-dependent analysis is pointless,
since no call-patterns can be excluded.
When computing this “overall” result for a benchmark, the presence of even a single precision loss for one of the measures overrides any precision improvement computed on the other
components.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
11
intrinsic computational cost of the tested domain combinations. Besides the lack
of widenings, which have a big impact on performance as can be observed by the
results reported in (Zaffanella et al. 1999), the reader should not forget that, for
ease of implementation, having targeted at precision we traded efficiency whenever
possible. Therefore, these tables provide, so to speak, upper-bounds: refined implementations can be expected to perform at least as well as those reported in the
tables.
As done for the precision results, the timings are summarized by partitioning
the suite into equivalence classes and reporting the cardinality of each class using
percentages. In the first table we consider the distribution of the absolute time
differences, that is we measure the slow-down and speed-up due to the incorporation
of the considered enhancement. Note that the class called ‘same time’ actually
comprises the benchmarks having a time difference below a given threshold, which
is fixed at 0.1 seconds. In the second table we show the distribution of the total
fixpoint computation times, both for the base analysis (in the columns labeled ‘%1’)
and for the enhanced one (in the columns labeled ‘%2’); the columns labeled ‘∆’
show how much each total time class grows or shrinks due to the inclusion of the
considered combination.
4 A Simple Combination with Pos
It is well-known that the domain Sharing (and thus also SFL) keeps track of ground
dependencies. More precisely, Sharing contains Def , the domain of definite Boolean
functions (Armstrong et al. 1998), as a proper subdomain (Cortesi et al. 1992; Zaffanella et al. 1999).
However, we consider here the combination of SFL with Pos, the domain of positive Boolean functions (Armstrong et al. 1998). There are several good reasons to
couple SFL with Pos:
1. Pos is strictly more expressive than Def in that it can represent (positive) disjunctive groundness dependencies that arise in the analysis of Prolog programs
(Armstrong et al. 1998). The ability to deal with disjunctive dependencies is
also needed for the precise approximation of the constraints of some CLP languages: for example, when using the finite domain solver of SICStus Prolog,
the user can write disjunctive constraints such as ‘X #= 4 #\/ Y #= 6’.
2. The increased precision on groundness propagates to the SFL component. It
can be exploited to remove redundant sharing groups and to identify more
linear variables, therefore having a positive impact on the computation of the
amgu operator of the SFL domain. Moreover, when dealing with sequences
of bindings, the added groundness information allows them to be usefully
reordered. In fact, while it has been proved that Sharing alone is commutative,
meaning that the result of the analysis does not depend on the ordering in
which the bindings are executed (Hill et al. 1998), the domain SFL does not
enjoy this property. In particular, even for the simpler combination of Sharing
with linearity it is known since (Langen 1990, pp. 66-67) that better results
12
R. Bagnara, E. Zaffanella, and P. M. Hill
3.
4.
5.
6.
are obtained if the grounding bindings are considered before the others.9 As
an example, consider the sequences
of unifications f (X, X, Y ) = A, X = a
and X = a, f (X, X, Y ) = A (Langen 1990, p. 66). The combination with
Pos is clearly advantageous in this respect.
Besides being useful for improving precision on other properties, disjunctive
dependencies also have a few direct applications, such as occurs-check reduction. As observed in (Crnogorac et al. 1996), if the groundness formula x ∨ y
holds, the unification x = y is occurs-check free, even when neither x nor y
are definitely linear.
Detecting the set of definitely ground variables through Pos and exploiting it to simplify the operations on SFL can improve the efficiency of the
analysis. In particular this is true if the set of ground variables is readily
available, as is the case, for instance, with the GER implementation of Pos
(Bagnara and Schachte 1999).
The combination with Pos is essential for the application of a powerful widening technique on SFL (Zaffanella et al. 1999). This is very important, since
analysis based on SFL is not practical without widenings.
In the context of the analysis of CLP programs, the notions of “ground variable” and the notion of “variable that cannot share a common variable with
other variables” are distinct. A numeric variable in, say, CLP(R), cannot
share with other numerical variables (not in the sense of interest in this paper) but is not ground unless it has been constrained to a unique value. Thus
the analysis of CLP programs with SFL alone either will lose precision on
pair-sharing (if arithmetic constraints are abstracted into “sharings” among
numeric variables in order to approximate the groundness of the latter) or
will be imprecise on the groundness of numeric variables (because only Herbrand constraints take part in the construction of sharing-sets). In the first
alternative, as we have already noted, the precision with which groundness
of numeric variables can be tracked will also be limited. Since groundness of
numeric variables is important for a number of applications (e.g., compiling
equality constraints down to assignments or tests in some circumstances), we
advocate the use of Pos and SFL at the same time.
Thus, as a first technique to enhance the precision of sharing analysis, we consider the simple propagation of the set of definitely ground variables from the Pos
component to the SFL component.10 We denote this domain by Pos × SFL.
As noted above, the GER implementation of (Bagnara and Schachte 1999), besides being the fastest implementation of Pos known to date, is the natural candidate for this combination, since it provides constant-time access to the set G of
the definitely ground variables. Note that the widenings on the Pos component
9
10
A binding x = t is grounding with respect to an abstract description if, in all the concrete
computation states approximated by the abstract description, either the variable x is ground
or all the variables in t are ground. For example, when considering an abstract description
sh ∈ SH , the binding x = t is grounding if rel({x}, sh) = ∅ or rel(vars (t), sh) = ∅.
A more precise combination will be considered in Section 7.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
Prec. class
5 < p ≤ 10
2<p≤5
0<p≤2
same precision
unknown
13
Goal Independent
Goal Dependent
O
I
G
F
L
O
I
G
F
L
—
—
—
—
— 0.5
— 0.5
—
—
0.3
— 0.3
—
—
—
—
—
—
—
0.6 0.6 0.6
— 0.6 3.2 3.6 2.3
— 2.7
95.8 96.1 95.8 96.7 96.1 92.8 92.8 93.7 96.4 93.7
3.3 3.3 3.3 3.3 3.3 3.6 3.6 3.6 3.6 3.6
Time difference class
% benchmarks
Goal Ind. Goal Dep.
degradation > 1
2.7
6.8
0.5 < degradation ≤ 1
1.5
0.5
0.2 < degradation ≤ 0.5
3.0
0.9
0.1 < degradation ≤ 0.2
5.7
5.0
both timed out
3.3
3.6
same time
81.6
81.9
0.1 < improvement ≤ 0.2
—
0.5
0.2 < improvement ≤ 0.5
0.9
0.5
0.5 < improvement ≤ 1
0.3
—
improvement > 1
0.9
0.5
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Ind.
Goal Dep.
%1 %2 ∆
%1 %2 ∆
3.3 3.3 — 3.6 3.6 —
8.4 9.0 0.6 7.2 7.2 —
0.6 0.3 -0.3 1.4 1.4 —
6.6 7.5 0.9 3.2 3.6 0.5
3.3 2.7 -0.6 5.4 5.4 —
7.2 8.4 1.2 10.4 13.1 2.7
70.5 68.7 -1.8 68.8 65.6 -3.2
Table 1. SFL2 versus Pos × SFL2 .
have been retained. The reason for this choice is that they fire for only a few benchmarks and, when coming into play, they rarely affect the precision of the groundness
analysis: by switching them off we would only obtain a few more time-outs.
In the SFL component, the set G of definitely ground variables is used
• to reorder the sequence of bindings in the abstract unification so as to handle
the grounding ones first;
• to eliminate the sharing groups containing at least one ground variable; and
• to recover from previous linearity losses.
The experimental results for Pos × SFL are compared with those obtained for the
domain SFL considered in isolation and reported in Table 1. It can be observed that
14
R. Bagnara, E. Zaffanella, and P. M. Hill
a precision improvement is observed in all of the measured quantities but freeness,
affecting up to 3.6% of the programs.
Note that there is a small discrepancy between these results and those of (Bagnara et al. 2000)
where more improvements were reported. The reason is that the current SFL implementation uses an enhanced abstract unification operator, fully exploiting the anticipation of the grounding bindings even on the base domain SFL itself. In contrast,
in the earlier SFL implementation used for the results in (Bagnara et al. 2000),
only the syntactically grounding bindings were anticipated.11
As for the timings, even if the figures in the tables seem to contradict what we
claimed in point 4 above, a closer inspection of the detailed results reveals that this
is only due to a very unfortunate interaction between the increased precision given
by Pos and the absence of widening operators on SFL. This state of affairs forces
the analyzer to compute a few, but very expensive, further iterations in the fixpoint
computation.
Because of the reasons detailed above, we believe Pos should be part of the global
domain employed by any “production analyzer” for CLP languages. That is why,
for the remaining comparisons, unless otherwise stated, this simple combination
with the Pos domain is always included.
5 Tracking Explicit Structural Information
A way of increasing the precision of almost any analysis domain is by enhancing it with structural information. For mode analysis, this idea dates back to
(Janssens and Bruynooghe 1992). A more general technique was proposed in (Cortesi et al. 1994),
where the generic structural domain Pat(ℜ) was introduced. A similar proposal,
tailored to sharing analysis, is due to (Bruynooghe et al. 1994a), where abstract
equation systems are considered. In the experimental evaluation the Pattern(·)
construction (Bagnara 1997a; Bagnara 1997b; Bagnara et al. 2000) is used. This
is similar to Pat(ℜ) and correctly supports the analysis of languages omitting the
occurs-check in the unification procedure as well as those that do not.
The construction Pattern(·) upgrades a domain D (which must support a certain
set of basic operations) with structural information. The resulting domain, where
structural information is retained to some extent, is usually much more precise
than D alone. There are many occasions where these precision gains give rise to
consistent speed-ups. The reason for this is twofold. First, structural information
has the potential of pruning some computation paths on the grounds that they
cannot be followed by the program being analyzed. Second, maintaining a tuple
of terms with many variables, each with its own description, can be cheaper than
computing a description for the whole tuple (Bagnara et al. 2000). Of course, there
is also a price to be paid: in the analysis based on Pattern(D), the elements of
11
A binding x = t is syntactically grounding if vars(t) = ∅. This “syntactic” definition differs
from the “semantic” one provided before in that it does not depend on the information provided
by an abstract description.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
15
D that are to be manipulated are often bigger (i.e., there are more variables of
interest) than those that arise in analyses that are simply based on D.
When comparing the precision results, the difference in the number of variables
tracked by the two analyses poses a non-trivial problem. How can we provide a
fair measure of the precision gain? There is no easy answer to such a question.
The approach chosen is simple though unsatisfactory: at the end of the analysis,
first throw away all the structural information in the results and then calculate
the cardinality of the usual sets. In other words, we only measure how the explicit
structural information in Pattern(D) improves the precision on D itself, which is
only a tiny part of the real gain in accuracy. As shown by the following example,
this solution greatly underestimates the precision improvement coming from the
integration of structural information.
Consider a simple but not trivial Prolog program: mastermind.12 Consider also
the only direct query for which it has been written, ‘?- play.’, and focus the
attention on the procedure extend code/1. A standard goal-dependent analysis of
the program with the Pos × SFL domain cannot say anything on the successes of
extend code/1. If the analysis is performed with Pattern(Pos × SFL) the situation
changes radically. Here is what such a domain allows China to derive:13
extend_code([([A|B],C,D)|E]) :list(B), list(E),
(functor(C,_,1);integer(C)),
(functor(D,_,1);integer(D)),
ground([C,D]), may_share([[A,B,E]]).
This means: “during any execution of the program, whenever extend code/1 succeeds it will have its argument bound to a term of the form [([A|B],C,D)|E],
where B and E are bound to list cells (i.e., to terms whose principal functor is either
’.’/2 or []/0); C and D are ground and bound to a functor of arity 1 or to an
integer; and pair-sharing may only occur among A, B, and E”. Once structural information has been discarded, the analysis with Pattern(Pos × SFL) only specifies
that extend code/1 may succeed. Thus, according to our approach to the precision
comparison, explicit structural information gives no improvements in the analysis
of extend code/1 (which is far from being a fair conclusion).
Of course, structural information is very valuable in itself. For example, when
exploited for optimized compilation it allows for enhanced clause indexing and simplified unification. Several other semantics-based program manipulation techniques
(such as debugging, program specialization, and verification) benefit from this kind
of information. However, the value of this extra precision could only be measured
from the point of view of the target application of the analysis.
12
13
This program which implements the game “Mastermind” was rewritten by
H. Koenig and T. Hoppe after code by M. H. van Emden and available at
http://www.cs.unipr.it/China/Benchmarks/Prolog/mastermind.pl.
Some extra groundness information obtained by the analysis has been omitted for simplicity:
this says that, if A and B turn out to be ground, then E will also be ground.
16
R. Bagnara, E. Zaffanella, and P. M. Hill
Prec. class
p > 20
10 < p ≤ 20
5 < p ≤ 10
2<p≤5
0<p≤2
same precision
unknown
p<0
Goal Independent
Goal Dependent
O
I
G
F
L
O
I
G
F
L
7.5 2.7 3.9 2.1 3.3 6.3 1.4 3.6 1.8 3.6
3.9 2.1 2.7
— 2.4 2.7 2.3 1.4
— 2.7
4.5 1.8 2.7 2.4 2.4 1.8 0.9 2.3 0.9 1.4
7.5 6.0 3.9 2.7 5.1 2.7 3.2 1.4 1.8 2.3
7.8 9.0 6.6 6.9 12.0 2.3 4.5 1.8 1.8 5.0
61.7 71.7 73.5 79.2 67.8 74.2 78.3 80.1 84.2 75.1
6.6 6.6 6.6 6.6 6.6 9.5 9.5 9.5 9.5 9.5
0.3
—
—
— 0.3 0.5
—
—
— 0.5
Time diff. class
% benchmarks
Goal Ind. Goal Dep.
degradation > 1
11.7
17.6
0.5 < degradation ≤ 1
1.2
0.9
0.2 < degradation ≤ 0.5
3.6
4.1
0.1 < degradation ≤ 0.2
1.5
4.1
both timed out
3.3
3.6
same time
70.8
66.5
0.1 < improvement ≤ 0.2
0.9
0.5
0.2 < improvement ≤ 0.5
1.5
—
0.5 < improvement ≤ 1
0.6
0.5
improvement > 1
4.8
2.3
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Ind.
Goal Dep.
%1 %2 ∆
%1 %2 ∆
3.3 6.6 3.3 3.6 9.5 5.9
9.0 8.4 -0.6 7.2 8.6 1.4
0.3 1.5 1.2 1.4 1.8 0.5
7.5 6.6 -0.9 3.6 5.0 1.4
2.7 3.3 0.6 5.4 3.2 -2.3
8.4 10.2 1.8 13.1 13.6 0.5
68.7 63.3 -5.4 65.6 58.4 -7.2
Table 2. Pos × SFL2 versus Pattern(Pos × SFL2 ).
Thus the precision of the domain Pos × SFL has been compared with that obtained using the domain Pattern(Pos × SFL) and the results reported in Table 2.
It can be seen that, for goal-independent analysis, on one third of the benchmarks
compared there is a precision improvement in at least one of the measured quantities; the same happens for one sixth of the benchmarks in the case of goal-dependent
analysis. Moreover, the increase in precision can be considerable, as testified by the
percentages of benchmarks falling in the higher precision classes.
The reader may be surprised, as the authors were, to see that in some cases the
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
17
precision actually decreased.14 Indeed, to the best of our knowledge, this possibility
has escaped all previous research work investigating this kind of abstract domain enhancement, including (Cortesi et al. 1994; Bruynooghe et al. 1994a; Bagnara 1997a).
The reason for these precision losses lies in a subtle interaction between the explicit
structural information and the underlying abstract unification operator.
When using the base domain Pos × SFL, the abstract evaluation of a single
syntactic binding, such as x = f (y, z), directly corresponds to a single application of
the amgu operator. In contrast, when computing on Pattern(Pos×SFL), it may well
happen that the computed abstract description already contains the information
that variable x is bound to a term, such as f g(w), w . As a consequence, after
peeling the principal functor f /2, the abstract computation should proceed by
evaluating, on the base domain Pos × SFL, the set of bindings y = g(w), z = w .
Here the problem is that, as already noted, the amgu operator on the base domain
Pos × SFL is not commutative. While this improvement in the data used by the
abstract computation very often allows for a corresponding increase in the precision
of the result, in rare situations it may happen that a sub-optimal ordering of the
bindings is chosen, incurring a precision loss.
It should be noted that such a negative interaction with the explicit structural information is only possible when the underlying domain implements noncommutative abstract operators. In particular, this phenomenon could not be observed when computing on Pattern(SH ) or Pattern(Pos).
One issue that should be resolved is whether the improvements provided by explicit structural information subsume those previously obtained for the simple combination with Pos. Intuitively, it would seem that this cannot happen, since these
two enhancements are based on different kinds of information: while the Pattern(·)
construction encodes some definite structural information, the precision gain due
to using Pos rather than just Def only stems from disjunctive groundness dependencies. However, the impact of these techniques on the overall analysis is really
intricate and some overlapping cannot be excluded a priori: for instance, both techniques affect the ordering of bindings in the computation of abstract unification on
SFL. In order to provide some experimental evidence for this qualitative reasoning,
the precision results are computed for the simpler domain Pattern(SFL) and then
compared with those obtained for the domain Pattern(Pos × SFL). Since the main
differences between Tables 1 and 3 can be explained by discrepancies in the numbers of programs that timed-out, these results confirm our expectations that these
two enhancements are effectively orthogonal.
Similar experimental evaluations, but based on the abstract equation systems of
(Bruynooghe et al. 1994a), were reported by A. Mulkers et al. in (Mulkers et al. 1994;
Mulkers et al. 1995). Here a depth-k abstraction (replacing all subterms occurring
at a depth greater than or equal to k with fresh abstract variables) is conducted
on a small benchmark suite (19 programs) for values of k between 0 and 3. The
domain they employed was not suitable for the analysis of real programs and, in
14
This happens for the program attractions2 in the case of goal-independent analysis and for
the program semi in the case of goal-dependent analysis.
18
R. Bagnara, E. Zaffanella, and P. M. Hill
Prec. class
5 < p ≤ 10
2<p≤5
0<p≤2
same precision
unknown
Goal Independent
Goal Dependent
O
I
G
F
L
O
I
G
F
L
—
—
—
—
— 0.5
— 0.5
—
—
0.3
— 0.3
—
—
— 0.5
—
—
—
—
—
—
—
— 3.2 3.2 2.7
— 2.7
93.1 93.4 93.1 93.4 93.4 86.4 86.4 86.9 90.0 87.3
6.6 6.6 6.6 6.6 6.6 10.0 10.0 10.0 10.0 10.0
Time diff. class
% benchmarks
Goal Ind. Goal Dep.
degradation > 1
5.7
7.7
0.5 < degradation ≤ 1
2.4
0.5
0.2 < degradation ≤ 0.5
3.6
5.4
0.1 < degradation ≤ 0.2
5.4
2.7
both timed out
6.6
9.5
same time
75.6
73.8
0.1 < improvement ≤ 0.2
—
—
0.2 < improvement ≤ 0.5
0.6
—
0.5 < improvement ≤ 1
—
—
improvement > 1
—
0.5
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Ind.
Goal Dep.
%1 %2 ∆
%1 %2 ∆
6.6 6.6 — 10.0 9.5 -0.5
8.1 8.4 0.3 7.7 8.6 0.9
1.5 1.5 — 2.3 1.8 -0.5
5.1 6.6 1.5 4.5 5.0 0.5
3.9 3.3 -0.6 3.2 3.2 —
7.2 10.2 3.0 10.9 13.6 2.7
67.5 63.3 -4.2 61.5 58.4 -3.2
Table 3. Pattern(SFL2 ) versus Pattern(Pos × SFL2 ).
fact, even the analysis of a modest-sized program like ann could only be carried out
with depth-0 abstraction (i.e., without any structural information). Such a problem
in finding practical analyzers that incorporated structural information with sharing
analysis was not unique to this work: there was at least one other previous attempt
to evaluate the impact of structural information on sharing analysis that failed
because of combinatorial explosion [A. Cortesi, personal communication, 1996].
What makes the more realistic experimentation now possible is the adoption
of the non-redundant domain PSD , where the exponential star-union operation is
replaced by the quadratic self-bin-union. Note that, even if biased by the absence
of widenings, the timings reported in Table 2 show that the Pattern(·) construction
is computationally feasible. Indeed, as demonstrated by the results reported in
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
19
(Bagnara et al. 2000), an analyzer that incorporates a carefully designed structural
information component, besides being more precise, can also be very efficient.
The results obtained in this section demonstrate that there is a relevant amount
of sharing information that is not detected when using the classical set-sharing
domains. Therefore, in order to provide an experimental evaluation that is as systematic as possible, in all of the remaining experiments the comparison is performed
both with and without explicit structural information.
6 Reordering the Non-Grounding Bindings
As already explained in Section 4, the results of abstract unification on SFL may
depend on the order in which the bindings are considered and will be improved if
the grounding bindings are considered first. This heuristic, which has been used for
all the experiments in this paper, is well-known: in the literature all the examples
that illustrate the non-commutativity of the abstract mgu on SFL use a grounding
binding. However, as observed in Section 5, the problem is more general than that.
To illustrate this, suppose that VI = {u, v, w, x, y, z} is the set of relevant variables, and consider the SFL element15
def
d = {vy, wy, xy, yz}, ∅, {u, x, z} ,
where no variable is free and u, x, and z are linear with the bindings v = w and
x = y. Then, applying amgu to these bindings in the given ordering, we have:
d1 = amgu(d , v = w)
= {vwy, xy, yz}, ∅, {u, x, z} ,
d1,2 = amgu(d1 , x = y)
= {vwxy, vwxyz, xy, xyz}, ∅, {u, z} .
Using the reverse ordering, we have:
d2 = amgu(d , x = y)
= {vwxy, vwxyz, vxy, vxyz, wxy, wxyz, xy, xyz}, ∅, {u, z} ,
d2,1 = amgu(d2 , v = w)
= {vwxy, vwxyz, xy, xyz}, ∅, {u} .
Thus d2,1 loses the linearity of z (which, in turn, could cause bigger precision losses
later in the analysis).
In principle, optimality can be obtained by adopting the brute-force approach:
trying all the possible orderings of the non-grounding bindings. However, this is
clearly not feasible. While lacking a better alternative, it is reasonable to look for
heuristics that can be applied in the context of a local search paradigm: at each
15
Elements
of SH are written in a simplified notation, omitting the inner braces. For instance,
the set {x}, {x, y}, {x, z}, {x, y, z} is written as {x, xy, xz, xyz}.
20
R. Bagnara, E. Zaffanella, and P. M. Hill
step, the next binding for the amgu procedure is chosen by evaluating the effect of
its abstract execution, considered in isolation, on the precision of the analysis.
Suppose the number of independent pairs is taken as a measure of precision. Then,
at each step, for each of the bindings under consideration, the new component sh ′ ,
as given by Definition 4, must be computed. However, because the computation
of sh ′ is the most costly operation to be performed in the computation of the
amgu operator, a direct application of this heuristic does not appear to be feasible.
As an alternative, consider a heuristic based on the number of star-unions that
have to be computed. Star-unions are likely to cause large losses in the number of
independent pairs that are found. As only non-grounding bindings are considered,
any binding requiring the computation of a star-union will need the star-union
even if it is delayed, although a binding that does not require the star-union may
require it if its computation is postponed: its variables may lose their freeness,
linearity or independence as a result of evaluating the other bindings. It follows
that one potential heuristic is: “delay the bindings requiring star-unions as much as
possible”. In the next example, by adopting this heuristic, the linearity of variable
y is preserved.
Consider the application of the bindings x = z and v = w to the following
abstract description:
def
d = {vw, wx, wy, z}, ∅, {u, v, x, y} .
Since x is linear and independent from z, computing amgu(d , x = z) requires one
star-union, while two star-unions are needed when computing amgu(d , v = w)
because v and w may share. Thus, with the proposed heuristic, x = z is applied
before v = w, giving:
d1 = amgu(d , x = z)
= {vw, wxz, wy}, ∅, {u, v, y} ,
d1,2 = amgu(d1 , v = w)
= {vw, vwxyz, vwxz, vwy}, ∅, {u, y} .
In contrast, if v = w is applied first, we have:
d2 = amgu(d , v = w)
= {vw, vwx, vwxy, vwy, z}, ∅, {u, x, y} ,
d2,1 = amgu(d2 , x = z)
= {vw, vwxyz, vwxz, vwy}, ∅, {u} .
Note that the same number of independent pairs is computed in both cases.
It should be noted that this heuristic, considered in isolation, is not a general
solution and can actually lead to precision losses. The problem is that, if a binding
that needs a star-union is delayed, then, when the star-union is computed, it may
be done on a larger sharing-set, forcing more (independent) pairs of variables into
the same sharing group.
Consider the application of the bindings u = x and v = w to the abstract
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
21
description
def
d = {u, uw, v, w, xy, xz}, {u, x}, {u, x} .
Since x and u are both free variables, no star-union is needed in the computation of
amgu(d , u = x), while two star-unions are needed when computing amgu(d , v = w).
d1 = amgu(d , u = x)
= {uwxy, uwxz, uxy, uxz, v, w}, {u, x}, {u, x} ,
d1,2 = amgu(d1 , v = w)
= {uvwxy, uvwxyz, uvwxz, uxy, uxz, vw}, ∅, ∅ .
Using the other ordering, we have:
d2 = amgu(d , v = w)
= {u, uvw, vw, xy, xz}, {x}, {x} ,
d2,1 = amgu(d2 , u = x)
= {uvwxy, uvwxz, uxy, uxz, vw}, ∅, ∅ .
Note that in d2,1 variables y and z are independent, whereas they may share in d1,2 .
Thus, in this example, by delaying the only binding that requires the star-unions,
v = w, the number of known independent pairs is decreased.
Another possibility is to consider a heuristic that uses the numbers of free and
linear variables as a measure of precision for local optimization. That is, it chooses
first those bindings for which these numbers are maximal. However, the last example
shown above is evidence that even such a proposal may also cause precision losses
(the binding u = x would be chosen first as it preserves the freeness of variable u).
In order to evaluate the effects of these two heuristics on real programs, we have
implemented and compared them with respect to the “straight” abstract computation, which considers the non-grounding bindings using the left-to-right order.16
The results reported in Tables 4 and 5 can be summarized as follows:
1. the precision on the groundness and freeness components is not affected;
2. the precision on the independent pairs and linearity components is rarely
affected, in particular when considering goal-dependent analyses;
3. even for real programs, as was the case for the artificial examples given above,
the precision can be increased as well as decreased.
Looking at Tables 4 and 5, it can be seen that the heuristic based on freeness and
linearity information is slightly better than the use of the straight order, which, in
its turn, is slightly better than the heuristic based on the number of star-unions.
Clearly, since these results could not be generalized to other orderings, our investigation cannot be considered really conclusive. Besides designing “smarter” heuristics, it would be interesting to provide a kind of responsiveness test for the underlying domain with respect to the choice of ordering for the non-grounding bindings: a
16
The base domain is Pos × SFL, both with and without structural information.
22
R. Bagnara, E. Zaffanella, and P. M. Hill
Goal Independent
without Struct Info
with Struct Info
Prec. class
O
I
G
F
L
O
I
G
F
L
0<p≤2
0.9
—
—
— 0.9
—
—
—
—
—
same precision
94.6 95.5 96.4 96.4 95.5 91.3 91.3 93.1 93.1 93.1
unknown
3.6 3.6 3.6 3.6 3.6 6.9 6.9 6.9 6.9 6.9
−2 ≤ p < 0
Goal Dependent
Prec. class
same precision
unknown
0.9
0.9
—
—
—
1.8
1.8
—
—
—
without Struct Info
with Struct Info
O
I
G
F
L
O
I
G
F
L
96.4 96.4 96.4 96.4 96.4 90.5 90.5 90.5 90.5 90.5
3.6 3.6 3.6 3.6 3.6 9.5 9.5 9.5 9.5 9.5
Time diff. class
Goal Ind.
Goal Dep.
w/o SI with SI w/o SI with SI
degradation > 1
4.5
3.0
7.2
4.1
0.5 < degradation ≤ 1
0.6
0.3
—
—
0.2 < degradation ≤ 0.5
2.4
0.9
0.5
0.5
0.1 < degradation ≤ 0.2
1.5
0.6
0.5
0.5
both timed out
3.0
6.3
3.6
9.5
same time
80.7
80.7
85.5
76.9
0.1 < improvement ≤ 0.2
1.5
1.2
0.5
0.5
0.2 < improvement ≤ 0.5
1.8
1.2
1.4
2.3
0.5 < improvement ≤ 1
0.9
0.6
—
0.9
improvement > 1
3.0
5.1
0.9
5.0
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Independent
Goal Dependent
without SI
with SI
without SI
with SI
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
3.3 3.3 — 6.6 6.6 — 3.6 3.6 — 9.5 9.5 —
9.0 8.1 -0.9 8.4 9.0 0.6 7.2 7.7 0.5 8.6 8.1 -0.5
0.3 0.9 0.6 1.5 1.2 -0.3 1.4 0.9 -0.5 1.8 2.7 0.9
7.5 7.5 — 6.6 6.3 -0.3 3.6 3.2 -0.5 5.0 4.1 -0.9
2.7 2.4 -0.3 3.3 3.0 -0.3 5.4 5.9 0.5 3.2 3.6 0.5
8.4 9.3 0.9 10.2 10.5 0.3 13.1 12.7 -0.5 13.6 13.1 -0.5
68.7 68.4 -0.3 63.3 63.3 — 65.6 66.1 0.5 58.4 58.8 0.5
Table 4. The heuristic based on the number of star-unions.
simple test consists in measuring how much the precision can be affected, in either
way, by the application of an almost arbitrary order. This is the motivation for the
comparison reported in Table 6, where the order is from right-to-left, the reverse
of the usual one. As for the results given in Tables 4 and 5, the number of changes
to the precision observed in Table 6 is small and all the observations made above
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
23
Goal Independent
without Struct Info
with Struct Info
Prec. class
O
I
G
F
L
O
I
G
F
L
5 < p ≤ 10
0.3
—
—
— 0.3 0.3
—
—
— 0.3
0<p≤2
0.9
—
—
— 0.9 2.7 2.4
—
— 0.3
same precision
94.3 95.5 96.4 96.4 95.2 89.5 90.1 93.4 93.4 92.8
unknown
3.6 3.6 3.6 3.6 3.6 6.6 6.6 6.6 6.6 6.6
−2 ≤ p < 0
0.6 0.6
—
—
— 0.9 0.9
—
—
—
p < −20
0.3 0.3
—
—
—
—
—
—
—
—
Goal Dependent
Prec. class
0<p≤2
same precision
unknown
−20 ≤ p < −10
without Struct Info
with Struct Info
O
I
G
F
L
O
I
G
F
L
0.5
—
—
— 0.5
—
—
—
—
—
94.6 95.0 95.5 95.5 95.0 89.6 89.6 89.6 89.6 89.6
4.5 4.5 4.5 4.5 4.5 10.4 10.4 10.4 10.4 10.4
0.5 0.5
—
—
—
—
—
—
—
—
Time diff. class
Goal Ind.
Goal Dep.
w/o SI with SI w/o SI with SI
degradation > 1
6.9
4.8
8.1
7.7
0.5 < degradation ≤ 1
2.1
1.5
1.8
0.5
0.2 < degradation ≤ 0.5
2.4
1.8
1.8
2.7
0.1 < degradation ≤ 0.2
1.2
3.3
2.3
3.2
both timed out
2.4
5.7
3.6
9.0
same time
77.4
73.5
78.7
71.9
0.1 < improvement ≤ 0.2
1.2
0.3
—
—
0.2 < improvement ≤ 0.5
0.6
1.8
0.9
0.9
0.5 < improvement ≤ 1
0.9
—
0.5
—
improvement > 1
4.8
7.2
2.3
4.1
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Independent
Goal Dependent
without SI
with SI
without SI
with SI
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
3.3 2.7 -0.6 6.6 5.7 -0.9 3.6 4.5 0.9 9.5 10.0 0.5
9.0 9.6 0.6 8.4 8.7 0.3 7.2 6.8 -0.5 8.6 7.7 -0.9
0.3 2.1 1.8 1.5 1.8 0.3 1.4 1.4 — 1.8 2.7 0.9
7.5 6.0 -1.5 6.6 6.9 0.3 3.6 4.5 0.9 5.0 5.0 —
2.7 3.0 0.3 3.3 3.9 0.6 5.4 4.1 -1.4 3.2 3.6 0.5
8.4 9.9 1.5 10.2 13.3 3.0 13.1 13.1 — 13.6 15.4 1.8
68.7 66.6 -2.1 63.3 59.6 -3.6 65.6 65.6 — 58.4 55.7 -2.7
Table 5. The heuristic based on freeness and linearity.
24
R. Bagnara, E. Zaffanella, and P. M. Hill
still hold. Surprisingly, this reversed ordering provides marginally better precision
results than those obtained using the considered heuristics.17
7 The Reduced Product between Pos and Sharing
The overlap between the information provided by Pos and the information provided
by Sharing mentioned in Section 4 means that the Cartesian product Pos × SFL
contains redundancy, that is, there is more than one element that can characterize
the same set of concrete computational states.
In (Bagnara et al. 2000), two techniques that are able to remove some of this
redundancy were experimentally evaluated. One of these aims at identifying those
pairs of variables (x, y) for which the Boolean formula of the Pos component implies the binary disjunction x ∨ y. In such a case, it is always safe to assume that
the variables x and y are independent.18 Since the number of independent pairs is
one of the quantities explicitly measured, this enhancement has the potential for
“immediate” precision gains. The other technique exploits the knowledge of the
sets of ground-equivalent variables: the variables in e ⊆ VI are ground-equivalent
in φ ∈ Pos if and only if, for each x, y ∈ e, φ |= (x ↔ y). For a description
of how these sets can be used to improve sharing analysis, the reader is referred
to (Bagnara et al. 2000). The main motivation for experimenting with this specific reduction was the ease of its implementation, since all the needed information
can easily be recovered from the already computed E component of the GER implementation of Pos (Bagnara and Schachte 1999). The experimental evaluation
results given in (Bagnara et al. 2000) for these two techniques show precision improvements with only three of the programs and, also, only with respect to the
number of independent pairs that were found. Those results just apply to these
limited forms of reduction, so could not be considered a complete account of all the
possible precision gains.
The full reduced product (Cousot and Cousot 1979) between Pos and Sharing has
been elegantly characterized in (Codish et al. 1999), where set-sharing à la Jacobs
and Langen is expressed in terms of elements of the Pos domain itself. Let [φ]VI
denote the set of all the models of the Boolean function φ defined over the set of
variables VI . Then, the isomorphism maps each set-sharing element sh ∈ SH into
the Boolean formula φ ∈ Pos such that
[φ]VI = { VI \ S | S ∈ sh } ∪ {VI }.
The sharing information encoded by an element (φg , φsh ) ∈ Pos × Pos can be improved by replacing the second component (that is, the Boolean formula describing
set-sharing information) with the conjunction φg ∧ φsh . The reader is referred to
17
18
It is worth noting that the only precision improvement reported in Table 6 for the goal-dependent
analysis with structural information (caused by the program semi) corresponds to the precision
decrease reported in Table 2. This confirms that, as informally discussed in Section 5, such a
precision decrease was due to the non-commutativity of the amgu operator on Pos × SFL.
Note that this observation dates back, at least, to (Crnogorac et al. 1996).
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
25
Goal Independent
without Struct Info
with Struct Info
Prec. class
O
I
G
F
L
O
I
G
F
L
5 < p ≤ 10
0.3
—
—
— 0.3 0.3
—
—
— 0.3
0<p≤2
0.9 0.3
—
— 0.6 4.2 3.0
—
— 1.2
same precision
94.3 95.2 96.4 96.4 95.5 87.7 89.2 93.4 93.4 91.9
unknown
3.6 3.6 3.6 3.6 3.6 6.6 6.6 6.6 6.6 6.6
−2 ≤ p < 0
0.6 0.6
—
—
— 1.2 1.2
—
—
—
p < −20
0.3 0.3
—
—
—
—
—
—
—
—
Goal Dependent
Prec. class
0<p≤2
same precision
unknown
−20 ≤ p < −10
without Struct Info
with Struct Info
O
I
G
F
L
O
I
G
F
L
0.5
—
—
— 0.5 0.5
—
—
— 0.5
95.5 95.9 96.4 96.4 95.9 90.0 90.5 90.5 90.5 90.0
3.6 3.6 3.6 3.6 3.6 9.5 9.5 9.5 9.5 9.5
0.5 0.5
—
—
—
—
—
—
—
—
Time diff. class
Goal Ind.
Goal Dep.
w/o SI with SI w/o SI with SI
degradation > 1
4.2
6.0
4.5
6.8
0.5 < degradation ≤ 1
0.6
0.6
—
—
0.2 < degradation ≤ 0.5
2.4
1.5
1.4
0.9
0.1 < degradation ≤ 0.2
1.8
0.9
0.5
—
both timed out
2.4
5.7
3.6
9.0
same time
78.3
76.2
82.8
74.2
0.1 < improvement ≤ 0.2
1.5
1.2
1.8
0.9
0.2 < improvement ≤ 0.5
1.8
0.3
1.4
1.8
0.5 < improvement ≤ 1
0.9
0.9
0.5
0.5
improvement > 1
6.0
6.6
3.6
5.9
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Independent
Goal Dependent
without SI
with SI
without SI
with SI
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
3.3 2.7 -0.6 6.6 5.7 -0.9 3.6 3.6 — 9.5 9.0 -0.5
9.0 8.7 -0.3 8.4 9.9 1.5 7.2 7.7 0.5 8.6 8.1 -0.5
0.3 1.8 1.5 1.5 1.5 — 1.4 0.5 -0.9 1.8 2.7 0.9
7.5 6.9 -0.6 6.6 6.0 -0.6 3.6 3.2 -0.5 5.0 4.5 -0.5
2.7 2.4 -0.3 3.3 2.7 -0.6 5.4 5.4 — 3.2 3.6 0.5
8.4 8.7 0.3 10.2 11.1 0.9 13.1 13.1 — 13.6 12.2 -1.4
68.7 68.7 — 63.3 63.0 -0.3 65.6 66.5 0.9 58.4 59.7 1.4
Table 6. Reversing the ordering of the non-grounding bindings.
26
R. Bagnara, E. Zaffanella, and P. M. Hill
Goal Independent
without Struct Info
with Struct Info
Prec. class
O
I
G
F
L
O
I
G
F
L
5 < p ≤ 10
—
—
—
—
— 0.3 0.3
—
—
—
2<p≤5
0.3 0.3
—
—
—
—
—
—
—
—
0<p≤2
2.7 2.7
—
— 0.6 3.9 3.9
—
— 0.6
same precision
86.1 86.1 89.2 89.2 88.6 80.7 80.7 84.9 84.9 84.3
unknown
10.8 10.8 10.8 10.8 10.8 15.1 15.1 15.1 15.1 15.1
Goal Dependent
Prec. class
p > 20
10 < p ≤ 20
5 < p ≤ 10
0<p≤2
same precision
unknown
without Struct Info
with Struct Info
O
I
G
F
L
O
I
G
F
L
0.5 0.5
—
—
—
—
—
—
—
—
—
—
—
—
— 0.5 0.5
—
—
—
—
—
—
—
— 0.5 0.5
—
—
—
2.7 2.7
—
—
— 2.7 2.7
—
—
—
89.1 89.1 92.3 92.3 92.3 77.8 77.8 81.4 81.4 81.4
7.7 7.7 7.7 7.7 7.7 18.6 18.6 18.6 18.6 18.6
Table 7. Pos × SFL2 versus Pos ⊗ SFL.
(Codish et al. 1999) for a complete account of this composition and a justification
of its correctness.
This specification of the reduced product can be reformulated, using the standard set-sharing representation for the second component, to define a reduction
procedure reduce : Pos × SH → SH such that, for all φg ∈ Pos, sh ∈ SH ,
reduce(φg , sh) = S ∈ sh (VI \ S) ∈ [φg ]VI .
The enhanced integration of Pos and SFL, based on the above reduction operator,
is denoted here by Pos ⊗ SFL. From a formal point of view, this is not the reduced
product between Pos and SFL: while there is a complete reduction between Pos
and SH , the same does not necessarily hold for the combination with freeness and
linearity information. Also note that the domain Pos ⊗ SFL is strictly more precise
than the domain ShPSh , defined in (Scozzari 2000) for pair-sharing analysis. This is
because the domain ShPSh is the reduced product of a strict abstraction of Pos and
a strict abstraction of SH .
When using the domain PSD in place of SH , the ‘reduce’ operator specified
above can interact in subtle ways with an implementation removing the ρ-redundant
sharing groups from the elements of PSD . The following is an example where such
an interaction provides results that are not correct.
Let VI = {x, y, z} and sh = {xy, xz, yz, xyz} ∈ PSD be the current set-sharing
description. Suppose that the implementation internally represents sh by using the
ρ-reduced element sh red = {xy, xz, yz}, so that sh = ρ(sh red ). Suppose also that
the groundness description computed
on the domain Pos is φg = (x ↔ y ↔ z).
Note that we have [φg ]VI = ∅, {x, y, z} . Then we have
sh ′ = reduce(sh, φg ) = {xyz};
sh ′red = reduce(sh red , φg ) = ∅.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
27
The two Pos-reduced elements sh ′ and sh ′red are not equivalent, even modulo ρ.
Note that the above example does not mean that the reduced product between
Pos and PSD yields results that are not correct; neither does it mean that it is less
precise than the reduced product between Pos and SH for the computation of the
observables. More simply, the optimizations used in our current implementation of
PSD are not compatible with the above reduction process. Therefore, in Table 7 we
show the precision results obtained when comparing the base domain Pos × SFL2
with the domain Pos ⊗ SFL: the implementation of Pos ⊗ SFL, by avoiding ρreductions, is not affected by the correctness problem mentioned above.
The precision comparison provides empirical evidence that Pos ⊗ SFL is more
effective than the combination considered in (Bagnara et al. 2000). However, as
indicated by the number of time-outs reported in Table 7, using Pos ⊗ SFL is not
feasible due to its intrinsic exponential complexity. We deliberately decided not to
include the time comparison, since it would have provided no information at all:
the efficiency degradations, which are largely caused by the lack of ρ-reductions,
should not be attributed to the enhanced combination with Pos. In this respect,
the reader looking for more details is referred to (Bagnara et al. 2002).
For the only purpose of investigating how many precision improvements may
have been missed in the previous comparison due to the high number of time-outs,
we have performed another experimental evaluation where we have compared the
base domain Pos × SFL2 and the domain Pos ⊗ SFL2 . We stress the fact that,
given the observation made previously, such a precision comparison provides an
over-estimation for the actual improvements that can be obtained by a correct
integration of the ρ-reduction and the ‘reduce’ operators. A detailed investigation
of the experimental data, which cannot be reported here for space reasons, has
shown that the number of precision improvements shown in Table 7 could at most
double. In particular, improvements are more likely to occur for goal-independent
analyses.
8 Ground-or-free Variables
Most of the ideas investigated in the present work are based on earlier work by other
authors. In this section, we describe one originally proposed in (Bagnara et al. 2000).
Consider the analysis of the binding x = t and suppose that, on a set of computation
paths, this binding is reached with x ground while, on the remaining computation
paths, the binding is reached with x free. In both cases x will be linear and this is all
that will be recorded when using the usual combination Pos×SFL. This information
is valuable since, in the case that x and t are independent, it allows the star-union
operation for the relevant component for t to be dispensed with. However, the information that is lost, that is, x being either ground or free, is equally valuable, since
this would allow the avoidance of the star-union of both the relevant components
for x and t, even when x and t may share. This loss has the disadvantages that
CPU time is wasted by performing unnecessary but costly operations and that the
precision is potentially degraded: not only are the extra star-unions useless for cor-
28
R. Bagnara, E. Zaffanella, and P. M. Hill
rectness but may introduce redundant sharing groups to the detriment of accuracy.
It is therefore useful to track the additional mode ‘ground-or-free’.
def
The analysis domain SFL is extended with the component GF = ℘(VI ) consisting of the set of variables that are known to be either ground or free. As for
freeness and linearity, the approximation ordering on GF is given by reverse subset
inclusion. When computing the abstract mgu on the new domain
def
SGFL = SH × F × GF × L,
the property of being ground-or-free is used and propagated in almost the same
way as freeness information.
Definition 6
(Improved abstract operations over SGFL.) Let d = hsh, f, gf , li ∈ SGFL. We
define the predicate gfree d : Terms → Bool such that, for each first order term t,
def
where Vt = vars(t) ⊆ VI ,
def
gfree d (t) = rel(Vt , sh) = ∅ ∨ (∃x ∈ VI . x = t ∧ x ∈ gf ).
Consider the specification of the abstract operations over SFL given in Definition 4.
The improved operator amgu : SGFL × Bind → SGFL is given by
def
amgu d, x = t = hsh ′ , f ′ , gf ′ , l′ i,
where f ′ and l′′ are defined as in Definition 4 and
sh ′ = rel(Vxt , sh) ∪ bin Sx , St ;
(
Rx , if gfree d (x) ∨ gfree d (t) ∨ lin d (t) ∧ ind d (x, t) ;
Sx =
Rx⋆ , otherwise;
(
Rt , if gfree d (x) ∨ gfree d (t) ∨ lin d (x) ∧ ind d (x, t) ;
St =
Rt⋆ , otherwise;
gf ′ = VI \ vars(sh ′ ) ∪ gf ′′ ;
gf ,
if gfree d (x) ∧ gfree d (t);
gf \ vars(R ),
if gfree d (x);
x
gf ′′ =
gf \ vars(Rt ),
if gfree d (t);
gf \ vars(Rx ∪ Rt ), otherwise;
l′ = gf ′ ∪ l′′ .
The computation of the set gf ′′ is very similar to the computation of the set
f ′ as given in Definition 4. The new ground-or-free component gf ′ is obtained by
adding to gf ′′ the set of all the ground variables: in other words, if a variable
“loses freeness” then it also loses its ground-or-free status unless it is known to
be definitely ground. It can be noted that, in the computation of this improved
amgu, the ground-or-free property takes the role previously played by freeness. In
particular, when computing sh ′ , all the tests for freeness have been replaced by
tests on the newly defined Boolean function gfree d ; similarly, in the computation
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
29
of the new linearity component l′ , the set f ′ has been replaced by gf ′ (since any
ground-or-free variable is also linear). It is also easy to generalize the improvement
for definitely cyclic bindings introduced in Definition 5 to the domain SGFL: as
before, the test free d (x) needs to be replaced with the new test gfree d (x).
To summarize, the incorporation of the set of ground-or-free variables is cheap,
both in terms of computational complexity and in terms of code to be written.
As far as computational complexity is concerned this extension looks particularly
promising, since the possibility of avoiding star-unions has the potential of absorbing
its overhead if not of giving rise to a speed-up.
Thus the domain Pos × SGFL was experimentally evaluated on our benchmark
suite, with and without the structural information provided by Pattern(·), both in
a goal-dependent and in a goal-independent way, and the results compared with
those previously obtained for the domain Pos × SFL. Note that the implementation
def
uses the non-redundant version SGFL2 = PSD × F × GF × L. In the precision
comparisons of Table 8, the new column labeled GF reports precision improvements
measured on the ground-or-free property itself.19
As far as the timings are concerned, the experimentation fully confirms our qualitative reasoning: efficiency improvements are more frequent than degradations and,
even with widening operators switched off, the distributions of the total analysis
times show minor changes only. As for precision, disregarding the many improvements in the GF columns, few changes can be observed, and almost all of these
concern just the linearity information.20
The results in Table 8, show that tracking ground-or-free variables, while being
potentially useful for improving the precision of a sharing analysis, rarely reaches
such a goal. In contrast, the precision gains on the ground-or-free property itself are
remarkable, affecting from 39% to 74% of the programs in the benchmark suite. It
is possible to foresee several direct applications for this information that, together
with the just mentioned negligible computational cost, fully justify the inclusion of
this enhancement in a static analyzer. In particular, there are at least two ways in
which a knowledge of ground-or-free variables could improve the concrete unification
procedure.
The first case applies in the context of occurs-check reduction (Søndergaard 1986;
Crnogorac et al. 1996), that is when a program designed for a logic programming
system performing the occurs-check is to be run on top of a system omitting this
test. In order to ensure correct execution, all the explicit and implicit unifications in
the program are treated as if the ISO Prolog built-in unify with occurs check/2
was used to perform them. In order to minimize the performance overhead, it is
important to detect, as precisely as possible and at compile-time, those NSTO
(short for Not Subject To the Occurs-check (Deransart et al. 1991; ISO/IEC 1995))
unifications where the occurs-check will not be needed. For these unifications, =/2
19
20
For this comparison, in the analysis using Pos × SFL, the number of ground-or-free variables is
computed by summing the number of ground variables with the number of free variables.
In fact the sole improvement to the number of independent pairs is due to a synthetic benchmark,
named gof, that was explicitly written to show that variable independence could be affected.
30
R. Bagnara, E. Zaffanella, and P. M. Hill
Goal Ind.
Prec. class
p > 20
10 < p ≤ 20
5 < p ≤ 10
2<p≤5
0<p≤2
same precision
unknown
Goal Dep.
without Struct
O
I
G
F
52.7 0.3
—
—
11.7
—
—
—
5.4
—
—
—
2.4
—
—
—
0.3
—
—
—
24.1 96.4 96.7 96.7
3.3 3.3 3.3 3.3
Info
with Struct Info
GF L
O
I
G
F GF L
52.7
— 48.5 0.3
—
— 48.5
—
11.7
— 16.0
—
—
— 16.0
—
5.4
— 7.5
—
—
— 7.5
—
2.4
— 1.8
—
—
— 1.8
—
0.3 1.5 0.6
—
—
— 0.6 1.5
24.1 95.2 19.0 93.1 93.4 93.4 19.0 91.9
3.3 3.3 6.6 6.6 6.6 6.6 6.6 6.6
without Struct Info
Prec. class
p > 20
10 < p ≤ 20
5 < p ≤ 10
2<p≤5
0<p≤2
same precision
unknown
with Struct Info
O
I
G
F GF L
O
I
G
F
5.9
—
—
— 5.9
— 5.9
—
—
—
4.5
—
—
— 4.5
— 5.4
—
—
—
7.7 0.5
—
— 7.7
— 5.4 0.5
—
—
13.1
—
—
— 13.1
— 12.2
—
—
—
8.1
—
—
— 8.1 0.5 10.0
—
—
—
57.0 95.9 96.4 96.4 57.0 95.9 51.6 90.0 90.5 90.5
3.6 3.6 3.6 3.6 3.6 3.6 9.5 9.5 9.5 9.5
GF L
5.9
—
5.4
—
5.4
—
12.2
—
10.0
—
51.6 90.5
9.5 9.5
Time diff. class
Goal Ind.
Goal Dep.
w/o SI with SI w/o SI with SI
degradation > 1
—
0.6
—
0.9
0.5 < degradation ≤ 1
0.3
—
0.5
—
0.2 < degradation ≤ 0.5
—
0.6
0.5
1.4
0.1 < degradation ≤ 0.2
0.3
—
—
0.5
both timed out
3.3
6.6
3.6
9.5
same time
88.6
85.2
87.3
82.8
0.1 < improvement ≤ 0.2
1.2
1.2
1.8
1.4
0.2 < improvement ≤ 0.5
2.4
2.4
1.8
0.9
0.5 < improvement ≤ 1
2.1
0.9
2.3
0.9
improvement > 1
1.8
2.4
2.3
1.8
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Independent
Goal Dependent
without SI
with SI
without SI
with SI
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
3.3 3.3 — 6.6 6.6 — 3.6 3.6 — 9.5 9.5 —
9.0 9.0 — 8.4 8.4 — 7.2 7.2 — 8.6 8.6 —
0.3 0.3 — 1.5 1.5 — 1.4 1.4 — 1.8 1.8 —
7.5 7.5 — 6.6 6.6 — 3.6 3.6 — 5.0 5.0 —
2.7 2.7 — 3.3 3.6 0.3 5.4 5.9 0.5 3.2 3.2 —
8.4 8.7 0.3 10.2 10.5 0.3 13.1 12.7 -0.5 13.6 14.0 0.5
68.7 68.4 -0.3 63.3 62.7 -0.6 65.6 65.6 — 58.4 57.9 -0.5
Table 8. P os × SFL2 versus Pos × SGFL2 .
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
31
can safely be used; for the remaining ones, the program will have to be transformed
so that unify with occurs check/2 is explicitly called to perform them. Groundor-freeness can be of help for this application, since a unification between two
ground-or-free variables is NSTO. Note that this is an improvement with respect
to the technique used in (Crnogorac et al. 1996), since it is not required that the
two considered variables are independent.
As a second application, ground-or-freeness can be useful to replace the full concrete unification procedure by a simplified version. Since a ground-or-free term is
either ground or free, a single run-time test for freeness will discriminate between
the two cases: if this test succeeds, unification can be implemented by a single assignment; if the test fails, any specialized code for unification with a ground term
can be safely invoked. In particular, when unifying two ground-or-free variables
that are not free at run-time, the full unification procedure can be replaced by a
simpler recursive test for equivalence.
9 More Precise Exploitation of Linearity
In (King 1994), A. King proposes a domain for sharing analysis that performs
a quite precise tracking of linearity. Roughly speaking, each sharing group in a
sharing-set carries its own linearity information. In contrast, in the approach of
(Langen 1990), which is the one usually followed, a set of definitely linear variables
is recorded along with each sharing-set. The proposal in (King 1994) gives rise to
a domain that is quite different from the ones presented here. Since (King 1994)
does not provide an experimental evaluation and we are unaware of any subsequent
work on the subject, the question whether this more precise tracking of linearity is
actually worthwhile (both in terms of precision and efficiency) seems open.
What interests us here is that part of the theoretical work presented in (King 1994)
may be usefully applied even in the more classical treatments of linearity such as
the one being used in this paper. As far as we can tell, this fact was first noted
in (Bagnara et al. 2000).
In (King 1994), point 3 of Lemma 5 (which is reported to be proven in (King 1993))
states that, if s is a linear term independent from a term t, then in the unifier for
s = t any sharing between the variables in s is necessarily caused by those variables
that can occur more than once in t.
This result can be exploited even when using the domain SFL. Given the abstract
element d = hsh, f, li, let x ∈ (l \ f ) be a non-free but linear variable and let t be
a non-linear term such that ind d (x, t). Let also Vx , Vt , Vxt , Rx and Rt be as given
in Definition 4. In such a situation, when abstractly evaluating the binding x = t,
the standard amgu operator gives the set-sharing component
sh ′ = rel(Vxt , sh) ∪ bin(Rx⋆ , Rt ).
Suppose the set Vt is partitioned into the two components Vtl and Vtnl , where Vtnl
is the set of the “problematic” variables, that is, those variables that potentially
32
R. Bagnara, E. Zaffanella, and P. M. Hill
Goal Independent
Goal Dependent
O
I
G
F
L
O
I
G
F
L
0.3 0.3
—
—
—
—
—
—
—
—
—
—
—
—
— 0.5 0.5
—
—
—
93.1 93.1 93.4 93.4 93.4 90.0 90.0 90.5 90.5 90.5
6.6 6.6 6.6 6.6 6.6 9.5 9.5 9.5 9.5 9.5
Prec. class
p > 20
2<p≤5
same precision
unknown
Time difference class
degradation > 1
0.5 < degradation ≤ 1
0.2 < degradation ≤ 0.5
0.1 < degradation ≤ 0.2
both timed out
same time
0.1 < improvement ≤ 0.2
0.2 < improvement ≤ 0.5
0.5 < improvement ≤ 1
improvement > 1
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
% benchmarks
Goal Ind. Goal Dep.
0.3
—
—
—
—
—
0.3
0.5
6.6
9.5
85.2
83.7
0.9
2.4
0.6
3.6
1.8
0.5
2.7
1.4
Goal Ind.
Goal Dep.
%1 %2 ∆
%1 %2 ∆
6.6 6.6 — 9.5 9.5 —
8.4 8.4 — 8.6 8.6 —
1.5 1.5 — 1.8 1.8 —
6.6 6.6 — 5.0 5.0 —
3.3 3.3 — 3.2 3.2 —
10.2 11.1 0.9 13.6 14.0 0.5
63.3 62.3 -0.9 58.4 57.9 -0.5
Table 9. The effect of enhanced linearity on Pattern(Pos × SFL2 ).
make t a non-linear term. Formally,
y∈l
l def
;
Vt = y ∈ vars(t) y A mvars(t) =⇒ y ∈
/ vars(sh)
∀z ∈ vars(t) : y = z ∨ ind (y, z)
d
def
Vtnl =
Vt \
Vtl .
Let Rtl = rel(Vtl , sh) and Rtnl = rel(Vtnl , sh). Note that Rtnl 6= ∅, because t is a nonlinear term. If also Rtl 6= ∅ then the standard amgu can be replaced by an improved
version (denoted by amguk ) computing the following set-sharing component:
sh ′k = rel(Vxt , sh) ∪ bin(Rx , Rtl ) ∪ bin(Rx⋆ , Rtnl ).
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
33
As a consequence of King’s result (King 1994, Lemma 5), only Rtnl (the relevant
component of sh with respect to the problematic variables Vtnl ) has to be combined
with Rx⋆ while Rtl can be combined with just Rx (without the star-union).
For a working example, suppose VI = {v, w, x, y, z} is the set of variables of
interest and consider the SFL element
def
d = {vx, wx, y, z}, {v, w, y}, {v, w, x, y}
with the binding x = f (y, z). Note that all the applicability conditions specified
above are met: in particular t = f (y, z) is not linear because z ∈
/ l. As Rx = {vx, wx}
and Rt = {y, z}, a standard analysis would compute
d ′ = amgu d , x = f (y, z)
= {vwxy, vwxz, vxy, vxz, wxy, wxz}, ∅, {y} .
On the other hand, since Vtl = {y} and Vtnl = {z}, the enhanced analysis would
compute
dk′ = amguk d , x = f (y, z)
= {vwxz, vxy, vxz, wxy, wxz}, ∅, {y} .
Note that dk′ does not include the sharing group vwxy. This means that, if in the
sequel of the computation variable z is bound to a ground term, then variables
v and w will be known to be definitely independent. This independence is not
captured when using the standard amgu since d ′ includes the sharing group vwxy,
and therefore the variables v and w will potentially share even after grounding z.
The experimental evaluation for this enhancement is reported in Table 9. The
comparison of times shows that the efficiency of the analysis, when affected, is
more likely to be improved than degraded. As for the precision, improvements are
observed for only two programs; moreover, these are synthetic benchmarks such as
the above example. Nevertheless, despite its limited practical relevance, this result
demonstrates that the standard combination of Sharing with linearity information
is not optimal, even when all the possible orderings of the non-grounding bindings
are tried.
10 Sharing and Freeness
As noted by several authors (Bruynooghe et al. 1994a; Bueno et al. 1994; Cabeza and Hermenegildo 1994),
the standard combination of Sharing and Free is not optimal. G. Filé (Filé 1994)
formally identified the reduced product of these domains and proposed an improved
abstract unification operator. This new operator exploits two properties that hold
for the most precise abstract description of a single concrete substitution:
1. each free variable occurs in exactly one sharing group;
2. two free variables occur in the same sharing group if and only if they are
aliases (i.e., they have become the same variable).
34
R. Bagnara, E. Zaffanella, and P. M. Hill
When considering the general case, where sets of concrete substitutions come into
play, property 1 can be used to (partially) recover disjunctive information. In particular, it is possible to decompose an abstract description into a set of (maximal)
descriptions that necessarily come from different computation paths, each one satisfying property 1. The abstract unification procedure can thus be computed separately on each component, and the results of each subcomputation are then joined
to give the final description. As such components are more precise than the original
description (they possibly contain more ground variables and less sharing pairs),
precision gains can be obtained.
Furthermore, by exploiting property 2 on each component, it is possible to correctly infer that for some of them the computation will fail due to a functor clash
(or to the occurs-check, if considering a system working on finite trees). Note that a
similar improvement is possible even without decomposing the abstract description.
As an example, consider an abstract element such as the following:
d = {xy, u, v}, {x, y}, {x, y} .
Since the sharing group xy is the only one where the free variables x and y occur,
property 2 states that x and y are indeed the same variable in all the concrete
computation states
described by d ∈ SFL. Therefore, when abstractly evaluating
the substitution x = f (u), y = g(v) , it can be safely concluded that its concrete
counterparts will result in failure due to the functor clash. In the same circumstances,
it can also be concluded that a concrete substitution corresponding to, say,
x = f (y) will cause a failure of the occurs-check, if this is performed.
As was the case for the reduced product between Pos and SH (see Section 7), the
interaction between the enhanced abstract unification operator and the elimination
of ρ-redundant elements can lead to results that are not correct.
To see this, let VI = {w, x, y, z} and consider the set of concrete substitutions
Σ = ℘(σ), where σ = {x 7→ v, y 7→ v, z 7→ v} (note that v ∈
/ VI ). The abstract
element describing Σ is d = hsh, f, li ∈ SFL, where sh = {w, x, xy, xyz, xz, y, yz, z}
and f = l = VI . Suppose that the implementation represents d by using the reduced
element dred = hsh red , f, li, where sh red = sh \ {xyz}, so that sh = ρ(sh red ).
According to the specification of the enhanced operator, dred can be decomposed
into the following four components:
c1 = {w, x, y, z}, f, l ,
c3 = {w, xz, y}, f, l ,
c2 = {w, x, yz}, f, l ,
c4 = {w, xy, z}, f, l .
Consider the binding x = f (y, w) and, for
each i ∈ {1, . . . , 4}, the computation
of c′i = hsh ′i , fi′ , li′ i = amgu ci , x = f (y, w) , where we have l1′ = l2′ = l3′ = VI and
l4′ = {w, z}. In all four cases, we have z ∈ li′ , so that z keeps its linearity even after
merging the results of the four subcomputations into a single abstract description.
In contrast, when performing the same computation with the original abstract
description d in the decomposition phase, we also obtain a fifth component,
c5 = {w, xyz}, f, l .
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
35
Goal Independent
without Struct Info
with Struct Info
Prec. class
O
I
G
F
L
O
I
G
F
L
p > 20
0.3 0.3
—
—
—
—
—
—
—
—
5 < p ≤ 10
—
—
—
—
— 0.3
—
—
— 0.3
0<p≤2
0.9 0.3
—
— 0.6 3.6 3.0
—
— 0.6
same precision
94.6 95.2 95.8 95.8 95.2 86.1 87.0 90.1 90.1 89.2
unknown
4.2 4.2 4.2 4.2 4.2 9.9 9.9 9.9 9.9 9.9
Goal Dependent
Prec. class
same precision
unknown
without Struct Info
with Struct Info
O
I
G
F
L
O
I
G
F
L
96.4 96.4 96.4 96.4 96.4 89.6 89.6 89.6 89.6 89.6
3.6 3.6 3.6 3.6 3.6 10.4 10.4 10.4 10.4 10.4
Time diff. class
degradation
0.5 < degradation
0.2 < degradation
0.1 < degradation
>1
≤1
≤ 0.5
≤ 0.2
both timed out
same time
0.1 < improvement ≤ 0.2
0.2 < improvement ≤ 0.5
0.5 < improvement ≤ 1
improvement > 1
Total time class
timed out
t > 10
5 < t ≤ 10
1<t≤5
0.5 < t ≤ 1
0.2 < t ≤ 0.5
t ≤ 0.2
Goal Ind.
Goal Dep.
w/o SI with SI w/o SI with SI
9.6
13.6
3.2
5.9
0.6
1.8
1.4
1.4
3.3
2.4
1.8
3.6
0.6
1.5
2.3
1.4
3.3
82.2
—
0.3
—
—
6.6
73.5
—
—
—
0.6
3.6
87.8
—
—
—
—
9.5
77.8
—
—
—
0.5
Goal Independent
Goal Dependent
without SI
with SI
without SI
with SI
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
%1 %2 ∆
3.3 4.2 0.9 6.6 9.9 3.3 3.6 3.6 — 9.5 10.4 0.9
9.0 9.6 0.6 8.4 8.4 — 7.2 7.2 — 8.6 8.1 -0.5
0.3 0.9 0.6 1.5 1.2 -0.3 1.4 1.4 — 1.8 1.8 —
7.5 6.9 -0.6 6.6 5.7 -0.9 3.6 3.6 — 5.0 4.5 -0.5
2.7 2.1 -0.6 3.3 4.5 1.2 5.4 5.9 0.5 3.2 3.2 —
8.4 8.4 — 10.2 12.0 1.8 13.1 12.7 -0.5 13.6 14.9 1.4
68.7 67.8 -0.9 63.3 58.1 -5.1 65.6 65.6 — 58.4 57.0 -1.4
Table 10. The effect of enhanced freeness on Pos × SFL2 .
When computing c′5 = hsh ′5 , f5′ , l5′ i = amgu c5 , x = f (y, w) , we obtain l5′ = {w},
so that z loses its linearity when merging the five results into a single abstract
description. Note that this is not an avoidable precision loss, since in the concrete
36
R. Bagnara, E. Zaffanella, and P. M. Hill
computation path corresponding to the substitution σ we would have computed
σ ′ = x 7→ f (x, w), y 7→ f (y, w), z 7→ f (z, w) ,
where z is bound to a non-linear term (namely, an infinite rational term with an
infinite number of occurrences of variable w). Therefore, the result obtained when
using the abstract description dred is not correct.
As already observed in Section 7, the above correctness problem lies not in the
SFL2 domain itself, but rather in our optimized implementation, which removes
the ρ-redundant elements from the set-sharing description.
We implemented the first idea by Filé (i.e., the exploitation of property 1) on
the usual base domain P os × SFL2 . As noted above, this implementation may yield
results that are not correct: the precision comparison reported in Table 10 provides
an over-estimation of the actual improvements that could be obtained by a correct
implementation. However, it is not possible to assess the magnitude of this overestimation, since our implementation of this enhancement on the domain Pos×SFL,
where no ρ-redundancy elimination is performed, times-out on a large fraction of
the benchmarks. The results in Table 10 show that precision improvements are only
observed for goal-independent analysis. When looking at the time comparisons,
it should be observed that the analysis of several programs had to be stopped
because of the combinatorial explosion in the decomposition, even though we used
the domain Pos × SFL2 . Among the proposals experimentally evaluated in this
paper, this one shows the worst trade-off between cost and precision.
Note that, in principle, such an approach to the recovery of disjunctive information can be pursued beyond the integration of sharing with freeness. In fact, by
exploiting the ground-or-free information as in Section 8, it is possible to obtain
decompositions where each component contains at most one occurrence (in contrast
with the exactly one occurrence of Filé’s idea) of each ground-or-free variable. In
each component, the ground-or-free variable could then be “promoted” as either a
ground variable (if it does not occur in the sharing groups of that component) or
as a free variable (if it occurs in exactly one sharing group).
It would be interesting to experiment with the second idea of Filé. However,
such a goal would require a big implementation effort, since at present there is no
easy way to incorporate this enhancement into the modular design of the China
analyzer.21
11 Tracking Compoundness
In (Bruynooghe et al. 1994a; Bruynooghe et al. 1994b), Bruynooghe and colleagues
considered the combination of the standard set-sharing, freeness, and linearity domains with compoundness information. As for freeness and linearity, compoundness
21
Roughly speaking, the SFL component should be able to produce some new (implicit) structural
information and notify it to the enclosing Pattern(·) component, which would then need to
combine this information with the (explicit) structural information already available. However,
in order to be able to receive notifications from its parameter, the Pattern(·) component, which
is implemented as a C++ template, would have to be heavily modified.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
37
was represented by the set of variables that definitely have the corresponding property.
As discussed in (Bruynooghe et al. 1994a; Bruynooghe et al. 1994b), compoundness information is useful in its own right for clause indexing. Here though, the
focus is on improving sharing information, so that the question to be answered is:
can the tracking of compoundness improve the sharing analysis itself? This question
is also considered in (Bruynooghe et al. 1994a; Bruynooghe et al. 1994b) where a
technique is proposed that exploits the combination of sharing, freeness and compoundness. This technique relies on the presence of the occurs-check.
Informally, consider the binding x = t together with an abstract description
where x is a free variable, t is a compound term and x definitely shares with t.
Since x is free, x is aliased to one of the variables occurring in t. As a consequence,
the execution of the binding x = t will fail due to the occurs-check. In a more
general case, when only possible sharing information is available, the precision of
the abstract description can be safely improved by removing, just before computing
the abstract binding, all the sharing groups containing both x and a variable in t.
In addition, if this reduction step removes all the sharing groups containing a free
variable, then it can be safely concluded that the computation will fail.
To see how this works in practice, consider the binding x = f (y, z) and the
def
description d1 = hsh 1 , f1 , l1 i ∈ SFL such that
def
sh 1 = {wx, xy, xz, y, z},
def
f1 = {x},
def
l1 = {w, x, y, z}.
Since x is free and f (y, z) is compound, the sharing-groups xy and xz can be
removed so that the amgu computation will give the set-sharing and linearity components
def
sh ′1 = {wxy, wxz},
def
l1′ = {w, x, y, z}
instead of the less precise
def
sh ′1 = {wxy, wxz, xy, xyz, xz},
def
l1′ = {w}.
Note that the precision improvement of this particular example could also be obtained by applying, in its full generality, the second technique proposed by Filé and
sketched in the previous section. This is because the term with which x is unified is
“explicitly” compound. However, if the term t was “implicitly” compound (i.e., if it
was an abstract variable known to represent compound terms) then the technique
by Filé would not be applicable. For example, consider the binding x = y and the
38
R. Bagnara, E. Zaffanella, and P. M. Hill
def
description d2 = hsh 2 , f2 , l2 i ∈ SFL such that
def
sh 2 = {wx, xyz, y},
def
f2 = {x},
def
l2 = {w, x, y, z}
supplemented by a compoundness component ensuring that y is compound. Then
the sharing-group xyz can be removed so that the amgu will compute
def
sh ′2 = {wxy},
def
l2′ = {w, x, y, z}
instead of
def
sh ′2 = {wxy, wxyz, xyz},
def
l2′ = {w}.
To see how a knowledge of the compoundness can be used to identify definite failure,
def
consider the unification x = f (y, z) and the description d3 = hsh 3 , f3 , l3 i ∈ SFL
such that
def
sh 3 = {wxy, wxz, x, y, z},
def
f3 = {w, x},
def
l3 = {w, x, y, z}.
def
As in the examples above, variable x is free and term t = f (y, z) is compound so
that, by applying the reduction step, we can remove the sharing groups wxy and
wxz. However, this has removed all the sharing groups containing the free variable
w, resulting in an inconsistent computation state.
We did not implement this technique, since it is only sound for the analysis
of systems performing the occurs-check, whereas we are targeting at the analysis
of systems possibly omitting it. Nonetheless, an experimental evaluation would
be interesting for assessing how much this precision improvement can affect the
accuracy of applications such as occurs-check reduction.
12 Conclusion
In this paper we have investigated eight enhanced sharing analysis techniques that,
at least in principle, have the potential for improving the precision of the sharing
information over and above that obtainable using the classical combination of setsharing with freeness and linearity information. These techniques either make a
better use of the already available sharing information, by defining more powerful
abstract semantic operators, or combine this sharing information with that captured
by other domains. Our work has been systematic since, to the best of our knowledge,
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
39
we have considered all the proposals that have appeared in the literature: that is,
better exploitation of groundness, freeness, linearity, compoundness, and structural
information.
Using the China analyzer, seven of the eight enhancements have been experimentally evaluated. Because of the availability of a very large benchmark suite,
including several programs of respectable size, the precision results are as conclusive as possible and provide an almost complete account of what is to be expected
when analyzing any real program using these domains.
The results demonstrate that good precision improvements can be obtained with
the inclusion of explicit structural information. For the groundness domain Pos,
several good reasons have been given as to why it should be combined with setsharing. As for the remaining proposals, it is hard to justify them as far as the
precision of the analysis is concerned.
Regarding the efficiency of the analysis, it has been explained why the reported
time comparisons can be considered as upper bounds to the additional cost required by the inclusion of each technique. Moreover, it has been argued that, from
this point of view, the addition of a ‘ground-or-free’ mode and the more precise
exploitation of linearity are both interesting: they are not likely to affect the cost
of the analysis and, when this is the case, they usually give rise to speed-ups.
No further positive indications can be derived from the precision and time comparisons of the remaining techniques. In particular, it has not been possible to
identify a good heuristic for the reordering of the non-grounding bindings. The experimentation suggests that sensible precision improvements cannot be expected
from this technique. When considering these negative results, the reader should be
aware that the precision gains are measured with respect to an analysis tool built on
the base domain Pos × SFL which, to our knowledge, is the most accurate sharing
analysis tool ever implemented.
The experimentation reported in this paper resulted in both positive and negative
indications. We believe that all of these will provide the right focus in the design
and development of useful tools for sharing analysis.
Acknowledgments
This paper is dedicated to all those who take a visible stance in favor of scientific integrity. In particular, it is dedicated to David Goodstein, for “Conduct and
Misconduct in Science”; to John Koza, for “A Peer Review of the Peer Reviewing Process of the International Machine Learning Conference”; to Krzsystof Apt,
Veronica Dahl and Catuscia Palamidessi for the Association for Logic Programming’s “Code of Conduct for Referees”; and to the large number of honest and
thorough referees who do so much to help maintain and improve the quality of all
publications.
References
Armstrong, T., Marriott, K., Schachte, P., and Søndergaard, H. 1998. Two
40
R. Bagnara, E. Zaffanella, and P. M. Hill
classes of Boolean functions for dependency analysis. Science of Computer Programming 31, 1, 3–45.
Bagnara, R. 1997a. Data-flow analysis for constraint logic-based languages. Ph.D. thesis,
Dipartimento di Informatica, Università di Pisa, Pisa, Italy. Printed as Report TD-1/97.
Bagnara, R. 1997b. Structural information analysis for CLP languages. In Proceedings of
the “1997 Joint Conference on Declarative Programming (APPIA-GULP-PRODE’97)”,
M. Falaschi, M. Navarro, and A. Policriti, Eds. Grado, Italy, 81–92.
Bagnara, R., Hill, P. M., and Zaffanella, E. 1997. Set-sharing is redundant for
pair-sharing. In Static Analysis: Proceedings of the 4th International Symposium, P. Van
Hentenryck, Ed. Lecture Notes in Computer Science, vol. 1302. Springer-Verlag, Berlin,
Paris, France, 53–67.
Bagnara, R., Hill, P. M., and Zaffanella, E. 2000. Efficient structural information
analysis for real CLP languages. In Proceedings of the 7th International Conference
on Logic for Programming and Automated Reasoning (LPAR 2000), M. Parigot and
A. Voronkov, Eds. Lecture Notes in Artificial Intelligence, vol. 1955. Springer-Verlag,
Berlin, Réunion Island, France, 189–206.
Bagnara, R., Hill, P. M., and Zaffanella, E. 2002. Set-sharing is redundant for
pair-sharing. Theoretical Computer Science 277, 1-2, 3–46.
Bagnara, R. and Schachte, P. 1999. Factorizing equivalent variable pairs in ROBDDbased implementations of Pos. In Proceedings of the “Seventh International Conference
on Algebraic Methodology and Software Technology (AMAST’98)”, A. M. Haeberer,
Ed. Lecture Notes in Computer Science, vol. 1548. Springer-Verlag, Berlin, Amazonia,
Brazil, 471–485.
Bagnara, R., Zaffanella, E., and Hill, P. M. 2000. Enhanced sharing analysis
techniques: A comprehensive evaluation. In Proceedings of the 2nd International ACM
SIGPLAN Conference on Principles and Practice of Declarative Programming, M. Gabbrielli and F. Pfenning, Eds. Association for Computing Machinery, Montreal, Canada,
103–114.
Blockeel, H., Demoen, B., Janssens, G., Vandencasteele, H., and Van Laer, W.
2000. Two advanced transformations for improving the efficiency of an ILP system. In
Work-in-Progress Reports, Tenth International Conference on Inductive Logic Programming, J. Cussens and A. Frisch, Eds. London, UK, 43–59.
Bourdoncle, F. 1993a. Efficient chaotic iteration strategies with widenings. In Proceedings of the International Conference on “Formal Methods in Programming and Their
Applications”, D. Bjørner, M. Broy, and I. V. Pottosin, Eds. Lecture Notes in Computer Science, vol. 735. Springer-Verlag, Berlin, Academgorodok, Novosibirsk, Russia,
128–141.
Bourdoncle, F. 1993b. Sémantiques des langages impératifs d’ordre supérieur et interprétation abstraite. PRL Research Report 22, DEC Paris Research Laboratory.
Bruynooghe, M. and Codish, M. 1993. Freeness, sharing, linearity and correctness
— All at once. In Static Analysis, Proceedings of the Third International Workshop,
P. Cousot, M. Falaschi, G. Filé, and A. Rauzy, Eds. Lecture Notes in Computer Science, vol. 724. Springer-Verlag, Berlin, Padova, Italy, 153–164. An extended version is
available as Technical Report CW 179, Department of Computer Science, K.U. Leuven,
September 1993.
Bruynooghe, M., Codish, M., and Mulkers, A. 1994a. Abstract unification for a
composite domain deriving sharing and freeness properties of program variables. In
Verification and Analysis of Logic Languages, Proceedings of the W2 Post-Conference
Workshop, International Conference on Logic Programming, F. S. de Boer and M. Gabbrielli, Eds. Santa Margherita Ligure, Italy, 213–230.
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
41
Bruynooghe, M., Codish, M., and Mulkers, A. 1994b. A composite domain for freeness, sharing, and compoundness analysis of logic programs. Technical Report CW 196,
Department of Computer Science, K.U. Leuven, Belgium. July.
Bueno, F., de la Banda, M. G., and Hermenegildo, M. V. 1994. Effectiveness of
global analysis in strict independence-based automatic program parallelization. In Logic
Programming: Proceedings of the 1994 International Symposium, M. Bruynooghe, Ed.
MIT Press Series in Logic Programming. The MIT Press, Ithaca, NY, USA, 253–268.
Bueno, F., de la Banda, M. G., and Hermenegildo, M. V. 1999. Effectivness of
abstract interpretation in automatic parallelization: a case study in logic programming.
ACM Transactions on Programming Languages and Systems 21, 2, 189–239.
Cabeza, D. and Hermenegildo, M. V. 1994. Extracting non-strict independent andparallelism using sharing and freeness information. In Static Analysis: Proceedings of the
1st International Symposium, B. Le Charlier, Ed. Lecture Notes in Computer Science,
vol. 864. Springer-Verlag, Berlin, Namur, Belgium, 297–313.
Chang, J.-H., Despain, A. M., and DeGroot, D. 1985. AND-parallelism of logic
programs based on a static data dependency analysis. In Digest of Papers of COMPCON
Spring’85. IEEE Computer Society Press, San Francisco, California, 218–225.
Codish, M., Dams, D., Filé, G., and Bruynooghe, M. 1993. Freeness analysis for
logic programs — and correctness? In Logic Programming: Proceedings of the Tenth
International Conference on Logic Programming, D. S. Warren, Ed. MIT Press Series
in Logic Programming. The MIT Press, Budapest, Hungary, 116–131. An extended
version is available as Technical Report CW 161, Department of Computer Science,
K.U. Leuven, December 1992.
Codish, M., Dams, D., and Yardeni, E. 1991. Derivation and safety of an abstract
unification algorithm for groundness and aliasing analysis. See Furukawa (1991), 79–93.
Codish, M., Søndergaard, H., and Stuckey, P. J. 1999. Sharing and groundness
dependencies in logic programs. ACM Transactions on Programming Languages and
Systems 21, 5, 948–976.
Cortesi, A. and Filé, G. 1999. Sharing is optimal. Journal of Logic Programming 38, 3,
371–386.
Cortesi, A., Filé, G., and Winsborough, W. 1992. Comparison of abstract interpretations. In Proceedings of the 19th International Colloquium on Automata, Languages
and Programming (ICALP’92), M. Kuich, Ed. Lecture Notes in Computer Science, vol.
623. Springer-Verlag, Berlin, Wien, Austria, 521–532.
Cortesi, A., Le Charlier, B., and Van Hentenryck, P. 1994. Combinations of
abstract domains for logic programming. In Conference Record of POPL’94: 21st ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages. ACM Press,
Portland, Oregon, 227–239.
Cousot, P. and Cousot, R. 1979. Systematic design of program analysis frameworks.
In Proceedings of the Sixth Annual ACM Symposium on Principles of Programming
Languages. ACM Press, New York, 269–282.
Crnogorac, L., Kelly, A. D., and Søndergaard, H. 1996. A comparison of three
occur-check analysers. In Static Analysis: Proceedings of the 3rd International Symposium, R. Cousot and D. A. Schmidt, Eds. Lecture Notes in Computer Science, vol. 1145.
Springer-Verlag, Berlin, Aachen, Germany, 159–173.
Deransart, P., Ferrand, G., and Téguia:, M. 1991. NSTO programs (Not Subject to
Occur-Check). In Logic Programming: Proceedings of the 1991 International Symposium,
V. A. Saraswat and K. Ueda, Eds. MIT Press Series in Logic Programming. The MIT
Press, San Diego, USA, 533–547.
Filé, G. 1994. Share × Free: Simple and correct. Tech. Rep. 15, Dipartimento di Matematica, Università di Padova. Dec.
42
R. Bagnara, E. Zaffanella, and P. M. Hill
Furukawa, K., Ed. 1991. Logic Programming: Proceedings of the Eighth International
Conference on Logic Programming. MIT Press Series in Logic Programming. The MIT
Press, Paris, France.
Hans, W. and Winkler, S. 1992. Aliasing and groundness analysis of logic programs
through abstract interpretation and its safety. Tech. Rep. 92–27, Technical University
of Aachen (RWTH Aachen).
Hermenegildo, M. V. and Greene, K. J. 1990. &-Prolog and its performance: Exploiting independent And-Parallelism. In Logic Programming: Proceedings of the Seventh
International Conference on Logic Programming, D. H. D. Warren and P. Szeredi, Eds.
MIT Press Series in Logic Programming. The MIT Press, Jerusalem, Israel, 253–268.
Hermenegildo, M. V. and Rossi, F. 1995. Strict and non-strict independent andparallelism in logic programs: Correctness, efficiency, and compile-time conditions. Journal of Logic Programming 22, 1, 1–45.
Hill, P. M., Bagnara, R., and Zaffanella, E. 1998. The correctness of set-sharing. In
Static Analysis: Proceedings of the 5th International Symposium, G. Levi, Ed. Lecture
Notes in Computer Science, vol. 1503. Springer-Verlag, Berlin, Pisa, Italy, 99–114.
ISO/IEC. 1995. ISO/IEC 13211-1: 1995 Information technology — Programming languages — Prolog — Part 1: General core. International Standard Organization.
Jacobs, D. and Langen, A. 1989. Accurate and efficient approximation of variable
aliasing in logic programs. In Logic Programming: Proceedings of the North American
Conference, E. L. Lusk and R. A. Overbeek, Eds. MIT Press Series in Logic Programming. The MIT Press, Cleveland, Ohio, USA, 154–165.
Jacobs, D. and Langen, A. 1992. Static analysis of logic programs for independent
AND parallelism. Journal of Logic Programming 13, 2&3, 291–314.
Janssens, G. and Bruynooghe, M. 1992. Deriving descriptions of possible values of
program variables by means of abstract interpretation. Journal of Logic Programming 13, 2&3, 205–258.
King, A. 1993. A new twist on linearity. Tech. Rep. CSTR 93-13, Department of Electronics and Computer Science, Southampton University, Southampton, UK.
King, A. 1994. A synergistic analysis for sharing and groundness which traces linearity.
In Proceedings of the Fifth European Symposium on Programming, D. Sannella, Ed.
Lecture Notes in Computer Science, vol. 788. Springer-Verlag, Berlin, Edinburgh, UK,
363–378.
King, A. and Soper, P. 1994. Depth-k sharing and freeness. In Logic Programming: Proceedings of the Eleventh International Conference on Logic Programming, P. Van Hentenryck, Ed. MIT Press Series in Logic Programming. The MIT Press, Santa Margherita
Ligure, Italy, 553–568.
Langen, A. 1990. Advanced techniques for approximating variable aliasing in logic programs. Ph.D. thesis, Computer Science Department, University of Southern California.
Printed as Report TR 91-05.
Mulkers, A., Simoens, W., Janssens, G., and Bruynooghe, M. 1994. On the practicality of abstract equation systems. Report CW 198, Department of Computer Science,
K. U. Leuven, Leuven, Belgium.
Mulkers, A., Simoens, W., Janssens, G., and Bruynooghe, M. 1995. On the practicality of abstract equation systems. In Logic Programming: Proceedings of the Twelfth
International Conference on Logic Programming, L. Sterling, Ed. MIT Press Series in
Logic Programming. The MIT Press, Kanagawa, Japan, 781–795.
Muthukumar, K. and Hermenegildo, M. V. 1991. Combined determination of sharing
and freeness of program variables through abstract interpretation. See Furukawa (1991),
49–63. An extended version appeared in (Muthukumar and Hermenegildo 1992).
Enhanced Sharing Analysis Techniques: A Comprehensive Evaluation
43
Muthukumar, K. and Hermenegildo, M. V. 1992. Compile-time derivation of variable
dependency using abstract interpretation. Journal of Logic Programming 13, 2&3, 315–
347.
Santos Costa, V., Srinivasan, A., and Camacho, R. 2000. A note on two simple
transformations for improving the efficiency of an ILP system. In Inductive Logic Programming: Proceedings of the 10th International Conference, ILP 2000, J. Cussens and
A. Frisch, Eds. Lecture Notes in Computer Science, vol. 1866. Springer-Verlag, Berlin,
London, UK, 397–412.
Scozzari, F. 2000. Abstract domains for sharing analysis by optimal semantics. In Static
Analysis: 7th International Symposium, SAS 2000, J. Palsberg, Ed. Lecture Notes in
Computer Science, vol. 1824. Springer-Verlag, Berlin, Santa Barbara, CA, USA, 397–
412.
Søndergaard, H. 1986. An application of abstract interpretation of logic programs:
Occur check reduction. In Proceedings of the 1986 European Symposium on Programming, B. Robinet and R. Wilhelm, Eds. Lecture Notes in Computer Science, vol. 213.
Springer-Verlag, Berlin, Saarbrücken, Federal Republic of Germany, 327–338.
Zaffanella, E., Bagnara, R., and Hill, P. M. 1999. Widening Sharing. In Principles
and Practice of Declarative Programming, G. Nadathur, Ed. Lecture Notes in Computer
Science, vol. 1702. Springer-Verlag, Berlin, Paris, France, 414–431.
Zaffanella, E., Hill, P. M., and Bagnara, R. 1999. Decomposing non-redundant
sharing by complementation. In Static Analysis: Proceedings of the 6th International
Symposium, A. Cortesi and G. Filé, Eds. Lecture Notes in Computer Science, vol. 1694.
Springer-Verlag, Berlin, Venice, Italy, 69–84.
| 6cs.PL
|
Path Computation in Multi-Layer Multi-Domain
Networks: A Language Theoretic Approach
arXiv:1512.06532v1 [cs.DS] 21 Dec 2015
Mohamed Lamine Lamalia,∗, Hélia Pouyllaua , Dominique Barthb
b Lab.
a Alcatel-Lucent Bell Labs, Route de Villejust, 91620 Nozay, France
PRiSM, UMR8144, Université de Versailles, 45, av. des Etas-Unis, 78035 Versailles
Cedex, France
Abstract
Multi-layer networks are networks in which several protocols may coexist at
different layers. The Pseudo-Wire architecture provides encapsulation and decapsulation functions of protocols over Packet-Switched Networks. In a multidomain context, computing a path to support end-to-end services requires the
consideration of encapsulation and decapsulation capabilities. It appears that
graph models are not expressive enough to tackle this problem. In this paper,
we propose a new model of heterogeneous networks using Automata Theory. A
network is modeled as a Push-Down Automaton (PDA) which is able to capture
the encapsulation and decapsulation capabilities, the PDA stack corresponding
to the stack of encapsulated protocols. We provide polynomial algorithms that
compute the shortest path either in hops or in the number of encapsulations
and decapsulations along the inter-domain path, the latter reducing manual
configurations and possible loops in the path.
Keywords: Multi-layer networks, Pseudo-Wire, Push-Down Automata
1. Introduction
Most carrier-grade networks comprise multiple layers of technologies (e.g.
Ethernet, IP, etc.). These layers are administrated by different control and/or
management plane instances. The Pseudo-Wire (PWE3) architecture [2] unifies
control plane functions in heterogeneous networks to allow multi-layer services
(e.g. Layer 2 VPN). To this end, it defines encapsulation (a protocol is encapsulated into another) and decapsulation (a protocol is unwrapped from another)
functions, called adaptation functions further in this paper. These functions
∗ Corresponding
author
Email addresses: [email protected] (Mohamed Lamine
Lamali), [email protected] (Hélia Pouyllau),
[email protected] (Dominique Barth)
c 2013. Licensed under the Creative Commons CC-BY-NC-ND 4.0 license
http://creativecommons.org/licenses/by-nc-nd/4.0/
c Published version: http: // dx. doi. org/ 10. 1016/ j. comcom. 2012. 11. 009
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
allow the emulation of services (e.g. Frame Relay, SDH, Ethernet, etc.) over
Packet-Switched Networks (PSN, e.g. IP or MPLS).
Prior to a service deployment in a multi-layer network, the resources must be
identified during the path computation process. The path computation must
take into account the adaptation function capabilities in order to explore all
resources and to ensure the end-to-end service deployment. The authors of [1]
defined the multi-segment Pseudo-Wire architecture for multi-domain networks.
In [4], the authors mention the problem of path determination over such an
architecture, stressing the importance of having path computation solutions.
In such an architecture, the path computation should comply with protocol
compatibility constraint: if a protocol is encapsulated in a node, it must be
decapsulated in another node; the different encapsulation processes should be
transparent to the source and target nodes. Thus, some nodes may be physically
connected but, due to protocol incompatibility, no feasible path (i.e., which comply with protocol compatibility) can be found between them. This constraint
leads to non trivial characteristics of a shortest path [7]: i) it may involve loops
(involving the same link several times but with different protocols); ii) its subpaths may not be feasible. Computing such a path is challenging and cannot
be performed by classical shortest path algorithms.
Currently, the configuration of these functions is manually achieved within
each network domain: when an encapsulation function is used, the corresponding decapsulation function is applied within the domain boundaries. In largescale carrier-grade networks or in multi-domain networks, restricting the location of the adaptation functions to the domain boundaries might lead to ignore
feasible end-to-end paths leading to a signaling failure. Hence, in the path
computation process, it must be possible to nest several encapsulations (e.g.
Ethernet over MPLS over SDH). A decapsulation should not be restricted to
the same domain as its corresponding encapsulation. This allows the exploration
of more possible paths and new resources in the path computation.
The problem we address is to compute the shortest feasible path either in
the number of nodes or in the number of involved (and possibly nested) adaptation functions. The latter is motivated by two goals: i) as such functions
are manually configured on router interfaces, minimizing their number would
simplify the signaling phase when provisioning the path; ii) as our algorithms
do not allow loops without adaptation functions (loops without encapsulations
or decapsulations are useless and can be deleted), reducing the number of adaptation functions leads to reducing the number of loops. Reducing the number
of loops is important because if there are several loops involving the same link,
the available bandwidth on this link may be exceeded.
The authors of [14, 6] focused on the problem of computing a path in a multilayer network under bandwidth constraints. In [16], we demonstrated that the
problem under multiple Quality of Service constraints is NP-Complete. In this
paper we demonstrate that the problem without QoS constraints is polynomial,
and we provide algorithms to compute the shortest path. These algorithms use
a new model of multi-layer networks based on Push-Down Automata (PDAs).
The encapsulation and decapsulation functions are designed as pushes and pops
2
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
in a PDA respectively, the PDA stack allowing to memorize the nested protocols.
If the goal is to minimize the number of adaptation functions, the PDA is transformed in order to bypass sub-paths without adaptation functions. The PDA or
transformed PDA is then converted into a Context-Free Grammar (CFG) using
a method of [13]. A shortest word, either corresponding to the shortest path in
nodes or in adaptation functions, is generated from this CFG.
This paper extends the work published in [15], providing the detailed proofs
of the correctness of the algorithms and including new algorithms (transforming
the PDA, converting the PDA to a CFG, generating the shortest word). Furthermore, the complexity analyses are refined and the total worst case complexity
of the path computation is significantly reduced.
This paper is organized as follows: Section 2 recalls the context of multilayer multi-domain networks and the related work on path computation in such
networks and presents the proposed approach; Section 3 provides a model of
multi-layer multi-domain networks and a formal definition of the problem; Section 4 explains how a network is converted into a PDA and provides the complexity of this transformation; finally, Section 5 gives the different methods that
compute the shortest path in nodes or in encapsulations/decapsulation.
In order to ease the paper reading, a table summarizing the used notations
is provided in Appendix A.
2. Path computation in Pseudo-Wire networks
Some standards define the emulation of lower layer protocols over a PSN
(e.g. Ethernet over MPLS, [19], Time-Division Multiplexing (TDM) over IP,
[21]). For instance, one node in the network encapsulates the layer 2 frames in
layer 3 packets and another node unwraps them. This allow to cross a part of
the network by emulating a lower layer protocol, and thus to overcome protocol
incompatibilities.
The PWE3 architecture [2], as well as the multi-layer networking description
of [20], assumes an exhaustive knowledge of the network states. This assumption
is not valid at the multi-domain scale. Thus, the authors of [1] defines an
architecture for extending the Pseudo-Wire emulation across multiple PSNs
segments. The authors of [4] stress the importance of path determination in
such a context and suggest it to be an off-line management task. They also
suggest to use the Path Computation Element architecture (PCE) [8], which is
adapted to the multi-domain context. It could be a control plane container for
solution detailed in this paper. It would require protocol and data structure
extensions in order to add encapsulation/decapsulation capabilities in the PCE
data model.
2.1. Related work on path computation in multi-layer networks
The problem of path computation in heterogeneous networks raised first at
the optical layer. Due to technology incompatibilities (different wavelengths,
different encodings, etc.), it soon became clear that classical graph models cannot capture these incompatibilities thus forbidding classical routing algorithms
3
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
to tackle this problem. The authors of [3] propose a Wavelength Graph Model
instead of the classical network graph models to resolve the problem of wavelength continuity. The authors of [23] propose an Auxiliary Graph Model to
resolve the problems of traffic grooming and wavelength continuity in heterogeneous WDM mesh networks, this model is later simplified in [22]. The common
feature of these works is that they model each physical device as several nodes,
each node being indexed by a technology. The existence of an edge depends
on the existence of a physical link, but also on the technology compatibility.
In [11], the authors take into account the compatibility constraints on several
layers: wavelength continuity, label continuity, etc. They propose a Channel
Graph Model to resolve the multi-layer path computation problem. The proposed models and algorithms take into account protocol and technology but
ignore the encapsulation/decapsulation (adaptation) capabilities. As they do
not have a stack to store the encapsulated protocols, they cannot model the
PWE3 architecture.
In the PWE3 architecture, the compatibility between two technologies on a
layer depends also on the encapsulated protocols on the lower layer. In [6], the
authors addressed the problem of computing the shortest path in the context
of the ITU-T G.805 recommendations on adaptation functions. They stress the
lack of solutions on path selection and the limitations of graph theory to handle
this problem. The authors of [7] present the specificities of a multi-layer path
computation taking into account the encapsulation/decapsulation (adaptation)
capabilities: the shortest path can contain loops and its sub-paths may not satisfy the compatibility constraints. They provide an example of topology where
classical routing algorithms cannot find the shortest path because it contains a
loop. In [14], the authors addressed the same problem in a multi-layer network
qualifying it as an NP-Complete problem. The NP-Completeness comes from
the problem definition as they consider that the loops in the path can exceed
the available bandwidth, because if the path involves the same link several times
it may overload this link. They aim to select the shortest path in nodes and
provide new graph models allowing to express the adaptation capabilities. They
propose a Breadth-First Search algorithm which stores all possible paths and
selects the shortest one. This algorithm has an exponential time complexity.
In the problem we consider, we exclude bandwidth constraints and propose a
solution for minimizing the number of encapsulations and decapsulations. Our
algorithm does not allow loops without adaptation functions, the only loops
that may exist involve encapsulations or decapsulations. Thus, minimizing the
number of adaptation functions in the path also leads to minimizing the number
of loops - and avoiding them if a loop-free feasible path with less encapsulations
exists.
2.2. Proposed approach
In this work, we propose a new multi-domain multi-layer network model
wich takes into account encapsulation and decapsulation capabilities. To the
best of our knowledge no previous work has considered this problem at the
multi-domain scale. It induces to go beyond the domain boundaries allowing
4
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
multi-domain compatibility to determine a feasible inter-domain path: when an
encapsulation for a given protocol is realized in one domain its corresponding
decapsulation must be done in another. It appears that PDAs are naturally
adapted to model the encapsulation and decapsulation capabilities, as push and
pop operations easily model encapsulations and decapsulations, and the PDA
stack can model the stack of encapsulated protocols. By using powerful tools
of Automata and Language Theory, we propose polynomial algorithms that
generate the shortest sequence of protocols of a feasible path. This sequence
allows to find the shortest feasible path.
Furthermore, we consider two kind of objectives: either the well-known objective of minimizing the number of hops or the objective of minimizing the
number of adaptation functions. The latter is motivated by the fact that it is
equivalent to minimize the number of configuration operations, which are often
done manually and can be quite complex. It is also motivated by reducing the
number of possible loops (as the number of adaptation functions involved in a
path is correlated with the number of loops), and thus avoiding to use the same
link several times and to exhaust the available bandwidth on it.
Figure 1 summarizes our proposed approach. It presents the different models
leading to the shortest path and the algorithms computing them.
I. Convert a multi-domain Pseudo-Wire network into a PDA,
i. If the goal is to minimize the number of adaptation functions, transform the PDA to bypass the “passive” functions (i.e. no protocol
adaptation),
ii. Else let the PDA as is,
II. Derive a CFG from the PDA or the transformed PDA,
III. Determine the “shortest” word generated by the CFG,
IV. Identify the shortest path from the shortest word.
Figure 1: Proposed approach to compute the shortest feasible path.
5
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Compared to the preliminary version of this work [15], we detail the proofs
of correctness and refine the complexity of our algorithms. We also provide
the detailed algorithm which converts the PDA to a CFG, and we propose a
new method to generate the shortest word which is linear in the length of the
shortest word.
Appendix A summarizes the notations used in this paper.
3. Multi-layer multi-domain network model
A multi-domain network having routers with encapsulation/decapsulation
capabilities can be defined as a 3-tuple: a directed graph G = (V, E) modeling
the routers of a multi-domain network, we consider a pair of vertices (S, D) in G
corresponding to the source and the destination of the path we focus on; a finite
alphabet A = {a, b, c, . . . } in which each letter is a protocol; an encapsulation
or a decapsulation function is a pair of different letters in the alphabet A:
• Figure 2(a) illustrates the encapsulation of the protocol x by the node U
in the protocol y;
• Figure 2(b) illustrates that the protocol x is unwrapped by the node U
from the protocol y;
• Figure 2(c) illustrates that the protocol x transparently crosses the node
U (no encapsulation or decapsulation function is applied). Such pairs are
referred as passive further in this paper.
We denote by ED and by ED the set of all possible encapsulation functions and
decapsulation functions respectively, (i.e., ED ⊂ A2 ). A subset P (U ) of ED∪ED
indicates the set of encapsulation, passive and decapsulation functions supported
by vertex U ∈ V. We define In(U ) = {a ∈ A s.t. ∃b ∈ A s.t. (a, b) or (b, a) ∈
P (U )} and Out(U ) = {b ∈ A s.t. ∃a ∈ A s.t. (a, b) or (b, a) ∈ P (U )}. The set
T (U ) = {a ∈ A s.t. (a, a) ∈ P (U )} is the set of protocols that can passively
cross the node U .
(a) Encapsulation
of
protocol x in protocol y,
(x, y) ∈ P (U )
(b) Decapsulation of protocol x from protocol y,
(x, y) ∈ P (U )
(c)
Passive
crossing,
(x, x) ∈ P (U )
Figure 2: Different transitions when a protocol crosses a node U
6
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Considering a network N = (G = (V, E), A, P ), a source S ∈ V, a destination D ∈ V and a path C = S, x1 , U1 , x2 , . . . , Un−1 , xn , D where each Ui is a
vertex in V and each xi ∈ A ∪ A (where A = {a : a ∈ A}).
• TC = x1 . . . xn represents the sequence of protocols which is used over the
path C. It is called the trace of C. For each xi :
– xi = a and xi+1 = b or b, means that Ui encapsulates the protocol a
in b ( a, b, b ∈ A ∪ A)
– xi = a and xi+1 = b or b means that Ui unwraps the protocol b from
a
– xi = a and xi+1 = a or a means that Ui passively transports a
• The transition sequence of C, denoted HC , is the sequence β1 , . . . , βn
obtained from C s.t. for i = 1..n:
– if xi = a ∈ A and xi+1 = b ∈ A or xi+1 = b ∈ A then βi = (a, b)
– if xi = b ∈ A and xi+1 = a ∈ A\{b} or xi+1 = a ∈ A\{b} then
βi = (a, b)
Note that the pair (a, a), a ∈ A can appear in a transition sequence,
as it represents a passive function. However, according the definition
above, a pair (a, a), a ∈ A cannot appear, as it is forbidden to
encapsulate (and thus to decapsulate) a protocol a in itself.
0
• The well-parenthesized sequence of C, denoted MC = β10 , . . . , βm
, is obtained from HC by deleting each passive transition βi s.t. βi = (a, a) and
a∈A
Example. Consider the path C = S, a, U, b, V, b, W, a, D in the network illustrated by Fig. 3(a). The transition sequence corresponding to C is HC =
(a, b), (b, b), (a, b) and its trace is TC = abba. The well-parenthesized sequence
from C is MC = (a, b), (a, b).
Let denotes the empty word, “•” indicates the concatenation operation,
and HC denotes the transition sequence obtained from a path C as explained
above. The following definitions give a formal characterization of the feasible
paths.
Definition 1. A sequence MC from HC is valid if and only if MC ∈ L, where
L is the formal language recursively defined by:
[
(x, y) • L • (x, y) • L
L=∪
(x,y)∈A2
Definition 2. A path C is a feasible path in N from S to D if:
• S, U1 , . . . , Un−1 , D is a path in G = (V, E),
7
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
• The well-parenthesized sequence MC from C is valid,
• For each i ∈ {1, . . . , n}:
– if xi = a and xi+1 = b or b then (a, b) ∈ P (Ui )
– if xi = a and xi+1 = b or b then (a, b) ∈ P (Ui )
– if xi = a and xi+1 = a or a then a ∈ T (Ui )
The language L of valid sequences is known as the generalized Dyck language
[17]. It is well-known that this language is context free but not regular. Thus,
push-down automata are naturally adapted to model this problem.
Example. The multi-domain network illustrated by Fig. 3(a) has 6 routers and
two protocols labeled a and b. Adaptation function capabilities are indicated
below each node. For example, the node U can encapsulate the protocol a in the
protocol b (function denoted by (a, b)) or can passively transmit the protocol
a (function denoted by (a, a)). In this multi-domain network, the only feasible
path between S and D is S, a, U, b, V, b̄, W, a, D and involves the encapsulation
of the protocol a in the protocol b by the node U , the passive transmission of
the protocol b by the node V and the decapsulation of the protocol a from b by
the node W . (functions (a, b), (b, b) and (a, b) respectively).
Problem definition. As explained above, our goal is to find a feasible multidomain path. Furthermore, we set as an objective function either the size of
the sequence of adaptation functions or the size of the path in number of nodes.
Hence, the problem we aim to solve can be defined as follows:
min |HC | or |MC |
C
s.t. C is a feasible path
4. From the network model to a PDA
In this section, we address the conversion from a network to a PDA. Algorithm 1 takes as input a network N = (G = (V, E), A, P ) and converts it to a
PDA AN = (Q, Σ, Γ, δ, SA , Z0 , {DA }), where Q is the set of states of the PDA,
Σ the input alphabet, Γ the stack alphabet, δ the transition relation, SA the
start state, Z0 the initial stack symbol and DA the accepting state, is the
empty string. The automaton AN from N is obtained as follows:
• Create a state Ux of the automaton for each U ∈ V and each x ∈ In(U ),
except for the source node S for which a single state SA is created,
• The top of the stack is the last encapsulated protocol,
• If the current state is Ux then the current protocol is x,
8
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
• The adaptation functions will be converted into transitions in the PDA.
A transition t ∈ δ is denoted (Ux , hx, α, βi, Vy ) where Ux ∈ Q is the state
of the PDA before the transition, x ∈ Σ is the input character, α ∈ Γ is
the top of the stack before the transition (it is popped by the transition),
β ∈ Γ∗ is the sequence of symbols pushed to the stack, and Vy is the
state of the PDA after the transition. Thus if β = α then t is a passive
transition, if β = xα then t is a push of x, and if β = ∅ then t is a pop,
• A passive transition of a protocol x across a node U is modeled as a
transition without push or pop between the state Ux and the following
state Vx . It is denoted (Ux , hx, α, αi, Vx ),
• An encapsulation of a protocol x in a protocol y by a node U is modeled
as a push of the character x in the stack between the state Ux and the
following state Vy . It is denoted (Ux , hx, α, xαi, Vy )2 ,
• A decapsulation of y from x by a node U is modeled as a pop of the
protocol y from the stack. It is denoted (Ux , hx, y, ∅i, Vy ).
Algorithm 1 Convert a network to a PDA
Input: a network N = (G = (V, E), A, P ), a source S and a destination D
Output: push-down automaton AN = (Q, Σ, Γ, δ, SA , Z0 , {DA })
(1) Σ ← A ∪ A ; Γ ← A ∪ {Z0 }
(2) Create a single state SA for the node S
(3) For each node U 6= S and each protocol x ∈ In(U ), create a state Ux
(4) For each state Ux s.t. (S, U ) ∈ E and x ∈ Out(S)
Create a transition (SA , h, Z0 , Z0 i, Ux )
(5) For each link (U, V ) ∈ E s.t. U 6= S and for each (x, y) ∈ P (U ) and each
α∈Γ\{x}
(5.1) If x ∈ T (U ) ∩ In(V ) Create a transition (Ux , hx, α, αi, Vx ){passive
trans.}
(5.2) If x 6= y and y ∈ In(V ) Create a transition (Ux , hx, α, xαi, Vy ){encap.}
(6) For each link (U, V ) ∈ E s.t. U 6= S and for each (y, x) ∈ P (U )
(6.1) If x ∈ In(V ) Create a transition (Ux , hx, y, ∅i, Vy ){decap.}
(7) Create a fictitious final state DA .
(8) For each x ∈ In(D) and each α ∈ Γ\{x} Create a transition
(Dx , hx, Z0 , ∅i, DA )
Complexity. Each node U from the graph generates |In(U )| states, except the
source node S. A fictitious final state is added. Thus, the number of states is
2 Note that, even if x = a ∈ A, the transition has the form (U , ha, α, aαi, V ). Characters
a
y
in A are only used as input characters. Characters indexing states and pushed characters in
the stack are their equivalent in A.
9
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
at worst 2 + (|V| − 1) × |A| so in O(|V| × |A|). The worst case complexity of
algorithm 1 is in O(max((|V| × |A|), (|E| × ((|A| × |ED|) + |ED|))))). We assume
that the network is connected, so |E| ≥ |V| − 1. Since ED is a subset of A2 ,
then |ED| < |A|2 and |ED| < |A|2 . Thus, the upper bound complexity is in
O(|A|3 × |E|), which is also an upper bound for the number of transitions.
Proposition 1. Considering a network N = (G = (V, E), A, P ), a source S ∈
V and a destination D ∈ V, the language recognized by AN is the set of traces
of the feasible paths from S to D in N .
Proof. Let a feasible path C = Sx1 U1 x2 . . . xi Ui xi+1 . . . Un−1 xn D in a network N and its trace TC = x1 . . . xi xi+1 . . . xn leading to a valid sequence MC .
We aim to demonstrate that: i) TC is recognized by AN and ii) U1 , . . . , Un−1
is a path in N and iii) There exists a feasible path corresponding to TC in
N if TC is recognized by the PDA AN . It is sufficient to show that C =
S, x1 , U1 , x2 , . . . , Un−1 , xn , D is a feasible path in N if and only if TC is recognized by AN (i.e the final state is reached and the stack is empty).
From C we deduce the following path in AN defined as a sequence of transitions. This path begins with transitions: (SA , h, Z0 , Z0 i, (Ui )x1 ) for each Ui
s.t. x1 ∈ In(Ui ). Hence, transition (SA , h, Z0 , Z0 i, (U1 )x1 ) belongs to this set.
Then, the path in AN follows the order induced by TC , for each element in
x1 , . . . , xn , we consider one or two transitions in AN as follows:
• if xi = a then
– if xi+1 = a or xi+1 = a consider the transition ((Ui )a , ha, α, αi, (Ui+1 )a )
where α ∈ Γ\{a},
– if xi+1 = b or xi+1 = b consider the transition ((Ui )a , ha, α, aαi, (Ui+1 )b ),
where α ∈ Γ\{a}
• else (i.e., if xi = a) then consider the transition ((Ui )a , ha, b, ∅i, (Ui+1 )b ).
Note that, if i = n then Ui+1 = D.
It is clear that this sequence of transitions is a path in AN recognizing the
trace TC . Since MC is a well parenthesized word, if xi = a then the head
of the stack when reaching state (Ui )xi contains aZ0 and when reaching state
(Ui+1 )xi+1 , the head of the stack is reduced to Z0 . Then, when reaching state
DA , the stack is empty.
Conversely, we can show with a similar construction that any well recognized
word on a path in AN induces one unique trace TC of feasible paths in G from
S to D.
2
Example. Figure 3(b) is an example of output of algorithm 1. The algorithm
transforms the network illustrated by Fig. 3(a) into a PDA of Fig 3(b). For
instance, the link (U, V ) is transformed into the transitions (Ua , ha, Z0 , aZ0 i, Vb )
and (Ua , ha, b, abi, Vb ) (pushes) because the node U should encapsulate the protocol a in the protocol b using the adaptation function (a, b) before transmitting to the node V . The link (W, D) is transformed into the transitions
10
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
(Wb , hb, Z0 , bZ0 i, Da ) and (Wa , hb, a, bai, Da ) (pushes) according to the adaptation function (a, b). It is also transformed into the transition (Wb , hb, a, ∅i, Da )
(pop) because the node W can decapsulate the protocol a from b using the
adaptation function (a, b) before transmitting to the node D. The link (V, W ) is
transformed into the transitions (Vb , hb, Z0 , Z0 i, Wb ) and (Vb , hb, a, ai, Wb ) (passive transitions) according the capability of the node V (adaptation function
(b, b)).
5. The shortest feasible path
In section 4, we provided a method to build a PDA allowing to determine
the feasible paths. The next step is to minimize either the number of nodes or
the number of adaptation functions. The method to minimize the number of
nodes uses directly the PDA as described in section 5.1. But to minimize the
number of adaptation functions, the PDA is transformed in order to bypass the
sub-paths without any adaptation function, as detailed in section 5.2. Then, a
CFG derived from the PDA (or the transformed PDA) generates words whose
length is equivalent to the number of nodes (or to the number of adaptations).
An algorithm browses the CFG to determine the shortest word. Finally, another
algorithm identifies the multi-domain path corresponding to this shortest word.
5.1. Minimizing the number of nodes
The number of characters in a word accepted by the automaton AN is the
number of links in the corresponding feasible path (each character is a protocol
used on a link). Thus the step of automaton transformation (section 5.2) should
be skipped. The automaton AN is directly transformed into a CFG, then the
shortest word is generated as described in section 5.3.2. The corresponding
feasible path is computed by algorithm 6 described in section 5.3.3.
5.2. Minimizing the number of adaptation functions
To enumerate only encapsulations and decapsulations in the length of each
word (and thus minimize adaptation functions by finding the shortest word
accepted), a transformed automaton A0N in which all sequences involving passive
transitions are bypassed must be determined. The length of the shortest word
accepted by A0N is the number of adaptation functions plus a fixed constant.
Let us define Qa (a ∈ A) as Qa = {Vx ∈ Q|x = a}, and let AaN be the subautomaton induced by Qa . By analogy with an induced subgraph, an induced
sub-automaton is a multigraph with labeled edges such that the set of vertices
is Qa and the set of edges is the set of transitions between elements of Qa . Since
there are only passive transitions between two states in Qa , all paths in the subautomaton are passive. Let define P (Ux , Vx ) as the shortest path length between
Ux and Vx . This length can be computed between all pairs of nodes in Qa using
the Floyd-Warshall algorithm. Let Succ(Vx ) be the set of successors of Vx in
the original automaton AN , i.e., Succ(Vx ) = {Wy ∈ Q|∃(Vx , hx, α, βi, Wy ) ∈ δ}.
11
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Algorithm 2 takes as input AN and computes the transformed automaton
A0N = (Q0 , Σ0 , Γ0 , δ 0 , SA , Z0 , {DA }). A0N is initialized with the values of AN .
Then, the algorithm computes the sub-automaton for each character x ∈ A
(step (1)) and the length values P (Ux , Vx ) for each pair of states in the subautomaton (step (2)). Each path between a pair of states is a sequence of passive
transitions. If such a path exists (step (3.1)), the algorithm adds transitions to
δ 0 from Ux to each state in Succ(Vx ) (steps (3.1.2) and (3.1.3)). These added
transitions are the same that those which connect Vx to its successors Wy , but
with an input character indexed by the number of passive transitions between Ux
and Vx , (i.e., P (Ux , Vx )) plus one (indicating that there is a transition sequence
which matches with a sequence of protocols xx . . . x of length P (Ux , Vx ) + 1).
The indexed character is added to the input alphabet Σ0 (step (3.1.1)).
Algorithm 2 Transform automaton AN
Input: push-down automaton AN = (Q, Σ, Γ, δ, SA , Z0 , {DA })
Output:
transformed
push-down
automaton
A0N
=
0
0
0 0
(Q , Σ , Γ , δ , SA , Z0 , {DA })
Q0 ← Q, Σ0 ← Σ, Γ0 ← Γ, δ 0 ← δ
For each x ∈ A
(1) Compute AxN
(2) Compute the distance P (Ux , Vx ) between all pairs of states Ux and Vx
in AxN
(3) For each Ux ∈ Qx and each Vx in Qx
(3.1) If P (Ux , Vx ) < ∞ {there is a path between Ux and Vx }
(3.1.1) Add xP (Ux ,Vx )+1 and xP (Ux ,Vx )+1 to Σ
(3.1.2) For each Wy ∈ Succ(Vx )\{Ux } and each (Vx , hx, α, βi, Wy ) ∈ δ
Add the transition (Ux , hxP (Ux ,Vx )+1 , α, βi, Wy ) to δ 0
(3.1.3) For each Wy ∈ Succ(Vx )\{Ux } and each (Vx , hx, α, βi, Wy ) ∈ δ
Add the transition (Ux , hxP (Ux ,Vx )+1 , α, βi, Wy ) to δ 0
Complexity. Computing the sub-automaton (step (1)) is done in O(|δ| + |Q|)
by checking all the nodes and all the transitions. Computing the shortest path
length between all pairs of states is done by the Floyd-Warshall algorithm in
O(|Qx |3 ). If x=y, there are at worst |A| − 1 transitions between Vx and Wy in Q
(transitions in the form (Ux , hx, α, αi, Wx ) for all possible values of α except x
— as the construction of the automaton forbids the encapsulation of a protocol
x in x, and thus it is impossible to have x as current protocol and at the top of
the stack in the same time). If x 6= y, there are at worst |A| transitions between
Vx and Wy in Q (the pop (Vx , hx, y, ∅i, Wy ) and the pushes (Vx , hx, α, xαi, Wy )
for all possible values of α except x). And as |Succ(Vx )| < |Q|, the steps
(3.1.2) and (3.1.3) are bounded by O(|Q| × |A|). However, a state belongs to
only one sub-automaton, the complexity of algorithm 2 is therefore bounded by
12
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
P
3
2
O
x∈A |δ| + |Q| + |Qx | + (|Qx | × |Q| × |A|) . And ∀x ∈ A, |Qx | ≤ |A|
(see algorithm 1 for the construction of the set of states Q). Thus, the complex
ity of algorithm 2 is in O max |A| × (|δ| + |Q|), |A| ×|V|3 , |A|2 × |V|2 × |Q| ,
which corresponds to O max(|A|4 × |E|, |A|3 × |V|3 ) in the network model.
P
2
The number of transitions in δ 0 is bounded by |δ|+O
x∈A (|Qx | × |Q| × |A|) ,
which corresponds to O(|A|3 × |V|3 ) in the network model.
Example. Algorithm 2 transforms the PDA in Fig. 3(b) into the PDA in
Fig. 3(c). For the protocol b, the algorithm computes the sub-automaton induced by the states Vb , Wb and Kb . The distance P (Vb , Wb ) = 1 is then computed. Thus, to bypass the sequence of transitions (Vb , hb, a, ai, Wb ) (Wb , hb, a, ∅i, Da ),
the transition (Vb , hb2 , a, ∅i, Da ) is added.
Let L(AN ) be the set of words accepted by AN , and let L(A0N ) be the set of
words accepted by A0N . Let f : Σ0 → Σ∗ be a function s.t.:
• if xi = ai ∈ A0 then f (xi ) = aa
. . aa}
| .{z
i occurrences
• if xi = ai ∈ A0 then f (xi ) = |aa .{z
. . aa}
i occurrences
∗
The domain of f is extended to (Σ0 ) :
∗
f : (Σ0 ) → Σ∗
w0 = x1i x2j . . . xnk → f (w0 ) = f (x1i )f (x2j ) . . . f (xnk )
For simplicity, we consider that x and x1 are the same character. f (L(A0N ))
denotes the set of words accepted by A0N transformed by f (i.e. f (L(A0N )) =
{f (w0 ) s.t. w0 ∈ L(A0N )}).
It is clear that f is not a bijection (f (xi xj ) = f (xi+j )). So to operate the
transformation between L(AN ) and L(A0N ), we define g : Σ∗ → (Σ0 )∗ s.t. :
yy . . . y . . . |zz {z
. . . z} ∈ Σ∗ , g(w) = xi yj . . . zk .
for each w = xx
. . . x}
| {z
| {z }
i occurrences j occurrences
k occurrences
In other words, w0 = g(w) is the shortest word in (Σ0 )∗ s.t. f (w0 ) = w.
The following lemmas and proposition show that the set of words accepted
by the transformed automaton is ‘equivalent’ to the set of words accepted by
the original automaton. Equivalent in the meaning that, using f , each word
accepted by the transformed automaton can be associated to a word accepted
by the original one. In addition, the transformed automaton has a new propriety: there is a linear relation between the length of an accepted word and the
minimum number of pushes (and pops) involved to accept it. This propriety
allows to find the word requiring the minimum number of pushes accepted by
the original automaton. This is done by finding the shortest word accepted by
the transformed automaton.
13
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Lemma 1. f (L(A0N )), the set of words accepted by A0N and transformed by f ,
is equal to L(AN ), the set of words accepted by AN .
Proof. We prove this by double inclusion:
1. L(AN ) ⊆ f (L(A0N )) : ∀w ∈ L(AN ), f (w) = w (remind that we consider
that x = x1 ). So f (L(AN )) = L(AN ) (1). On the other hand, the construction of A0N by algorithm 2 does not delete any character, transition,
state or set of final states from AN . So L(AN ) ⊆ L(A0N ). And thus
f (L(AN )) ⊆ f (L(A0N )) (2)
By (1) and (2), L(AN ) ⊆ f (L(A0N )).
2. L(AN ) ⊇ f (L(A0N )) : Let w0 = x1i x2j . . . xnl be a word in L(A0N ) and let
t01 t02 . . . t0n+1 be a sequence of transitions accepting w0 . For each transition t0m = (Uxm−1 , hxkm−1 , α, βi, Wxm ) (1 < m < n) in this sequence, s.t.
t0m ∈ δ 0 \δ, there is a sequence of k − 1 passive transitions in AN (because the creation of t0m in A0N requires the existence of such a sequence
in AN ). This sequence begins from the state Uxm−1 and is followed by a
transition (Vxm−1 , hxm−1 , α, βi, Wxm ). Thus, this sequence in AN matches
with f (xm−1
) And for each t0m in A0N , either t0m ∈ δ so t0m matches with
k
m+1
xk
(if k = 1), or t0m ∈ δ 0 \δ and there is a sequence of transitions in
AN which matches with f (xm+1
) (if k > 1). And since, by definition,
k
f (w0 ) = f (x1i )f (x2j ) . . . f (xnl ), f (w0 ) ∈ L(AN ). Thus f (L(A0N )) ⊆ L(AN ).
2
A0
N
Let w0 be the shortest word accepted by A0N , let N bpush
(w0 ) be the minimum
number of pushes in a sequence of transitions accepting w0 in A0N , and let
A0N
N bpop
(w0 ) be the minimum number of pops in a sequence of transitions accepting
0
w in A0N .
Lemma 2. The length of the shortest path (sequence of transitions) accepting
A0N
A0N
A0N
A0N
w0 in A0N is 1 + N bpush
(w0 ) + N bpop
(w0 ) and N bpop
(w0 ) = N bpush
(w0 ) + 1.
m+1
Proof. First we prove that w0 cannot be in the form of . . . xm
. . . with
i xj
m
m+1
0
x =x
. Let Uxm be the state in which AN is before reading the character
0
m
xm
i . Let Vxm+1 be the state in which AN is after reading xi and before reading
m+1
0
xj . And let Wy be the state in which AN is after reading xm+1
. If xm =
j
m+1
m
x
, then, by construction, there is a transition (Uxm , hxi+j , α, βi, Wy ) (where
m+1
α, β is a pop of y or a push of xm if xm 6= y). So, if xm
is replaced by
i xj
m
0
0
xi+j in w , the word obtained is shorter then w and is also accepted by A0N .
Thus, the shortest word accepted by A0N is w0 = x1i x2j . . . xnk where xm 6= xm+1
for (1 ≤ m < n).
0
By construction of A0N , for each character xm
i in w , there is a push transition
m
m
m
0
m
(Uxm , hxi , α, x αi, Wy ) if xi ∈ A or a transition (Uxm , hxm
i , y, ∅i, Wy ) if xi ∈
0
0
A . So all transitions in the shortest sequence that accepts w are pops or
pushes, except the first transition from the initial state (SA , h, Z0 , Z0 i, Ux1 )
and the final pop (Vxn , hxnk , Z0 , ∅i, DA ). The number of other pops is equal to
14
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
the number of pushes (in order to have Z0 at the top of the stack before the
final transition).
2
N
Now let w be a word accepted by AN , let N bA
push (w) be the minimum number
N
of pushes in a sequence of transitions accepting w in AN , and let N bA
pop (w) be
the minimum number of pops in a sequence of transitions accepting w in AN .
Lemma 3. For each w ∈ L(AN ):
A0
N
N
• N bA
pop (w) = N bpop (g(w))
A0
N
N
• N bA
push (w) = N bpush (g(w))
Proof. Let t1 t2 . . . tn be the sequence of transitions accepting w in AN with
minimum number of pushes and pops. For each sequence ti ti+1 . . . tj of passive
transitions, followed by a push or a pop transition, there is a transition with
the same push or pop and the same input character indexed by j − i + 1. So
g(w) ∈ L(A0N ).
Let t01 t02 . . . t0n0 be the sequence of transitions with minimum pushes and pops
accepting g(w) in A0N . It is clear that f (g(w)) = w. So if there is a sequence
t001 t002 . . . t00n0 that accepts g(w) with less pops and pushes than t01 t02 . . . t0n0 , then
each t00i ∈ δ 0 \δ can be replaced by a sequence of passive transitions followed by
a pop or a push in AN . And theses sequences accept f (g(w)) (which is w) with
less pops and pushes than in the sequence t1 t2 . . . tn , which contradicts the fact
that it minimizes the number of pops and pushes to accept w.
2
Proposition 2. The word accepted by AN which minimizes the number of
pushes is f (w0 ), where w0 is the shortest word (i.e., with minimum number
of characters) accepted by A0N .
Proof. By lemma 2, the shortest word accepted by A0N minimizes the number
of pushes in A0N (since the number of pushes grows linearly as a function of the
length of the word). Let w0 be this word.
A0N
0
N
Suppose that there is a word w accepted by AN such that N bA
push (w) < N bpush (w ).
A0
A0
N
N
N
By lemma 3, N bpush
(g(w)) = N bA
push (w), which leads to N bpush (g(w)) <
A0
A0
N
N
N bpush
(w0 ), while w0 = arg min N bpush
. Ad absurdum, f (w0 ) is the word which
minimizes the number of pushes in AN .
2
5.3. The shortest path as a shortest word
In order to find the shortest word accepted by AN (resp. A0N ), the CFG GN
such that L(GN ) = L(AN ) (resp. L(A0N )) is computed. The shortest word in
L(GN ) is then generated.
15
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
5.3.1. From the PDA to the CFG.
The transformation of a PDA into a CFG is well-known. We adapted a
general method described in [13] to transform AN (resp. A0N ) into a CFG. The
output of algorithm 3 is a CFG GN = (N , Σ, [SG ], P) (resp. (N , Σ0 , [SG ], P))
where N is the set of nonterminals (variables), Σ (resp. Σ0 ) is the input alphabet, [SG ] is the initial symbol (initial nonterminal) and P is the set of production
rules. Except [SG ], nonterminals are in the form [U XV ] where U, V ∈ Q and
X ∈ Γ (resp. Q0 and Γ0 ). The demonstration of the correctness of this transformation is also in [13].
Algorithm 3 Converting a PDA to a CFG
Input: PDA AN = (Q, Σ, Γ, δ, SA , Z0 , {DA }) (resp. Transformed PDA A0N =
(Q0 , Σ0 , Γ0 , δ 0 , SA , Z0 , {DA }))
Output: a CFG GN = (N , Σ, [SG ], P) (resp. (N , Σ0 , [SG ], P))
(1) Create the nonterminal [SG ]
(2) For each state Ux ∈ Q
(2.1) create a nonterminal [SA Z0 Ux ] and a production [SG ] → [SA Z0 Ux ]
(3) For each transition (Ux , hx, α, βi, Vy )
(3.1) If β = ∅ (pop), create a nonterminal [Ux αVy ] and a production
[Ux αVy ] → x
(3.2) If β = α (passive transition), create for each state W ∈ Q (resp. Q0 )
(3.2.1) Nonterminals [Ux αW ] and [Vy αW ]
(3.2.2) A production [Ux αW ] → x[Vy αW ]
(3.3) If β = xα, x ∈ Γ (push), create for each states (W, W 0 ) ∈ Q2 (resp.
02
Q )
(3.3.1) Nonterminals [Ux αW 0 ], [Vy αW ] and [W αW 0 ]
(3.3.2) A production [Ux αW 0 ] → x[Vy xW ][W αW 0 ]
Complexity. The number of production rules in GN is bounded by 1 + |Q| +
(|δ| × |Q|2 ) (resp. 1 + |Q0 | + (|δ 0 | × |Q0 |2 ). As all the nonterminals (except [SG ])
are in the form [Ux αVy ] with Ux , Vy ∈ Q and α ∈ Γ, the number of non terminals
is bounded by O(|A| × |Q|2 ) (resp. O(|A| × |Q|2 )). The worst case complexity
of algorithm 3 is in O(|δ| × |Q|2 ) (resp. O(|δ 0 | × |Q0 |2 )). W.r.t. the definition
of the network, the upper bound is in O(|A|5 ×|V|2 ×|E|) (resp. O(|A|5 ×|V|5 )).
Example. This method transforms the PDA in Fig. 3(c) into a CFG. Figure 3(d) is a subset of production rules of the obtained CFG. This subset allows
generating the shortest trace of a feasible path in the network in Fig. 3(a). For
instance, the transition (Vb , hb2 , a, ∅i, Da ) gives the production rule [Vb aDa ] →
b2 . The transition (Ua , ha, Z0 , aZ0 i, Vb ) gives all the production rules [Ua Z0 X 0 ] →
a[Vb aX][XZ0 X 0 ] where X, X 0 ∈ Q0 , including the production [Ua Z0 DA ] →
a[Vb aDa ][Da Z0 DA ] .
Remark. There are two mechanisms of acceptance defined for PDAs: an input
word can be accepted either by empty stack (i.e., if the stack is empty after
reading the word) or by final state (i.e., if a final state is reached after reading the
16
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
word - even if the stack is not empty). Algorithm 3 takes as input an automaton
which accepts words by empty stack. AN and A0N accept words by empty stack
and by final state, because the only transitions that empty the stack are those
that reach the final state (the transitions in the form (Dx , hx, Z0 , ∅i, DA ), x ∈
In(D)). Thus, Algorithm 3 performs correctly with AN or A0N as input.
5.3.2. The shortest word generated by a CFG.
To find the shortest word generated by GN , a function ` associates each
nonterminal to the length of the shortest word that it generates.
∗
∗
More formally, ` : {N ∪ Σ ∪ {}} or {N ∪ Σ0 ∪ {}} → N ∪ {∞} s.t.:
• if w = then `(w) = 0,
• if w ∈ Σ or Σ0 then `(w) = 1,
• if
= α1 . . . αn (with αi ∈ {N ∪ Σ ∪ {}} or {N ∪ Σ0 ∪ {}}) then `(w) =
Pw
n
i=1 `(αi ).
Algorithm 4 computes the value of `([X]) for each [X] ∈ N .
Algorithm 4 Compute the values `([X]) for each nonterminal [X] ∈ N
Input: GN = (N , Σ, [SG ], P) or (N , Σ0 , [SG ], P)
Output: `([X]) for each nonterminal [X]
(1) Initialize each `([X]) to ∞
(2) While there is at least one `([X]) updated at the previous iteration do
(2.1) For each production rule [X]
α1 . . . αn in P
P→
n
(2.1.1) `([X]) ← min{`([X]), i=1 `(αi )}
Proposition 3. Algorithm 4 terminates at worst after |N | iterations, and each
`([X]) ([X] ∈ N ) obtained is the length (number of characters) of the shortest
word produced by [X].
Proof. We prove that at each iteration (except the last one), there is a least
one nonterminal [X] s.t. `([X]) is updated with the length of the shortest word
that [X] produces.Thus, the update of all `([X]) with their correct values is done
at worst in |N | iterations. We proceed by induction on the number of iterations:
Basis: There is no -production in GN , and there is at least one production in
the form [X] → x where [X] ∈ N and x ∈ Σ (resp. Σ0 ) (see algorithm 3 for the
construction of GN ). For each [X] ∈ N s.t. {[X] → x} ∈ P, `([X]) = 1. And
algorithm 4 assigns these values to each `([X]) at the first iteration.
Induction: Suppose that at iteration n, there are at least n0 (n ≤ n0 < |N |)
nonterminals [X] s.t. the algorithm has assigned the correct value to `([X]). So
there are |N | − n0 nonterminals which have not the correct `-value yet.
17
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
• Either there is a nonterminal [Y ] s.t. `([Y ]) has not already the correct
value (i.e., it can already be updated), but for all productions [Y ] →
γ1 | . . . |γm (each [Y ] → γi is a production) all nonterminals in γ1 , . . . , γm
have their correct `-values. And thus, `([Y ]) will be updated with the
correct value at the end of the following iteration.
• Or, for each nonterminal [Y ] with a wrong `-value, there is at least a
nonterminal in each of its productions [Y ] → γ1 | . . . |γm which has not its
correct `-value yet.
– Either each γi contains a nonterminal which generates no word. Thus,
[Y ] generates no word and the value of `([Y ]) = ∞ is correct.
– Or all nonterminals in each γi generate a word. However, in the
derivation of the shortest word generated by any nonterminal, no
nonterminal is used twice or more (otherwise a shorter word can be
generated by using this nonterminal once). Thus, among all nonterminals with a wrong `-value, there is one which does not use others
to generate its shortest word. So its `-value is already correct or it
will be updated to the correct value at the following iteration.
2
Complexity. As there is at worst |N | iteration of the while loop, the complexity
of algorithm 4 is in O(|N | × |P|) which is O(|A|8 × |V |4 × |E|) (resp. O(|A|8 ×
|V |7 )) in the network model
There are several algorithms which allow the generation of a uniform random
word of some length from a CFG [12, 9, 18], or, more recently, a non-uniform
random generation of a word according to some targeted distribution [5, 10].
For instance, the boustrophedonic and the sequential algorithms described in
[9] generate a random labeled combinatorial object of some length from any
decomposable structure (including CFGs). The boustrophedonic algorithm is
in O(n log n) (where n is the length of the object) and the sequential algorithm is in O(n2 ) but may have a better average complexity. Both algorithms
use a precomputed table of linear size. This table can be computed in O(n2 ).
These algorithms require an unambiguous CFG, and the CFG computed by
algorithm 3 can be inherently ambiguous depending on the input network (as
several feasible paths can use the same sequence of protocols and thus have the
same trace TC ). However, the unambiguity requirement is only for the randomness of the generation. Recall that our goal is to generate the trace of the
shortest feasible path. Thus, we do not take into consideration the randomness and the distribution over the set of shortest traces. In order to generate
the shortest word in L(GN ), the boustrophedonic algorithm can take GN and
`([SG ]) as input (recall that `([SG ]) is the length of the shortest word generated
by GN ). Thus, the generation of the shortest word w (resp. w0 ) would have been
in O(|w|2 ) (resp. O(|w0 |2 )), where |w| denotes the length (number of characters)
of w. This complexity includes the precomputation of the table. However, this
complexity hides factors depending of the size of the CFG, the construction of
the precomputed table requires at least one pass over all the production rules.
18
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Actually, as the `-values are already computed, the generation can simply be
done in linear time in the length of the shortest word. The standard algorithm 5
takes as input the CFG and the `-values. Each nonterminal is replaced by the
left part with minimal `-value among its productions.
Algorithm 5 Computing the shortest word generated by GN
Input: GN = (N , Σ0 , [SG ], P) and all `-values
Output: w, the shortest word generated by GN
{[SG ] → γ1 |γ2 | . . . |γm are all production rules with [SG ] as left part}
(1) w ← arg min{`(γ1 ), . . . , `(γm )}
(2) While there is at least one nonterminal in w
(2.1) For each nonterminal [X] in w do
0
(2.1.1) Replace [X] in w by arg min{`(γ10 ), . . . , `(γm
0 )}
0
0
{[X] → γ1 | . . . |γm0 are all production rules with [X] as left part}
Complexity. Since there is no derivation of the form [X] → in GN (see
algorithm 3), all branches in the derivation tree end with a character from Σ
(resp. Σ0 ). The length of each branch is at worst |N | (as a nonterminal does not
appears twice in the same branch, otherwise a shorter word could be derivated
by using this nonterminal once). Thus, the number of derivations and the complexity of algorithm 5 are bounded by O(|N | × |w|) (resp. O(|N | × |w0 |)), which
corresponds to O(|A|3 × |V|2 × |w|) (resp. O(|A|3 × |V|2 × |w0 |)) in the network
model.
Example. Algorithm 4 gives `([SG ]) = 3. Algorithm 5 computes the shortest
word using the production rules in Fig. 3(d). The derivation is:
(1)
(2)
(3)
(4)
(5)
[SG ] ` [SA Z0 DA ] ` [Ua Z0 DA ] ` a[Vb aDa ][Da Z0 DA ] ` ab2 [Da Z0 DA ] ` ab2 a
Thus, the shortest word accepted by the transformed PDA is ab2 a. And the
shortest trace of a feasible path is f (ab2 a) = abba.
5.3.3. From the shortest word to the path.
If the goal is to minimize the number of nodes in the path, algorithm 6 takes
as input the shortest word w accepted by AN . Otherwise, as w0 is the shortest
word accepted by A0N and generated by GN , according to proposition 2, f (w0 )
is the word which minimizes the number of pops and pushes in AN . In such a
case it is the trace TC of the shortest feasible path C in the network N . It is
possible that several paths match with the trace TC = w (resp. f (w0 )). In such
a case, a load-balancing policy can choose a path.
Algorithm 6 is a dynamic programming algorithm that computes C. It starts
at the node S and takes at each step all the links in E which match with the
current character in TC . Let TC = x1 x2 . . . xn (xi ∈ A ∪ A). At each step i, the
algorithm starts from each node U in N odes[i] and adds to Links[i] all links
(U, V ) which match with xi . It also adds each V in N odes[i+1]. When reaching
D, it backtracks to S and selects the links from D to S.
19
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Algorithm 6 Find the shortest path
Input: Network N and TC
Output: Shortest path C
(1) N odes[1] ← S ; i ← 1
(2) While D is not reached do
(2.1) for each U ∈ N odes[i] and each V ∈ V s.t. (U, V ) ∈ E do
(2.1.1) If xi ∈ A, xi ∈ Out(U ), xi ∈ In(V ) and (xi−1 , xi ) ∈ P (U )
(2.1.1.1) Add (U, V ) to Links[i] and V to N odes[i + 1]
(2.1.2) If xi ∈ A, xi ∈ Out(U ), xi ∈ In(V ) and (xi , xi−1 ) ∈ P (U )
(2.1.2.1) Add (U, V ) to Links[i] and V to N odes[i + 1]
(2.2) i + +
(3) Backtrack from D to S by adding each covered link in the backtracking
to C.
Algorithm
Algo.
Algo.
Algo.
Algo.
Algo.
Algo.
1:
2:
3:
4:
5:
6:
Network to PDA
Transform PDA
PDA to CFG
Shortest word length
Shortest word
Shortest path
Upper-Bound Complexity
Minimizing hops
Minimizing enc./dec.
O(|A|3 × |E|)
/
O(max(|A|4 × |E|, |A|3 × |V|3 )
O(|A|5 × |V|2 × |E|)
O(|A|5 × |V|5 )
O(|A|8 × |V|4 × |E|)
O(|A|8 × |V|7 )
O(|A|3 × |V|2 × |w|)
O(|A|3 × |V|2 × |w0 |)
O(|TC | × |V| × |E|)
Table 1: Algorithms and their complexities
Complexity. The while loop stops exactly after TC steps, because it is sure
that there is a feasible path of length |TC | if TC is accepted by the automaton
AN . At each step, all links and nodes are checked in the worst case. Thus,
algorithm 6 is in O(|TC | × |V| × |E|) in the worst case.
Example. From the shortest trace abba, algorithm 6 computes the only feasible
path in the network on Fig. 3(a), which is S, a, U, b, V, b, W, a, D.
6. Conclusion
The problem of path computation in a multi-layer network has been studied
in the field of intra-domain path computation but is less addressed in the interdomain field with consideration of the Pseudo-Wire architecture. There was no
polynomial solution to this problem and the models used were not expressive
enough to capture the encapsulation/decapsulation capabilities described in the
Pseudo-Wire architecture.
In this paper, we provide algorithms that compute the shortest path in
a multi-layer multi-domain network, minimizing the number of hops or the
number of encapsulations and decapsulations. The presented algorithms involve
Automata and Language Theory methods. A Push-Down Automaton models
20
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
the multi-layer multi-domain network. It is then transformed in order to bypass
passive transitions and converted into a Context-Free Grammar. The grammar
generates the shortest protocol sequence, which allows to compute the path
matching this sequence.
The different algorithms of our methodology have polynomial upper-bound
complexity as summarized by Table 1. Compared to the preliminary version
of this work, the proofs of correctness of the algorithms are detailed and the
complexity analysis is significantly refined (the highest algorithm complexity is
in O(|A|8 × |V|7 ) instead of O(|A|11 × |V|7 × |E|2 )).
In order to figure out the whole problem of end-to-end service delivery, we
plan to extend our solution to support end-to-end Quality of Service constraints
and model all technology constraints on the different layers (conversion or “mapping” of protocols). As a future work, we also plan to investigate which part of
our algorithms can be distributed (e.g., Can the domains publish their encapsulation/decapsulation capabilities without disclosing their internal topology?)
and how such a solution can be established on today’s architectures.
Acknowledgment
The first author would like to thank Kaveh Ghasemloo for his help about
Algorithm 4. This work is partially supported by the ETICS-project, funded
by the European Commission. Grant agreement no.: FP7-248567 Contract
Number: INFSO-ICT-248567.
Appendix A. List of notations
In order to facilitate the paper reading, Table A.2 summarizes the symbols
used in the paper.
References
[1] M. Bocci and S. Bryant. RFC5659 - An Architecture for Multi-Segment
Pseudowire Emulation Edge-to-Edge, 2009.
[2] S. Bryant and P. Pate. RFC3985 - Pseudo Wire Emulation Edge-to-Edge
(PWE3) Architecture, 2005.
[3] I. Chlamtac, A. Faragó, and T. Zhang. Lightpath (Wavelength) Routing
in Large WDM Networks. IEEE Journal on Selected Areas in Communications, 14(5):909–913, 1996.
[4] H. Cho, J. Ryoo, and D. King.
Stitching dynamically
and
statically
provisioned
segments
to
construct
end-toend
multi-segment
pseudowires.
http://www.ietf.org/id/
draft-cho-pwe3-mpls-tp-mixed-ms-pw-setup-01.txt, 2011.
21
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
(a) Network
(b) Corresponding PDA
(c) Transformed PDA
(d) Subset of GN which generates TC
Figure 3: Example
22
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
Symbols and their signification
V: The set of nodes of the network
G = (V, E): The graph modeling the
network topology
D: The destination node
A: {a s.t. a ∈ A}
ED: The set of all possible encapsulations and passive functions
In(U ): The set of protocols that node
U can receive
T (U ): The set of protocols that can passively cross node U
(a, b): Encapsulation of protocol a in b
TC : The sequence of protocols used over
the path C (trace of C)
MC : The well-parenthesized sequence
of C
Q: The set of states of AN
Γ: The stack alphabet of AN
SA : Initial state of AN
DA : Final state of AN
Qa : Sub-automaton indexed by a
L(AN ): Language accepted by AN
GN : Context-free grammar corresponding to network N
[SG ]: Initial nonterminal (axiom) of GN
`([X]): Length of the shortest word generated by nonterminal [X]
Symbols and their signification
E: The set of links of the network
S: The source node
A: The alphabet (set of protocol)
P(U): The set of adaptation functions of
node U
ED: The set of all possible decapsulations
Out(U ): The set of protocols that node
U can send
(a, a): Passive function
(a, b): Decapsulation of protocol a from
b
HC : Sequence of adaptation functions
involved in path C (transition sequence
of C)
AN : Automaton corresponding to network N
Σ: The input alphabet of AN
δ: The set of transitions of AN
Z0 : Initial stack symbol
(Ux , hx, α, βi, Vy ): A transition between
states Ux and Vy , where x is read, α is
popped from the stack and β is pushed
on the stack
A0N : Transformed PDA
N
N bA
push (w): Minimum number of pushes
in a sequence of transitions accepting
the word w in AN
N : The set of nonterminals of GN
P: The set of production rules of GN
Table A.2: List of symbols used in the paper.
23
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
[5] A. Denise, Y. Ponty, and M. Termier. Controlled non-uniform random generation of decomposable structures. Theoretical Computer Science, 411(4042):3527–3552, 2010.
[6] F. Dijkstra, B. Andree, K. Koymans, J. van der Ham, P. Grosso, and
C. de Laat. A multi-layer network model based on ITU-T G.805. Comput.
Netw., 2008.
[7] F. Dijkstra, J. Van der Ham, P. Grosso, and C. de Laat. A path finding
implementation for multi-layer networks. Future Generation Comp. Syst.,
25(2):142–146, 2009.
[8] A. Farrel, JP. Vasseur, and J. Ash. RFC4655 - A Path Computation Element (PCE)-Based Architecture, 2006.
[9] Ph. Flajolet, P. Zimmermann, and B. Van Cutsem. A calculus for the random generation of labelled combinatorial structures. Theoretical Computer
Science, 1994.
[10] D. Gardy and Y. Ponty. Weighted random generation of context-free languages: Analysis of collisions in random urn occupancy models. GASCom,
2010.
[11] S. Gong and B. Jabbari. Optimal and Efficient End-to-End Path Computation in Multi-Layer Networks. In ICC, pages 5767–5771, 2008.
[12] T. Hickey and J. Cohen. Uniform random generation of strings in a contextfree language. SIAM J. Comput., 12(4):645–655, 1983.
[13] J. E. Hopcroft, R. Motwani, and J. D. Ullman. From PDA’s to Grammars.
In Introduction to Automata Theory, Languages, and Computation, chapter 6.3.2, pages 247–251. Addison-Wesley Longman Publishing Co., Inc.,
Boston, MA, USA, 2006.
[14] F. A. Kuipers and F. Dijkstra. Path selection in multi-layer networks.
Computer Communications, 2009.
[15] M. L. Lamali, H. Pouyllau, and D. Barth. Path computation in Multi-Layer
multi-domain networks. In IFIP/TC6 Networking 2012 (NETWORKING
2012), Prague, Czech Republic, May 2012.
[16] M.L. Lamali, H. Pouyllau, and D. Barth. End-to-End Quality of Service
in Pseudo-Wire Networks. In ACM CoNEXT Student Workshop, 2011.
[17] Jens Liebehenschel. Lexicographical Generation of a Generalized Dyck
Language. SIAM J. Comput., 2003.
[18] H. G. Mairson. Generating Words in a Context-Free Language Uniformly
at Random. Inf. Process. Lett., 49(2):95–99, 1994.
24
Author’s personal copy
Published version doi:10.1016/j.comcom.2012.11.009
[19] L. Martini, E. Rosen, N. El-Aawar, and G. Heron. RFC4448 - Encapsulation Methods for Transport of Ethernet over MPLS Networks, 2008.
[20] K. Shiomoto, D. Papadimitriou, JL. Le Roux, M. Vigoureux, and D. Brungard. RFC5212 - Requirements for GMPLS-based multi-region and multilayer networks (MRN/MLN), 2008.
[21] Y(J). Stein, R. Shashoua, R. Insler, and M. Anavi. RFC5087 - Time Division Multiplexing over IP (TDMoIP), 2007.
[22] W. Yao and B. Ramamurthy. A Link Bundled Auxiliary Graph Model for
Constrained Dynamic Traffic Grooming in WDM Mesh Networks. IEEE
Journal on Selected Areas in Communications, 23(8):1542–1555, 2005.
[23] H. Zhu, H. Zang, K. Zhu, and B. Mukherjee. A novel generic graph model
for traffic grooming in heterogeneous WDM mesh networks. IEEE/ACM
Trans. Netw., 11(2):285–299, 2003.
25
| 8cs.DS
|
An efficient Search Tool for an Anti-Money Laundering
Application of an Multi-national Bank's Dataset
Nhien-An Le Khac1, Sammer Markos1, Michael O'Neill1, Anthony Brabazon1 and M-Tahar Kechadi1
1
School of Computer Science & Informatics, University College Dublin, Dublin, Ireland
Abstract - Today, money laundering (ML) poses a serious
threat not only to financial institutions but also to the nations.
This criminal activity is becoming more and more
sophisticated and seems to have moved from the cliché of drug
trafficking to financing terrorism and surely not forgetting
personal gain. Most of the financial institutions
internationally have been implementing anti-money
laundering solutions (AML) to fight fraud activities. However,
the AML systems are so complicated that simple query tools
provided by current DBMS may produce incorrect and
ambiguous results and they are also very time-consuming due
to the complexity of the database system architecture.
In this paper, we present a new approach for identifying
customers quickly and easily as part of an AML application.
This will help AML experts to identify quickly customers who
are managed independently across separate databases of the
organization. This approach is tested on large and real-world
financial datasets. Some preliminary experimental results
show that this new approach is efficient and effective.
Keywords: Anti-Money laundering, customer identification,
search algorithms, tree topology, inverted list.
1
Introduction
Money laundering (ML) is a process of disguising the
illicit origin of "dirty" money and makes them appear
legitimate. It has been defined by Genzman as an activity that
"knowingly engage in a financial transaction with the
proceeds of some unlawful activity with the intent of
promoting or carrying on that unlawful activity or to conceal
or disguise the nature location, source, ownership, or control
of these proceeds" [7]. Through money laundering, criminals
try to convert monetary proceeds derived from illicit activities
into “clean” funds using a legal medium such as large
investment or pension funds hosted in retail or investment
banks. This type of criminal activity is getting more and
more sophisticated and seems to have moved from the cliché
of drug trafficking to financing terrorism and surely not
forgetting personal gain. Today, ML is the third largest
“Business” in the world after Currency Exchange and Auto
Industry. According to the United Nations Office on Drug
and Crime, worldwide value of laundered money in a year
ranges from $500 billion to $1 trillion [1] and from this
approximately $400-450 Billion is associated with drug
trafficking. These figures are at times modest and are partially
fabricated using statistical models, as no one exactly knows
the true value of money laundering, one can only forecast
according to the fraud that has already been exposed.
Nowadays, it poses a serious threat not only to financial
institutions but also to the nations. Some risks faced by
financial institutions can be listed as reputation risk,
operational risk, concentration risk and legal risk. At the
society level, ML could provide the fuel for drug dealers,
terrorists, arms dealers and other criminals to operate and
expand their criminal enterprises. Hence, the governments,
financial regulators require financial institutions to implement
processes and procedures to prevent/detect money laundering
as well as the financing of terrorism and other illicit activities
that money launderers are involved in. Therefore, anti-money
laundering (AML) is of critical significance to national
financial stability and international security..
Typically, an AML system is composed of some components
such as customer identification, transaction monitoring, case
management, reporting system, etc. Among them, customer
identification is one of the most important tasks as it assists
AML experts in monitoring customer behaviours;
transactions that they are involved in, their frequencies,
values, etc. Fundamentally, a customer is identified by
searching customer databases using query tools provided by
DBMS. However, in the case where a specific customer is
stored in separate databases that are managed independently,
this will require a very large processing time due the search
operations initiated over all the databases. Users need firstly
to login to different databases, run the same query repeatedly,
get the results separately, and displayed independently.
Furthermore, in large financial institutions, these databases
are heterogeneous and have very complex designs. This sort
of approach allows great flexibility, however it has poor
performance. In addition, data quality is also another factor
that makes this naïve approach becoming unfeasible.
In this paper, we present a new approach for identifying
customers in an international investment bank BEP1. This
approach provides a global view of customer information and
1
Real name of the bank can not be disclosed because of
confidential agreement of the project.
it is developed as a tool that allows the users to quickly and
efficiently identify customers who are managed
independently across separate databases. This tool is a
component of an AML solution developed for BEP.
The rest of this paper is organised as follows: the section 2
presents a background highlighting the current status of
BEP’s datasets and their customer search problems within the
AML context. Some indexing approaches are also discussed
in this section. We present our new approach that is a global
indexing based on word-ordered grouping and inverted list in
the Section 3. We describe the implementation of this
approach in Section 4. Section 5 presents preliminary
experimental results. Finally, we conclude in section 6.
2
Meanwhile, the data quality is also another impact that affects
the searching task. BEP’s input GUI is not efficient and its
databases design is cumbersome. Each customer database is
“identical”, i.e. the customer identification (CID) is only
unique in this database but the CID is not unique in all
databases. For instance, we can have (name= “John Smith”,
CID= “12345”) in database A vs. (name= “Peter Chang”,
CID= “12345”) in database B. Briefly; there is a uniqueness
violation at the global level. Furthermore, each database has a
different set of quality problems at the instance level. Some
problems can be listed as:
Missing values, dummy values or null. These
appear in most of the data fields in all databases
except the CID, the customer type (corporate,
individual and joint) and the fund name.
•
Misspellings; usually typos and phonetic errors.
For instance, we have “MACAO” vs. “MACAU”,
“11 1101” vs. “11-1101”, “Bloggs Corporation A/C
001” vs. “Bloggs Corporation 001”, etc.
•
Abbreviations; e.g. “A/C” vs. “AC” and “Account”
•
Word transpositions; e.g. “John Smith” vs. “Smith
John”
•
Duplicated records e.g. “John Smith” vs. “J.
Smith”
Backgrounds
We start this section with a brief presentation of an
AML project at BEP and then we will discuss on customer
search problems in its current environment. We finish this
section by reviewing some indexing approaches for data
search in the literature.
2.1
•
AML in BEP
Similar to any banking institution, BEP is required by
law to conduct strict AML governance on all transactions.
The BEP AML Unit does not have an automated solution to
support pattern recognition and detection of suspicious
activities. The purpose of this project is to apply new
principles and methodologies to build an AML framework in
order to detect suspicious customer transactions and
behaviour for the AML Unit. In this framework, one of the
important components is customer identification. Before
launching any customer transactional investigation, the
customer should be identified in all customer databases of
BEP. The structure of the BEP databases is complicated and
there are many problems with data quality that will can be
extracted and analysed, which are discussed in the following
paragraphs.
BEP datasets are divided into different environments
corresponding to sixteen clients with multiple funds per client
and managed hence by sixteen independent databases. When
a new customer or an investor X want to invest into a specific
fund (client specific), the AML team would request certain
documentation and will always treat him as a new customer
even though he could already invest into one/more of the
other fifteen clients, i.e. already exist in another databases.
The purpose of the customer search is to verify and identify a
customer’s profile in all invested funds. The AML Unit is
currently applying a manual search based on DBMS queries.
However, this is a time-consuming task because users should
login separately to each database and carry out repeated
queries. Moreover, each database contains not only data but
also its meta-data, so many joint operations are needed to
retrieve the information required.
Moreover, the names of some corporate customers are
normally not identical even though they are the same
company. For instance, “First Commercial Bank Ltd”2, “First
Commercial Bank Ltd OBB Account”, “First Commercial
Bank Ltd Trust Account TA 101010”, “First Commercial
Bank Ltd Trust Account TA 505055”, etc. We call this a
“company name group” property. Besides, some customer
databases also have the problem of incoherent data in address
data fields. The address information includes the following
data fields: “Street”, “Town”, “Zip”, “Country Code” and
“Country”. And then, for example, the “Zip” field contains
information about the street, house number and/or town, city
instead of its zip code.
Because of the customer datasets quality as well as its
complicated design, the manual customer search task by
DBMS queries currently takes more than two hours to
identify a customer.
2.2
Indexing
Fundamentally, search engines index the data in order to
facilitate fast and accurate information retrieval. Some
indexing methods in literature are tree-based, suffix tree,
2
Again, due to the confidential agreement, all examples
presented in this paper do not use the real customer names,
company names and address
inverted list, citation index, ngram index, term document
matrix, etc. The tree-based index would be the most popular
method where the search operations are linked with tree
nodes. The tree topology can be varied from a binary [10] to a
B-Tree family such as B/B*/ B+Tree [2] [3] [12]. For
instance, some DBMS implement an index structure based on
B-Tree such as MySQL, SQL Server [11]. Nevertheless, this
topology is not efficient enough for indexing complex,
heterogeneous, and bad quality data fields.
a
n
a
$
5
$
3
b
a
n
a
n
a
$
n
a
$
na
$
n
a
$
4
2
0
1
suffix link
internal node
leaf node
1 starting position
as they depend strongly on the quality of the data sets.
Therefore, the current quality of BEP’s customer data sets
should be improved before running any query. For instance,
in order to correct the misspelling problem, a spelling module
should be implemented to deal with typos and phonetic
errors. Similarly, abbreviated words must be uniform across
all separate databases; e.g. “A.C”, “Account”, “AC.” are
transformed to “A/C”. Meanwhile, data mining techniques
such as decision tree induction, regression, and inferencebased tools can be applied to fill missing values (tuples that
contain missing value fields cannot be ignored because all
customer information are important). Indeed, in some cases,
we should fill the missing value manually. Similarity, the
word transposition and duplicated records often need manual
intervention. Besides, the incoherent data problem in address
data fields (ref. Section 2.1) can only be manually corrected
but it is an unfeasible task with large datasets. Briefly, a
general solution for improving efficiently the quality of
BEP’s customer data sets is still an open question. Finally yet
importantly, the execution of DBMS queries on sixteen
independent BEP’s customer databases is also a very timeconsuming task. Next, we present our approach, which can
overcome the quality and design problems of BEP’s customer
databases
Figure 1. A suffix tree for “banana$”
Suffix tree [8], so-called PAT tree or position tree, is a data
structure that presents a given string in a suffix way (Figure
1). The suffix tree for a string S is a tree whose edges are
labelled with strings such that each suffix of S corresponds
exactly to one path from the tree’s root to a leaf. The
advantage of suffix tree is that operations on S and its
substring can be performed quickly. However, the
constructing suffix tree takes time and storage space linear in
the length of S.
3.1
Basic concepts
In this new approach, we aim to provide a global view
of information about all customers managed independently
across the BEP’s customer databases. Concretely, we build a
global index of these customers and provide a search engine
for AML users. Firstly, by analysing BEP’s customer
datasets, some important features can be summarised as:
•
There are two main types of customer: individual
and corporate. The individual customer has two
name fields: “First Name” and “Last Name”. In
some records, “First Name” (resp. “Last Name”)
field stores all parts of customer name; e.g. in a
record X, “First Name” field stores “John Smith”
and its “Last Name” is empty. This is a special kind
of missing value. The corporate customer has only
one name field: “Company Name” and most of them
have a “company name group” property as
mentioned in the Section 2.1 above.
•
The “Country” field is the most popular, i.e. its
missing value is less than 1%.
Inverted list [13] is another kind of index where each entry in
the index table includes two elements: an atomic search item
and a list of occurrences of this item in the whole search
space. For example, the index of a book lists every page on
which certain important words appear. This approach is
normally implemented by the hash table [4] or the binary tree
[10]. Inverted list is one of the most efficient index structures
[14].
Citation index approach [6] stores the citation or hyperlinks
between documents to support citation analysis. This
approach is normally applied in the Bibliography domain.
Ngram index [9] stores sequences of length of data and term
document matrix stores the occurrences of words in the
documents in a two-dimensional sparse matrix. The last two
index methods are mainly used in information retrieval or text
mining [5].
3
Customer Search approaches
As mentioned above, the current techniques based
DBMS queries are not suitable for the BEP’s AML system,
In addition, we also build a summary of all abbreviation cases
after this pre-processing process. We build our solution based
on these features and assuming that all abbreviation words are
expressed in a uniform way as well as all minor data
preparation is performed.
In order to deal with the incoherence of the address data
(Section 2.1), we merge all of the address data fields into one
field called “Address”. Similarity, we merge the “First
Name” and “Last Name” of individual customer into one
field named “Customer Name” to solve the missing value
problem. For each data record, word order in “Address” and
“Customer Name” can vary. For instance, we can have “John
Smith” vs. “Smith John” or “123, Main Street” vs. “Main
Street, 123” (word transposition problem). Therefore, the
inverted list technique can be used in this case to index
“Address” and “Customer Name”.
Meanwhile, the word order in the “Company Name” field is
important because of the “company name group” property.
Therefore, we need another type of index in this case. We can
address the “company name group” as a kind of suffix
problem and we can then use the suffix-tree topology.
Nevertheless, the implementation of this topology is
complicated. Hence, we rely on this topology to build an
index tree, which is simpler than the suffix tree for the
“Company Name” field of BEP’s customer datasets. So, our
global index is composed of three main parts: “Company
Name” index, “Customer Name” and “Address” index and
they are based on tree topology (the first index) and inverted
list (the last two ones), which are detailed in the following
section.
3.2
•
∀ element em ∈ T at the level l (l > 0), there is a
link, so-called node-link, between em and a node pem
∈ T at the level l+1. The node pem contains the first
word of all suffixes of the word wj stored in em.
•
A path from the root to the leaf by following nodelinks will create a specific company name. Each
element at the leaf level does not have a node-link
but a list of {client identification (FID), customer
identification (CID)} of the company name that
creates this path.
Level
0 (root)
ABC BANK FIRST INTERNATIONAL
1
2
3
CAPITAL
GROUP NEW
{Skada,B123}
YORK
4
BRANCH
5
{Abba,566}
OF
UBUBA
{Abba,470}
AMERICA BANK COMMERCIAL
BANK
LTD
BANK
INVESTMENT
LTD
OBB
LTD
CORP
TRUST
ACCOUNT
ACCOUNT {Abba,77}
The main architecture of our index consists of
“Company Name” tree, “Customer Name” and “Address”
inverted lists (Figure 2, 3 and 4). Furthermore, the whole
customer datasets are grouped by the customer type
(corporate or individual) and by the country.
Company Name Tree (CN tree) design. This is a suffix tree
based topology. Generally, the first word of a company name
appears at the root level (level 0) and its last word is at the
leaf level (Figure 2). The CN tree includes a set of nodes.
Each node contains one or many elements and each element
at level l links to only one node at level l+1 or to NULL if l is
at the leaf level. Each element has a key, which is a word
from the “company name” string. Hence, each node p at the
level l (l>0) contains all words derived from their prefix word
at the level l-1 and so on. Formally, supposing that a customer
datasets includes a set of n company names N and each
company name cni ∈ N (i=1, 2 ... n) is a string composed of a
set of words wj and the number of words in a cni is noted by
Ci, a CN tree T of N from customer datasets is defined by the
following:
•
•
•
•
The height of T is h, h =max(Ci), i =1,2...n
∀ node pk ∈ T, pk ⊇ set of elements {em}:
Card{em}>0, em ≡ wj.
The level 0 has at most one node p0.
Each element em0i ∈ p0, em0i contains the first
word of each company name cni.
7
{Merlu,888},
{Skada, B341}
{Abba,555}
Index architecture
6
DDD
TA
Legend
OF
101010 505055
LTD
{Abba,555}
{Merlu,1024}, {Skada,B987}
{Abba,392}
node
element
element at leaf level
information at each
leaf-level element, a
list of {FID, CID}
Figure 2. An example of CN tree index
For instance, as shown in Figure 2, we have the following
company name: “FIRST COMMERCIAL BANK LTD”,
“FIRST BANK LTD OBB ACCOUNT”, “FIRST AMERICA
BANK LTD TRUST ACCOUNT TA 101010”, “FIRST
AMERICA BANK LTD TRUST ACCOUNT TA 505055”,
“ABC CAPITAL GROUP”, “ABC CAPITAL NEW YORK
BRANCH”, “BANK OF UBUBA” and “INTERNATIONAL
DDD INVEST CORP”. Hence, the root node has four
elements: em00 = “ABC”, em01 = “BANK”, em02 = “FIRST”
and em03 = “INTERNATIONAL”. The element em00 links to
a node that has one element “CAPITAL” at the level 1. Then,
this element links to another node that has two elements:
“GROUP” and “NEW”. The element “GROUP” is at the leaf
level, it then contains {“Skada”, “B123”} (list of {FID, CID}
has only one element). Otherwise, the element “NEW” links
to another node at the level 2, etc. The path from the root
node with the element “ABC” following its node-links creates
two company names “ABC CAPITAL GROUP” and “ABC
CAPITAL NEW YORK BRANCH”. Similarity, the element
em02 links to a node with three elements at the level 1:
“AMERICA”, “BANK” and “COMMECIAL”. The element
“AMERICA” links to another node at the level 2 and so on, at
the level 7, the element “101010” is at the leaf level and
contains [{Merlu, 1024}, {Abba, 392}] (list of {FID, CID}
has two elements).
Table are used for an individual customer. Advantages and
problems of this approach will be discussed in section 5.
Based on this CN Tree, the search engine can find all {FID,
CID} of a requested “company name”. For instance, if the
query is “ABC CAPITAL GROUP” then the result is
{“Skada”, “B123”}, etc. Indeed, this index also supports an
approximate search i.e. users might not know the sufficient
name of a company so they just input its first few worlds, e.g.
the query is “ABC CAPITAL” then the result list will be
{“Skada”, “B123”} and {“Abba”, “566”}. Next, we can
retrieve all details of customers whose {“FID”, “CID”} are in
the result list.
Items
(words)
…...
John
…...
Murphy
…...
Smith
…...
Items
(words)
List of {Fund ID, Customer ID}
…...
…...
123
{“Abba”, “1234”}, {“Skada”, “347”}
…...
…...
Avenue
{“Merlu”, “112”}
…...
Sunset
…...
{“Abba”, “1234”}
…...
…...
Figure 4. An example of Address Index
List of {Fund ID, Customer ID}
…...
{“Abba”, “1234”}, {“Merlu”, “112”}
…...
Country
{“Merlu”, “112”}
Index
Afghanistan
…...
...
{“Abba”, “1234”}
Belarus
…...
...
...
...
...
Zimbabwe
Figure 3. An example of Customer Name Index
“Customer Name Index” is an index table based on the
inverted list technique. This index table consists of two parts:
items and a collection of lists; one list per item (Figure 3). An
item is a word from the “Customer Name” i.e. each customer
name in customer datasets is parsed into a set of separate
words. For instance, the customer name “John Smith” is split
into two words: “John” and “Smith”. A list Li of a word wi
records tupes of {FID, CID} of customers whose names
contain the word wi. We have, for example, the customer
“John Smith” with {FID= “ABBA”, CID= “1234”} and
“Murphy John” with {FID= “MERLU”, CID= “112”}.
Hence, the index table has three elements: [“John”:
{“ABBA”, “1234”}, {“MERLU”, “112”}], [“Murphy”:
{“MERLU”, “112”}] and [“Smith”: {“ABBA”, “1234”}].
“Address Index” is also an index table based on inverted list
technique. Similar to the “Customer Name Index”, its index
table consists of two parts: items and a collection of lists, one
list per item (Figure 4). An item is a word from the “Address”
i.e. each address in customer datasets is also parsed into a set
of separate words. For instance, we have an “address index”
table as shown in Figure 4.
The whole index structure is shown in Figure 5. In order to
limit the search space, these datasets are also grouped by
country (the most popular data field, ref. section 3.1).
Therefore, our search engine allows users to launch requests
on customer information managed in different databases only
through their name and address. If a customer is a corporate
then the search process will scan the CN Tree and Address
Index table. Meanwhile, Customer Index Table and Address
Figure 5. Index structure
4
Implementation
We developed our approach as a search tool based on a
distributed paradigm and this tool is implemented as web
services that can support 2-tier or 3-tier application model
(Figure 6). We implemented services for two kinds of users:
end-users and administrative users. There are indexing,
updating services for administrators. End-users exploit this
system through searching and extracting services.
4.1
Indexing service
This service scans all the BEP’s databases once and
builds indexes of “Company Name”, “Customer Name” and
“Address”. Elements in each node of “Company Name”
index as well as items in “Customer Name” and “Address”
index table are sorted by lexical order. Indexing service also
builds a country list and each element in this list stores
information (hash code) about appropriate entries of three
indexes above. These indexes are organised is main memory
and this service allows them to be saved in secondary
memory. Hence, the administrator only needs to create
indexes once and stores them in databases (index databases)
and then each time she/he reloads it to the main memory on
application launch. In the real world of banking application,
these indexes are loaded permanently in the main memory of
the servers and are synchronised periodically with their
databases (index databases). The customer information is not
real-time data processing i.e. when a customer opens an
account s/he always has to wait for a certain period of time
for all the security checks to be carried out (7 days, for
example) before the account is activated to perform her/his
first transaction (or Subscription in term of the investment
banking). Therefore, if indexes in the main memory are
damaged due to system halt, the electric cut, etc., the
administrators can reload them from index databases without
losing any information.
Databases server +
Application server
Front-end user
2-tier
Databases server
Application server
Front-end user
3-tier
Figure 6. Application models
4.2
Updating service
When indexes are being exploited, new customer
profiles are added in the customer datasets. Therefore, this
service allows updating new customer information into
indexes. It updates firstly in the main memory of the servers
and it then synchronises this information with the index
databases. The update service is automatically performed at a
predefined time by scanning all the databases (all update
information of customer is always stored for auditing).
4.3
Searching and extracting service
A search request submitted by the users includes
customer/company name and its address. Searching service
uses this information to look for a set of related {CID, FID}
on “Company Name”/ “Address” indexes (corporate
customer) or on “Customer Name”/ ”Address” (individual
customer). This is followed by performing queries on
databases to retrieve related customer information. The users
can choose which information to extract or perform further
investigations.
5
Evaluation and discussion
We implemented and tested our approach on real-world
customer datasets. The database architecture is similar to
BEP’s databases. The hardware platform for testing includes
1 Pentium Dual Core 3.4Ghz 2Gb RAM Windows Server
2003 (database server), 1 Pentium 4 Hyper Threading 3.4Ghz
1Gb Windows XP SP2 (application server), 1 Pentium 4
2.7Ghz Windows XP SP2 512Mb (front-end user). This web
service-based tool is developed in C#/Visual Studio 2005 and
we use SQL Server 2005. All services are implemented at the
application server and the database server manages the
datasets (3-tier model, Figure 6). The number of records is
approximately 32000 for all the databases.
We ran different tests on this platform and took the average
results. The indexing time I is about 17 seconds. The total
search time S is about 15 seconds (15s 40ms) for one request.
The search time S is composed of local search on the indexes,
query process by SQL server and communication overheads;
among them, the local search on the indexes only takes about
2 milliseconds on the application server. We also launched a
customer
search
by
SQL
queries
with
exact
“customer/company name” and “address” automatically on
all the databases and it takes about 3 minutes for one request.
We can see that our technique is much more efficient.
The approach presented in this paper has many advantages.
First, it solves the problem of access and querying
independent and separate databases by providing a global
view of all customer information without changing the
current architecture of BEP’s databases system. Then, this
approach also overcomes the data quality problems that
normally take an important time to pre-process, especially the
manual correction of incoherent address data. The preliminary
tests show that it is efficient. It is about 10 times faster than
the normal approach by DBMS queries and exhibits better
accuracy than the traditional approach.
Moreover, our approach also supports parallel processing
where two threads can be launched to search independently
on “Company/Customer Name” index and “Address” index.
It can benefit from the multi-core architecture of BEP’s
servers.
Besides, two main aspects of this approach need to be
improved. The first one is memory consumption because all
index structures are stored in the main memory. However,
each item stored is not a word but its hash code and it uses a
small amount of memory in our experiments. Furthermore, in
the real BEP’s servers, the main memory space is greater than
100 Gb and the whole indexes take less than 0.2%. Besides,
loading index in the main memory is normal practice of many
current DBMSs to exploit it efficiently. Another aspect is the
replication of items in the “Company Name” tree, e.g. the
word “BANK” exists in many nodes (Figure 2). This problem
can be improved by replacing this tree with a graph where
each item appears once as a node and the edges linking these
items represent the paths.
6
Conclusions and Future Works
In this paper, we have presented an approach for
identifying specific customer patterns in an investment bank.
This approach has been developed as a tool, which is a set of
web services on a distributed platform. The contribution of
this research is to provide a set of indexes combined with
suffix-tree based on an inverted list in order to overcome the
problem of database design and data quality of BEP’s
customer datasets. Our experimental results obtained on parts
of the BEP’s customer datasets, we can conclude that
approach is very efficient tool and it satisfies the needs of an
AML task.
Experimental results on real-platforms of BEP are also being
produced and these will allow us to test and evaluate the tool
robustness. We are currently working on the graph index
approach that takes in account the memory consumption issue
to tackle huge datasets. A multithreading version is also under
development.
7
References
[1] R. Baker, “The biggest loophole in the free-market
system”, Washington Quarterly, 1999, 22, pp. 29-46.
[2] R. Bayer, “Binary B-Trees for Virtual Memory”,
Proceedings of 1971 ACM-SIGFIDET Workshop on Data
Description, Access and Control, San Diego, California,
November 11-12, 1971.
[3] R. Bayer and E.M. McCreight “Organization and
Maintenance of Large Ordered Indices”. Acta Informatica 1,
1972: 173-189
[4] T. H. Cormen, C.E Leiserson, R.L. Rivest, C. Stein,
Clifford “Introduction to Algorithms” MIT Press and
McGraw-Hill 2nd Ed. 2001. pp. 222.
[5] R. Feldman and J. Sanger “The Text mining handbook:
Advanced Approaches in Analyzing Unstructed Data”,
Cambridge University Press, 2007.
[6] E. Garfield, A. I. Pudovkin, V. S. Istomin. “Algorithmic
Citation-Linked Historiography-Mapping the Literature of
Science”. ASIS&T 2002: Information, Connections and
Community. 65th Annual Meeting of ASIST, Philadelphia,
PA. November 18-21, 2002.
[7] L. Genzman, “Responding to organized crime: Laws
and law enforcement”, Organized crime, In H.Abadinsky
(Ed.) Belmont, CA: Wadsworth, pp. 342.
[8] R. Giegerich and S. Kurtz "From Ukkonen to McCreight
and Weiner: A Unifying View of Linear-Time Suffix Tree
Construction". Algorithmica 19 (3): 331—353, 1997
[9] http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?
catalogId =LDC2006T13
[10] D. E. Knuth “The Art of Computer Programming:
Fundamental Algorithms”, Addison Wesley, 3rd Ed. Volume
1, Chapter 2, 1997
[11] http://www.simple-talk.com/sql/learn-sql-server/sqlserver-index-basics/
[12] A.A. Toptsis. "B**-tree: a data organization method for
high storage utilization". Computing and Information, 1993.
Proceedings ICCI '9, 1993.: pages 277–281
[13] J. Zobel, A. Moffat, R. Sacks-Davis. “An efficient
indexing technique for full-text database systems” Proceeding
of the 18th VLDB Conference Vancouver, British Columbia,
Canada, 1992, 352-362
[14] J. Zobel, A. Moffat, R. Sacks-Davis. “Searching Large
Lexicons for Partially Specified Terms using Compressed
Inverted Files”, Proceeding of the 19th VLDB Conference
Dublin, Ireland, 1993, 290-301
| 5cs.CE
|
C-VQA: A Compositional Split of the
Visual Question Answering (VQA) v1.0 Dataset
arXiv:1704.08243v1 [cs.CV] 26 Apr 2017
∗
Aishwarya Agrawal∗ , Aniruddha Kembhavi† , Dhruv Batra‡ , Devi Parikh‡
Virginia Tech, † Allen Institute for Artificial Intelligence, ‡ Georgia Institute of Technology
[email protected], [email protected], {dbatra, parikh}@gatech.edu
Abstract
Visual Question Answering (VQA) has received a lot of attention over the past
couple of years. A number of deep learning models have been proposed for this
task. However, it has been shown [1–4] that these models are heavily driven
by superficial correlations in the training data and lack compositionality – the
ability to answer questions about unseen compositions of seen concepts. This
compositionality is desirable and central to intelligence. In this paper, we propose
a new setting for Visual Question Answering where the test question-answer pairs
are compositionally novel compared to training question-answer pairs. To facilitate
developing models under this setting, we present a new compositional split of the
VQA v1.0 [5] dataset, which we call Compositional VQA (C-VQA). We analyze
the distribution of questions and answers in the C-VQA splits. Finally, we evaluate
several existing VQA models under this new setting and show that the performances
of these models degrade by a significant amount compared to the original VQA
setting.
1
Introduction
Automatically answering questions about visual content is considered to be one of the holy grails
of artificial intelligence research. Visual Question Answering (VQA) poses a rich set of challenges
spanning various domains such as computer vision, natural language processing, knowledge representation and reasoning. VQA is a stepping stone to visually grounded dialog and intelligent agents
[6–9]. In the past couple of years, VQA has received a lot of attention. Various VQA datasets have
been proposed by different groups [2, 3, 5, 10–15] and a number of deep-learning models have been
developed [5, 16–33].
However, it has been shown that despite recent progress, today’s VQA models are heavily driven
by superficial correlations in the training data and lack compositionality [1–4] – the ability to
answer questions about unseen compositions of seen concepts. For instance, a model is said to be
compositional if it can correctly answer [“What color are the safety cones?”, “green”] without seeing
this question-answer (QA) pair during training, but perhaps having seen [“What color are the safety
cones?”, “orange”] and [“What color are the plates?”, “green”] during training.
In order to evaluate the extent to which existing VQA models are compositional, we create a
compositional split of the VQA v1.0 dataset [5], called Compositional VQA (C-VQA). This new
dataset is created by re-arranging the train and val splits of the VQA v1.0 dataset in such a way that
the question-answer (QA) pairs in C-VQA test split are compositionally novel with respect to those
in C-VQA train split, i.e., QA pairs in C-VQA test split are not present in C-VQA train splits but
most concepts constituting the QA pairs in test split are present in the train split. Fig. 1 shows some
examples from our C-VQA splits. Since, C-VQA test split contains the QA pair [“What is the color
of the plate?”, “red”], similar QA pairs such as [“What color is the plate?”, “red”] are not present in
C-VQA train split. But C-VQA train split contains other QA pairs consisting of the concepts “plate”,
Figure 1: Examples from our Compositional VQA (C-VQA) splits. Words belonging to same
concepts are highlighted with same color to show the training instances from which the model can
learn those concepts.
“red” and “color” such as [“What color is the plate?”, “green”] and [“What color are stop lights?”,
“red”].1
Evaluating a VQA model under such setting helps in testing – 1) whether the model is capable
of learning disentangled representations for different concepts (e.g., “plate”, “green”, “stop light”,
“red”), 2) whether the model can compose the concepts learned during training to correctly answer
questions about novel compositions at test time. Please see Section 3 for more details about C-VQA
splits.
To demonstrate the difficulty of our C-VQA splits, we report the performance of several existing
VQA models [17, 20, 23, 27, 34] on our C-VQA splits. Our experiments show that the performance
of the VQA models drops significantly (with performance drop being smaller for models which are
compositional by design such as the Neural Module Networks [20]) when trained and evaluated
on train and test splits (respectively) of C-VQA, compared to when these models are trained and
evaluated on train and val splits (respectively) of the original VQA v1.0. Please see Section 4 for
more details about these experiments.
2
Related Work
Visual Question Answering. Several papers have proposed visual question answering datasets to
train and test machines for the task of visual understanding [2, 3, 5, 10–15]. Over the span of time, the
size of VQA datasets has become larger and questions have becomes more free form and open-ended.
For instance, one of the earliest VQA datasets [12] considers questions generated using templates and
consists of fixed vocabulary of objects, attributes, etc. [13] also consider questions whose answers
come from a closed world. [15] generate questions automatically using image captions and their
1
It should be noted that in the VQA v1.0 splits, a given Image, Question, Answer (IQA) triplet is not shared
across splits but a given QA pair could be shared across splits.
2
answers belong to one of the following four types – object, number, color, location. [5, 10, 11, 14]
consist of free form open-ended questions. Of these datasets, the VQA v1.0 dataset [5] has been used
widely to train deep models. Performance of such models has increased steadily over the past two
years on the test set of VQA v1.0 which has a similar distribution of data points as its training set.
However, careful examination of the behaviors of such models reveals that these models are heavily
driven by superficial correlations in the training data and lack compositionality [1]. This is partly
because the training set of VQA v1.0 contains strong language priors which data-driven models can
learn easily and can perform well on the test set which consists of similar priors as the training set,
without truly understanding the visual content in images [3], because it is easier to learn the biases of
the data (or even our world) than to truly understand images.
In order to counter the language priors, Goyal et al. [3] balance every question in the VQA v1.0
dataset by collecting complementary images for every question. Thus, for every question in the
VQA v2.0 dataset, there are two similar images that have different answers to the question. Clearly,
language priors are significantly weaker in the VQA v2.0 dataset. However, such balancing does
not test for compositionality because the train and test distributions are similar. So, in order to test
whether models can learn each concept individually irrespective of the correlations in the data and can
perform well on a test set which has a different distribution of correlations compared to the training
set, we propose a compositional split of the VQA v1.0 dataset, which we call Compositional-VQA
(C-VQA).
Compositionality. The ability to generalize to novel compositions of concepts learned during training
is desirable from any intelligent system. Compositionality has been studied in various forms in the
vision community. Zero-shot object recognition using attributes is based on the idea of composing
attributes to detect novel object categories [35, 36]. More recently, [37] have studied compositionality
in the domain of image captioning by focusing on structured representations (subject-relation-object
triplets). We study compositionality for visual question answering where the questions and answers
are open-ended and in free-form natural language. The work closest to us is [4] where they study
compositionality in the domain of VQA. However, their dataset (images as well as questions) is
synthetic and has only limited number of objects and attributes. On the contrary, our C-VQA splits
consist of real images and questions (asked by humans) and hence involve a variety of objects and
attributes, as well as activities, scenes, etc. Andreas et al. [20, 24] have developed compositional
models for VQA that consist of different modules each specialized for a particular task. These
modules can be composed together based on the question structure to create a model architecture
for the given question. Although, compositional by design, these models have not been evaluated
specifically for compositionality. Our C-VQA splits can be used to evaluate such models to test the
degree of compositionality. In fact, we report the performance of Neural Module Networks on VQA
v1.0 and C-VQA splits (Section 4).
The compositionality setting we are proposing is one type of zero-shot VQA where test QA pairs are
novel. Other types of zero-shot VQA have also been explored. [38] propose a setting for VQA where
the test questions (the question string itself or the multiple choices) contain atleast one unseen word.
[39] propose answering questions about unknown objects (e.g., “Is the dog black and white?” where
“dog” is never seen in training (neither in questions, nor in answers)).
3
Compositional Visual Question Answering (C-VQA)
3.1
C-VQA Creation
The C-VQA splits are created by re-arranging the training and validation splits of the VQA v1.0
dataset [5]2 . These splits are created such that the question-answer (QA) pairs in the C-VQA test split
(e.g., Question: “What color is the plate?”, Answer: “green”) are not seen in the C-VQA train split,
but in most cases, the concepts that compose the C-VQA test QA pairs (e.g., “plate”, “green”) have
been seen in the C-VQA train split (e.g., Question: “What color is the apple?”, Answer: “Green”,
Question: “How many plates are on the table?”, Answer: “4”).
The C-VQA splits are created using the following procedure –
2
We can not use the test splits from VQA v1.0 because creation of C-VQA splits requires access to test
annotations which are not publicly available.
3
Question Reduction: Every question is reduced to a list of concepts needed to answer the question.
For instance,
“What color are the cones?” is reduced to [“what”, “color”, “cone”].
We do this in order to reduce similar questions to the same form. For instance, “What color are the
cones?” and “What is the color of the cones?” both get reduced to the same form – [“what”, “color”,
“cone”]’. This reduction is achieved using simple text processing such as removal of stop words and
lemmatization.
Reduced QA Grouping: Questions having the same reduced form and the same ground truth answer
are grouped together. For instance,
[“What color are the cones?”, “orange”] and [“What are the color of the cones?”, “orange”] are
grouped together whereas [“What color are the cones?”, “green”] is put in a different group.
This grouping is done after merging the QA pairs from the VQA v1.0 train and val splits.
Greedily Re-splitting: A greedy approach is used to redistribute data points (image, question,
answer) to the C-VQA train and test splits so as to maximize the coverage of the test concepts in the
C-VQA train split while making sure QA pairs are not repeated between test and train splits. In this
procedure, we loop through all the groups created above, and in every iteration, we add the current
group to the C-VQA test split unless the group has already been assigned to the C-VQA train split.
We always maintain a set of concepts3 belonging to the groups in the C-VQA test split that have not
yet been covered by the groups belonging to the C-VQA train split. From the groups that have not
yet been assigned to either of the splits, we find the group that covers majority of the concepts (in the
list) and add that group to the C-VQA train split.
The above approach results in 73.5% of the unique C-VQA test concepts to be covered in the C-VQA
train split. The coverage is 98.8% when taking into account the frequency of occurrence of each
concept in C-VQA test split.
Table 1 shows the number of questions, images and answers in the train and val splits of the VQA
v1.0 dataset and those in train and test splits of the C-VQA dataset. We can see that the number of
questions and the number of answers in the C-VQA splits is similar to that in the VQA v1.0 splits.
However, the number of images in the C-VQA splits is more than that in the VQA v1.0 splits. This is
because in the C-VQA splits, the same image can be present in both the train and the test sets. Note
that there are three questions for every image in the VQA v1.0 dataset and the splitting for C-VQA
is done based on QA pairs, not based on images. Consider the following two questions associated
with the same image in VQA v1.0 – “What color are the cones?” (with the answer “orange”) and
“What time of day is it?” (with the answer “afternoon”). It is possible that “What color are the cones?”
(along with the image and the ground-truth answers) gets assigned to C-VQA train split and “What
time of day is it?” gets assigned to C-VQA test split. As a result, the image corresponding to these
questions would be present in both the train and test splits of C-VQA.4
Dataset
Split #Questions #Images #Answers
Train
248,349
82,783
2,483,490
Val
121,512
40,504
1,215,120
Train
246,574
118,663 2,465,740
Test
123,287
86,700
VQA (v1.0) [5]
C-VQA (proposed)
1,232,870
Table 1: Statistics of the VQA v1.0 and our C-VQA splits.
3
For a given group, concepts are the set of all unique words present in the reduced question and the ground
truth answer belonging to that group
4
To verify that sharing of images across splits does not make the problem easier, we randomly split the VQA
v1.0 train+val into random-train and random-val. We then trained and evaluated the deeper LSTM Q + norm I
model from [5] on these new splits. We saw that the this new setup leads to only ∼1% increase in the model
performance compared to the VQA v1.0 train and val setup.
4
3.2
C-VQA Analysis
In this section we analyze how the distributions of questions and answers in the C-VQA train and test
splits differ from those in the VQA v1.0 train and val splits.
Question Distribution. Fig. 2 shows the distribution of questions based on the first four words
of the questions for the train (left) and test (right) splits of the C-VQA dataset. We can see that
splitting the dataset compositionally (as in C-VQA) does not lead to significant differences in the
distribution of questions across splits, keeping the distributions qualitatively similar to VQA v1.0
splits [5]. Quantitatively, 46.06% of the question strings in the VQA v1.0 val split are also present in
the VQA v1.0 train split, whereas this percentage is 37.76 for the C-VQA splits.
Figure 2: Distribution of questions by their first four words for a random sample of 60K questions for
C-VQA train split (left) and C-VQA test split (right). The ordering of the words starts towards the
center and radiates outwards. The arc length is proportional to the number of questions containing
the word. White areas are words with contributions too small to show.
Answer Distribution. Fig. 3 shows the distribution of answers for several question types such as
“what color”, “what sport”, “how many”, etc. for the train (top) and test (bottom) splits of the C-VQA
dataset. We can see that the distributions of answers for a given question type is significantly different.
However, for VQA v1.0 dataset, the distribution for a given question type is similar across train and
val splits [5]. For instance, “tennis” is the most frequent answer for the question type “what sport”
in C-VQA train split whereas “skiing” is the most frequent answer for the same question type in
C-VQA test split. However, for the VQA v1.0 splits, “tennis” is the most frequent answer for both the
train and val splits. Similar differences can be seen for other question types as well – “what animal”,
“what brand”, “what kind”, “what type”, “what are”. Quantitatively, 32.49% of the QA pairs in the
VQA v1.0 val split are also present in the VQA v1.0 train split, whereas this percentage is 0 for the
C-VQA splits (by construction).
4
Baselines
We report the performances of the following VQA models when trained on C-VQA train split and
evaluated on C-VQA test split and compare this with the setting when these models are trained on
VQA v1.0 train split and evaluated on VQA v1.0 val split (Table 2).
Deeper LSTM Question + normalized Image (deeper LSTM Q + norm I) [34]: This model was
proposed in [5]. It is a two channel model – one channel processes the image and the other channel
processes the question. For each image, the image channel extracts the activations (4096-dim) of
the last hidden layer of the VGGNet [40] and normalizes them. For each question, the question
channel extracts the hidden state and cell state activations of the last hidden layers of 2-layered LSTM,
resulting in a 2048-dim encoding of the question. The image features (4096-dim) obtained from the
image channel and the question features (2048-dim) obtained from the question channel are linearly
5
Figure 3: Distribution of answers per question type for a random sample of 60K questions for C-VQA
train split (top) and C-VQA test split (bottom).
transformed to 1024 dimensions each and fused together via element-wise multiplication. This fused
vector is then passed through one more fully-connected layer in a Multi-Layered Perceptron (MLP),
which finally outputs a 1000-way softmax score over the 1000 most frequent answers from the
training set. The entire model, except the CNN (which is not fine-tuned) is learned end-to-end with a
cross-entropy loss.
Neural Module Networks (NMN) [20]: This model is designed to be compositional in nature. The
model consists of composable modules where each module has a specific role (such as detecting
a dog in the image, counting the number of dogs in the image, etc.). Given an image and the
natural language question about the image, NMN first decomposes the question into its linguistic
substructures using a parser. These structures determine which modules need to be composed together
in what layout to create the network for answering the question. The resulting compound networks
are jointly trained. At test time, the image and the question are forward propagated through the
dynamically composed network which outputs a distribution over answers. In addition to the network
composed using different modules, NMN also uses an LSTM to encode the question which is then
added elementwise to the representation produced by the last module of the NMN. This combined
representation is passed through a fully-connected layer to output a softmax distribution over answers.
The LSTM encodes priors in the training data and models syntactic regularities such as singular
vs. plural (“what is flying?” should be answered with “kite” whereas “what are flying?” should be
answered with “kites”).
Stacked Attention Networks (SAN) [17]: This is one of the widely used models for VQA. This
model is different from other VQA models in that it uses multiple hops of attention over the image.
Given an image and the natural language question, SAN uses the question to obtain an attention map
over the image. The attended image is combined with the encoded question vector which becomes
the new query vector. This new query vector is used again to obtain a second round of attention
6
over the image. The query vector obtained from the second round of attention is passed through a
fully-connected layer to obtain a distribution over answers.5
Hierarchical Question-Image Co-attention Networks (HieCoAtt) [23]: This is one of the top
performing models for VQA. In addition to modeling attention over image, this model also models
attention over question. Both image and question attention are computed in a hierarchical fashion.
The attended image and question features obtained from different levels of the hierarchy are combined
and passed through a fully-connected layer to obtain a softmax distribution over the space of answers.
Multimodal Compact Bilinear Pooling (MCB) [27]: This model won the real image track of the
VQA Challenge 2016. MCB uses multimodal compact bilinear pooling to predict attention over
image features and also to combine the attended image features with the question features. These
combined features are passed through a fully-connected layer to obtain a softmax distribution over
the space of answers.
Model
Dataset
Yes/No Number Other Overall
VQA v1.0 val
79.81
33.26
40.35
54.23
C-VQA test
70.60
29.76
31.83
46.69
VQA v1.0 val
80.39
33.45
41.07
54.83
C-VQA test
72.96
31.02
34.49
49.05
VQA v1.0 val
78.54
33.46
44.51
55.86
C-VQA test
66.96
24.30
33.19
45.25
VQA v1.0 val
79.81
34.93
45.64
57.09
C-VQA test
71.11
30.48
38.31
50.12
VQA v1.0 val
81.62
34.56
52.16
60.97
C-VQA test
71.33
24.90
47.84
54.15
deeper LSTM Q + norm I [34]
NMN [20]
SAN [17]
HieCoAtt [23]
MCB [27]
Table 2: Accuracies of existing VQA models on the VQA v1.0 val split when trained on VQA v1.0
train split and those on C-VQA test split when trained on C-VQA train split.
From Table 2, we can see that the performance of all the existing VQA models drops significantly in
the C-VQA setting compared to the VQA v1.0 setting. Note that even though the Neural Module
Networks architecture is compositional by design, their performance suffers on C-VQA. We posit
this may be because they use an additional LSTM encoding of the question to encode priors in the
dataset. In C-VQA, the priors learned from the train set are unlikely to generalize to the test set. Also
note that other models suffer a larger drop in performance compared to Neural Module Networks.
Another interesting observation from Table 2 is that the ranking of the models based on overall
performance changes from VQA v1.0 to C-VQA. For VQA v1.0, SAN outperforms deeper LSTM Q
+ norm I and NMN, whereas for C-VQA, these two models outperform SAN. Also note the change in
ranking of the models for different types of answers (“yes/no”, “number”, “other”). For instance, for
“number” questions, MCB outperforms all the models except HieCoAtt for VQA v1.0. However, for
C-VQA, all the models except SAN outperform MCB.
Examining the accuracies of these models for different question types shows that the performance
drop from VQA v1.0 to C-VQA is larger for some question types than the others. For Neural Module
Networks (NMN), Stacked Attention Networks (SAN) and Hierarchical Question-Image Co-attention
Networks (HieCoAtt), questions starting with “what room is” (such as “What room is this?”) have
the largest drop – 33.28% drop for NMN, 40.73% drop for SAN and 32.56% drop for HieCoAtt.
For such questions in the C-VQA test split, one of the correct answers is “living room” which is
not one of the correct answers to such questions in the C-VQA train split (the correct answers in
the C-VQA train split are “kitchen”, “bedroom”, etc.). So, models tend to answer the C-VQA test
5
We use a torch implementation of SAN, available at https://github.com/abhshkdz/
neural-vqa-attention, for our experiments.
7
questions with what they have seen during training (such as “kitchen”). Note that “living room” is
seen during training for questions such as “Which room is this?”. For deeper LSTM + norm I model
and Multimodal Compact Bilinear Pooling (MCB) model, the largest drop is for “is it” questions
(such as “Is it daytime?”) – 29.52% drop for deeper LSTM Q + norm I and 30.77% drop for MCB
model. For such questions in the C-VQA test split, the correct answer is “yes” whereas the correct
answer for such questions in C-VQA train split is “no”. Again, models tend to answer the C-VQA
test questions with “no”. Other question types resulting in significant drop in performance (more
than 10%) for all the models are – “what is the color of the”, “how many people are in”, “are there”,
“is this a”.
5
Conclusion
In conclusion, we introduce a novel setting for Visual Question Answering – Compositional Visual
Question Answering. Under this setting, the question-answer pairs in the test set are compositionally
novel compared to the question-answer pairs in the training set. We create a compositional split of the
VQA (v1.0) dataset [5], called C-VQA, which facilitates training compositional VQA models. We
show the similarities and differences between the VQA v1.0 and C-VQA splits. Finally, we report
performances of several existing VQA models on the C-VQA splits and show that the performance
of all the models drops significantly compared to the original VQA v1.0 setting. This suggests that
today’s VQA models do not handle compositionality well and that C-VQA splits can be used as a
benchmark for building and evaluating compositional VQA models.
References
[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering
models. In EMNLP, 2016.
[2] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and Yang: Balancing
and answering binary visual questions. In CVPR, 2016.
[3] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter:
Elevating the role of image understanding in visual question answering. arXiv preprint arXiv:1612.00837,
2016.
[4] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross
Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. arXiv
preprint arXiv:1612.06890, 2016.
[5] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,
and Devi Parikh. VQA: Visual Question Answering. In ICCV, 2015.
[6] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M.F. Moura, Devi Parikh,
and Dhruv Batra. Visual Dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2017.
[7] Abhishek Das, Satwik Kottur, José M.F. Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual
dialog agents with deep reinforcement learning. arXiv preprint arXiv:1703.06585, 2017.
[8] Harm de Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville.
Guesswhat?! visual object discovery through multi-modal dialogue. arXiv preprint arXiv:1611.08481,
2016.
[9] Nasrin Mostafazadeh, Chris Brockett, Bill Dolan, Michel Galley, Jianfeng Gao, Georgios P. Spithourakis,
and Lucy Vanderwende. Image-grounded conversations: Multimodal context for natural question and
response generation. CoRR, abs/1701.08251, 2017.
[10] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen,
Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision
using crowdsourced dense image annotations. arXiv preprint arXiv:1602.07332, 2016.
[11] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering
in images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
4995–5004, 2016.
8
[12] Donald Geman, Stuart Geman, Neil Hallonquist, and Laurent Younes. Visual turing test for computer
vision systems. Proceedings of the National Academy of Sciences, 112(12):3618–3623, 2015.
[13] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world
scenes based on uncertain input. In Advances in Neural Information Processing Systems, pages 1682–1690,
2014.
[14] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a
machine? dataset and methods for multilingual image question. In Advances in Neural Information
Processing Systems, pages 2296–2304, 2015.
[15] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering.
In Advances in Neural Information Processing Systems, pages 2953–2961, 2015.
[16] Kan Chen, Jiang Wang, Liang-Chieh Chen, Haoyuan Gao, Wei Xu, and Ram Nevatia. ABC-CNN: an
attention based convolutional neural network for visual question answering. CoRR, abs/1511.05960, 2015.
[17] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alexander J. Smola. Stacked attention networks
for image question answering. In CVPR, 2016.
[18] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for
visual question answering. In ECCV, 2016.
[19] Aiwen Jiang, Fang Wang, Fatih Porikli, and Yi Li. Compositional memory for visual question answering.
CoRR, abs/1511.05676, 2015.
[20] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Deep compositional question answering
with neural module networks. In CVPR, 2016.
[21] Peng Wang, Qi Wu, Chunhua Shen, Anton van den Hengel, and Anthony R. Dick. Explicit knowledgebased reasoning for visual question answering. CoRR, abs/1511.02570, 2015.
[22] Kushal Kafle and Christopher Kanan. Answer-type prediction for visual question answering. In CVPR,
2016.
[23] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for
visual question answering. In NIPS, 2016.
[24] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural networks
for question answering. In NAACL, 2016.
[25] Kevin J. Shih, Saurabh Singh, and Derek Hoiem. Where to look: Focus regions for visual question
answering. In CVPR, 2016.
[26] Jin-Hwa Kim, Sang-Woo Lee, Dong-Hyun Kwak, Min-Oh Heo, Jeonghee Kim, Jung-Woo Ha, and
Byoung-Tak Zhang. Multimodal residual learning for visual QA. In NIPS, 2016.
[27] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach.
Multimodal compact bilinear pooling for visual question answering and visual grounding. In EMNLP,
2016.
[28] Hyeonwoo Noh and Bohyung Han. Training recurrent answering units with joint loss minimization for
vqa. CoRR, abs/1606.03647, 2016.
[29] Ilija Ilievski, Shuicheng Yan, and Jiashi Feng. A focused dynamic attention model for visual question
answering. CoRR, abs/1604.01485, 2016.
[30] Qi Wu, Peng Wang, Chunhua Shen, Anton van den Hengel, and Anthony R. Dick. Ask me anything:
Free-form visual question answering based on knowledge from external sources. In CVPR, 2016.
[31] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual
question answering. In ICML, 2016.
[32] Xiang Sean Zhou and Thomas S. Huang. Relevance feedback in image retrieval: A comprehensive review.
Proceedings of ACM Multimedia Systems, 2003.
[33] Kuniaki Saito, Andrew Shin, Yoshitaka Ushiku, and Tatsuya Harada. Dualnet: Domain-invariant network
for visual question answering. CoRR, abs/1606.06108, 2016.
9
[34] Jiasen Lu, Xiao Lin, Dhruv Batra, and Devi Parikh. Deeper lstm and normalized cnn visual question
answering model. https://github.com/VT-vision-lab/VQA_LSTM_CNN, 2015.
[35] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes
by between-class attribute transfer. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE
Conference on, pages 951–958. IEEE, 2009.
[36] Dinesh Jayaraman, Fei Sha, and Kristen Grauman. Decorrelating semantic visual attributes by resisting the
urge to share. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
1629–1636, 2014.
[37] Yuval Atzmon, Jonathan Berant, Vahid Kezami, Amir Globerson, and Gal Chechik. Learning to generalize
to new compositions in image understanding. arXiv preprint arXiv:1608.07639, 2016.
[38] Damien Teney and Anton van den Hengel. Zero-shot visual question answering. arXiv preprint
arXiv:1611.05546, 2016.
[39] Santhosh K Ramakrishnan, Ambar Pal, Gaurav Sharma, and Anurag Mittal. An empirical evaluation of
visual question answering for novel objects. arXiv preprint arXiv:1704.02516, 2017.
[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014.
10
| 2cs.AI
|
Energetic Galerkin Projection of Electromagnetic
Fields between Different Meshes
Z. Wang1 , Z. Tang2 , T. Henneron2 , F. Piriou2 and J. C. Mipo1
arXiv:1505.01118v1 [cs.CE] 5 May 2015
1
Valeo electrical systems, 2 rue André Boulle, BP 150, 94017 Créteil, France
2
L2EP, University of Lille 1, 59655 Villeneuve d’ascq, France aa
[email protected]
Abstract—In order to project electromagnetic fields between
different meshes with respect to the conservation of energetic
values, Galerkin projection formulations based on the energetic
norm are developed in this communication. The proposed formulations are applied to an academic example.
Index Terms—Finite element methods, Galerkin method, Interpolation, Modeling, Projection.
field H or the magnetic flux density B is conformed with
physical properties. We denote by H s and B s the fields
obtained on the source mesh and by Ht and Bt the fields to
be calculated on the target mesh. The energetic norms of the
interpolation error are defined as:
Z
εH =
µ |Ht − H s |2 dτ
(1)
I. Introduction
In recent years, the numerical studies of coupled problems
are more and more investigated. These studies deal with
the interactions between different physical phenomena, e.g.
electromagnetic - thermal or magnetic - mechanic. According
to the importance of the interaction, coupled problems can
be treated with different strategies. One possibility is that the
study domain is discretized on different meshes for different
problems. In this case, it is necessary to communicate and
transfer fields between different meshes. In the literature, the
concept of Galerkin projection based on the L2 error norm
provides a very convenient tool to carry out this transfer [1],
[2], [3]. In comparison to the direct interpolation, Galerkin
projection enjoys several advantages, especially in terms of
precision. However, using this method, the conservation of
the energy is not assured between the original and target
meshes [4]. To tackle this issue, formulations deduced from
the minimization of the energetic norm can be used.
In this communication, energetic approaches for Galerkin
projection are developed in order to conserve the magnetic
energy and electric power between different meshes. Firstly,
the numerical models developed from the minimization of
the energetic norm are presented for magneto-static and eddy
current problems. Secondly, the obtained formulations are
applied to an academic example.
II. Energetic Galerkin projection formulations
Given a solution of electromagnetic computation on a source
mesh, we aim to project this result onto a target mesh which
differs. The energetic Galerkin projection formulations are
investigated for magneto-static problems as well as eddy
current problems.
A. Magneto-static problems
In order to solve magneto-static problems, different formulations such as the formulations based on scalar or vectorial
potentials can be employed. As a result, either the magnetic
D
εB =
Z
D
1
|Bt − B s |2 dτ
µ
(2)
with µ the linear magnetic permeability.
Using Whitney elements in 3D, Ht and Bt are discretized
in theX
edge and facet element
spaces respectively such that
X
Ht =
wei hi and Bt =
wif bi . Here wei (resp. wif ) and hi
(resp. bi ) are the shape functions and the values of Ht (resp.
Bt ) associated with the i−th edge (resp. facet). To project fields
onto a target mesh with respect to the magnetic energy, weak
formulations based on the minimization of the energetic norm
are developed. In (1), the energetic error norm is minimized
when its derivatives with respect to all degrees of freedom are
equal to zero, thus for each edge i:
Z
∂
µ |Ht − H s |2 dτ = 0
(3)
∂h
i
D
Z
µ(wei · Ht − wei · H s )dτ = 0
(4)
D
Finally, the matrix equation to solve can be written:
where Ci j =
Z
D
[C] [h] = [F]
(5)
Z
µwei · wej dτ, Fi =
µwei · H s dτ and [h] is the
D
vector of degrees of freedom to be calculated.
A similar demonstration can be applied to (2) for the
projection of B s with respect to the magnetic energy.
B. Eddy current problems
For eddy current problems, either magnetic or electric
harmonic formulations can be used in order to calculate
harmonic fields. We obtain either H s or E s conformed with
physical properties. Thus, depending on the used formulation,
the energetic error norm to minimize can be defined [5]:
Z
Z
1
εH =
µ |Ht − H s |2 dτ+
|curl Ht − curl H s |2 dτ (6)
D
Dc σω
εE =
Z
Dc
σ
|Et − E s |2 dτ +
ω
Z
D
1 curl Et − curl E s 2
dτ (7)
µ
ω
S
with σ the electrical conductivity and ω the pulsation.
In order to project fields onto a target mesh with respect
to the magnetic energy in D and electric power in conducting
domain Dc , weak formulations can be obtained by the minimization of the energetic norm. Equation (6) gives rise to
matrix equation:
([C] + [Ce ]) [h] = [F] + [Fe ]
(8)
where [C] and [F] are
Z the same matrix as in magneto-static
1
problems, Cei j =
curl wei · curl wej dτ and Fei =
σω
D
c
Z
1
curl wei · curl H s dτ. Here [C] + [Ce ] is a positiveDc σω
definite matrix.
A similar matrix system can be obtained from (7) in order
to project E s .
III. Application
The proposed projection formulations are applied in an
academic example. This example is composed of a magnetic
core cylinder and an excitation coil (Fig. 1). Two meshes are
considered: Ms (306K elements) and Mt (60K elements). In
order to carry out projections, Ms is used as the source mesh
for the first computation and Mt is considered as the target
mesh for the projection.
Firstly, a constant current is applied in the coil. The
magneto-static problem is solved on Ms or Mt using the scalar
potential formulation. The obtained field H s on Ms is then
projected to Mt using (5). The corresponding magnetic flux
density on the clipping plane (S in Fig. 1) is presented in
Fig. 2a. In order to illustrate the advantage of the projection
method, the field H s is also projected using a classical L2 norm
projection (Fig. 2b). In comparison to the energetic projection,
the L2 projection fails to provide a correct distribution of B at
the boundary of the cylinder. Table I presents the max values of
fields and the magnetic energy calculated on Ms and Mt alone
for reference, and then from Ms to Mt projections. With the
formulation deduced from the minimization of the energetic
norm, the magnetic energy and the max values of fields are
close to the reference results (calculated on Mt alone).
Secondly, a sinusoidal current is applied as excitation. In
this case, eddy currents appear in the conductor cylinder.
The magneto harmonic problem is solved with the magnetic
formulation. Table II presents the reference results (calculated
on Ms and Mt alone) for the max values of fields, the magnetic
energy in the domain and the ohmic losses in the cylinder. As
the L2 projection fails to conserve the magnetic energy and
the ohmic losses, only the results obtained using the energetic
projection are shown. They are obtained using (8). In this
example, the magnetic energy as well as the ohmic losses are
well conserved using the energetic Galerkin projection.
Coil
Iron core
Figure 1: Geometry of the example
0.0887
4.18
0.0798
3.76
0.0709
3.34
0.0621
2.92
0.0532
2.51
0.0443
2.09
0.0355
1.67
0.0266
1.25
0.0177
0.836
0.00887
0.418
5.5e-06
5.5e-06
(b) L2 Galerkin projection
(a) Energetic Galerkin projection
Figure 2: Distribution of B (T) on a clipping plane, calculated
from the projected Ht field.
Table I: Validation of the energetic projection approach
(magneto-static problem)
magnetic energy
(mJ)
|H|max
(kA/m)
|B|max
(T )
calculated on Ms
calculated on Mt
8.01
8.25
44.1
49.0
0.116
0.096
L2 proj. Ms→Mt
energetic proj. Ms→Mt
10.28
7.32
42.9
45.3
4.17
0.089
Table II: Validation of the energetic projection approach (eddy
current problem)
magnetic
energy (mJ)
ohmic losses
(µW)
|H|max
(kA/m)
|J|max
(kA/m2 )
calculated on Ms
calculated on Mt
8.05
8.18
104
104
44.9
49.7
21.1
20.2
energetic proj.
Ms→Mt
7.54
98
46.8
19.6
References
[1] C. Geuzaine, B. Meys, F. Henrotte, P. Dular, and W. Legros, “A galerkin
projection method for mixed finite elements,” IEEE Transactions on
Magnetics, vol. 35, no. 3, pp. 1438 –1441, may 1999.
[2] P. Farrell, M. Piggott, C. Pain, G. Gorman, and C. Wilson, “Conservative
interpolation between unstructured meshes via supermesh construction,”
Computer Methods in Applied Mechanics and Engineering, vol. 198, no.
33–36, pp. 2632 – 2642, 2009.
[3] Z. Wang, T. Henneron, N. Nemitz, J.-C. Mipo, and F. Piriou, “Electromagnetic field projection on finite element overlapping domains,” IEEE
Transactions on Magnetics, vol. 49, no. 4, pp. 1290–1298, April 2013.
[4] Z. Wang, T. Henneron, J.-C. Mipo, and F. Piriou, “Projection des
champs entre deux maillages par une approche énergétique,” in Numélec,
Marseille, France, 2012, poster presentation.
[5] E. Creusé, S. Nicaise, Z. Tang, Y. Le Menach, N. Nemitz, and F. Piriou,
“Residual-based a posteriori estimators for the A/φ magnetodynamic
harmonic formulation of the maxwell system,” Mathematical Models and
Methods in Applied Sciences, vol. 22, no. 05, p. 1150028, 2012.
| 5cs.CE
|
Better Labeling Schemes for Nearest Common Ancestors
through Minor-Universal Trees
Paweł Gawrychowski1,2 and Jakub Łopuszański2
1
arXiv:1707.06011v1 [cs.DS] 19 Jul 2017
2
University of Haifa, Israel
University of Wrocław, Poland
Abstract
Preprocessing a tree for finding the nearest common ancestor of two nodes is a basic
tool with multiple applications. Quite a few linear-space constant-time solutions are known
and the problem seems to be well-understood. This is however not so clear if we want to
design a labeling scheme. In this model, the structure should be distributed: every node
receives a distinct binary string, called its label, so that given the labels of two nodes (and
no further information about the topology of the tree) we can compute the label of their
nearest common ancestor. The goal is to make the labels as short as possible. Alstrup,
Gavoille, Kaplan, and Rauhe [Theor. Comput. Syst. 37(3):441-456 2004] showed that
O(log n)-bit labels are enough, with a somewhat large constant. More recently, Alstrup,
Halvorsen, and Larsen [SODA 2014] refined this to only 2.772 log n, and provided a lower
bound of 1.008 log n.
We connect the question of designing a labeling scheme for nearest common ancestors
to the existence of a tree, called a minor-universal tree, that contains every tree on n nodes
as a topological minor. Even though it is not clear if a labeling scheme must be based
on such a notion, we argue that all already existing schemes can be reformulated as such.
Further, we show that this notion allows us to easily obtain clean and good bounds on
the length of the labels. As the main upper bound, we show that 2.318 log n-bit labels are
enough. Surprisingly, the notion of a minor-universal tree for binary trees on n nodes has
been already used in a different context by Hrubes et al. [CCC 2010], and Young, Chu, and
Wong [J. ACM 46(3):416-435, 1999] introduced a very closely related (but not equivalent)
notion of a universal tree. On the lower bound side, we show that any minor-universal tree for
trees on n nodes must contain at least Ω(n2.174 ) nodes. This highlights a natural limitation
for all approaches based on defining a minor-universal tree. Our lower bound technique also
implies that a universal tree in the sense of Young et al. must contain at least Ω(n2.185 )
nodes, thus dramatically improves their lower bound of Ω(n log n). We complement the
existential results with a generic transformation that allows us, for any labeling scheme for
nearest common ancestors based on a minor-universal tree, to decrease the query time to
constant, while increasing the length of the labels only by lower order terms.
1
Introduction
A labeling scheme assigns a short binary string, called a label, to each node in a network,
so that a function on two nodes (such as distances, adjacency, connectivity, or nearest common
ancestors) can be computed by examining their labels alone. We consider designing such scheme
for finding the nearest common ancestor (NCA) of two nodes in a tree. More formally, given
the labels of two nodes of a rooted tree, we want to compute the label of their nearest common
ancestor (for this definition to make sense, we need to explicitly require that the labels of all
nodes in the same tree are distinct).
Computing nearest common ancestors is one of the basic algorithmic questions that one can
consider for trees. Harel and Tarjan [23] were the first to show how to preprocess a tree using
a linear number of words, so that the nearest common ancestor of any two nodes can be found
in constant time. Since then, quite a few simpler solutions have been found, such as the one
described by Bender and Farach-Colton [12]. See the survey by Alstrup et al. [7] for a more
detailed description of these solutions and related problems.
While constant query time and linear space might seem optimal, some important applications
such as network routing require the structure to be distributed. That is, we might want to
associate some information with every node of the tree, so that the nearest common ancestor of
two nodes can be computed using only their stored information. The goal is to distribute the
information as evenly as possible, which can be formalized by assigning a binary string, called a
label, to every node and minimizing its maximum length. This is then called a labeling scheme
for nearest common ancestors. Labeling schemes for multiple other queries in trees have been
considered, such as distance [4, 8, 19, 20, 27], adjacency [5, 11, 13], ancestry [2, 17], or routing [31].
While we focus on trees, such questions make sense and have been considered also for more
general classes of graphs [1–6, 8–10, 16, 17, 21, 22, 26, 29]. See [30] for a survey of these results.
Looking at the structure of Bender and Farach-Colton, converting it into a labeling scheme
with short label is not trivial, as we need to avoid using large precomputed tables that seem
essential in their solution. However, Peleg [28] showed how to assign a label consisting of
O(log2 n) bits to every node of a tree on n nodes, so that given the labels of two nodes we can
return the predetermined name of their nearest common ancestor. He also showed that this is
asymptotically optimal. Interestingly, the situation changes quite dramatically if we are allowed
to design the names ourselves. That is, we want to assign a distinct name to every node of a tree,
so that given the names of two nodes we can find the name of their nearest common ancestor
(without any additional knowledge about the structure of the tree). This is closely connected
to the implicit representations of graphs considered by Kannan et al. [25], except that in their
case the query was adjacency. Alstrup et al. [7] showed that, somewhat surprisingly, this is
enough to circumvent the lower bound of Peleg by designing a scheme using labels consisting of
O(log n) bits. They did not calculate the exact constant, but later experimental comparison by
Fischer [16] showed that, even after some tweaking, in the worst case it is around 8. In a later
paper, Alstrup et al. [9] showed an NCA labeling scheme with labels of length 2.772 log n+O(1)1
and proved that any such scheme needs labels of length at least 1.008 log n − O(1). The latter
non-trivially improves an immediate lower bound of log n + Ω(log log n) obtained from ancestry.
They also presented an improved scheme for binary trees with labels of length 2.585 log n+O(1).
The scheme of Alstrup et al. [9] (and also all previous schemes) is based on the notion of
heavy path decomposition. For every heavy path, we assign a binary code to the root of each
subtree hanging off it. The length of a code should correspond to the size of the subtree, so
that larger subtrees receive shorter codes. Then, the label of a node is the concatenation of
the codes assigned to the subtrees rooted at its light ancestors, where an ancestor is light if
1
In this paper, log denotes the logarithm in base 2.
1
it starts a new heavy path. These codes need to be appropriately delimited, which makes the
whole construction (and the analysis) somewhat tedious if one is interested in optimizing the
final bound on the length.
1.1
Our Results
Our main conceptual contribution is connecting labeling schemes for nearest common ancestors
to the notion of minor-universal trees, that we believe to be an elegant approach for obtaining
simple and rather good bounds on the length of the labels, and in fact allows us to obtain
significant improvements. It is well known that some labeling problems have a natural and
clean connection to universal trees, in particular these two views are known to be equivalent for
adjacency [25] (another example is distance [18], where the notion of a universal tree gives a quite
good but not the best possible bound). It appears that no such connection has been explicitly
mentioned in the literature for nearest common ancestors so far. Intuitively, a minor-universal
tree for trees on n nodes, denoted Un , is a rooted tree, such that the nodes of any rooted tree
T on n nodes can be mapped to the nodes of Un as to preserve the NCA relationship. More
formally, T should be a topological minor of Un , meaning that Un should contain a subdivision
of T as a subgraph, or in other words there should exists a mapping f : T → Un such that
f (NCA(u, v)) = NCA(f (u), f (v)) for any u, v ∈ T . This immediately implies a labeling scheme
for nearest common ancestors with labels of length log |Un |, as we can choose the label of a node
u ∈ T to be the identifier of the node of Un it gets mapped to (in a fixed mapping), so small Un
implies short labels. In this case, it is not clear if a reverse connection holds. Nevertheless, all
previously considered labeling schemes for nearest common ancestors that we are aware of can
be recast in this framework.
The notion of a minor-universal tree has been independently considered before in different
contexts. Hrubes et al. [24] use it to solve a certain problem in computational complexity, and
construct a minor-universal tree of size n4 for all ordered binary trees on n nodes (we briefly
discuss how our results relate to ordered trees in Appendix A). Young et al. [32] introduce a
related (but not equivalent) notion of a universal tree, where instead of a topological minor
we are interested in minors that preserve the depth modulo 2, to study a certain question on
boolean functions, and construct such universal tree of size O(n2.376 ) for all trees on n nodes.
Our technical contributions are summarized in Table 1. The upper bounds are presented in
Section 3 and should be compared with the labeling schemes of Alstrup et al. [9], that imply
a minor-universal tree of size O(n2.585 ) for binary trees, and O(n2.772 ) for general (without
restricting the degrees) trees, and O(n2.585 ), and the explicit construction of a minor-universal
tree of size O(n4 ) for binary trees given by Hrubes et al. [24]. The lower bounds are described in
Section 4. We are aware of no previously existing lower bounds on the size of a minor-universal
tree, but in Appendix B we show that our technique implies a lower bound of Ω(n2.185 ) on the
size of a universal tree in the sense of Young et al. [32], which dramatically improves their lower
bound of Ω(n log n).
The drawback of our approach is that a labeling scheme obtained through a minor-universal
tree is not necessarily effective, as computing the label of the nearest common ancestor might
Trees
Binary
General
Lower bound
O(n1.728 )
O(n2.174 )
Upper bound
Ω(n1.894 )
Ω(n2.318 )
Figure 1: Summary of the new bounds on the size of minor-universal trees.
2
require inspecting the (large) minor-universal tree. However, in Section 5 we show that this
is, in fact, not an issue at all: any labeling scheme for nearest common ancestors based on a
minor-universal tree with labels of length c log n can be converted into a scheme with labels
of length c log n + o(log n) and constant query time. This further strengthens our claim that
minor-universal trees are the right approach for obtaining a clean bound on the size of the labels,
at least from the theoretical perspective (of course, in practice the o(log n) term might be very
large).
1.2
Our Techniques
Our construction of a minor-universal tree for binary trees is recursive and based on a generalization of the heavy path decomposition. In the standard heavy path decomposition of a tree T ,
the top heavy path starts at the root and iteratively descends to the child corresponding to the
largest subtree. Depending on the version, this either stops at a leaf, or at a node corresponding
to a subtree of size less than |T |/2. After some thought, a natural idea is to introduce a parameter α and stop after reaching a node corresponding to a subtree of size less than α · |T |. This is
due to a certain imbalance between the subtrees rooted at the children of the node where we stop
and all subtrees hanging off the top heavy path. Our minor-universal tree for binary trees on n
nodes consists of a long path to which the top heavy path of any T consisting of n nodes can
be mapped, and recursively defined smaller minor-universal trees for binary trees of appropriate
sizes attached to the nodes of the long path. Hrubes et al. [24] also follow the same high-level
idea, but work with the path leading to a centroid node. In their construction, there is only one
minor-universal tree for binary trees on 2/3n nodes attached to the last node of the path. We
attach two of them: one for binary trees on α · n nodes and one for binary trees on n/2 nodes.
We choose the minor-universal trees attached to the other nodes of the long path using the same
reasoning as Hrubes et al. [24] (which is closely connected to designing an alphabetical code with
codewords of given lengths used in many labeling papers, see for example [31]), except that
we can use a stronger bound on the total size of all subtrees hanging off the top path and not
attached to its last node than the one obtained from the properties of a centroid node. Finally,
we choose α as to optimize the whole construction. Very similar reasoning, that is, designing
a decomposition strategy by choosing the top heavy path with a cut-off parameter α and then
choosing α as to minimize the total size, has been also used by Young et al. [32], except that
their definition of a universal tree is not the same as our minor-universal tree and they do not
explicitly phrase their reasoning in terms of a heavy path decomposition, which makes it less
clear.
To construct a minor-universal tree for general trees on n nodes we need to somehow deal
with nodes of large degree. We observe that essentially the same construction works if we use
the following standard observation: if we sort the children of the root of T by the size of their
subtrees, then the subtree rooted at the i-th child is of size at most |T |/i.
To show a lower bound on the size of a minor-universal tree for binary trees on n leaves, we
also apply a recursive reasoning. The main idea is to consider s-caterpillars, which are binary
trees on s leaves and s − 1 inner nodes. For every node u in the minor-universal tree we find the
largest s, such that an s-caterpillar can be mapped to the subtree rooted at u. Then, we use the
inductive assumption to argue that there must be many such nodes, because we can take any
binary tree on bn/sc leaves and replace each of its leaves by an s-caterpillar. For general trees,
we consider slightly more complex gadgets, and in both cases need some careful calculations.
To show that any labeling scheme based on a minor-universal tree can be converted into
a labeling scheme with roughly the same label length and constant decoding time, we use a
recursive tree decomposition similar to the one used by Thorup and Zwick [31], and tabulate all
3
possible queries for tiny trees.
2
Preliminaries
We consider rooted trees, and we think that every edge is directed from a parent to its child.
Unless mentioned otherwise, the trees are unordered, that is, the relative order of the children is
not important. NCA(u, v) denotes the nearest common ancestor of u and v in the tree. deg(u)
denotes the degree (number of children) of u. A tree is binary if every node has at most two
children. For a rooted tree T , |T | denotes its size, that is, the number of nodes. In most cases,
this will be denoted by n. The whole subtree rooted at node u ∈ T is denoted by T u . If we say
that T 0 is a subtree of T , we mean that T 0 = T u for some u ∈ T , and if we say that T 0 is a
subgraph of T , we mean that T 0 can be obtained from T by removing edges and nodes.
If s is a binary string, |s| denotes its length, and we write s <lex t when s is lexicographically
less than t. is the empty string.
Let T be a family of rooted trees. An NCA labeling scheme for T consists of an encoder and
a decoder. The encoder takes a tree T ∈ T and assigns a distinct label (a binary string) `(u) to
every node u ∈ T . The decoder receives labels `(u) and `(v), such that u, v ∈ T for some T ∈ T ,
and should return `(NCA(u, v)). Note that the decoder is not aware of T and only knows that u
and v come from the same tree belonging to T . We are interested in minimizing the maximum
length of a label, that is, maxT ∈T maxu∈T |`(u)|.
3
NCA Labeling Schemes and Minor-Universal Trees
We obtain an NCA labeling scheme for a class T of rooted trees by defining a minor-universal
tree T for T . T should be a rooted tree with the property that, for any T 0 ∈ T , T 0 is a topological
minor of T , meaning that a subdivision of T 0 is a subgraph of T . In other words, it should be
possible to map the nodes of T 0 to the nodes of T as to preserve the NCA relationship: there
should exist a mapping f : T 0 → T such that f (NCA(u, v)) = NCA(f (u), f (v)) for any u, v ∈ T 0 .
We will define a minor-universal tree for trees on n nodes, denoted by Un , and a minor-universal
tree for binary trees on n nodes, denoted by Bn . Note that Bn does no have to be binary.
A minor-universal tree Un (or Bn ) can be directly translated into an NCA labeling scheme
as follows. Take a rooted tree T on n nodes. By assumption, there exists a mapping f : T → Un
such that f (NCA(u, v)) = NCA(f (u), f (v)) for any u, v ∈ T (if there are multiple such mappings,
we fix one). Then, we define an NCA labeling scheme by choosing, for every u ∈ T , the label
`(u) to be the (binary) identifier of f (u) in Un . The maximum length of a label in the obtained
scheme is dlog |Un |e. In the remaining part of this section we thus focus on defining small
minor-universal trees Bn and Un .
3.1
Binary Trees
Before presenting a formal definition of Bn , we explain the intuition.
Consider a binary rooted tree T . We first explain the (standard) notion of heavy path
decomposition. For every non-leaf u ∈ T , we choose the edge leading to its child v, such that
|T v | is the largest (breaking ties arbitrarily). We call v the heavy child of u. This decomposes
the nodes of T into node-disjoint heavy paths. The topmost node of a heavy path is called its
head. All existing NCA labeling schemes are based on some version of this notion and assigning
variable length codes to the roots of all subtrees hanging off the heavy path, so that larger
subtrees receive shorter codes. Then, the label of a node is obtained by concatenating the codes
of all of its light ancestors, or in other words ancestors that are heads of their heavy paths.
4
Ba(1)−1
...
Ba(2)−1
Ba(k)−1
Bbα·nc
Bb n−1 c
2
Figure 2: A schematic illustration for the recursive construction of Bn .
There are multiple possibilities for how to define the codes (and how to concatenate them while
making sure that the output can be decoded). Constructing such a code is closely connected
to the following lemma used by Hrubes et al. [24] to define a minor-universal tree for ordered
binary trees. The lemma can be also extracted (with some effort, as it is somewhat implicit)
from the construction of Young et al. [32].
Lemma 1 (see Lemma 8 of [24]). Let aN be a sequence recursively defined as follows: a1 = (1),
and aN = abN/2c ⊕ (N ) ⊕ abN/2c , where ⊕ denotes concatenation. Then, for any sequence
b = (b(1), b(2), . . . , b(k)) consisting of positive integers summing up to at most N , there exists a
subsequence a0 of aN that dominates b, meaning that the i-th element of a0 is at least as large as
the i-th element of b, for every i = 1, 2, . . . , k.
The sequence defined in Lemma 1 contains 1 copy of N , 2 copies of bN/2c, 4 copies of bN/4c,
and so on. In other words, there are 2i copies of bN/2i c, for every i = 0, 1, . . . , blog N c there.
To present our construction of a minor-universal binary tree we need to modify the notion of
heavy path decomposition. Let α ∈ (1/2, 1) be a parameter to be fixed later. We define α-heavy
path decomposition as follows. Let T be a rooted tree. We start at the root of T and, as long as
possible, keep descending to the (unique) child v of the current node u, such that |T v | ≥ α · |T |.
This defines the top α-heavy path. Then, we recursively decompose every subtree hanging off
the top α-heavy path.
Now we are ready to present our construction of the minor-universal binary tree Bn , that
immediately implies an improved nearest common ancestors labeling scheme for binary trees on
n nodes as explained in the introduction. B0 is the empty tree and B1 consists of a single node.
For n ≥ 2 the construction is recursive. We invoke Lemma 1 with N = b(1 − α)nc to obtain a
sequence ab(1−α)nc = (a(1), a(2), . . . , a(k)). Then, Bn consists of a path u1 − u2 − . . . − uk+1 .
We attach a copy of Ba(i)−1 to every ui , for i = 1, 2, . . . , k. Additionally, we attach a copy of
Bbα·nc and Bb(n−1)/2c to uk+1 . See Figure 2. Note that a(i) − 1 < n, bα · nc < n for α < 1,
and b(n − 1)/2c < n, so this is indeed a valid recursive definition. We claim that Bn is a
minor-universal tree for all binary trees on n nodes.
Lemma 2. For any binary tree T on n nodes, Bn contains a subgraph isomorphic to a subdivision
of T .
Proof. We prove the lemma by induction on n.
Consider a binary tree T on n ≥ 2 nodes and let v1 − v2 − . . . − vs be the path starting at
the root in the α-heavy path decomposition of T . Then, |T vs | ≥ α · n, but for every child u of vs
5
we have that |T u | < α · n. Consequently, the total size of all subtrees hanging off the path and
attached to v1 , v2 , . . . , vs−1 , increased by s−1, is at most (1−α)n. Also, denoting by u1 and u2 the
children of vs and ordering them so that |T u1 | ≥ |T u2 |, we have |T u1 | < α·n and |T u2 | ≤ (n−1)/2
(we assume that vs has two children, otherwise we can think that the missing children are of
size 0). Then, by the inductive assumption, a subdivision of T u1 is a subgraph of Bbα·nc , and
a subdivision of T u2 is a subgraph of Bb(n−1)/2c . Further, denoting by b(i) − 1 the size of the
P
subtree hanging off the path and attached to vi , we have s−1
i=1 b(i) ≤ m, where m = b(1 − α)nc,
and every b(i) is positive, so b is dominated by a subsequence of am = (a(1), a(2), . . . , a(k)). This
means that we can find indices 1 ≤ j(1) < j(2) < . . . < j(s − 1) ≤ k, such that b(i) ≤ a(j(i)),
for every i = 1, 2, . . . , s − 1. But then a subdivision of the subtree hanging off the path attached
to vi is a subgraph of Ba(j(i))−1 attached to uj(i) in Bn . Together, all these observations imply
that a subdivision of the whole T is a subgraph of Bn .
Finally, we analyze the size of Bn . Because Bn consists of a copy of Bbα·nc , a copy of
Bb(n−1)/2c , and 2i copies of Bb(1−α)n/2i c−1 attached to the path, for every i = 0, 1, . . . , blog(1 −
α)nc, we have the following recurrence:
blog(1−α)nc
X
|Bn | = 1 + |Bbα·nc | + |Bb(n−1)/2c | +
2i · (1 + |Bb(1−α)n/2i c−1 |)
i=0
We want to inductively prove that |Bn | ≤ nc , for some (hopefully small) constant c > 1. To
this end, we introduce a function b(x) = xc that is defined for any real x, and try to show that
|Bn | ≤ b(n) by induction on n. Using the inductive assumption, 1 + |Bm | ≤ b(m + 1) holds for
any m < n by applying the Bernoulli’s inequality and checking m = 0 separately. We would
also like to use 1 + |Bb(n−1)/2c | ≤ b(n/2) for n ≥ 2, which requires additionally verifying that
1 + k c ≤ (k + 1/2)c for any k ≥ 1. For c > 1.71, this holds for k ≥ 2 by the Bernoulli’s inequality
and can be checked for k = 1 separately. These inequalities allow us to upper bound |Bn | as
follows:
X
|Bn | ≤ b(α · n) + b(n/2) +
2i · b((1 − α)n/2i )
i≥0
To conclude that indeed |Bn | ≤ b(n), it suffices that the following inequality holds:
X
αc + (1/2)c +
2i ((1 − α)/2i )c ≤ 1
i≥0
c
c
α + (1/2) + (1 − α)c
c
c
α + (1 − α) · 2
X
(1/2c−1 )i ≤ 1
i≥0
c−1
c−1
/(2
− 1) ≤ 1 − (1/2)c
Minimizing f (x) = xc + (1 − x)c · 2c−1 /(2c−1 − 1) we obtain x = A/(1 + A), where A =
(2c−1 /(2c−1 −1))1/(c−1) . Thus, it is enough that (A/(1+A))c +(1/A)c ·2c−1 /(2c−1 −1) ≤ 1−(1/2)c .
This can be solved numerically for the smallest possible c and verified to hold for c = 1.894 by
choosing A = 2.372 and α = 0.704.
3.2
General Trees
To generalize the construction to non-binary trees, we use the same notion of α-heavy path
decomposition. Again, we invoke Lemma 1 with N = (1 − α)n to obtain a sequence ab(1−α)nc =
(a(1), a(2), . . . , a(k)). Un consists of a path u1 − u2 − . . . − uk+1 . For every i = 1, 2, . . . , k, we
6
...
Ua(1)−1
Uba(1)/2c
...
...
Ua(2)−1
Uba(2)/2c
...
Ua(k)−1
Uba(k)/2c
...
Ubα·nc
Ub n−1 c
2
Ub n−1 c
3
Figure 3: A schematic illustration for the recursive construction of Un .
attach a copy of Ua(i)−1 and, for every j ≥ 2, a copy of Uba(i)/jc to ui . Additionally, we attach
a copy of Ubα·nc to uk+1 and also, for every j ≥ 2, a copy of Ub(n−1)/jc . See Figure 3. We claim
that Un is indeed a minor-universal tree for all trees on n nodes.
Lemma 3. For any tree T on n nodes, Un contains a subgraph isomorphic to a subdivision of
T.
Proof. The proof is very similar to the proof of Lemma 2.
Consider a tree T on n nodes and let v1 − v2 − . . . − vs be the path starting at the root
in the α-heavy path decomposition of T . Again, the total size of all subtrees hanging off the
path and attached to v1 , v2 , . . . , vs−1 , increased by s − 1, is less than (1 − α)n, and if we denote
by u1 , u2 , u3 , . . . the children of vs and order them so that |T u1 | ≥ |T u2 | ≥ |T u3 | ≥ . . . then
|T u1 | < α · n and |T uj | ≤ (n − 1)/j for every j ≥ 2. Then, a subdivision of T u1 is a subgraph
of Ubα·nc , and a subdivision of T uj is a subgraph of Ub(n−1)/jc , for every j = 2, 3, . . .. Denoting
by b(i) − 1 the total size of all subtrees hanging off the path and attached to vi , we can find
indices 1 ≤ j(1) < j(2) < . . . < j(s − 1), such that b(i) ≤ a(j(i)), for every i = 1, 2, . . . , s − 1,
where ab(1−α)nc = (a(1), a(2), . . . , a(k)). Let v(i, 1), v(i, 2), . . . be the children of vi ordered so
that |T v(i,1) | ≥ |T v(i,2) | ≥ . . .. Then a subdivision of T v(i,1) is a subgraph of Ua(j(i))−1 attached
to uj(i) in Un and, for every k ≥ 2, a subdivision of T v(i,k) is a subgraph of Uba(j(i))/kc attached
to the same uj(i) in Un . This all imply that a subdivision of the whole T is a subgraph of Un .
To analyze the size of Un , observe that |Un | can be bounded by
1 + |Ubα·nc | +
n−1
X
i=2
b(1−α)n/2i c
blog(1−α)nc
|Ub(n−1)/ic | +
X
2i · (1 + |Ub(1−α)n/2i c−1 | +
i=0
X
|Ub(1−α)n/(2i ·j)c |)
j=2
We want to inductively prove that |Un | ≤ s(n), where s(n) = nc , for some constant c > 1. By
the same reasoning as the one used to bound |Bn |:
X
X X
|Un | ≤ s(α · n) +
s(n/i) +
2i
s((1 − α)n/(2i · j))
i≥2
i≥0
7
j≥1
For the inductive step to hold, it suffices that:
X X
X
((1 − α)/(2i · j))c ≤ 1
2i
(1/i)c +
αc +
i≥0
i≥2
j≥1
X
X
2i ((1 − α)/2i )c · ζ(c) ≤ 2
(1/i)c +
α +
c
i≥1
c
i≥0
c
α + (1 − α) · ζ(c)2c−1 /(2c−1 − 1) ≤ 2 − ζ(c)
P
c
where ζ(c) = ∞
i=1 1/i is the standard Riemann zeta function. Recall that our goal is to make
c as small as possible, and we can adjust α. By approximating ζ(c), we can verify that, after
choosing α = 0.659, the above inequality holds for c = 2.318.
4
Lower Bound for Minor-Universal Trees
In this section, we develop a lower bound on the number of nodes in a minor-universal tree for
binary trees on n nodes, and a minor-universal tree for general trees on n nodes. In both cases,
it is convenient to lower bound the number of leaves in a tree, that contains as a subgraph a
subdivision of any binary tree (or a general tree) T , such that T contains n leaves and no degree1 nodes. This is denoted by b(n) and u(n), respectively. Because we do not allow degree-1 nodes
in T , it has at most 2n − 1 nodes, thus b(n) is a lower bound on the size of a minor-universal
tree for binary trees on 2n − 1 nodes, and similarly u(n) is a lower bound on the size of a
minor-universal tree for general trees on 2n − 1 nodes.
4.1
Binary Trees
We want to obtain a lower bound on the number of leaves b(n) in a tree, that contains as a
subgraph a subdivision of any binary tree T on n leaves and no degree-1 nodes.
P
Lemma 4. b(n) ≥ 1 + s≥2 b(bn/sc).
Proof. For any s ≥ 2, we define an s-caterpillar to be a binary tree on s leaves and s − 1 inner
nodes creating a path. Consider a tree T , that contains as a subgraph a subdivision of any
binary tree on n leaves and no degree-1 nodes. For a node v ∈ T , let s(v) be the largest s, such
that T v contains a subdivision of an s-caterpillar as a subgraph. We say that such v is on level
s(v). We observe that s(v) has the following properties:
1. For every child u of v, s(u) ≤ s(v).
2. If the degree of v is 1 then, for the unique child u of v, s(u) = s(v).
3. If the degree of v is 2 then, for some child u of v, s(u) = s(v) − 1.
Choose a parameter s ≥ 2 and consider any binary tree on bn/sc leaves and no degree-1
nodes. By replacing all of its leaves by s-caterpillars we obtain a binary tree on at most n leaves
and still no degree-1 nodes. A subdivision of this new binary tree must be a subgraph of T .
The leaves of the original binary tree must be mapped to nodes on level at least s in T . Thus,
by removing all nodes on level smaller than s from T we obtain a tree T 0 that contains as a
subgraph a subdivision of any binary tree on bn/sc leaves and no degree-1 nodes, and so there
are at least b(bn/sc) leaves in T 0 . By the properties of s(u), a leaf of T 0 corresponds to a node
u ∈ T on level s (as otherwise u has a child on level at least s), and furthermore the degree of
8
u must be at least 2 (as otherwise the only child of u is on the same level). Because the level of
every u ∈ T is unambiguously defined, the total number of degree-2 nodes in T is at least:
X
b(bn/sc)
s≥2
To complete the proof, observe that in any tree the number of leaves is larger than the number
of degree-2 nodes.
We want to extract an explicit lower bound on b(n) from Lemma 4.
Theorem 5. For any c > 1 such that ζ(c) > 2 we have b(n) = Ω(nc ).
P
c
Proof.
We assume that ζ(c) > 2, so ∞
s=2 (1/s) > 1. Then there exists > 0 and t, such that
Pt
c
s=2 (1/s) = 1 + .
Now consider a function f (x) = xc . We claim that there exists x0 , such that for all x ≥ x0
we have f (bxc) ≥ f (x − 1) ≥ f (x)/(1 + ). This is because of the following transformations:
f (x − 1) ≥ f (x)/(1 + )
(x − 1)c ≥ xc /(1 + )
1 − 1/x ≥ 1/(1 + )1/c
where the right side is smaller than 1, so the inequality holds for any sufficiently large x.
We are ready to show that b(n) ≥Pa · nc for some constant a. We proceed by induction on n.
By Lemma 4, we know that b(n) ≥ s≥2 b(bn/sc). By adjusting a, it is enough to show that,
for sufficiently large values of n, b(N ) ≥ a · N c holding for all N < n implies b(n) ≥ a · nc . We
lower bound b(n) as follows:
n/x0
b(n) >
X
b(bn/sc) ≥
X
s≥2
c
a · (bn/sc) ≥ a
X
s=2
s≥2
n/x0
c
c
(bn/sc) ≥ a · n ·
X
(1/s)c /(1 + )
s=2
where in the last inequality we used that, as explained in the previous paragraph, (bn/sc)c ≥
(n/s)/(1 + ) for n/s ≥ x0 . By restricting n to be so large that n/x0 ≥ t, i.e., n ≥ t · x0 , we
further lower bound b(n) as follows:
b(n) ≥ a · nc ·
t
X
(1/s)c /(1 + ) ≥ a · nc · (1 + )/(1 + ) = a · nc
s=2
To apply Theorem 5, we verify with numerical calculation that ζ(1.728) > 2, and so b(n) =
Ω(n1.728 ).
4.2
General Trees
Now we move to general trees. We want to lower bound the number of leaves u(n) in a tree,
that contains as a subgraph a subdivision of any tree on n leaves and no degree-1 nodes.
We start with lower bounding the number of nodes of degree at least d in such a tree, denoted
u≥ (n, d). Similarly, u(n, d) denotes the number of nodes of degree exactly d.
P
Lemma 6. For any d ≥ 2, we have u≥ (n, d) ≥ s≥2 u(bn/((s − 1)(d − 1) + 1)c).
9
Proof. Fix d ≥ 2. For any s ≥ 2, we define an (s, d)-caterpillar to consist of path of length s − 1,
where we connect d − 1 leaves to every node except for the last, where we connect d leaves. The
total number of leaves in an (s, d)-caterpillar is hence (s − 1)(d − 1) + 1. Consider a tree T , that
contains as a subgraph a subdivision of any tree on n nodes and no degree-1 nodes. For any
node v ∈ T , let s(v) be the largest s, such that T v contains a subdivision of an (s, d)-caterpillar
as a subgraph. This is a direct generalization of the definition used in the proof of Lemma 4,
and so similar properties hold:
1. For every child u of v, s(u) ≤ s(v).
2. If the degree of v is less than d then, for some child u of v, s(u) = s(v).
3. If the degree of v is at least d then, for some child u of v, s(u) = s(v) − 1.
Choose any s ≥ 2 and consider a tree on bn/((s − 1)(d − 1) + 1)c leaves. By replacing all of
its leaves by (s, d)-caterpillars, we obtain a tree on at most n leaves, so subdivision of this new
tree must be a subgraph of T . The leaves of the original tree must be mapped to nodes on level
at least s in T , and by the same reasoning as in the proof of Lemma 4 this implies that there
are at least u(bn/((s − 1)(d − 1) + 1)c) nodes on level s and of degree at least d in T , making
the total number of nodes of degree d or more at least:
X
u(bn/((s − 1)(d − 1) + 1)c)
s≥2
Proposition
7. Let n(d) denote the number of nodes of degree d, then the number of leaves is
P
1 + d≥1 n(d) · (d − 1).
Proof. We apply induction on the size of the tree. If the tree consists of only one node, then
the claim holds. Assume that the root is of degree a ≥ 1. Then, by applying the inductive
assumption on every subtree attached to the root and denotingPby n0 (d) the number of non-root
nodes of degree d, we obtain that the number of leaves is a + d n0 (d) · (d − 1). Finally, n(d) is
n0 (d) if d 6= a, and n(a) = n0 (a) + 1, so the number of leaves is in fact:
X
X
a + n0 (a) · (a − 1) +
n0 (d) · (d − 1) = 1 +
n(d) · (d − 1)
d6=a
d
By combining Proposition 7 with u≥ (n, d) = u(n, d)+u≥ (n, d+1) and telescoping, we obtain
that the number of leaves is at least:
X
X
X
1+
u(n, d) · (d − 1) = 1 +
(u≥ (n, d) − u≥ (n, d + 1)) · (d − 1) = 1 +
u≥ (n, d)
d≥2
d≥2
d≥2
Finally, by substituting Lemma 6 we obtain:
XX
u(n) ≥ 1 +
u(bn/((s − 1)(d − 1) + 1)c)
d≥2 s≥2
· y + 1)c > 1 we have u(n) = Ω(nc ).
P
Proof. We extend the proof of Theorem 5. From x,y≥1 1/(x · y + 1)c > 1 we obtain that there
P
P
exists > 0 and t, such that tx=1 ty=1 1/(x · y + 1)c = 1 + .
Theorem 8. For any c > 1 such that
P
x,y≥1 1/(x
10
P
We know that u(n) ≥ s,d≥2 u(bn/((s − 1)(d − 1) + 1)c). As in the proof of Theorem 5, we
only need to show that u(n) ≥ a · nc for sufficiently large n. We lower bound u(n):
X
u(n) ≥
u(bn/((s − 1)(d − 1) + 1)c)
s,d≥2
≥
t X
t
X
a · (bn/(x · y + 1)c)c
x=1 y=1
≥ a · nc ·
0 ·x)
X n/(x
X
x≥1
(1/(x · y + 1))c /(1 + )
y=1
We choose n so large that n/(x0 · t) ≥ t and further lower bound u(n):
u(n) ≥ a · nc ·
t X
t
X
(1/(x · y + 1))c /(1 + )
x=1 y=1
c
≥ a · n · (1 + )/(1 + )
= a · nc
P
We verify with numerical calculations that x,y≥1 1/(x · y + 1)2.174 > 1 by computing the
P
P1000
2.174 and conclude that u(n) = Ω(n2.174 ).
sum 1000
x=1
y=1 1/(x · y + 1)
5
Complexity of the Decoding
In this section we present a generic transformation, that converts our existential results into
labeling schemes with constant query time in the word-RAM model with word size Ω(log n).
Our goal is to show the following statement for any class T of rooted trees closed under taking
topological minors: if, for any n, there exists a minor-universal tree of size nc , then there exists
a labeling scheme with labels consisting of c log n + o(log n) bits, such that given the labels of
two nodes we can compute the label of their nearest common ancestor in constant time. We
focus on general trees, and leave verifying that the same method works for any such T to the
reader.
Before proceeding with the main part of the proof, we need a more refined method of converting a minor-universal tree into a labeling scheme for nearest common ancestors. Intuitively,
we would like some nodes to receive shorter labels. This can be enforced with the following
lemma.
Lemma 9. For any tree T , it is possible to assign a distinct label `(u) to every node u ∈ T ,
such that |`(u)| ≤ 2 + log(|T |/(1 + deg(u))).
Proof. We partition the nodes of T into classes. For every k = 0, 1, . . . , blog(|T | + 1)c, the k-th
class contains all nodes with degree from [2k −1, 2k+1 −1). Observe that the sum of degrees of all
nodes of T is |T |−1, and consequently the k-th class consists of at most (2|T |−1)/2k nodes. Thus,
we can assign a distinct binary code of length dlog((2|T | − 1)/2k )e ≤ 2 + log(|T |/(1 + deg(u)))
as the label of every node u in the k-th class. The length of the code uniquely determines k so
the labels are indeed distinct.
Let b be a parameter to be fixed later. To define the labels of all nodes of a tree T , we
recursively decompose it into smaller trees as follows, similarly to [31]. First, we call a node
u such that |T u | ≥ n/b big, and small otherwise. Let T 0 be the subgraph of T consisting of
11
Figure 4: A schematic illustration of a partition by choosing O(b) interesting nodes. Circles
represent big nodes, and filled circles represent interesting big nodes.
all big nodes (notice that if u is big, then so is its parent). Then there are at most b leaves
in T 0 , as each of them corresponds to a disjoint subtree of size at least n/b. Therefore, there
are less than b branching nodes of T 0 . We call all leaves, all branching nodes, all big children
of branching nodes and, if the root of T 0 has exactly one big child, also the root and its only
big child, interesting. The total number of nodes designated as interesting so far is O(b). We
additionally attach virtual interesting nodes to some interesting nodes as follows. For a big
node u, weight(u) is defined as the total size of all small subtrees attached to it. If u is the
root, a leaf, or a branching node of T 0 , then we attach dweight(u)/(n/b)e virtual interesting
nodes as its children. If u is a big child of an interesting node, then there is a unique path
v = v0 − v1 − . . . vs = u from a leafPor a branching node of T 0 to u that do not contain any
other interesting nodes. We attach d si=1 weight(vi )/(n/b)e virtual interesting nodes as children
of u. We denote by T c the tree induced by all interesting nodes (including the virtual ones),
meaning that its nodes are all interesting nodes and the parent of a non-root interesting node
is its nearest interesting ancestor in T . Because the sum of weight(u) over all big nodes u is at
most n, the total number of virtual interesting nodes is O(b). Thus, the total number of nodes
in T c is O(b) = a · b. Further, any interesting node has at most one big non-interesting child. See
Figure 4 for a schematic illustration of such a partition. Every subtree rooted at a small node
such that its parent is big is then decomposed recursively using the same parameter b. Observe
that the depth of the recursion is logb n = log n/ log b.
We are ready to define a query-efficient labeling scheme. The label of every node consists
of O(log n/ log b) variable-length nonempty fields f1 , f2 , . . . , fs and some shared auxiliary information which will be explained later. The fields are simply concatenated together, and hence
also need to separately store |f1 |, |f2 |, |f3 |, . . . , |fs−1 |. This is done by extending a standard
construction as explained in Appendix C.
Lemma 10. Any set of at most s integers from [1, M ], such that M = O(log n), can be encoded
with O(s · max{1, log M
s }) bits, so that we can implement the following operations in constant
time:
1. extract the k th integer,
2. find the successor of a given x,
12
3. construct the encoding of a new set consisting of the smallest k integers.
The encoding depends only on the stored set (and the values of s and M ) and not on how it was
obtained.
Pi
Lemma 10 is applied to the set containing all numbers of the form
j=1 |fj |, for i =
1, 2, . . . , s − 1. Then, given a position in the concatenation we can determine in constant time
which field does it belong to, or find the first position corresponding to a given field. We can
also truncate the concatenation to contain only f1 , f2 , . . . , fi in constant time.
The i-th step of the recursive decomposition corresponds to three fields f3i−2 , f3i−1 , f3i ,
except that the last step corresponds to between one and two fields. Below we describe how the
fields corresponding to a single step are defined.
Consider a node u ∈ T and let u0 be its nearest interesting ancestor. The first field is the
label of u0 obtained from by applying Lemma 9 on T c . If u = u0 then we are done. Otherwise,
u 6= u0 , and we have two possibilities. If u0 is a leaf or a branching node of T 0 , the second field
contains a single 0. Otherwise, u0 is a big child of a branching node or the root of T 0 , and the
nearest big ancestor of u is some vj on a path v0 − v1 − . . . − vs = u0 , where j ∈ {1, 2, . . . , s}
and v0 is a leaf or a branching node of T 0 . In such case, we assign binary codes to all nodes
v1 , v2 , . . . , vs and choose the second field to contain
P the code `j assigned to node vj . The codes
should have the property that |`j | ≤ 1 + log(( si=1 weight(vi ))/weight(vj )), and furthermore
`1 <lex `2 <lex . . . <lex `s . Such codes can be obtained by the following standard lemma, that
essentially follows by the reasoning from Lemma 1. This is almost identical to Lemma 2.4 in [31]
(or Lemma 4.7 of [9]), except that we prefer the standard lexicographical order.
Pm
Lemma 11. Given positive integers b1 , b2 , . . . , bm and denoting B =
i=1 bi , we can find
nonempty binary strings s1 <lex s2 <lex . . . <lex sm , such that |si | ≤ 1 + log(B/bi ).
P
Proof. We choose the largest j, such that i<j bj ≤ bB/2c. We set sj+1 = 1. We recursively
define the binary strings for b1 , b2 , . . . , bj and prepend 0 to each of them. Then, we recursively
define the binary strings for bj+2 , . . . , bm and prepend 1 to each of them. To verify that si ≤
1 + log(B/bi ) holds, observe that the sum decreases by a factor of at least 2 in every recursive
call, and the length of the binary strings increases by 1.
Let u00 be the nearest big ancestor of u. If u = u00 then we are done. Otherwise, let
v1 , v2 , . . . , vd be all the small children of u00 . We order them so that |T vi | ≥ |T vi+1 | for i =
1, 2, . . . , d − 1. Then, u belongs to the subtree T vk , for some k ∈ {1, 2, . . . , d}. The third and
final field is simply the binary encoding of k consisting of 1 + blog kc ≤ 1 + log k bits. This
completes the description of the fields appended to the label in a single step of the recursion.
In every step of the recursion, the size of the current tree decreases by at least a factor of
b. If we could guarantee that the total length of all fields appended in a single step is at most
c log(a·b)+o(log b), this would be enough to bound the total length of a label by c log n+o(log n)
as desired. However, it might happen that the fields appended in the same step consist of even
log n bits. We claim that in such case the size of the current tree decreases more significantly,
similarly to the analysis of the labeling scheme for routing given in [31]. The following lemma
captures this property.
Lemma 12. Let t denote the total length of all fields corresponding to a single step of the
recursion, and s = 4 + c log(a · b). Then the size of the current tree decreases by at least a factor
of b · 2max{0,t−s} .
Proof. If only the first field is defined, the claim is trivial, as its length is always at most
s. Observe that a node of degree d must be mapped to a node of degree at least d in the
13
minor-universal tree, and consequently by Lemma 9 the length of the first field is at most
2 + c log(a ·P
b) − log(1 + deg(u0 )), where u0 is the nearest interesting node. By construction,
there
d si=1 weight(vi )/(n/b)e virtual nodes attached to u0 , P
and so log(1 + deg(u0 )) ≥
Pare
s
log( i=1 weight(vi )/(n/b)). The length of the second field is 1+log(( si=1 weight(vi ))/weight(vj ))
(if u0 is a leaf or a branching node of T 0 , we define s = 1 and v1 = u0 ). Finally, the length of the
third field is 1 + log k (or there is no third field). All in all, we have the following:
s
s
X
X
t − s ≤ − log(
weight(vi )/(n/b)) + log((
weight(vi ))/weight(vj )) + log k
i=1
i=1
≤ log(n/b) − log(weight(vj )) + log k
= log((n/b)k/weight(vj ))
Observe that weight(vj ) ≤ n/b, so b · 2max{0,t−s} ≤ b · (n/b)k/weight(vj ). Finally, the size of
the current tree changes to at most weight(vj )/k due to the ordering of the children of vj , or in
other words decreases at least by a factor of b · (n/b)k/weight(vj ) ≥ b · 2max{0,t−s} .
Now we analyze the total contribution of all steps to the total length of the label. Let ti
be the total length of all fields added in the i-th step, r denote the number of steps, and s be
defined as in Lemma 12. Using c ≥ 1, the total length of a label is then:
r
X
ti ≤
i=1
r
X
s + max{0, ti − s} = r(4 + c log a) + c
i=1
r
X
log b + max{0, ti − s}
i=1
Because in every step the size of the current tree decreases at least by a factor of b · 2max{0,ti −s} ,
the product of such expressions is at most n, and so the total length of a label can be upper
bounded by:
r
X
ti ≤ log n/ log b · (4 + c log a) + c log n = O(log n/ log b) + c log n
i=1
As long as b = ω(1), this is c log n + o(log n) as required.
We move on to explaining how to implement a query in constant time given the labels of u and
v. By considering the parts containing the concatenated fields, finding the first position where
they differ, and querying the associated rank/select structure, we can determine in constant time
the first field that is different in both labels. This gives us the step of the recursive decomposition,
such that u and v belong to different small subtrees, or at least one of them is a big node (and
thus does not participate in further steps). Observe that the nearest common ancestor of u and
v must a big node. Its label can be found by, essentially, truncating the label of u and v and
possibly appending a label obtained from the non-efficient scheme. We now describe the details
of this procedure.
Let u0 and v 0 denote the nearest interesting ancestor of u and v, respectively. We would like
to find the nearest common ancestor w of u0 and v 0 . Note that, by construction, w must be
an interesting node. Thus, we can use the minor-universal tree to obtain its label. However,
the minor-universal tree does not allow us to answer a query efficiently by itself. Thus, we
preprocess all such queries in a table NCA[x][y], where x and y are labels consisting of at most
2 + c log(a · b) bits, and every entry also consists of at most 2 + c log(a · b) bits (to facilitate
constant-time access, NCA is stored as (2 + c log(a · b))2 separate tables of the same size, one
for each possible combination of |x| and |y|, and every entry is encoded with Elias γ code [15]
and stored in a field of length 2(2 + c log(a · b))). This lookup table is shared between all steps
14
of the recursion. Now, if w ∈
/ {u0 , v 0 } then w is the sought nearest common ancestor. Its label
can be obtained by truncating the label of, say, u, and appending a field storing the label of w
in the minor-universal tree. We also need to update the rank/select structure. This can be also
done by truncating and does not require adding a new integer to the set, because we do not
store the length of the last field explicitly. Hence, the label of w can be obtained in constant
time with the standard word-RAM operations. Otherwise, assume without losing the generality
that w = v 0 . If w = u0 also holds, then we look at the second field of both labels (if there is
none in one of them then again w is the sought nearest common ancestor). If there are equal
then the nearest big ancestor of u and v is the same, and should be returned as the nearest
common ancestor. Its label can be obtained by truncating the label of either u or v. Otherwise,
recall that w has at most one big child, and so there is a path v0 − v1 − . . . vs = w between two
interesting nodes, such that u belongs to a small subtree attached to some vi and v is in a small
subtree attached to some vj , where i, j ∈ {1, 2, . . . , s}. Because the binary codes assigned to the
nodes of the path preserve the bottom-top order, we can check whether i < j, i = j, or i > j. If
i > j (i < j), then vi (vj ) is the sought nearest common ancestor, and its label can be obtained
by truncating the label of u (v). Finally, if i = j, then the nearest common ancestor must be
vi = vj , because we know that u and v do not belong to the same small subtree, and truncate
the label of either u or v.
We analyze the total length of a label. It consists of 1) the concatenated fields, 2) rank/select
structure encoding the lengths of the fields, 3) a lookup table for answering queries in the minoruniversal tree. The total length of all the fields is L = O(log n/ log b) + c log n. The rank/select
structure from Lemma 10 is built for a set of at most log n/ log b integers from [L], and so takes
O(log n/ log b · log(L · log b/ log n)) bits of space. The lookup table uses (a · b)2c · 2(2 + c log(a · b))3
bits of space, making the total length:
c log n + O(log n · log log b/ log b + (a · b)2c log3 b)
By setting b = 1/a·(log n)1/(4c) we obtain labels of length c log n+o(log n) and constant decoding
time.
Theorem 13. Consider any class T of rooted trees closed under taking topological minors. If,
for any n, there exists a minor-universal tree of size nc then there exists a labeling scheme for
nearest common ancestors with labels consisting of c log n+o(log n) bits and constant query time.
References
[1] Amir Abboud, Pawel Gawrychowski, Shay Mozes, and Oren Weimann. Near-optimal compression for the planar graph metric. CoRR, abs/1703.04814, 2017.
[2] Serge Abiteboul, Stephen Alstrup, Haim Kaplan, Tova Milo, and Theis Rauhe. Compact
labeling scheme for ancestor queries. SIAM Journal on Computing, 35(6):1295–1309, 2006.
[3] Noga Alon and Rajko Nenadov. Optimal induced universal graphs for bounded-degree
graphs. In 28th SODA, pages 1149–1157, 2017.
[4] Stephen Alstrup, Philip Bille, and Theis Rauhe. Labeling schemes for small distances in
trees. SIAM Journal on Discrete Mathematics, 19(2):448–462, 2005.
[5] Stephen Alstrup, Søren Dahlgaard, and Mathias Bæk Tejs Knudsen. Optimal induced
universal graphs and adjacency labeling for trees. In 56th FOCS, pages 1311–1326, 2015.
15
[6] Stephen Alstrup, Cyril Gavoille, Esben Bistrup Halvorsen, and Holger Petersen. Simpler,
faster and shorter labels for distances in graphs. In 27th SODA, pages 338–350, 2016.
[7] Stephen Alstrup, Cyril Gavoille, Haim Kaplan, and Theis Rauhe. Nearest common ancestors: a survey and a new distributed algorithm. In 14th SPAA, pages 258–264, 2002.
[8] Stephen Alstrup, Inge Li Gørtz, Esben Bistrup Halvorsen, and Ely Porat. Distance labeling
schemes for trees. In 43rd ICALP, pages 132:1–132:16, 2016.
[9] Stephen Alstrup, Esben Bistrup Halvorsen, and Kasper Green Larsen. Near-optimal labeling schemes for nearest common ancestors. In 25th SODA, pages 972–982, 2014.
[10] Stephen Alstrup, Haim Kaplan, Mikkel Thorup, and Uri Zwick. Adjacency labeling schemes
and induced-universal graphs. In 47th STOC, pages 625–634, 2015.
[11] Stephen Alstrup and Theis Rauhe. Small induced-universal graphs and compact implicit
graph representations. In 43rd FOCS, pages 53–62, 2002.
[12] Michael A. Bender and Martin Farach-Colton. The LCA problem revisited. In 4th LATIN,
pages 88–94, 2000.
[13] Nicolas Bonichon, Cyril Gavoille, and Arnaud Labourel. Short labels by traversal and
jumping. Electronic Notes in Discrete Mathematics, 28:153–160, 2007.
[14] David Richard Clark. Compact Pat Trees. PhD thesis, University of Waterloo, 1998.
[15] Peter Elias. Universal codeword sets and representations of the integers. IEEE Transactions
on Information Theory, 21(2):194–203, 1975.
[16] Johannes Fischer. Short labels for lowest common ancestors in trees. In 17th ESA, pages
752–763, 2009.
[17] Pierre Fraigniaud and Amos Korman. Compact ancestry labeling schemes for xml trees. In
21st SODA, pages 458–466, 2010.
[18] Ofer Freedman, Pawel Gawrychowski, Patrick K. Nicholson, and Oren Weimann. Optimal
distance labeling schemes for trees. CoRR, abs/1608.00212, 2016.
[19] Cyril Gavoille and Arnaud Labourel. Distributed relationship schemes for trees. In 18th
ISAAC, pages 728–738, 2007. Announced at PODC’07.
[20] Cyril Gavoille, David Peleg, Stéphane Pérennes, and Ran Raz. Distance labeling in graphs.
Journal of Algorithms, 53(1):85–112, 2004. A preliminary version in 12th SODA, 2001.
[21] Pawel Gawrychowski, Adrian Kosowski, and Przemyslaw Uznanski. Sublinear-space distance labeling using hubs. In 30th DISC, pages 230–242, 2016.
[22] Pawel Gawrychowski and Przemyslaw Uznanski. A note on distance labeling in planar
graphs. CoRR, abs/1611.06529, 2016.
[23] Dov Harel and Robert Endre Tarjan. Fast algorithms for finding nearest common ancestors.
SIAM J. Comput., 13(2):338–355, 1984.
[24] Pavel Hrubes, Avi Wigderson, and Amir Yehudayoff. Relationless completeness and separations. In 25th CCC, pages 280–290, 2010.
16
[25] Sampath Kannan, Moni Naor, and Steven Rudich. Implicit representation of graphs. SIAM
Journal on Discrete Mathematics, 5(4):596–603, 1992.
[26] Michal Katz, Nir A. Katz, Amos Korman, and David Peleg. Labeling schemes for flow and
connectivity. SIAM J. Comput., 34(1):23–40, 2004.
[27] David Peleg. Proximity-preserving labeling schemes. Journal of Graph Theory, 33(3):167–
176, 2000.
[28] David Peleg. Informative labeling schemes for graphs. Theor. Comput. Sci., 340(3):577–593,
2005.
[29] Casper Petersen, Noy Rotbart, Jakob Grue Simonsen, and Christian Wulff-Nilsen. Nearoptimal adjacency labeling scheme for power-law graphs. In 43rd ICALP, pages 133:1–
133:15, 2016.
[30] Noy Galil Rotbart. New Ideas on Labeling Schemes. PhD thesis, University of Copenhagen,
2016.
[31] Mikkel Thorup and Uri Zwick. Compact routing schemes. In 13th SPAA, pages 1–10, 2001.
[32] Fung Yu Young, Chris C. N. Chu, and D. F. Wong. Generation of universal series-parallel
boolean functions. J. ACM, 46(3):416–435, 1999.
17
A
Ordered Trees
Hrubes et al. [24] consider ordered binary trees and construct a minor-universal tree of size
O(n4 ) for ordered binary trees on n nodes. We modify their construction to obtain a smaller
minor-universal tree for ordered binary trees Bn0 as described below.
We invoke Lemma 1 with N = b(1−α)nc to obtain a sequence ab(1−α)nc = (a(1), a(2), . . . , a(k)).
Then, Bn0 consists of a path u1 − v1 − u2 − v2 − . . . − uk+1 − vk+1 − w. For every i = 1, 2, . . . , k,
0
0
we attach a copy of Ba(i)−1
as the left child of ui , and we also attach a copy of Ba(i)−1
as the
0
right child of vi . Additionally, we attach a copy of Bbα·nc as the left child of w, and another
0
copy of Bbα·nc
as the right child of w. By a similar argument to the one used to argue that Bn is
a minor-universal tree for all binary trees on n nodes we can show that Bn0 is a minor-universal
tree for all ordered binary trees on n nodes if α ≥ 0.5. Its size can be bounded as follows:
blog(1−α)nc
|Bn0 |
=1+
0
2|Bbα·nc
|
+2
X
0
2i · (1 + |Bb(1−α)n/2
i c−1 |)
i=0
To show that |Bn0 | ≤ nc it is enough that the following inequality holds:
2αc + (1 − α)c · 2c /(2c−1 − 1) ≤ 1
So it is enough that 2(A/(1+A))c +(1/A)c ·2c /(2c−1 −1) ≤ 1, where A = (2c−1 /(2c−1 −1))1/(c−1) .
This can be verified to hold for c = 2.331 by choosing A = 1.463 and α = 0.594.
B
Universal Trees of Young et al.
To present the definition of a universal tree in the sense of Young et al. [32] we first need to
present their original definition of two operations on trees:
Cutting. Two nodes a and b, such that a is a child of b, are selected. The entire subtree rooted
at a and the edge between a to b are removed.
Contraction. An internal node b, which has parent a and a single child c, is selected. Node b
is removed. If c is internal node, the children of c are made children of a and c is removed.
If c is a leaf, it becomes a child of a.
Then, tree T implements tree T 0 if T 0 can be obtained by applying a sequence of cutting and
contraction operations to T . Finally, T is an n-universal tree if it can implement any tree T 0
with at most n leaves and no degree-1 nodes. Notice that the degrees of the nodes of T 0 are
not bounded in the original definition. However, for our purposes it will be enough to consider
binary trees. We want to prove a lower bound on the number of leaves of an n-universal tree.
We introduce the notion of parity-preserving minor-universal trees. We say that T is a
parity-preserving minor-universal tree for a class T of rooted trees if, for any T 0 ∈ T , the nodes
of T 0 can be mapped to the nodes of T as to preserve the NCA relationship and the parity of
the depth of every node.
Lemma 14. An n-universal tree is a parity-preserving minor-universal tree for binary trees on
n leaves and no degree-1 nodes.
Proof. We first observe that the definition of contraction can be changed as follows:
Contraction. An internal node b, which has parent a and a single child c, is selected. Node b
is removed. The children of c are made children of a and c is removed.
18
This is because if c is a leaf, reattaching c to a and removing b is equivalent to cutting c.
Let T be an n-universal tree and consider any binary tree T 0 on n leaves and no degree-1
nodes. By assumption, T implements T 0 , so we can obtain T 0 from T by a sequence of cutting
and contraction operations. We claim that if T implements T 0 then the nodes of T 0 can be
mapped to the nodes of T as to preserve the NCA relationship and the parity of the depth
of every node. We prove this by induction on the length of the sequence. If the sequence is
empty, the claim is obvious. Otherwise, assume that T1 is obtained from T by a single cutting
or contraction, and by the inductive assumption the nodes of T 0 can be mapped to the nodes of
T1 as to preserve the NCA relationship and the parity of the depth of every node. For cutting,
the claim is also obvious, as we can use the same mapping. For contraction, we might need to
modify it. Observe that at most one node u ∈ T 0 is mapped to the node a. If there is no such
node, or the degree of u in T 0 is 1, we are done because the parity of the depth of every node
that appears in both T and T1 is the same and the NCA relationship in T restricted to the nodes
that appear in T1 is also identical. Otherwise, let v1 and v2 be the children of u in T 0 . v1 (v2 )
is mapped to a node in the subtree rooted at a child a1 (a2 ) of a in T1 . If both a1 and a2 are
children of c in T then we modify the mapping so that u is mapped to c in T , and otherwise u
is mapped to the original a in T . Mapping of other nodes of T 0 remains unchanged. It can be
verified that the obtained mapping indeed preserves the NCA relationship and the parity of the
depth of every node. Thus, T is indeed a parity-preserving minor-universal tree for binary trees
on n leaves and no degree-1 nodes.
We need one more definition. Let inner(T 0 ) be the set of inner nodes of a tree T 0 . A tree
T is a parity-constrained minor-universal tree for a class T of rooted trees if, for any T 0 ∈ T
and any assignment c : inner(T 0 ) → {0, 1}, the nodes of T 0 can be mapped to the nodes of T as
to preserve the NCA relationship and, for any v ∈ inner(T 0 ), if c(v) = 0 then v is mapped to a
node at even depth and if c(v) = 1 then v is mapped to a node at odd depth. c is called the
parity constraint.
Lemma 15. A (2n − 1)-universal tree is a parity-constrained minor-universal tree for binary
trees on n leaves and no degree-1 nodes.
Proof. Given a tree T 0 on n leaves and no degree-1 nodes, and an assignment c : inner(T 0 ) →
{0, 1}, we will construct a tree T 00 on at most 2n − 1 leaves (and also no degree-1 nodes), such
that if the nodes of T 00 can be mapped to the nodes of T as to preserve the NCA relationship and
the parity of the depth of every node, then the nodes of T 0 can be mapped to the nodes of T as
to preserve the NCA relationship and respect the parity constraint. Together with Lemma 14,
this proves the lemma.
We transform T 0 into T 00 as follows. We consider all inner nodes of T 0 in the depth-first
order. Let r be the root of T 0 . If c(r) = 1, then we create a new root r0 , make r a child of r0 ,
and attach a new leaf as another child of r0 . For a node v with parent u in T 0 , if c(u) 6= c(v)
then we do nothing. If c(u) = c(v), then we attach a new child v 0 to u, make v a child of v 0 ,
and attach a new leaf as another child of v 0 . The total number of new leaves created during
the process is at most the number of inner nodes of T 0 , so the total number of leaves in T 00 is
at most 2n − 1. It is easy to see that preserving the parity of the depth of every node of T 00
implies respecting the parity constraint for the original nodes of T 0 , and the NCA relationship
restricted to the original nodes in T 00 is the same as in T 0 .
We are ready to proceed with the main part of the proof. Our goal is to lower bound the
number of leaves b(n) in a parity-constrained minor-universal tree for binary trees on n leaves and
no degree-1 nodes. By Lemma 15, this also implies a lower bound on the number of leaves (and
thus the size) of a universal tree in the sense of Young et al. [32]. Because a parity-constrained
19
minor-universal tree is a minor-universal tree, we could simply apply Lemma 4 and conclude
that b(n) = Ω(n1.728 ). Our goal is to obtain a stronger lower bound by exploiting the parity
constraint.
P
Lemma 16. b(n) ≥ 1 + 2 s≥2 b(bn/sc).
Proof. Let T be a parity-constrained minor-universal tree for binary trees on n leaves and no
degree-1 nodes. We choose d ∈ {0, 1} such that at most half of nodes of degree 2 or more in T
is at depth congruent to d modulo 2.
For any s ≥ 2, we define an s-caterpillar and s(v) for any node v ∈ T as in the proof of
Lemma 4, except that now we require that all inner nodes of the s-caterpillar should be mapped
to nodes at depth congruent to d modulo 2. This changes the properties of s(v) as follows:
1. For every child u of v, s(u) ≤ s(v).
2. If the degree of v is 1 then, for the unique child u of v, s(u) = s(v).
3. If the degree of v is at least 2 and the depth of v is not congruent to d modulo 2 then, for
some child u of v, s(u) = s(v).
4. If the degree of v is at least 2 and the depth of v is congruent to d modulo 2 then, for
some child u of v, s(u) = s(v) − 1.
Then, for any s ≥ 2, we consider any binary tree on bn/sc leaves and no degree-1 nodes, and
any choice of the parities for all of its inner nodes. We replace all leaves of the original binary
tree by s-caterpillars and require that their inner nodes are mapped to nodes at depth congruent
to d modulo 2, while for the original inner nodes the required parity remains unchanged. By
assumption, it must be possible to map the nodes of the new binary tree to the nodes of T as
to preserve the NCA relationship and respect the parity constraint. The leaves of the original
binary tree must be mapped to nodes on level at least s in T . As in the proof of Lemma 4, we
obtain a tree T 0 by removing all nodes on level smaller than s from T . From the properties of
s(v) it is clear that every leaf of T 0 is on level exactly s and of degree at least 2. We claim that,
additionally, all leaves of T 0 are at depth congruent to d modulo 2. This is because if a node
v ∈ T is at depth not congruent to d modulo 2 then, for some child u of v, s(u) = s(v), so in
fact v cannot be a leaf in T 0 . For any binary tree on bn/sc leaves and no degree-1 nodes and
any choice of the parities for the inner nodes, the nodes of the binary tree can be mapped to
the nodes of T 0 as to preserve the NCA relationship and respect the parity constraint. By lower
bounding the number of leaves in T 0 we thus obtain that the number of degree-2 nodes on level
s and at depth congruent to d modulo 2 in T is at least b(bn/sc). Thus, the total number of
degree-2 nodes at depth congruent to d modulo 2 in T is
X
b(bn/sc)
s≥2
Finally, by the choice of d the total number of degree-2 nodes is at least twice as large, and so
the total number of leaves exceeds
X
2
b(bn/sc)
s≥2
To extract an explicit lower bound from Lemma 16, we proceed as in Theorem 5. It is
straightforward to verify that the same reasoning can be used to show that, if ζ(c) > 1.5 then
b(n) = Ω(nc ). We verify that ζ(2.185) > 1.5, and so b(n) = Ω(n2.185 ).
20
C
Missing Proofs
Lemma 10. Any set of at most s integers from [1, M ], such that M = O(log n), can be encoded
with O(s · max{1, log M
s }) bits, so that we can implement the following operations in constant
time:
1. extract the k th integer,
2. find the successor of a given x,
3. construct the encoding of a new set consisting of the smallest k integers.
The encoding depends only on the stored set (and the values of s and M ) and not on how it was
obtained.
Proof. The encoding is similar to Lemma 2.2 of [18], except that we cannot use a black box
0
predecessor structure. Let L = s · max{1, log M
s } and the set consists of x1 < x2 < . . . < xs ,
where s0 ≤ s.
We partition the universe [1, M ] into blocks of length b = M
s . The encoding starts with b
encoded with the Elias γ code [15]. Then we store every xi mod b using 2 + log b bits. The
encodings of xi mod b are separated by single 1s. This takes O(log b + s + s log b) = L bits so far.
We need to also store every yi = xi div b. We observe that 0 ≤ y1 ≤ y2 ≤ . . . ys0 ≤ s. Hence, we
can encode them with a bit vector of length at most 2s, which is a concatenation of 0yi −yi−1 1 for
i = 1, 2, . . . , s0 . The bit vector is augmented with a select structure of Clark [14, Chapter 2.2],
which uses o(s) additional bits and allows us to extract the ith bit set to 1 in constant time.
This all takes O(L) bits of space and allows us to decode any xi in constant time by extracting
xi mod b and xi div b.
To find the successor of x, we first compute y = x div b. Then, using the bit vector we can
find in constant time the maximal range of integers xi , xi+1 , . . . , xj such xk div b = y for every
k = i, i + 1, . . . , j. The successor can be then found by finding the successor of x mod b among
xi mod b, xi+1 mod b, . . . , xj mod b and, if there is none, returning xj+1 . To find the successor
of x mod b in the range, we use the standard method of repeating the encoding of x mod b
separating by single 0s (j −i+1) times by multiplying with an appropriate constant (that can be
computed with simple arithmetical operations in constant time, assuming that we can multiply
and compute a power of 2 in constant time), and then subtracting the obtained bit vector from
a bit vector containing the encodings of xi mod b, xi+1 mod b, . . . , xj mod b separated by single
1s (that is obtained from the stored encoding with standard bitwise operations). The bit vectors
fit in a constant number of words, and hence all operations can be implemented in constant
time.
Finally, we describe how to truncate the encoding. The only problematic part is that we
have used a black box select structure. Now, we want to truncate the stored bit vector, and this
might change the additional o(s) bits. We need to inspect the internals of the structure.
Recall that the structure of Clark [14, Chapter 2.2] for selecting the k th occurrence of 1
partitions a bit vector of length m into macroblocks by choosing every tth
1 such occurrence,
where t1 = log m log log m. We encode every macroblock separately and concatenate their
encodings. Additionally, for every i we store the starting position of the i-th macroblock in the
bit vector and the starting position of its part of the encoding in an array using O(m/t1 ·log m) =
O(m/ log log m) bits. Now consider a single macroblock and let r be its length. If r > t21 , we
store the position of every 1 inside the macroblock explicitly. This is fine because there can be
at most m/t21 such blocks, so this takes O(m/t21 · t1 · log m) = O(m/ log log m) bits. Otherwise,
we will encode the relative position of every 1, but not explicitly. We further partition such
21
2
macroblock into blocks by choosing every tth
2 occurrence of 1, where t2 = (log log m) . We
encode every block separately and concatenate their encodings, and for every i store the relative
starting position of the i-th block (in its macroblock) and the relative starting position of its
part of the encoding (in the encoding of the macroblock) in an array using O(m/t2 · log log m)
bits (we will make sure that the encoding of any macroblock takes only O(polylog m) bits).
Then, let again r be length of a block. If r > t22 , we can store the relative position of every 1
(in its macroblock) inside the block explicitly. Otherwise, the whole block is of length less than
t22 < 21 log m, and we can tabulate. In more detail, for every bit vector of length at most 12 log m
√
(there are O( m) of them), we store the positions of the at most 12 log m 1s explicitly. This
precomputed table takes o(m) bits, so can be stored as a part of the structure. Then, given a
block of length less than 12 log m, we extract its corresponding fragment of the bit vector using
the standard bitwise operations, and use the precomputed table.
We are now ready to describe how to update the select structure after truncating the bit
vector after the k th occurrence of 1. We first determine the macroblock containing this occurrence, say that it is the ith macroblock. We can easily discard all further macroblocks by
checking where the encoding of the (i + 1)th macroblock starts and erasing everything starting
from there. We also erase the starting positions stored for all further macroblocks, and move
the encoding just after the remaining starting positions. This can be done in constant time
using standard bitwise operations. Then, we inspect the ith macroblock. If the positions of all
1s are stored explicitly, we erase a suffix of this sequence. This is now problematic, because
maybe after erasing a suffix r becomes at most t22 and we actually need the other encoding. We
overcome this difficulty by changing the definition: a macroblock is partitioned into a prefix of
length t22 and the remaining suffix. The occurrences of all 1s in the suffix are stored explicitly,
and we also store the number of occurrences in the prefix. Then, the prefix is partitioned into
blocks by choosing every tth
2 occurrence. To truncate the prefix, we need to completely erase a
suffix of blocks, which can be done in constant time, and modify the last remaining block. If the
encoding of the last block consists of explicitly stored relative positions, we just need to erase
its suffix, which again can be done in constant time. Otherwise, there is actually nothing to do.
Additionally, we need to make sure that the precomputed table does not have to be modified.
To this end, instead of tabulating every bit vector of length at most 21 log m, we tabulate every
bit vector of length at most 12 log s (instead of 12 log m).
22
| 8cs.DS
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.